2009:073
M A S T E R ' S T H E S I S
Prediction of Vehicle Trajectories with Map Data for Cooperative Systems
Diego Fernandes Boesel
Luleå University of Technology Master Thesis, Continuation Courses
Space Science and Technology Department of Space Science, Kiruna
2009:073 - ISSN: 1653-0187 - ISRN: LTU-PB-EX--09/073--SE
Diego Fernandes Boesel
Prediction of Vehicle Trajectories with Map Data for Cooperative Systems
Thesis submitted in partial ful llment of the requirements for the degree of Master of Science in Technology
Espoo August 9, 2009
Thesis Work carried out at Volvo Technology Corporation
Supervisors:
Professor Aarne Halme Professor Kalevi Hyyppä Helsinki University of Technology Luleå University of Technology
Instructor:
Erik Nordin
Volvo Technology Corporation
Preface
This work nalizes my trajectory within the SpaceMaster program. The work, however, focus on terrestrial applications of intelligent vehicles. In this way, I combined two of my passions.
For this work, I have to thank to Erik, for this fantastic months at Volvo, speaking from the professional and personal sides. You, Johan and Andreas, were also direct responsibles for exactly the same. However, all my colleages there had their contribution to my great time in Gothenburg.
In the other side, I should also acknowledge the academia. They were two special and busy years. I appreciate specially my student colleages and the sta of the SpaceMaster, which assisted me to go abroad for the thesis.
Finally, the most important: my family. You gave me the support to keep my studies and projects, as well as to overcome these di cult months we are passing.
Espoo, August 18, 2009
Diego F. Boesel
ii
I dedicate this work to my father, which still feels like present everyday.
Diego
iii
Helsinki University of Technology Abstract of the Master's Thesis
Author: Diego Fernandes Boesel
Title of the thesis: Prediction of Vehicle Trajectories with Map Data for Cooperative Systems
Date: August 18, 2009 Number of pages: 92
Faculty: Faculty of Electronics, Communications and Automation Department: Automation and System Technology
Program: Master's Degree Programme in Space Science and Technology Professorship: Automation Technology (Aut-84)
Supervisors: Professor Aarne Halme (TKK) Professor Kalevi Hyyppä (LTU) Instructor: Erik Nordin (VTEC)
Cooperative systems are being investigated to increase road safety. Here, vehicles and road infrastructure will establish a wireless network to communicate safety-related in- formation. A vehicle will obtain a perception of the environment extended in time and space, when compared with the perception achieved by the driver or by the current active safety systems. Applications, such as support to lane keeping and frontal colli- sion warning, will be able to provide an earlier anticipation of hazards. For that, the predicted trajectories of vehicles should span longer time ahead than todate.
This work aimed to develop an algorithm to predict trajectories of vehicles for cooper- ative systems. The speci c objectives were the prediction up to 5s ahead of (1) vehicle trajectories in straight and curved roads of one lane, and (2) vehicle route in urban junctions. The algorithm should be suitable for any vehicle in a cooperative system.
It was implemented an algorithm that fuses trajectories based on the vehicle motion and on the vehicle motion constrained to the road. An innovative model to predict trajectories of vehicles based on their motion was proposed. The model is the most accurate and among the fastest, when compared with models in the literature. In addition, a breakthrough algorithm was implemented to predict vehicle route in urban junctions. The algorithm achieves a high rate of correct predictions up to 5s ahead of the junction, while keeping the rate of false predictions near zero.
Prediction of trajectories were analyzed in dierent driving cases. Finally, the algo- rithms implemented were demonstrated in a cooperative scenario. Data was acquired on real time with vehicles that were exchanging information during the drive. The algorithm runs o ine, but no design change is needed for real time execution.
Keywords: prediction vehicle trajectories map cooperative systems iv
Contents
1 Introduction 1
1.1 Preface . . . 1
1.2 Motivation . . . 4
1.3 Objective . . . 5
1.4 Application in Space . . . 6
1.5 Report Structure . . . 7
2 Background in Cooperative Systems, Vehicle Trajectories, and in Prediction of Vehicle Trajectories 9 2.1 SafeSpot Project . . . 9
2.1.1 Cooperative Systems Requirements . . . 10
2.1.2 Innovative Technologies in SafeSpot Integrated Project (SafeSpot) . . . 11
2.1.3 Architecture of SafeSpot Vehicles . . . 13
2.2 Trajectories of Vehicles . . . 15
2.2.1 Vehicle Factors . . . 15
2.2.2 Environment Factors . . . 16
2.2.3 Driver Factors . . . 16
2.3 Prediction of Trajectories of Vehicles . . . 17
2.3.1 Applications using Future Trajectories of Vehicles . . . . 17
2.3.2 Use of Trajectory Factors for Trajectory Prediction . . . 18
2.4 Summary of Trajectories and Prediction . . . 24
3 Preliminaries to the Algorithm Developments 26 3.1 Proposed Algorithm . . . 26
3.2 Reference Frames . . . 28
3.3 Platforms for Acquisitions and Tests . . . 30 v
3.3.1 Vehicles . . . 30
3.3.2 Notation . . . 31
3.3.3 Data Acquired and Calculated . . . 32
3.4 Typical Driving Scenarios . . . 33
3.4.1 Driving Case - Straight Road Section with Free Flow . . 34
3.4.2 Driving Case - Curved Road Section . . . 35
3.4.3 Driving Case - Road Section with Obstructions . . . 35
3.5 Error Calculation . . . 36
4 Motion-based Prediction of Trajectories 39 4.1 Motion-based Prediction of Vehicles Trajectories . . . 39
4.2 Models for Vehicle Motion . . . 40
4.2.1 Classic Bicycle (Classic Bicycle (CB)) . . . 41
4.2.2 Constant Acceleration (Constant Acceleration (CA)) . . 41
4.2.3 Constant Turn Rate (Constant Turn Rate (CTR)) . . . . 42
4.2.4 Constant Turn Rate and Acceleration (Constant Turn Rate and Acceleration (CTRA)) . . . 43
4.2.5 Kinematic Unicycle (Kinematic Unicycle (KU)) . . . 43
4.2.6 Kinematic Unicycle - Extended (Kinematic Unicycle - Extended (KU-X)) . . . 44
4.2.7 Kinematic Unicycle - Extended and Modi ed (Kinematic Unicycle - Extended and Modi ed (KU-XM)) . . . 44
4.2.8 Kinematic Unicycle - Extended and Generalized (Kine- matic Unicycle - Extended and Generalized (KU-XG)) . 45 4.2.9 Kinematic Bicycle (KB) . . . 45
4.3 Comparison of Performance of Dierent Models for Motion-based Prediction . . . 45
4.4 Algorithm and KU-X Model Optimization . . . 47
4.4.1 Heading Calculation . . . 47
4.4.2 Insertion of Yaw Acceleration . . . 48
4.4.3 Insertion of Jerk . . . 48
4.4.4 Switching Models . . . 48
4.4.5 Insertion of Error in the Prediction . . . 49
4.5 Analysis of the Motion-based Prediction of Trajectories . . . 49 vi
4.5.1 Error Analysis . . . 50
4.5.2 Real Time Prediction of the Error . . . 51
4.5.3 Sensibility to the Observations Accuracy . . . 54
5 Prediction of Trajectories Aided by Map Data 56 5.1 Introduction . . . 56
5.2 Static Map . . . 57
5.3 Prediction of Vehicle Route . . . 59
5.3.1 Performance of the Prediction of Vehicle Route . . . 61
5.4 Road-based Prediction of Vehicle Trajectories . . . 63
5.4.1 Performance of the Road-based Prediction of Trajectories 64 5.4.2 Limitations of the Road-based Prediction . . . 67
