Dalarna University – SE-791 88 Falun – Phone +4623-77 80 00
Degree Project
Master thesis in Microdata Analysis
A contemporary machine learning approach to detect transportation mode - A case study of Borlänge, Sweden
Authors: Arman Golshan
Supervisor: Xiaoyun Zhao, Paria Sadeghian Examiner: Siril Yella
Subject/main field of study: Master's Thesis in Microdata Analysis Course code: MI4001
Credits: 30 credits
Date of examination: 20th January 2021
At Dalarna University it is possible to publish the student thesis in full text in DiVA. The publishing is open access, which means the work will be freely accessible to read and download on the internet. This will significantly increase the dissemination and visibility of the student thesis.
Open access is becoming the standard route for spreading scientific and academic information on the internet. Dalarna University recommends that both researchers as well as students publish their work open access.
I give my/we give our consent for full text publishing (freely accessible on the internet, open access):
Yes ☒ No ☐
Abstract
Understanding travel behavior and identifying the mode of transportation are essential for adequate urban devising and transportation planning. Global positioning systems (GPS) tracking data is mainly used to find human mobility patterns in cities. Some travel information, such as most visited location, temporal changes, and the trip speed, can be easily extracted from GPS raw tracking data. GPS trajectories can be used as a method to indicate the mobility modes of commuters. Most previous studies have applied traditional machine learning algorithms and manually computed data features, making the model error-prone. Thus, there is a demand for developing a new model to resolve these methods' weaknesses. The primary purpose of this study is to propose a semi-supervised model to identify transportation mode by using a contemporary machine learning algorithm and GPS tracking data. The model can accept GPS trajectory with adjustable length and extracts their latent information with LSTM Autoencoder.
This study adopts a deep neural network architecture with three hidden layers to map the latent information to detect transportation mode. Moreover, different case studies are performed to evaluate the proposed model's efficiency. The model results in an accuracy of 93.6%, which significantly outperforms similar studies.
Keywords: GPS data, Semi-supervised learning, Transport mode detection, LSTM Autoencoder, Deep Neural Network
Table of Contents
1. Introduction ... 1
2. Literature Review ... 4
2.1. Traditional methods to detect transportation modes ... 4
2.1.1. Rule-based methods ... 4
2.1.2. Statistical methods ... 5
2.1.3. Traditional machine learning methods (T-ML methods) ... 6
2.2. Contemporary machine learning methods ... 8
3. Methodology... 10
3.1. Dataset ... 10
3.2. Data Cleaning ... 11
3.3. Trip Segmentation ... 11
3.4. Feature extraction ... 12
3.5. Data Labeling ... 13
3.6. Extracting the latent information ... 14
3.7. Single Classifier ... 15
4. Result and Discussion ... 16
4.1. Dataset pre-processing ... 16
4.2. LSTM Autoencoder performance... 17
4.3. Model Accuracy ... 17
4.4. Model evaluation based on the validation dataset ... 21
4.5. Comparison with similar studies ... 21
5. Conclusion ... 24
References ... 25
List of Tables
Table 1. Summary of traditional approaches in transportation mode studies ... 4
Table 2. Summary of contemporary approaches in transportation mode studies ... 8
Table 3. Number of point data and segment for each transportation mode ... 16
Table 4. Model performance with different partition of labeled data ... 17
Table 5. Confusion matrix of proposed model with respect to proportation of labeled data ... 20
Table 6. Confusion matrix of the validation dataset ... 21
List of Figures
Figure 1. Model flow to enable transport mode detection ... 10Figure 2. Left: raw GPS trajectories Right: Public transportation network ... 11
Figure 3. A simple LSTM Autoencoder Architecture ... 15
Figure 4. Architecture of proposed model for detecting transport modes ... 15
Figure 5. Transportation modes per segment in new labeled dataset ... 16
Figure 6. Changes in loss function over numbers of iteration ... 17
Figure 7. Loss and accuracy over iterations per each fold ... 19
Figure 8. Comparison of transportation mode identification accuracy ... 22
1
1. Introduction
Human mobility pattern is closely connected to geographical patterns and spatial distributions (Ebrahimpour et al., 2020). Understanding the human mobility pattern indicates significant insights for various applications such as migration studies, disease spread, transport planning, tourism, and guidance systems (Ebrahimpour et al., 2020). The investigation of people's mobility behavior is crucial for urban planning and infrastructure facilities (Noulas et al., 2012).
Congestion prediction, fare pricing, new policies, and new plans for the transportation network can be considered as a result of studying people's travel habits (Rezaie et al., 2017).