5.5 Fused Prediction of Vehicle Trajectories . . . 70
6 Evaluation and Discussion 72 6.1 Evaluation of the Algorithms - Introduction . . . 72
6.2 Evaluation of the Prediction of the Vehicle Route . . . 73
6.3 Evaluation of the Fused Prediction of Vehicle Trajectories . . . . 74
6.3.1 Case Urban Streets . . . 75
6.3.2 Case Highway . . . 76
6.3.3 Case Limitation: Curves . . . 77
6.4 Application in Cooperative Systems . . . 79
6.4.1 Onboard and Remote Prediction of the Trajectory for a Vehicle . . . 80
6.5 Limitations . . . 82
6.6 Future Work . . . 82
7 Summary and Conclusions 85
References 86
vii
List of Tables
2.1 Theoretical Analysis of Vehicular Trajectories and its Prediction. 25 3.1 Characterization of the Base Acquisition and its Characteristics
Road Sections. . . 38 4.1 Comparison of Performance of Models of the Vehicle Motion in
Prediction of Trajectories. . . 46 4.2 RMS Error [m] in the Base Acquisition and its Characteristics
Road Sections. . . 51 4.3 RMS Error [m] in the Motion-based Prediction for the Car and
Truck in the same Road. . . 55 5.1 Algorithm for the prediction of the vehicle route in junctions of
urban roads. . . 59 5.2 Frequency of prediction of vehicle route in dierent horizons in
straight driving (SD) and in round-about (RA) scenarios in ur- ban areas. . . 63 5.3 Algorithm for the road-based prediction of trajectories. . . 64 5.4 Algorithm for the fusion of the motion-based and road-based
prediction of trajectories. . . 71 6.1 Frequency of prediction of vehicle route in dierent horizons. . . 74
viii
List of Figures
1.1 View of the environment perceived by automated vehicles (Re-
produced from (Han et al., 2007)) . . . 3
2.1 Dierent views in the Local Dynamic Map (LDM) (Reproduced from (SafeSpot, 2009)) . . . 13
3.1 Architecture of the proposed algorithm . . . 27
3.2 Reference frames adopted (exagerated): Geodetic, Local East, North, Up Coordinate system (ENU), Vehicle-attached, and Road- attached . . . 28
3.3 Vehicle eet utilized in this work for data acquisition and tests: Volvo FH16 truck (in the left of the gure) and Volvo S80 car (in the middle). The Volvo FH16 truck (in the right) was left as spare. . . 31
3.4 Vehicle trajectory encompassing three typical driving scenarios in particular road sections . . . 34
3.5 Trajectory in urban straight section with free ow . . . 35
3.6 Trajectory in urban curved section . . . 36
3.7 Obstacles and obstructions in the city street section of the base acquisition . . . 37
3.8 Trajectory in urban straight section with obstructions . . . 37
4.1 Trajectory Prediction with KU-X model . . . 50
4.2 Error per horizon of prediction using KU-X model . . . 52
4.3 Prediction error using KU-X model along a croped section of the base acquisition . . . 52
4.4 Vehicle acceleration during the section of the trajectory in Figure 4.3 . . . 53
ix
4.5 Anticipation of prediction error using KU-X model along a cropped section of the base acquisition . . . 54 4.6 Vehicles hardware comparison . . . 55 5.1 Vehicle trajectory and the position of the rst correct prediction
of the vehicle route in the junctions 5s ahead of the vehicle . . . 62 5.2 Prediction of trajectories by the motion- and road-based ap-
proaches in an urban area . . . 65 5.3 Components of the error of prediction by the motion-based and
road-based approaches for a trajectory in an urban area . . . 66 5.4 Lateral error of prediction by the motion-based and road-based
approaches for a straight driving . . . 67 5.5 Lateral error of prediction by the motion-based and road-based
approaches for a trajectory in a curved section in an urban road 67 5.6 Inaccurate representation of the road geometry by the road cen-
terline of the map . . . 69 5.7 Inaccurate representation of the trajectory of a vehicle in a cor-
ner by the road-based prediction of trajectories algorithm . . . . 70 6.1 Vehicle trajectory in an urban area and the position of the rst
prediction of the vehicle route in the junctions 5s ahead of the vehicle (vehicle travelling in the direction bottom-up in the gure) 73 6.2 Snapshots of Trajectories Predicted with Dierent Approaches . 75 6.3 Evaluation in an Urban Street Section of the Dierent Approach-
es to Predict Trajectories . . . 76 6.4 Evaluation in a Highway Section of the Dierent Approaches to
Predict Trajectories . . . 77 6.5 Snapshots of Trajectories Predicted with Dierent Approaches
in an Urban Street Curve . . . 78 6.6 Evaluation in an Urban Street Curve of the Dierent Approaches
to Predict Trajectories . . . 78 6.7 Snapshot in Dierent Situations of the Environment Perceived
by Ego Vehicle in Cooperative Systems . . . 80
x
6.8 Error of the predicted trajectory of a vehicle, when the calcula- tion is performed in the own vehicle (Prediction as Ego Vehicle) and in other vehicle in the area (Prediction as Obstacle Vehicle). 81
xi
Symbols and Abbreviations
ADAS Advanced Driver Assistance Systems AI Arti cial Intelligence
CA Constant Acceleration CAN Controller Area Network CB Classic Bicycle
CT Coordinated Turn
CTR Constant Turn Rate
CTRA Constant Turn Rate and Acceleration CWS Collision Warning System
DOF Degree-of-Freedom ECU Electronic Control Unit
ENU Local East, North, Up Coordinate system GNSS Global Navigation Satellite System
GPS Global Positioning System HMI Human-Machine Interface HMM Hidden Markov Model
ISO International Organization for Standardization xii
KB Kinematic Bicycle
KF Kalman Filter
KU Kinematic Unicycle
KU-X Kinematic Unicycle - Extended
KU-XG Kinematic Unicycle - Extended and Generalized KU-XM Kinematic Unicycle - Extended and Modi ed LAN Local Area Network
LDM Local Dynamic Map
OEM Original Equipment Manufacturer RMS Root Mean Square
SAE Society of Automotive Engineering SafeSpot SafeSpot Integrated Project
SI International System of units SVM Support Vector Machine TLC Time To Lane Change TTE Time To Establish
UTM Universal Transverse Mercator UWB Ultra Wide Band
V2I Vehicle-To-Infrastructure V2V Vehicle-To-Vehicle
WLAN Wireless Local Area Network
xiii
Chapter 1 Introduction
1.1 Preface
Automobiles have become an essential element of the society we live in today.
Automobiles were introduced in the end of the 19th century as the evolution of horse-drawn carriages. They were well accepted and extensively adopted.
Collisions are one of the major drawbacks of the use of automobiles. Automo- tive collisions lead to signi cant losses, including a expressive quantity of lives as well as nancial costs.
A number of technologies were adopted to mitigate the eects of accidents on the passengers of vehicles. In the automotive industry, technologies with this purpose are referred to as passive safety systems. Products of these technologies can comprise devices, such as seat belt, as well as special vehicle materials and geometry design, e.g., for the vehicle chassis and structure.
Later, passive safety systems incorporated electronic parts. These technologies are called here electronic safety systems, althought sometimes they are not regarded as safety systems. They can both contribute to mitigate the eects of accidents, involving devices as the air-bag, and also to assist the vehicle control, including systems as Anti-lock Braking System (ABS), Electronic Brakeforce
1.1 Preface 2
Distribution (EBD), Electronic Stability Control (ESC), as well as the rst cruise control systems.
Over the last decades, technologies to prevent accidents have received an in- tense development eort. Such technologies are called active safety systems (shang Liu and Peng, 1996). Typically, active systems obtain information of the environment through sensors and process it. This is in contrast to all ear- lier safety systems, where perception was basically limited to the vehicle states and its subsystems. If the active system anticipates a forthcoming harzard sit- uation, the driver is alerted through a Human-Machine Interface (HMI). This concept is implemented, for example, in parking sensors.
The accumulation of a number of safety systems in the vehicles originated the term Advanced Driver Assistance Systems (ADAS). Currently, vehicles feature a variety of safety functionalities (Daimler, 2008; Toyota, 2009; Volvo, 2009).
These functionalities dier signi cantly in a number of aspects. Many of the systems are not visible to the driver. Further, typically, each system leads to a dierent eect. Some of the systems support the driver in the basic vehi- cle guidance operations (Polychronopoulos et al., 2004). On the other hand, others units augment the driver capabilities by providing perception of the en- vironment or control of the vehicle that can not be achieved by the human driver (Polychronopoulos et al., 2004). Hence, the combined bene ts of the safety systems provide to the driver a sophisticated support to guide the vehi- cle. Additionally, these units are relatively integrated in terms of information exchange, especially sensing data (Polychronopoulos et al., 2004). Because of the combined bene t and relative integration, safety systems are now referred as an unique system.
Further, ADAS systems incorporated a number of innovations. An ellaborated representation of the environment is acquired through the perception of several properties of the vehicle surroundings (Polychronopoulos et al., 2004; Weiss et al., 2004; Eidehall et al., 2007). This is achieved by dierent sensors, which typically include cameras, infrared cameras, radars, and lidars (Polychronopou- los et al., 2004; Ferguson et al., 2008). Units of software extract information of
1.1 Preface 3
the environment from data of these sensors by the use of complex algorithms (Polychronopoulos et al., 2004; Ferguson et al., 2008). Hence, in order to per- mit the execution in real time of these algorithms, more powerfull processing units were also introduced (Polychronopoulos et al., 2004). Additionally, some ADAS systems provide navigation assistance to the driver with the utilization of maps and acquisition of the vehicle geodetic coordinates with the Global Positioning System (GPS) (Obradovic et al., 2007). Vehicles featuring ADAS systems are frequently called automated cars or intelligent vehicles (Ferguson et al., 2008).
Figure 1.1: View of the environment perceived by automated vehicles (Repro- duced from (Han et al., 2007))
A similar evolution of concepts is also seen among the road-based safety tech- nologies. Static approaches, such as signalization and the design of roads for speci c purposes, were the rst to be implemented to increase the tra c safety.
Later, these approaches were extended with the use of electric, electronic and optical devices, e.g., tra c lights and cat's eye, as well as construction of roads segments with ellaborated geometrical design and selection of materials, just to name a few technologies. Recently, sophisticated roads became adaptative.
They can actively change attributes, such as speed limit, tra c lights sequence, and direction of ow, according to environmental and tra c conditions. Road side signalization provides to the driver relevant instantaneous information, in- cluding these tra c changes. Some of these roads are connected with safety centrals, which typically monitor a network of roads.
1.2 Motivation 4
Some research projects currently under way on vehicular safety adopt the con- cept of cooperative safety systems. These systems converge vehicle and road- based active systems. The foundation of the cooperative systems is the commu- nication among vehicles traversing the same area (Caveney, 2007). The com- munication is established through an ad-hoc wireless network (SafeSpot, 2009).
In this network, each vehicle sends information such as its precise dynamic state (Caveney, 2007) as well as relevant status of the environment obtained through the vehicle sensors (SafeSpot, 2009). The advantages of the system are further improved inserting the road infrastructure in the communications (Tan and Huang, 2006; Caveney, 2007).
Cooperative systems provide important bene ts over autonomous-based ADAS systems vehicles. Typically, the network will supply more precise information on other vehicles states than if it was obtained with onboard sensors (Tan and Huang, 2006). Moreover, through the network will be possible to obtain information beyond the line-of-sight of the vehicle sensors (Tan and Huang, 2006) and through occluding objects (Lytrivis et al., 2008). As a result, these cooperative systems extend vehicle and, typically, also driver awareness of the environment, both in time and in space (SafeSpot, 2009).
1.2 Motivation
The prediction of the trajectory of vehicles is used in a number of driver as- sistance systems (Lin et al., 2000; Huang and Tan, 2006). The software unit of these vehicular applications computes possible positions of the vehicle in a future time frame (Huang and Tan, 2006). This information is employed in the calculations of the particular purpose of the assistance system. Lane keep- ing/departure assistants, for example, compare the vehicle trajectory with the geometry of the road (Lin et al., 2000), whereas the collision warning system analyzes the own vehicle trajectory with others vehicles trajectories (Tan and Huang, 2006).
With cooperative systems, the prediction of trajectories can span a longer time
1.3 Objective 5
horizon ahead. In previous autonomous-based ADAS systems, the line-of-sight of onboard sensors permited the projection of trajectories a few seconds ahead only (Polychronopoulos et al., 2004, 2007). Through the communication with other vehicles and with the road infrastructure, a vehicle acquires a perception of the environment that is extended in time and in space. To obtain full bene t of this extended perception, the time span of the trajectories predicted can be prolongated to several seconds ahead. The longer time span of the prediction favors an early detection of hazardous situations. Therefore, it contributes to the e cacy and smoothness of the driver assistance systems (Lidstrom and Larsson, 2008).
Nonetheless, a prediction with a long time frame should cover the impact of a larger range of elements in the trajectory of a vehicle. The vehicle dynamics is the dominant origin of the motion of a vehicle in short time frames (Mad- havan et al., 2006). However, in long time frames, both the driver actions as well as road tra c and environmental factors also become important sources (Madhavan et al., 2006).
1.3 Objective
The goal of this project is to develop an algorithm to predict trajectories of vehicles. The algorithm should take bene t of the extended perception of the environment provided in cooperative systems. Conversely, it has to comply with the features needed in these systems. These general statements result in the following speci c objectives:
O1 Develop an algorithm to predict trajectories of vehicles in straight roads of one lane
O2 Develop an algorithm to predict trajectories of vehicles in curved roads of one lane
O3 Develop an algorithm to predict vehicle route in urban junctions
1.4 Application in Space 6
The objectives above shall be achieved meeting the following requirements:
R1 A new predicted trajectory should be generated at each 0.1s,
R2 The predicted trajectory should span 5s ahead of the prediction instant.
A position at each 0.1s ahead is required.
R3 The algorithm should be applied generically to any vehicle in the cooper- ative system,
R4 A con dence level should be stated for the trajectory,
R5 Algorithms have to be implemented in Matlab/Simulink, in Windows en- vironment.
Further, the development work carried out in this project assumed the following situation:
A1 Real-time issues are not a priority, however, they should be observed.
A2 Road geometry and topology are available to the algorithm.
A3 Submeter positioning on the road reference frame are also available.
The project can be understood as a proof-of-concept. It is one of the pioneer works in the integration of trajectory and route prediction. Further, it is also one of the few works that incorporates road information with the aim to pre- dict trajectories. The developments are done from scratch, i.e., no previous development work will be continued. Hence, the project does not intend to demonstrate or optimize a technology for production, but only to produce an understanding of its principles.
1.4 Application in Space
Current missions for space robotic rovers encounter an environment which is shaped by the natural geography of the surface of the planet in exploration.
1.5 Report Structure 7
These missions are limited in basic geology and astronomy science, as well as engineering proof-of-concept of the vehicles (Biesiadecki and Maimone, 2006;
Bajracharya et al., 2008).