Human mobility data are essential to understand people's mobility patterns. Mobility data is useful for identifying the trip purpose and classifying the frequently visited locations in the city to identify the routing path, green transportation, infrastructure services, and transportation network (Manzoni et al., 2010). Collecting travel data regularly is one of the most significant challenges in urban transportation planning and transportation system management (Rezaie et al., 2017). Surveys and questionnaires are traditional tools for collecting human mobility data (Maat et al., 2005). However, inaccurate time reports, duplicated reports, inaccurate reports, and forgetting to report personal trips are disadvantages of these traditional data collection methods (McGowen & McNally, 2007). Using new technologies such as the global positioning system (GPS), geographical movement datasets' availability is facilitated (Siła-Nowicka et al., 2016). In collecting data with GPS, volunteers are asked to record their coordinates in a set of intervals of times. Although the number of sample populations and accurate geographic coordinates are the advantages of this method, privacy limitations and missing some human mobility patterns due to less sample size are the disadvantages of this method (Stolf Jeuken, 2017). Daily travel data and positional information can be recorded by using GPS devices.
Nevertheless, it does not have explicit and precise information regarding transportation mode, and a sort of techniques are needed to detect the mode of transportation from the GPS trajectories (James, 2020).
Identifying transportation mode is essential for establishing future smart cities and intelligent transportation systems (Adler & Blue, 1998; Scheiner & Holz-Rau, 2007). Accurate transport mode data is vital for stakeholders and research institute since aligning them with characteristics of the trajectory can lead to solving transport-related problems and a better understanding of human behaviors (James, 2018; James & Lam, 2017; Scheiner & Holz-Rau, 2007; B. Wang et al., 2017; F.-Y. Wang, 2010). Moreover, analyzing the trajectory data and travel mode choice
2 can improve urban planning and predicting travel demands (L. Gong et al., 2014; Lin & Hsu, 2014). GPS data parameters such as speed, acceleration, and direction help detecting transportation modes either by using these parameters or in combination with GIS data (Dabiri
& Heaslip, 2018; H. Gong et al., 2012).
Understanding travel mobility mode can also resolve a city's requirements, followed by citizens' happiness (Shin et al., 2015). Moreover, it can allow people to have a custom-made recommendation. For instance, for an individual driving in a car, advertisements related to car service can be shown, or if a person travels on a bus, recommendations related to books or tablets can be shown (Stenneth et al., 2011).
The majority of the contemporary literature has proposed a transport mode detection model based on some manually calculated features such as maximum velocity and acceleration (Xiao et al., 2017; Zheng, Li, et al., 2008; Zheng, Liu, et al., 2008). After creating hand-crafted features, various traditional supervised machine learning algorithms, including rule-based methods, fuzzy logic, decision tree, Bayesian belief network, multilayer perceptron, and support vector machine, have been applied to do the classification task (Wu et al., 2016).
Dabiri et al. (2019) have recently introduced the first semi-supervised travel mode identifier based on convolution neural networks and an auto-encoder. However, the approach depends excessively on the sample size and density of labeled data to train the encoder as a supervised classifier. This feature limits the identifier from acquiring accurate travel mode information when labeled data is limited. Hence, a new model to detect the transport mode is essential.
Moreover, Vu et al. (2016) have applied different types of Recurrent Neural Network (RNN) to raw data collected by accelerometer and sensor hub and without feature extraction and indicated that Long Short-Term Memory (LSTM) does not perform significantly well.
However, the strength of LSTM to acquire patterns in data over long sequences makes them suitable for time series sequential data (Nguyen et al., 2020). Furthermore, autoencoders are unsupervised machine learning algorithms that can encode and represent a compressed version of input data. Thus, this study is encouraged to apply LSTM Autoencoder on a new set of GPS data inputs for predicting transportation modes.
This study tries to obtain the following aims:
1. Developing a semi-supervised model to detect transportation modes with higher accuracy in comparison with previous studies
3 2. Whether applying LSTM Autoencoder can reduce the complexity of the model and at
the same time result in higher accuracy
As the main contribution of this study, the potential of applying a combination of LSTM Autoencoder and deep neural network on GPS mobility data is investigated. This attempt reveals the possibility of applying a more straightforward and efficient model to detect transport mode.
This study is structured as follows. Section 2 describes the relevant previous studies. Section 3 explains in detail the pre-processing methods, LSTM Autoencoder, and deep neural network.
Section 4 presents and discusses the empirical analysis of the GPS mobility data and the corresponding results. Section 5 concludes the thesis with the main takeaways and future work directions.
4
2. Literature Review
The related literature identifies transport mode by using two approaches: traditional approaches and modern approaches. Rule-based methods, statistical methods, traditional machine learning algorithms are the methods of traditional approaches. Table 1 briefly lists the most recent studies that detected transport modes by the traditional approach. Therefore, this chapter focuses on reviewing relevant studies and identifying the key characteristics and applied techniques used.