The high autonomy achieved by rovers will permit that they obtain ellaborated missions in the future. Robotic missions will be extended from those early exploration duties to site preparation for human presence, and, even further, to support long duration habitation (Bradley et al., 2005). Tasks in these future missions will include construction, assembly, inspection, maintenance, reconnaissance, and transportation (Mankins, 2004; Bluck, 2005; Stroupe et al., 2005; Fong et al., 2008). These tasks will be accomplished by a number of robots, which will estabilish teams and work in cooperation (Mankins, 2004;
Bradley et al., 2005). In addition, crewman and robots will work in the same site simultaneously, and, eventually, cooperate one to the other (Mankins, 2004;
Bluck, 2005; Stroupe et al., 2005; Mishkin et al., 2007; Fong et al., 2008).
The presence of crewman and other vehicles will raise the need to predict tra- jectories to navigate autonoumosly and safely. Currently, obstacles for rovers are rocks, slippery terrain, among other static hazards. In the future, however, objects such as astronauts and vehicles may change their position over time.
Hence, in order to avoid collision, it will be necessary to determine their tra- jectory. This problem of avoiding collision with moving obstacles in planetary vehicles resembles the same problem in automotive vehicles.
1.5 Report Structure
The remainder of this document is organized into Chapters. Chapter 2 in- troduces the literature review for the SafeSpot system as well as for vehicular trajectories and their prediction. In Chapter 3 is introduced the methodology and platforms utilized in this project. This chapter also provides an analysis of vehicle trajectories from acquired vehicle drive data. Chapter 4 adresses the development and test of the motion-based algorithm to predict vehicle trajecto- ries. Chapter 5.1 introduces the algorithms developed in this work that utilizes
1.5 Report Structure 8
map data: road-based prediction of trajectories, fused prediction of trajecto- ries, and the prediction of vehicle route in urban junctions. Chapter 6 presents the evaluation of the fused prediction of trajectories, the prediction of vehicle route in urban junctions, as well as the application of the fused prediction in cooperative systems. In that chapter, it is indicated the limitations of this work and possible directions for future work. Finally, conclusions drawn from this project are given in Chapter 7.
Chapter 2
Background in Cooperative
Systems, Vehicle Trajectories, and in Prediction of Vehicle
Trajectories
This chapter introduces concepts utilized in this project which were taken from the literature. Firstly, cooperative systems will be introduced through the de- scription of the SafeSpot project. The description will focus on those concepts that directly aect the development of this project. Later, the trajectories of vehicles and prediction of these trajectories are presented through a uni ed view.
2.1 SafeSpot Project
The SafeSpot project is a research work that develops cooperative systems for road safety (SafeSpot, 2009). Here, cooperative systems include intelligent vehicles, road infrastructure. Each unit will be able to identify potentially dangerous situations and then forward this information to the other elements
2.1 SafeSpot Project 10
in the system (SafeSpot, 2009). This scheme will make it possible to augment the drivers' awareness of the surrounding environment in space and in time.
SafeSpot is an integrated research project carried out by a consortium involving several segments of the European transportation industry (SafeSpot, 2009). It is co-funded by the European Comission. The consortium is constituted by ap- proximately 50 partners throughout Europe (SafeSpot, 2009). They represent dierent segments of the transportation industry, including Original Equipment Manufacturer (OEM), suppliers, road operators, and research centers. Several global players are among the partners of the project, such as Fiat, Daimler, Volvo, Bosch, Renault, and Magneti Marelli (SafeSpot, 2009). Projects that investigate similar concept comprehend the Crash Avoidance Metrics Partner- ship (CAMP), in North America, and the Advanced Safety Vehicle Project (ASP), in Japan (Caveney, 2007).
The remaining part of this section will be based on (Brignolo et al., 2008), unless otherwise stated.
2.1.1 Cooperative Systems Requirements
The cooperation of systems is a new a concept to the transportation and auto- motive industry (SafeSpot, 2009). Up to date, road safety has been developed with autonomous-based safety systems in intelligent vehicles. By establishing Vehicle-To-Vehicle (V2V) and Vehicle-To-Infrastructure (V2I) communication, the concept of road safety moves from autonomous vehicles to cooperative sys- tems (SafeSpot, 2009).
To achieve eective bene ts, cooperative systems should operate with new re- quirements. Currently, each vehicle obtain a position and time reference from onboard sensors (Huang and Tan, 2006). Absolute position and a master clock will be mandatory for safety applications running distributively on multiple vehicles and infrastructure (Tan and Huang, 2006; Caveney, 2007). This will eliminate di culties arising from the use of dierent reference frames. The requirement is adressed by the adoption of the Global Navigation Satellite
2.1 SafeSpot Project 11
System (GNSS), such as GPS. This system provides a common Earth inertial reference frame both for space and time (Tan and Huang, 2006; Caveney, 2007).
Further, the technology for intervehicle on-the- y communication is under de- velopment. Currently, this communication possesses a limited reliability, qual- ity and bandwidth. Owing to these limitations, messages broadcasted by vehi- cles should be small. They contain vehicle estimates of its own states with the respective timestamps (Tan and Huang, 2006). With the current communica- tion technology, however, the projection of the vehicle future trajectory will likely not be transmitted.
2.1.2 Innovative Technologies in SafeSpot
The extended use of telematic technologies, such as telecommunication, infor- matics, and GNSS, constitute the foundation of three innovative technologies in SafeSpot. They will be shortly introduced in the coming subsections.
Wireless Ad-hoc Dynamic Network
The rst innovative technology is a wireless ad-hoc dynamic network. This network will provide a communication channel in which the cooperating ele- ments will exchange information. Vehicles and infrastructure will constitute the nodes in this network. Typically, messages with information classi ed as
always useful will be transmited to direct neighbors by a simple broadcast mechanism at a 2-Hz rate. The header of this message will contain informa- tion such as node identi er, position, timestamp, velocity, and heading of the transmitting node. Additional elds in the message can provide information about node status or node environment.
2.1 SafeSpot Project 12
Sub-meter Positioning Relative to the World Coordinates
The second innovative technology is the positioning relative to the world coor- dinates. The positioning is required to feature sub-meter accuracy and preci- sion. In addition, it requires such a high level of accuracy also for heading and parameters related to the motion of the vehicle. This highly accurate position- ing will be obtained by the fusion of information acquired by several dierent techniques, including:
1. measurements from GNSS systems
2. measurements from inertial sensors, such as accelerations and angular velocity
3. map matching algorithms
4. track with laser scanner of landmarks with known position 5. track with camera of landmarks with known position
6. measurement of Time Dierence of Arrival (TDOA) from Ultra Wide Band (UWB) and Wireless Local Area Network (WLAN) signals
7. relative positioning of two or more vehicles
The fusion of the rst three technologies constitutes the core method of obtain- ing the position of the vehicle. It should typically be always available. The last four techonologies work in parallel to the core positioning method by improv- ing the position yielded by the core method. Nonetheless, these enhancement technologies are not always available, as a consequence of speci c requirements for their operation.
Local Dynamic Map (LDM)
The LDM is the third innovative technogy introduced in SafeSpot. It is a database which stores information acquired through the in-unit (vehicle or
2.1 SafeSpot Project 13
infrastructure) sensors and through the ad-hoc network. By storing this infor- mation, the LDM becomes the view of the world from the perspective of the unit. The database also retrieves information for other SafeSpot components.
As a result, the LDM becomes a central repository of information regarding the physical environment around the unit. Figure 2.1 provides a graphical representation of the concepts implemented in the LDM.
Figure 2.1: Dierent views in the LDM (Reproduced from (SafeSpot, 2009)) All objects in the database are associated to a geometry. In the LDM, spatially static objects are referred as features. Features may include the road network, tra c lights and signs, landmarks, as well as SafeSpot road-side units. In other side, dynamic objects refer to moving objects. This encompasses objects moving on or near the roads, such as vehicles and pedestrians. In addition, dynamic objects can also represent environmental or weather related objects, e.g., regions of fog, ice, heavy congestion, as well as road works.
2.1.3 Architecture of SafeSpot Vehicles
Vehicles in SafeSpot are equipped with a number of devices to acquire, process and communicate information. These vehicles are called within the project as intelligent or SafeSpot compatible vehicles. In-vehicle sensing comprehends information acquired mostly from o-the-shelf devices, i.e., sensors that are currently already installed in the cars during the production. Exceptions to this model may include the positioning system, laser scanner and radar.
2.1 SafeSpot Project 14
There will be a number of vehicular platforms in use during the project, such as Volvo trucks, Fiat, Renault, and other OEM passenger cars as well as other vehicles. These platforms dier in many perspectives, in aspects such as phys- ical dimensions, space available for SafeSpot hardware, and the availability of electrical power. However, the principal dierence in the project is the het- erogeneity in the sensor con guration and availability. Therefore, the concepts adopted in SafeSpot should be exible enough to cope with this dierence.
The hardware architecture consists basically of multiple Personal Computer units (PCs). The use of PCs provides the rapid prototyping of the development work, in opposite to the use of Electronic Control Unit (ECU), which would yield optimization of high volume product. A Local Area Network (LAN) con- nects the PCs in the vehicular platform. Data from the o-the-shelf sensors are accessible via a gateway, while a router supplies data from the wireless ad-hoc network. The hardware platform consists of the following main components:
• OEM Gateway - it provides the data available in the vehicle data network (typically a Controller Area Network (CAN) bus) and sensors, such as radars, lidars, and cameras, for the remaining sub-systems of the SafeSpot platform. The gateway is a logical and physical separation of the vehi- cle data network, which is typically proprietary, to the standard de ned within SafeSpot for its sub-systems
• Main Platform PC - this PC hosts the data fusion module and the LDM
• Application and HMI PC
• Positioning System - ideally, a separate PC runs the positioning system due to the high computational cost of tracking landmarks with camera
• Router - this device performs the communication with the other elements in the SafeSpot cooperative system
• Laser Scanner PC
• Esposytor PC - this PC executes an unobtrusive monitoring of the plat- form and its subsystems
2.2 Trajectories of Vehicles 15
• Switch - this device connects the hardware via an Ethernet LAN
As with the sensing architecture, the hardware architecture may vary consid- erably among the dierent vehicular platforms. It depends specially on the availability of physical space, electrical power, and installed sensors.
2.2 Trajectories of Vehicles
The trajectory of a vehicle can be regarded as the sequence in time of the vehicle positions (Huang and Tan, 2006; Lytrivis et al., 2008). Therefore, trajectory is a spatio-temporal concept (Spaccapietra et al., 2008), a notion which contrasts with a path, which is exclusively a spatial concept.
The trajectory performed by a vehicle is an intricated result of many factors.
In the literature, factors in uencing a trajectory are typically classi ed into three groups: vehicle, environment and driver (Lidstrom and Larsson, 2008).
The classi cation is merely academic and, with certain frequency, factors can not be clearly assigned to a single group. A short description of each of them is introduced in the following paragraphs.
2.2.1 Vehicle Factors
This group embraces the physical interactions of a moving body, i.e., a vehicle in this case, with the environment. These interactions arise primarily from the inertial characteristics of the vehicle. The vehicle inertia is the most signi cant contribution for the short term trajectory of the vehicle (Madhavan et al., 2006;
Lytrivis et al., 2008).
2.2 Trajectories of Vehicles 16
2.2.2 Environment Factors
The environment group comprises aspects that constrain the possible motion in the environment where the vehicle and driver are passing. This includes a broad range of aspects. The aspects could be man made, such as the road topology, or natural, e.g., the presence of fog. They can also be either static, including position of crossings, or dynamic, such as road accidents signalized by the local road tra c authorities. They dier as well between visible aspects, e.g., tra c lights, and implicit factors, which include tra c conventions. The in uence of these factors on the vehicle trajectory varies signi cantly. However, they typically aect the trajectory of a vehicle in the longer term, when compared with the vehicle factors (Lytrivis et al., 2008).
2.2.3 Driver Factors
Within the driver group lie factors related with the intentions, actions, behav- ior, and skills of the person driving a vehicle. These human factors are subject to numerous research works in the literature. A few conclusions from these studies that are relevant for this project will be introduced here. Nevertheless, a deep coverage of these aspects is beyond the scope of this project.
Factors might feature coupled characteristic. This implies that their eects depend on or aect other elements. In general, factors from the environment group can impel changes both to driver factors, e.g., reduced visibility due to fog, as well as vehicle factors, such as icy roads. Likewise, driver factors do in uentiate vehicle factors, as will be shown later. In contrast, the vehicle factors may aect the driver, but in a weaker form. Similarly, driver does not change the environment properties, but he/she changes the environment where the vehicle traverses.
2.3 Prediction of Trajectories of Vehicles 17
2.3 Prediction of Trajectories of Vehicles
2.3.1 Applications using Future Trajectories of Vehicles
The anticipation of the trajectory that a vehicle will follow nds application for numerous purposes. Within the vehicle, it can be used for confort and safety (Lin et al., 2000; Huang and Tan, 2006; Lytrivis et al., 2008). For infrastructure, such information can be adopted for an improved dynamic regulation of the road network (Chen et al., 2006). The actual list of use is wide and goes far beyond the functions listed in this paragraph. Moreover, new applications are under development. Below, three applications will be shortly introduced as an example.
Collision Warning System (CWS)
An automotive collision is yielded by the intersection or con ict in space and in time of two or more vehicle trajectories (Tan and Huang, 2006). CWS anticipates such critical situations. It calculates the distance among the own vehicle trajectory and others vehicles in the its surroundings, typically referred to as obstacles (Polychronopoulos et al., 2004). This calculation spans a certain time frame (Tan and Huang, 2006). If the distance calculated to a certain obstacle is below a safety threshold, and this persists for a certain amount of time, the potential for collision is high. In this case, the vehicle either warns the driver through an HMI or automatically drives around the critical region (Tan and Huang, 2006).
Lane Keeping/Lane Change Systems
Lane Keeping and Lane Change systems diverge in their purpose. Nonetheless, their working principle is similar in many aspects. They obtain an estimation of the lane in which the vehicle lies (Lin et al., 2000). Afterwards, it is cal- culated the vehicle short term future trajectory. With this information, the
2.3 Prediction of Trajectories of Vehicles 18
Time To Lane Change (TLC) metric is calculated (Lin et al., 2000). The TLC metric is used then to infer whether the driver should be assisted with the Lane Keeping or Lange Change systems. Further information on these systems can be found in the literature (shang Liu and Peng, 1996; Mammar et al., 2006;
Polychronopoulos et al., 2005; Eidehall et al., 2007).
Advertisement
There are ongoing research projects that apply trajectories for commercial pur- poses. The infotainment system of the car will tailor advertisements according to the predicted trajectory of the car. This will expose the driver to the business in the area where the vehicle will pass.
2.3.2 Use of Trajectory Factors for Trajectory Prediction
The future trajectory of a vehicle is typically approximated with a prediction of the trajectory. The prediction yields possible positions of the vehicle in a future time horizon.