Table 1. Summary of traditional approaches in transportation mode studies
Category Study No. algorithms Best Algorithm Model Accuracy No. Features Dataset type No. Volunteer
Rule-Based Chen et al., (2010) 1 GIS Information Layer 79% 3 GPS device 25 Rule-Based H. Gong et al. (2012) 1 GIS Information Layer 83% 5 GPS device 35 Statistics Sauerländer-Biebl et al., (2017) 1 Fuzzy Rules 75% 7 GPS application 1
Statistics Rasmussen et al. (2015) 1 Fuzzy Rules 91% 2 GPS device 183
Statistics Biljecki et al. (2013) 1 Fuzzy Rules 92% 4 Two datasets 1000 T-ML Rezaie et al., (2017) 3 Label Propagation 81% 3 GPS application 796 T-ML Feng & Timmermans, (2016) 7 Bayesian Network 99.6% 6 GPS device 8 T-ML Lari & Golroo, (2015) 1 Random Forest 96% 6 GPS application 35 T-ML G. Xiao et al. (2015) 1 Bayesian Network 94% 6 GPS application 30
2.1. Traditional methods to detect transportation modes 2.1.1. Rule-based methods
In the rule-based approach, modes of transport are detected based on some criteria and GIS layer information. To detect the transport mode and daily trip purpose in New York City's urban area, Chen et al. (2010) developed a multi rule model and applied it on two different datasets.
This study identified six transportation modes: walk, subway, rail, car, bus, and underground transfer. Among all transportation modes, the car achieved the highest accuracy, with 95.8%, while rail achieved the lowest success rate with 28.6%, resulting in a low number of rail trips.
Moreover, they suggested using the passive data collection method instead of the conventional travel survey in future research.
5 H. Gong et al. (2012) identified five different transportation modes, including walk, car, bus, subway, and commuter rail, through a set of criteria and GIS information layer. The two different datasets used in this study were collected by GPS devices and complete through a paper-based survey or an online survey. This study used the same dataset as is applied in Chen et al. (2010). The average accuracy in detecting the modes of transportation in this study was around 82%. Walk mode and train mode are detected with the highest accuracy and the lowest accuracy, respectively. According to the success rate as a finding of this study, it proposed adding the internet and smartphones with GPS and accelerometer to improve the success rate of identification of transport modes.
2.1.2. Statistical methods
In the statistical approach, fuzzy logic rule and probability matrix were used to detect the transport mode. To begin with, Sauerländer-Biebl et al. (2017) have developed a three-phase algorithm in order to derive the transport mode from trip trajectories that were collected through smartphones. In the first step, the authors specified the beginning and end of each trip and, based on them, divided the trips into different segments. In the second phase, they have applied a fuzzy logic rule to derive the mobility modes, including walk, bike, bus, car, and train. The third step was a correction phase. For instance, a walk segment was considered a car segment if the walk segment's speed is less than 30 km/h, but the walk segment was between two car segments. This study considered the turning angle as one of the beneficial features to identify the transport mode. This study applied extra information such as average speed, maximum speed, average acceleration, active time, and deceleration and acceleration length. The total accuracy of the proposed model in this study was 75%, while the model could predict the car mode with 98% accuracy. The authors suggested using more data and trips to confirm the finding since only 11 trips were analyzed in this study.
As the second study, Rasmussen et al. (2015) has developed a combined fuzzy logic and GIS- based algorithm to deduce the five transport modes, including walk, bike, car, bus, and train, from raw GPS data. The model detected trip legs and differentiated between the modes. They used speed and acceleration with GIS software as their step-wise process. Train mode was identified in the first step based on the rail network alignment and GIS layer information. In the second step, fuzzy logic rules have been applied to detect walk, bike, and either car or car/bus.
In the third and last step of this model, the authors used bus line alignment and GIS layer information to check whether a certain number of GPS records are close to bus stops. If yes,
6 the trip was assigned to be travel by bus. The study proved that their model could detect bus and train at a significant rate. Furthermore, the accuracy across the five transport modes is 92%.
In the third study, Biljecki et al. (2013) have used both GPS devices and smartphones to collect the data. The movements of 1000 users have been recorded for seven days in this study. To detect the transport modes, the authors used extracted information from GPS devices and GIS layers. Longitude, latitude, speed, and time were employed from GPS devices. Furthermore, the geographical information, including railways, bus lines, metro, tramlines, roads, and water surface locations, were employed from GIS layer information. They have modeled a combination of fuzzy logic rules and GIS layer information to detect ten different transport modes, including walk, bike, car, bus, train, tram, metro, airplane, boat, and ferry. This study used a probability score to classify the trajectories, and the result showed that maximum speed and average speed were vital for identifying transport mode. They reported that the accuracy level experienced significant growth after they used GIS layer information. Moreover, they concluded that fuzzy logic rules rely on many criteria and require more human interference than machine learning algorithms.
2.1.3. Traditional machine learning methods (T-ML methods)
In the traditional machine learning approaches, researchers used supervised and unsupervised methods to identify transport modes such as random forest, decision tree, and Bayesian network, etc. The pre-processing phase of the following studies is similar. As Initializing, the trajectories were divided into different segments based on the specific criteria of each studies.
Different traditional machine learning algorithms, including Random Forest, Decision Tree, Logistic Regression, Naïve Bayes, Multinominal Logistic Regression, Support Vector Machine, and Bayesian Network, were applied to the data.
To begin, Rezaie et al. (2017) have proposed label propagation as a semi-supervised algorithm and compared the accuracy of the model with two supervised methods: decision tree and random forest. The study used speed, duration, length of trip, and start and endpoint to the transit network to identify the transport mode. Three different datasets from different data sources have been used in this study. The study demonstrates that the random forest success rate is positively correlated with the number of labeled and unlabeled data, while label propagation performed the same and is not dependent on labeled data. The authors suggested using other semi-supervised algorithms in future research.
7 Feng & Timmermans (2016) has compared seven different algorithms, including Naïve Bayes, Bayesian Network, Logistic regression, Multilayer perceptron, support vector machine, decision table, and C4.5 for the same data. Eight individuals were asked to carry out a GPS device for a six to eight weeks period. They configured the GPS devices to record data every three seconds. Besides, the authors have collected travel data for the tram and metro in order to detect more transport modes. Moreover, they have asked the individuals to answer some prepared questions regarding the time, location of activities and trips, and transportation mode to label data and make the truth table. In this study, average speed, maximum speed, average acceleration, trip distance, distance to transportation networks, satellite information, and vehicle ownership were used to detect the transport mode. They concluded that Bayesian Network and C4.5 algorithms had better performance with 99% accuracy. However, the Bayesian Network was more robust. Since the dataset was created based on only eight people, the authors suggested using a larger sample of truth data associated with GPS traces in future research.
In the third study of this section, Lari & Golroo (2015) has developed a model that identifies walk, car, and bus as transport modes. Train and the rails trips were not detected in this study.
The authors collected the data via an application over two weeks in Tehran, and participants were asked to report their mode of transportation in order to label the data. The authors used some extra features and attributes that were recorded by the application to detect the transport mode. The application recorded time, date, instant speed, accuracy, bearing, altitude, latitude, and longitude. Moreover, delta bearing, delta speed, acceleration, and delta acceleration are extracted through application records. The random forest was used as a machine learning algorithm and showed 96% accuracy to detect transport mode. Instant speed and accuracy of GPS tracking data were the most important factors for identifying the transport mode.
In the last study, G. Xiao et al. (2015) distinguished walk, bike, e-bike, car, and bus modes from GPS data. 202 individuals recorded their data through an application, and in total, 1,248 data streams were stored after monitoring. Bayesian network, Multinominal logistic regression, support vector machine, and artificial neural network were employed to detect transport mode.
This study indicated that the Bayesian network was an appropriate algorithm for detecting the transport mode from GPS data based on a smartphone. Moreover, the authors showed that low- speed rate and average heading change significantly impact detecting transport mode. For future works, they suggested importing GIS layers into the methodology in order to differentiate
8 between car and bus. Furthermore, the features that can contribute to distinguishing between bike and e-bike were required.
2.2. Contemporary machine learning methods
In the modern machine learning approach, in order to detect the transport mode, a Multilayer perceptron, convolutional neural network, recurrent neural network, and long short-term memory were used. In a review paper by Nikolic & Bierlaire (2017), the authors mentioned that by increasing the number of transport modes to detect, the success rate decrease. Hence, deep learning approaches were introduced in order to improve accuracy. For instance, the accuracy level of transport mode detection Fang et al., (2017) was increased from 83.57% to 95% by employing a deep neural network. Table 2 summarized the most recent research that applied modern algorithms to identify modes of transportation.
Table 2. Summary of contemporary approaches in transportation mode studies
Study Algorithms Accuracy No. Features Dataset Type No. Volunteer James, (2020) Deep ensemble
method 91.5% 4 GPS 182
Asci & Guvensan
(2019) LSTM 96.8% NA
accelerometer, gyroscope, and magnetometer
8
Dabiri et al.
(2019)
Convolutional
Autoencoder 79.8% 4 GPS 182
Vu et al. (2016) CGRNN 94.7% NA Accelerometer 8
In the first study, James, (2020) proposed a model to detect transport mode even with few labeled data. This study introduced a semi-supervised deep ensemble learning method to use a minimal number of annotated data to identify the transport mode. The same dataset that was used in Zheng, Li, et al. (2008) and Dabiri et al. (2019) was applied in this study. The dataset presented the movement of 182 individuals over five years. Moreover, 69 users provided the truth data and reported the mode of transport. This study detected the five same transport mode as the two mentioned paper. Speed, acceleration, the difference of acceleration, and turn rate time were used as extra features in this work. The results of this study were compared with the result of Dabiri et al. (2019) and Zheng, Li, et al. (2008), and it showed that the proposed model performed better than the other two studies in terms of accuracy. Furthermore, the authors proved that the accuracy of GPS devices had no impact on the proposed model's performance.
9 In the second study, Asci & Guvensan, (2019) has introduced a novel input set for transportation mode to increase the detection rate. They took advantage of a frame-based approach to train the LSTM and prevent overlapping. The data was collected through an accelerometer, gyroscope, and magnetometer. They detected ten different transport modes with an overall accuracy of 96.82%. The authors suggested elaborating on the drop rates, learning rates, optimization, and error calculation as future work.