The factors in uencing trajectories of vehicles are frequently employed to pre- dict the trajectories. Some of these elements are dynamic, i.e., their state changes over time. This suggests that they aect trajectories in a dierent manner in dierent time instants. Hence, their instantaneous state should be obtained in real time. Conversely, there are factors that remain static, such as road geometry and topology. Thus, from these factors can be obtained a priori by the algorithm and be stored for later usage. An important dierence among the elements regards their continuity. It involves both continuously changing elements either in time or in space as well as elements that change at speci c instants or positions. The former includes road geometry and vehicle position, while the latter is illustrated by tra c lights and turning indicators. Likewise, factors can be directly measured, including most vehicle properties such as speed and yaw rate, whereas others are not measurable, e.g. humans factors as
2.3 Prediction of Trajectories of Vehicles 19
intention to turn or driving skills. Nonetheless, the latter factors are frequently estimated.
The literature describes numerous methods to predict the trajectories of ve- hicles. The importance and complexity of this task might explain this mul- tiplicity. In addition, the methodologies described feature dierent purposes.
Some are designed for the subject vehicle, while others may apply for obstacle vehicles. Moreover, algorithms can be proposed for tracking problems or for prediciton problems. They might employ explicit knowledge and models or exploit a model of the solution, e.g., Arti cial Intelligence (AI).
The following subsections introduces the approaches to predict trajectories found in the literature. The discussion is intended to give a general view, however some speci c matters are included. Algorithms designed for track- ing/navigation problems will not be dierentiated from methods for prediction.
Vehicle
Measurable vehicle factors include kinematics properties such as vehicle accel- erations, speed, and yaw rate.
A second group of measurements can be obtained in the vehicle. This group comprehends the state of the input controls of the vehicle, such as steering wheel angle and percentage of acceleration pedal usage. These parameters require the mathematical model of the dynamics of vehicle subsystems in order to obtain the vehicle kinematics properties. These subsystems might include the vehicle engine, transmission, drivetrain, brake, and suspension (Caveney, 2007). Because these parameters are dependent on the vehicle, they are called vehicle parameters (Tan and Huang, 2006).
The use of the dynamic model of these subsystems suers from a number of drawbacks. Firstly, they feature a high-frequency dynamics, which is irrele- vant for the time horizon of the intended predicted trajectory (Caveney, 2007).
These mathematical models typically dier in dierent vehicle models. In addi-
2.3 Prediction of Trajectories of Vehicles 20
tion, the models are typically uncertain, nonlinear, and computationally heavy (Huang and Tan, 2006). Moreover, vehicle subsystems models will not be avail- able nor transmitted in the cooperative system (Tan and Huang, 2006). Beside these issues, input controls may suer small variations without aecting the vehicle motion (Huang and Tan, 2006). Owing to these causes, the use of the dynamic model of these subsystems is typically neglected in cooperative sys- tems. As a result, only direct measurements of vehicle motion parameters will provide information on the kinematic state of the vehicle.
The prediction of trajectories utilizes the vehicle inertial characteristics to prop- erly project its short horizon trajectory (Madhavan et al., 2006; Lidstrom and Larsson, 2008). This is done by employing Newtonian physics, e.g., Dynam- ics and Kinematics (Lidstrom and Larsson, 2008). Such approaches describe the motion caused by the vehicle inertia through a mathematical model (Tan and Huang, 2006; Huang and Tan, 2006; Lidstrom and Larsson, 2008). Vehicle states are projected ahead in time using this model (Tan and Huang, 2006). An alternative approach is the numerical integration of the mathematical model (Caveney, 2007).
Environment
Considering that a vehicle traversing a road will follow the tra c conventions, tra c features, such as road geometry, topology and tra c conventions, provide relevant information about a vehicle future position (Chen et al., 2006; Ferguson et al., 2008).
The employment of road geometry, e.g., road centerline, road width, road cur- vature, for trajectory prediction is described only in relatively recent litera- ture. Initially, the geometry was estimated from onboard vehicle sensors, such as cameras and radars, in the ADAS systems of vehicles (Lin et al., 2000;
Polychronopoulos et al., 2004). However, lately, road geometry is also being retrieved from digital maps, tracks of obstacle vehicles, as well as other sources (Polychronopoulos et al., 2005; Shen et al., 2006; Tsogas et al., 2007; Eidehall et al., 2007; Ferguson et al., 2008; Li et al., 2009).
2.3 Prediction of Trajectories of Vehicles 21
Tra c conventions, i.e., the set of rules that drivers are suppose to follow when utilizing the roads, are another relevant factor in the environment group. They de ne allowed and prohibited behaviors. Assuming these conventions will be followed, they permit to predict the driver intention and the vehicle motion in a number of situations (Lidstrom and Larsson, 2008).
The road tra c, i.e., the presence and ow of obstacle vehicles, possess relevant in uence in the trajectory followed by a car. Environment factors, such as distance to leading or following vehicles, aect the driver actions (Sato and Akamatsu, 2008; Lidstrom and Larsson, 2008). In addition, tra c possess typically a stochastic, dynamic and fuzzy nature (Chen et al., 2006).
Road geometry is used in combination with a trajectory predicted using the vehicle parameters. Typically, a mathematical model of the road geometry is obtained. In the rst approaches, the vehicle was constrained either to a
xed lateral position along the horizon of prediciton (Scott and Drane, 1995;
Shen et al., 2006) or to follow the heading of the road (Weiss et al., 2004).
In recent methods, conversely, the prediction with vehicle parameters is fused with the road model, either by a function of the con dence in motion-based prediction (Lytrivis et al., 2008) or by a function of the prediction horizon (Polychronopoulos et al., 2007). Vehicle motion and the road network can be explored in a integrated fashion to predict trajectories (Chen et al., 2006;
Salmond et al., 2007).
On the other hand, road topology is employed to predict the route of a vehicle (Krumm, 2008), i.e., the sequence of streets that the driver will pass. The road topology is represented through graphs. The route can be predicted ap- plying AI methods, such as learning algorithms (Krumm, 2008). Vehicle route is also predicted in a single corner utilizing an approach similar to an interact- ing multiple-model particle lter and the topology of the corner, in (Lidstrom and Larsson, 2008). In some approaches, the driver nal destination is known (Krumm, 2008). The route prediction may span a much longer time frame than a trajectory prediction.
Finally, tra c information is seldomly integrated in a trajectory prediction
2.3 Prediction of Trajectories of Vehicles 22
algorithm. It could be used to predict vehicle speed and acceleration based on a leading vehicle (Lidstrom and Larsson, 2008).
Driver
The majority of the driver factors comprises parameters that can not be mea- sured. Instead, they should be estimated, or further, predicted. Estimation of human factors is a large research area, with several applications in many elds (Lidstrom and Larsson, 2008). In the automotive eld, driver actions are in u- enced by both environmental factors, such as distance to leading or following vehicles, as well as by vehicle factors, such as the vehicle speed when breaking to handle a turn (Sato and Akamatsu, 2008).
Estimation of driver intentions uses frequently the state of the input controls of the vehicle (Huang and Tan, 2006). Typical measurements are steering wheel angle, acceleration pedal percentage, brake pedal usage, turning indicators, gear level and other vehicle controls. The history of these driver inputs can also be used to predict driver factors (Tan and Huang, 2006). Other approaches employ tra c factors of the vehicle for this prediction (Lidstrom and Larsson, 2008).
Likewise, driver typical performance can also provide information on driver factors. Driver-vehicle dynamics is comprised mostly by low frequency com- ponents, bellow 1Hz (Tan and Huang, 2006). Therefore, changes in driver intentions possess a persistent eect, and they do not change dramatically in a short time (Huang and Tan, 2006). In addition, the preparations for turning, such as indicator light on and cover the break pedal, depend on the distance to the intersection rather than the time to it (Sato and Akamatsu, 2008). Other performance metrics are related to the lane change intentions of the driver.