Since most previous research tried to identify transportation mode through a set of hand-craft features, Dabiri et al. (2019) proposed a semi-supervised model called Semi-Supervised Convolutional Autoencoder (SECA). The model can extract useful information from unlabeled data and calculate the relevant features from raw GPS data. The study aimed to detect walk, bike, car, bus, and train as transportation modes. Dabiri et al. (2019) used the same dataset that was presented in the Zheng, Liu, et al. (2008). The authors compared the result of the proposed model with 1) supervised algorithms, including K-Nearest Neighbor, Support Vector Machine, Decision Tree, Multilayer Perceptron, and 2) semi-supervised algorithms, including Semi-Two- Step and Semi-Pseudo-Label. Their extensive experience showed the dominance of the SECA model, their trip segmentation algorithm, and the configuration of the model architecture compared to other baseline and alternatives.
Vu et al. (2016) developed a control gate recurrent neural network (CGRNN) to detect transport mode. The sequential data collected through an accelerometer was manipulated directly by the model. After comparing the model performance with LSTM and gate recurrent unit (GRU), the results showed that the proposed model not only could perform well at the window-based level but also at the frame-based level. The model was not as complicated as LSTM and GRU, besides predicting the transport mode with higher accuracy.
10
3. Methodology
The model flow that is applied in enabling the transport modes detection is illustrated in Figure 1. The model flow begins by taking raw GPS data as input and producing identical travel modes by a sequence of processing steps. Datapoints are first cleaned through specific steps and thresholds. Second, trips are segmented into different GPS trajectories so that each segment to a specific travel mode. While the time series sequential GPS data have no sufficient information to obtain travel modes, relevant features are extracted in the third step. To expose more relevant data properties, an LSTM Autoencoder is employed in the fourth step. These four steps provide the fundamental inputs of the classifier as the final step. In the end, the output of the transportation modes is acquired.
3.1. Dataset
This study uses the raw GPS trajectories dataset from the Urban Building Energy Modeling (UBEM) project in Borlänge city for examination, which contains the transportation of 91 volunteers in one year, September 2019 till September 2020. Each point data is displayed in the form of GPS sequence data in Figure 2. The date, time, latitude, longitude, altitude, and speed were recorded every five seconds. All the volunteers were requested to switch on the GPS device when they start a trip and switch off the device after they reached their destination and their trip finished. Over four million point data and 11,415 trips were recorded during the data collection period.
Figure 1. Model flow to enable transport mode detection GPS Tracking Data
Data Cleaning
Trip Segmentation Single Classifier
Latent Information Feature Extraction
Data labeling Extracting
Transport modes
11
3.2. Data Cleaning
To feed the model with correct data, data cleaning is an essential part of this study. The data cleaning section of this study has three sections. All the point data with a speed over 300 km/h are removed in the first step. All the point data that have a greater timestamp than the next point data are removed. All the trips with no more than three point data or the total duration of the trip were less than 120 seconds were excluded.
3.3. Trip Segmentation
One trip may include different types of transportation and therefore have different segments.
For instance, a trip can be started with walking, then taking a bus or driving a car, and walking again. Hence, the trip has three segments and two transportation modes. In general, partitioning a trip into different segments so that each segment has only one transport mode is called trip segmentation. It is essential to break a trip into a different segment where each segment has only one type of transportation in detecting transportation modes.
Pruned Exact Linear Time (PELT) is used for distinguishing the points within a dataset where the statistical properties change. The PELT model is based on the algorithm (Jackson et al., 2005) but involves a pruning step within the dynamic program (Frappart & Bourrel, 2018).
PELT model can do segmentation and detecting the changepoints in a dataset more accurately Figure 2. Left: raw GPS trajectories Right: Public transportation network
12 than binary segmentation (Killick et al., 2012). The PELT method adjusts the optimal partitioning method of A by pruning. It consolidates optimal partitioning and pruning to achieve exact and practical computational cost, which is linear in a number of point data.
𝐹(𝑛) = 𝑚𝑖𝑛
𝜏𝑚 { ∑ [𝐶(𝑦𝜏𝑖−1+1: 𝑦𝜏𝑖) + 𝛽] }
𝑖=𝑚+1
𝑖=1
Eq. 1
Where n is the number of point data, 𝛽 is the penalty to control the overfitting, and 𝐶 is a cost function for the 𝑖𝑡ℎ segment. 𝑚 is all number of changes in their positions, 𝜏.
Since PELT aims to find the optimal number of change points, the model needs to calculate the optimal number of breakpoints or optimal segmentation until that changes point minimizes the 𝐹(𝑛). This iterative nature can lead to an inner minimization, which is shown by 𝐹(𝜏𝑚) in Eq.
2.
𝐹(𝑛) = 𝑚𝑖𝑛 𝜏𝑚 { 𝑚𝑖𝑛
𝜏|𝜏𝑚∑[𝐶(𝑦𝜏𝑖−1+1: 𝑦𝜏𝑖) + 𝛽] + 𝐶(𝑦𝜏𝑚+1: 𝑦𝑛) }
𝑖=𝑚
𝑖=1
𝐹(𝑛) = 𝑚𝑖𝑛
𝜏𝑚 { 𝐹(𝜏𝑚) + 𝐶(𝑦𝜏𝑚+1: 𝑦𝑛)
Eq. 2
3.4. Feature extraction
The features like date, time, longitude, latitude, speed, and elevation of each point data were recorded in the raw GPS dataset. However, to identify transportation modes, more features are needed to be calculated. Hence, the distance between two point data, total distance, bearing rate, turning change rate, time difference, total duration, and features related to speed, such as average speed, maximum speed, minimum speed, acceleration, and jerk for each segment are calculated. To calculate the distance between two point data, the Vincenty formula is in Eq. 3 is employed
𝑑𝑖 = 𝑉𝑖𝑛𝑐𝑒𝑛𝑡𝑦(𝑙𝑎𝑡𝑖, 𝑙𝑛𝑔𝑖, 𝑙𝑎𝑡𝑖+1, 𝑙𝑛𝑔𝑖+1) Eq. 3 Where 𝑑𝑖 represents the distance between datapoint 𝑖𝑡ℎ and 𝑖 + 1𝑡ℎ. 𝑙𝑎𝑡𝑖 and 𝑙𝑛𝑔𝑖 are the geographical coordinates value of 𝑖𝑡ℎ datapoint.
Also, the total distance of the segment is calculated by applying the below equation
𝑇𝑜𝑡𝑎𝑙 𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒 = ∑(𝑑𝑖)
𝑛
𝑖=1
Eq. 4
In Eq. 4, n states for the number of point data in a segment.
13 According to the criteria for trip segmentation, the angle between each two point data, also known as bearing, is calculated in each segment. The range of changes in the bearing angle is different in various transportation modes. For instance, people can often alter their direction when they walk or ride a bike, while cars or buses can move alongside the existing street (Zheng, Liu, et al., 2008). The following equation is used to calculate the bearing rate:
𝑏𝑖 = 𝑡𝑎𝑛−1𝑉𝑖𝑛𝑐𝑒𝑛𝑡𝑦(𝑙𝑎𝑡𝑖 𝑙𝑛𝑔𝑖 𝑙𝑎𝑡𝑖+1 𝑙𝑛𝑔𝑖) 𝑉𝑖𝑛𝑐𝑒𝑛𝑡𝑦(𝑙𝑎𝑡𝑖 𝑙𝑛𝑔𝑖 𝑙𝑎𝑡𝑖 𝑙𝑛𝑔𝑖+1)
Eq. 5
Also, since the time difference is one of the monotone criteria for partitioning a trip into different segments, the time difference between every two points in a segment and each segment's total duration is calculated.
∆𝑡𝑖 = 𝑡𝑖+1− 𝑡𝑖 Eq. 6
𝑇𝑜𝑡𝑎𝑙 𝐷𝑢𝑟𝑎𝑡𝑖𝑜𝑛 = ∑(∆𝑡𝑖)
𝑛
𝑖=1
Eq. 7
In Eq. 7, n states for the number of point data in a segment, while 𝑡𝑖 indicates the time of ith point data.
A new speed, acceleration, turning rate, and jerk are calculated and added to the dataset.
𝑠𝑖 = 𝑑𝑖
∆𝑡𝑖 1 ≤ 𝑖 ≤ 𝑁; 𝑠𝑁 = 𝑠𝑁−1 Eq. 8
𝑎𝑖 = 𝑠𝑖+1− 𝑠𝑖
∆𝑡𝑖 1 ≤ 𝑖 ≤ 𝑁; 𝑎𝑁 = 0 Eq. 9
𝑗𝑖 = 𝑎𝑖+1− 𝑎𝑖
∆𝑡𝑖 1 ≤ 𝑖 ≤ 𝑁; 𝑗𝑁 = 0 Eq. 10
𝑡𝑟𝑖 = 𝑏𝑖+1− 𝑏𝑖
∆𝑡𝑖 1 ≤ 𝑖 ≤ 𝑁; 𝑡𝑟1 = 𝑡𝑟𝑁 = 0 Eq. 11 Where 𝑠𝑖, 𝑎𝑖, 𝑗𝑖, and 𝑡𝑟𝑖 are speed, acceleration, jerk (difference of acceleration), and turning rate values, respectively.
3.5. Data Labeling
There was a lack of labeled point data in the collected dataset initially. Integration of the rule- based method and GIS information layer is applied in order to label segments.
Walking data can be recognized quickly based on the following rules. The first rule is that the maximum speed of each segment needs to be less than or equal to 12 km/h (H. Gong et al.,
14 2012). The second rule is that the average speed of each segment needed to be below or equal to 6 km/h (Stopher et al., 2008).
In order to detect the bike mode, two rules are applied. First, the average speed and the maximum speed of the segment should be less than 25 km/h and 40 km/h, respectively.
Moreover, the total distance of the segment should not exceed 20 kilometers.