Lane departures typically occur with a relative yaw angle bellow 6o, whereas the driver accuracy in the lane keeping maneuver is about 20cm (Mammar et al., 2006). The lateral dynamics of the vehicle in a road is also frequently used to predict lane change/departure intention (Schön et al., 2006; Shen et al., 2006).
2.3 Prediction of Trajectories of Vehicles 23
In real situations, drivers do not possess any accurate information on these hu- man factors of the other drivers transiting on the roads (Lidstrom and Larsson, 2008). Nevertheless, the tra c ow is well coordinated (Lidstrom and Larsson, 2008). This could suggest that these elements are not fundamental to predict the motion of vehicles.
However, a human driver creates expectations on the motion of other vehicles based on the vehicle and environment factors listed above, as well as some visible driver intentions signs, such as turning indicators (Madhavan et al., 2006). In addition, these factors can be used to anticipate a maneuver intended to be executed by the driver, e.g., turn at a crossing or an intense acceleration.
Hence, driver factors can improve the precision of the trajectory predicted using vehicle and environmental factors.
Driver factors are frequently predicted with the use of AI techniques. AI meth- ods are typically applied to predict possible maneuvers that the driver will perform in the near future. Among the AI techniques, machine learning al- gorithms are largely employed in the automotive eld (Lidstrom and Larsson, 2008).
Machine learning can be used to predict the future lateral position of a vehicle on a road (Batavia et al., 1998). In addition, the identi cation of a lane change maneuver at subject and obstacle vehicle is achieved using the Support Vector Machine (SVM) (Mandalia and Salvucci, 2005; Han et al., 2007). A further application of machine learning is with the use of the Hidden Markov Model (HMM) to yield a probable trajectory that the driver will follow, which is retrieved from a learned set of trajectories (Nishiwaki et al., 2008).
The performance of machine learning methods depends on the training data.
They yield good results in general driving situations, in which it is feasible to collect data. However, the e ciency of the algorithm may degrade in situations that are di cult to acquire training data, such as accident scenarios (Lidstrom and Larsson, 2008).
Approaches to predict driver factors other than the use of AI techniques are also
2.4 Summary of Trajectories and Prediction 24
are described in the literature. The prediction of lane changes, and maneuvers in general, are also addressed with the assignment of a cost for these maneuvers (Gunnarsson et al., 2006; Madhavan et al., 2006; Ferguson et al., 2008). De- tection of lane departure or lane change of an obstacle vehicle is inferred from the error in the estimation of lateral position of these vehicles. This concept is implemented in dierent manners in practice (Weiss et al., 2004; Schön et al., 2006; Shen et al., 2006). Lane change can also be predicted by the across road speed of the vehicle (Polychronopoulos et al., 2007). Estimation of driver in- tentions can be accomplished by the tracking of some vehicle parameters that are dependent on the environment (Lidstrom and Larsson, 2008).
In practice, calculations to predict trajectories use only a subset of the factors determining a trajectory. The reason for such limitation is the likely high com- plexity and computational costs of an algorithm considering all those factors.
2.4 Summary of Trajectories and Prediction
The Table 2.1 summarizes the most relevant points of the discussion in Sections 2.2 and 2.3.
2.4 Summary of Trajectories and Prediction 25
Table 2.1: Theoretical Analysis of Vehicular Trajectories and its Prediction.
Group Vehicle Environment Driver
Trajectories of Vehicles
De nition Physical interaction of the
vehicle
Constrains to the permited and possible vehicle
motion
Driver intentions, actions, behavior, and
skills
Examples (1) Vehicle inertia (2) Vehicle skid
(1) Tra c conventions
(2) Road Geometry
(1) Lane Change Intention
(2) Skills to negotiate a curve
Eect Short Term Long Term Maneuvers
(longitudinal and lateral)
Prediction of Vehicle Trajectories
Information Sources
Measurements of Vehicle Kinematics Parameters
(1) Tracking of road and obstacles (2) Road data retrieved from map
Current and past measurements of:
(1) Vehicle Kinematics (2) Environment (3) Vehicle input controls
Approaches for Prediction
Propagation of vehicle kinematics
and dynamics
Fusion of road geometry with the
vehicle-based prediction
(1) Prediction using AI methods
(2) Track maneuvers
Chapter 3
Preliminaries to the Algorithm Developments
Chapter 3 introduces the methodology, conventions, and tools utilized in this work. In addition, it analyzes vehicular trajectories utilizing this framework.
The chapter begins with a proposal for an algorithm to meet the objectives.
Then, it introduces the reference frames and the symbols utilized in the work.
Later, the vehicles utilized are presented, together with the data that was observed during their cruises. Chapter 3 ends stating the method adopted to the calculate error.
3.1 Proposed Algorithm
The algorithm developed in this work is divided into three stages. The rst stage of the algorithm is referred to as motion-based prediction of trajectories.
At each time instant, the algorithm obtains the vehicle motion state, such as acceleration and yaw rate. A model of the vehicle motion propagates in time this vehicle motion state. This is implemented in a loop. This prediction addresses the objective O1 of the work (see Section 1.3).
3.1 Proposed Algorithm 27
The second stage utilizes road geometry to produce an independent prediction of the vehicle trajectory. This prediction is referred as road-based prediction.
At each time instant, the algorithm retrieves from the LDM the vehicle position on road and the road geometry ahead of the vehicle. With them, the vehicle position is projected in future time instants, with a simpler motion model than utilized in the motion-based approach. The projection constrains the vehicle motion to follow the road geometry. The road-based prediction should achieve the objective O2 of the work (see Section 1.3).
In junctions, the road that the vehicle will follow should be estimated in advance in order that the road-based approach can obtain the geometry of the road.
To this end, the vehicle route in junctions should be predicted. The third stage implements an algorithm that infers the driver intentions in junctions, analyzing the vehicle heading and yaw rate. Then, the algorithm predicts the vehicle route by matching the inferred driver intention with the allowed roads in the junction. The prediction of vehicle route should accomplish the objective
O3 of the work (see Section 1.3).
Finally, it is performed a fusion of the motion-based and road-based approaches to predict trajectories. The fusion was designed to combine the bene ts of each of these methods.
The Figure 3.1 shows a schematic of the architecture of the algorithm.
Figure 3.1: Architecture of the proposed algorithm
3.2 Reference Frames 28
3.2 Reference Frames
Four dierent reference frames were utilized throughout this work. Figure 3.2 shows a graphical view of the reference frames adopted in this work: (1) Geodetic system, (2) ENU, (3) Vehicle-attached system, and (4) Road-attached system. The dierent views captured by each of them can be compared in the
gure. They are shortly introduced in this section.
Figure 3.2: Reference frames adopted (exagerated): Geodetic, ENU, Vehicle- attached, and Road-attached
Geodetic System
Vehicle positions are obtained in cooperative systems in a geodetic coordinate system. A geodetic system is xed to the Earth (Wikipedia, 2009). An unique position on the Earths surface is speci ed by longitude and latitude (in de- grees) as well as by the position high (typically, in meters). The WGS-84 is the datum utilized in this project. The geodetic system oers a common reference frame for all vehicles in a cooperative system.
3.2 Reference Frames 29
Local East, North, Up (ENU) System
An ENU system is a Cartesian coordinate reference frame. It is obtained from a tangent plane to the Earth's surface in a speci ed position. Hence, this refer- ence frame is also xed to the Earth. Typically, the Earth's eastward direction is represented by the positive x-axis in the ENU system. Correspondingly, the northward direction by the positive y-axis and upwards by the positive z-axis.