The GIS layer information is used to detect the train mode. Two rules are employed; the first rule is all points of segments need to follow the railway networks. Moreover, the distance of the first and endpoint of each segment to the train stations should be less than 75 meters.
Distinguishing between car and bus is the challenging part of labeling data since the bus and car have a similar city pattern, such as speed and road network. Therefore, bus timetables, bus lanes, and bus stations are used in order to detect bus mode. The first and endpoint of each segment should be within 75 meters of a bus stop. As the second rule, the average speed of the segment should not be more than 88 km/h, and the maximum speed of the segment should be below 100 km/h (H. Gong et al., 2012). Furthermore, in the third condition, the point data of a segment should follow the bus lane.
To detect car mode, the segment's datapoints need to follow the road network, and the maximum speed of the segment should not exceed 180 km/h (Biljecki et al., 2013).
3.6. Extracting the latent information
The extracted features are widely used in previous studies. However, it is not clear which of these extracted features play significant roles in identifying transportation modes. Hence, an LSTM autoencoder was employed to extract latent information from the extracted features.
LSTM Autoencoder is an implementation of an autoencoder that accepts time series sequential data and uses LSTM architecture in its encoder and decoder. In this model, the input sequence is read by the encoder LSTM, and the model saves the latent information as a vector, which represents the entire original data. In order to regenerate the input data, the vector of hidden information can be passed as the input to the decoder part of the LSTM Autoencoder (Li et al., 2020). Figure 3 illustrates the architecture of a simple LSTM Autoencoder and the data flow within it.
15
3.7. Single Classifier
After extracting the latent information, it is possible to build a fully connected neural network that takes the inputs and detects the transport mode. The fully connected neural network can be trained by using categorical cross-entropy as a loss function since the purpose of the model is multiclass classification. Moreover, in order to optimize the model, Adam optimizer is used.
Figure 4 presents the architecture of the model. Three fully connected layers are employed, which takes latent information as input for the network. 35 and 18 neurons are employed in the first two ReLU activated layer. In the last layer with 𝑋 number of neurons, the softmax activation function is employed to identify the travel modes, where 𝑋 is the number of transportation modes.
Figure 3. A simple LSTM Autoencoder Architecture Latent Vector
LSTM LSTM LSTM
LSTM LSTM LSTM
Encoder Dencoder
Input Sequence Output Sequence
Figure 4. Architecture of proposed model for detecting transport modes 𝐿𝑎𝑡𝑒𝑛𝑡 𝐼𝑛𝑓𝑜𝑟𝑚𝑎𝑡𝑖𝑜𝑛
𝐿𝑆𝑇𝑀: 𝑂𝑢𝑡𝑝𝑢𝑡 128 × 𝑁
𝐿𝑆𝑇𝑀: 𝑂𝑢𝑡𝑝𝑢𝑡 64 × 𝑁
𝐿𝑆𝑇𝑀: 𝑂𝑢𝑡𝑝𝑢𝑡 64 × 𝑁
𝐿𝑆𝑇𝑀: 𝑂𝑢𝑡𝑝𝑢𝑡 128 × 𝑁 𝑅𝑒𝑝𝑒𝑎𝑡𝑒𝑑 𝑉𝑒𝑐𝑡𝑜𝑒𝑟: 𝑂𝑢𝑡𝑝𝑢𝑡 64 × 𝑁
𝐴𝑙𝑙 𝑙𝑎𝑏𝑒𝑙𝑒𝑑 𝑎𝑛𝑑 𝑢𝑛𝑙𝑎𝑏𝑒𝑙𝑒𝑑 𝑑𝑎𝑡𝑎
FCL: ReLU 35 × 1
FCL: ReLU 18 × 1
FCL: ReLU X × 1 Travel modes
16
4. Result and Discussion
This study proposes semi-supervised deep learning in order to identify transportation modes.
To examine the output of this study, first, the LSTM Autoencoder performance is assessed.
Second, the performance of the model and the accuracy of data labeling are evaluated by comparing the output of the model with the truth table of the study provided by 15 volunteers.
Third, the results of this study have been compared with the three recent transportation mode identifiers in the literature, which are almost close to this study in terms of methodology.
4.1. Dataset pre-processing
Among all volunteers, 15 volunteers provided their trajectories' transportation mode, resulting in 576 labeled trips. Point data and trips are identified and removed in this study as explained in detail in Section 3.2. Moreover, the PELT algorithm is applied on the speed and heading change rate to find the index of change points in a trip and resulted in 585,608 segments.
The volunteers labeled only five percent (268,518 point data) of all point data. These five percent is considered as the validation dataset to evaluate the efficiency of the proposed model.
Table 3 shows the number of point data and segments for each mode of transportation.
Table 3. Number of point data and segment for each transportation mode
Bike Bus Car Train Walk
Point data 19,568 4,115 54,964 21,434 168,438
Segment 2,271 816 5,757 3,372 26,485
To train the proposed model in section 3.6, more labeled data is required. These 268,518 point data are assumed as validation dataset. Since 80% of the dataset is required for training the model, 1,347,696 more point data are considered in order to apply labeling rules in section 3.4.