This reference frame will be utilized for most of the calculations in this project.
Its Cartesian and inertial characteristics are more intuitive for both calculations and analysis of the prediction of trajectories. This scheme diers from previous approaches, where a vehicle-attached reference frame was used for predictions of trajectories. In those approaches, obstacles vehicles were tracked from on- board sensors. Hence, measurements were obtained from the vehicle-attached reference frame. This could explain the use of such a reference. Here, for a cooperative system, information from vehicles motion is obtained from an in- ertial reference frame. Since calculations are notably simpli ed when executed in a xed coordinate frame, the developments in this project will adopt a ENU system.
The errors in the adoption of an ENU system should increase with the distance to the origin of the system. The cause of this error is the non-linear conversion of positions from geodetic to the ENU coordinates. The performance of such conversion should degrade the farther the position lies from the reference origin of the ENU system. In practical applications, the origin of the ENU system could be renewed at each arbitrary time elapsed or distance traveled. For simplicity, however, this concept was not be implemented in this project.
Vehicle-attached System
A Cartesian coordinate reference frame is attached to the vehicle. The direc- tion of movement of the vehicle is de ned the positive longitudinal axis. Using the International Organization for Standardization (ISO) right-hand rule con- vention, the direction perpendicular to the vehicle movement is labeled the
3.3 Platforms for Acquisitions and Tests 30
positive lateral axis. The vehicle-attached frame is dynamic in time, i.e., it is non-inertial. It characterizes the motion from the vehicle/driver perspective.
Road-attached System
The road-attached reference frame is a coordinate system that is attached and aligned with a road segment. The along road axis refers to the axis that de nes the distance driven by a vehicle along the road segment. Conversely, utilizing the Society of Automotive Engineering (SAE) right-hand rule convention, the the across road axis tags the axis that speci es the vehicle lateral position on the road segment. Notice that this reference frame changes at each road segment (which does not correspond to each street, see Section 5.2). The road-attached frame is employed for the incorporation of environmental information, such as road geometry and distance to junctions.
3.3 Platforms for Acquisitions and Tests
This section introduces the vehicle platforms available for data acquisition and tests during the development of the work. In the sequence, the data obtained and calculated in each acquisition is presented.
3.3.1 Vehicles
The project utilizes two vehicles to acquire data: a Volvo S80 car and a Volvo FH16 truck. These vehicles are SafeSpot compatible and used exclusively for research tests purposes. All sensors in these vehicles are o-the-shelf units, which were installed in their production line. The exceptions are the GPS receivers and the devices for the wireless communication, which were installed in both vehicles for the sake of the SafeSpot project. In addition, there are three laser scanners, installed in the frontal part of the truck only, also for the SafeSpot. Figure 3.3 shows these vehicles.
3.3 Platforms for Acquisitions and Tests 31
Figure 3.3: Vehicle eet utilized in this work for data acquisition and tests:
Volvo FH16 truck (in the left of the gure) and Volvo S80 car (in the middle).
The Volvo FH16 truck (in the right) was left as spare.
3.3.2 Notation
Position, speed, and acceleration along the x-axis are represented by the x, vx, and ax symbols, respectively. In a similar manner, position, speed, and acceleration along the y-axis are represented by the y, vy, and ay symbols, respectively. When a model is de ned in a generic coordinate direction, the symbols x, v, and a will be used to designate position, speed, and acceleration in that axis. In addition, it is also de ned the vectors position ~p = [x y], velocity ~v = [vx vy], and acceleration ~a = [ax ay].
The vehicle heading is represented by the symbol h. The heading is assumed to be in radians ([rad]), positive counter-clockwise, with the positive x-axis as reference. The yaw rate, also called angular or turn rate, will be designated with w. It is supposed to be in radians per second ([rad/s]).
Some parameters will be de ned in the vehicle-attached reference frame. They are the longitudinal speed (vlong), longitudinal acceleration (along), and lat- eral acceleration, (alat). In addition, ~ulong and ~ulat are unitary vectors of the longitudinal and lateral directions, respectively, written in the ENU reference frame.
Finally, the symbol k stands for the iteration of the prediction step, while T
3.3 Platforms for Acquisitions and Tests 32
represents the length in time of the predicted step. Time, position, velocity, and acceleration are utilized in the International System of units (SI) for cal- culations.
3.3.3 Data Acquired and Calculated
Observations from the vehicle state are obtained from the LDM, and therefore they are not direct measurements from sensors. Observations are obtained typ- ically at a 13Hz rate (i.e., at each 0.075s). Longitude and latitude positions are obtained in geodetic coordinates. Details of the conversion from geodetic coordinates to the ENU coordinates will be skipped. Longitudinal speed and acceleration as well as lateral acceleration are also obtained from the LDM, in SI units. Vehicle heading is provided in degrees, positive counter-clockwise, with the positive North as reference. The heading is transformed into radians and shifted by π2 to meet the convention used in this work. The phase shift neglects any further corrections needed due to the misalignment between ref- erence frames. The yaw rate is given in degrees per second ([◦/s]), positive counter-clockwise. It is converted to radians per second ([rad/s]) to meet the convention in this work.
Althought the lateral acceleration can be obtained from the LDM, the observa- tions were very noisy. Therefore, the algorithm calculates lateral acceleration by the relationship in Equation 3.1.
alat = vlong· w (3.1)
Equation 3.2 de nes ~ulong and ~ulat, the unitary vectors in the longitudinal and lateral directions, respectively. To obtain the value of, say, the longitudinal acceleration in the ENU frame, it is performed the multiplication alat · ~ulong. This can be generalized to any parameter in the vehicle-attached frame.
3.4 Typical Driving Scenarios 33
~ ulong =
"
cos(h) sin(h)
#
for longitudinal direction, and
~ ulat =
"
− sin(h) cos(h)
#
for lateral direction
(3.2)
Finally, equations utilized in the algorithms are always de ned in discrete time.
3.4 Typical Driving Scenarios
The dierent situations encountered in a vehicle cruise can be summarized ap- proximately by three scenarios. They are: (1) free ow in straight road sections, (2) curved road section, and (3) straight road section with obstructions, such as leading vehicles or stop signs. A cruise that covers all these three scenarios in a roughly even form was performed with the truck. This cruise will be the base of most of the analysis in Chapter 4. The cruise was performed during o ce hours on a weekday.
The roads traversed by the truck during this acquisition are viewed in Figure 3.4. In the particular case of this acquisition, the highway section of the cruise accounts for the situation with a free ow in a straight road section. The ramp section covers the driving scenario in curved sections. Finally, the section in a city street provides a scenario with a number of obstructions for the vehicle cruise.
The following subsections analyze in more detail each of these pairs road sec- tions/drive scenario. Table 3.1 summarizes basic aspects of the cruise, in gen- eral, and of these sections, in speci c.
3.4 Typical Driving Scenarios 34
Figure 3.4: Vehicle trajectory encompassing three typical driving scenarios in particular road sections
3.4.1 Driving Case - Straight Road Section with Free Flow
A scenario with a long straight section with free ow is typically found in highways. The kinematic evolution of the vehicle in such a situation is very particular. This can be observed in Figure 3.5. The vehicle motion is straight and, hence, vehicle heading is roughly constant during the traverse of this highway section. Consequently, vehicle yaw rate is near zero. Vehicle speed changes smoothly, and only in the end of the section (which coincides with a tra c light), it decreases in a more moderate rate. At this point is noticed the increase of longitudinal acceleration intensity. Notice that in some parts of the trajectory, the lateral acceleration is more intense than the longitudinal. Since the vehicle drives straight, this is possibly a result of the high vehicle speed in this section.