Figure 5 illustrates the number of newly labeled segments per transportation modes.
Figure 5. Transportation modes per segment in new labeled dataset
7,157 900 11,872 621 133,231
Bike Bus Car Train Walk
17
4.2. LSTM Autoencoder performance
A simple LSTM Autoencoder with the architecture illustrated in Figure 3 is trained with all labeled and unlabeled point data. Figure 6 illustrates the changes in the loss over the 200 iterations. It shows that the LSTM Autoencoder predicts the input values with less than 0.1 error. Therefore, it can be concluded that the obtained latent information can be reliable and a good representative of the dataset as the input values.
4.3. Model Accuracy
A single deep neural network with three hidden layers and two input and output layers is trained to detect transportation mode. K-fold cross-validation with 5 number of splits and only one iteration is applied to obtain a less biased result. Since the dataset is not balanced, the F1 score is employed to measure the model's performance to evaluate the model's accuracy in identifying transportation mode. Table 4 lists the model accuracy concerning the proportion of labeled data.
Table 4. Model performance with different partition of labeled data
Partition % Fold 1 Fold 2 Fold 3 Fold 4 Fold 5 F1 Score
10% 93.38% 93.01% 93.26% 93.07% 93.31% 93.21%
20% 93.39% 93.14% 93.30% 93.48% 93.21% 93.31%
50% 93.43% 93.44% 93.54% 93.49% 93.33% 93.45%
100% 93.67% 93.53% 93.67% 93.48% 93.63% 93.60%
As shown in Table 4, for the case with 10% of the labeled data, the obtained F1 score is 93.21%, while this value reaches 93.31%, 93.45%, and 93.60% for 20%, 50%, and 100% of labeled data respectively. The result indicates that even with 10 percent labeled data, over 93% of
Figure 6. Changes in loss function over numbers of iteration
18 transportation modes can be predicted correctly. Thus, the obtained result shows that the proposed model generates a significant transportation mode detection result. Moreover, the result confirmed that the proposed deep neural network does not suffer from over-fitting issues prominently.
Figure 7 illustrates the loss and accuracy of training the deep neural network on a full labeled dataset over 100 iterations for each fold. At the end of 100 iterations per each fold, the loss value is below 0.22, which shows how the model performed well in predicting the transportation modes. Moreover, the initial value of deep neural network accuracy per each fold is around 93% and after 100 iterations experience a gradual increase to 93.6%.
19
Fold 1 Fold 2
Fold 3 Fold 4
Fold 5
Figure 7. Loss and accuracy over iterations per each fold
20 Besides the statistical results showed in Table 4, the confusion matrix of predicted transportation mode is presented in Table 5.
Table 5. Confusion matrix of proposed model with respect to proportion of labeled data
10% of labeled data Predicted Mode Recall %
Real Mode Bike Bus Car Train Walk
Bike 6,632 45 589 0 2,922 65.09%
Bus 333 139 303 1 510 10.80%
Car 575 1 15,048 20 185 95.06%
Train 61 0 336 139 15 25.22%
Walk 1,245 53 133 0 78,633 98.21%
Precision % 74.97% 58.40% 91.70% 86.87% 95.58% 93.21%
20% of labeled data Predicted Mode Recall %
Real Mode Bike Bus Car Train Walk
Bike 12,896 123 1,284 1 6,145 63.06%
Bus 566 319 653 1 1,072 12.21%
Car 1,014 5 30,379 22 385 95.51%
Train 100 0 574 287 47 28.47%
Walk 2,029 130 284 2 157,518 98.47%
Precision % 77.66% 55.28% 91.57% 91.69% 95.36% 93.31%
50% of labeled data Predicted Mode Recall %
Real Mode Bike Bus Car Train Walk
Bike 32,292 259 3,191 5 14,880 63.78%
Bus 1,514 589 1,623 3 2,645 9.24%
Car 2,400 5 76,245 105 937 95.67%
Train 281 0 1,491 761 89 29.02%
Walk 4,996 194 714 4 394,366 98.52%
Precision % 77.84% 56.25% 91.57% 86.87% 95.50% 93.45%
100% of labeled data Predicted Mode Recall %
Real Mode Bike Bus Car Train Walk
Bike 67,640 467 6,004 7 26,767 67.04%
Bus 3,284 1,220 3,217 5 4,864 9.69%%
Car 4,961 5 152,487 125 1,815 95.66%
Train 555 0 3,017 1,527 177 28.94%
Walk 11,921 443 1,396 13 787,261 98.28%
Precision % 76.54% 57.14% 91.79% 91.05% 95.90% 93.60%
As listed in Table 5, the proposed model predicts 152,487 car, and 787,261 walk correctly, while it fails to precisely predict the bus mode in all proportion of the labeled dataset. The precision values are 91.79% and 95.90% for car and walk, respectively. Moreover, the recall values are 95.66% (car) and 98.28% (walk) for these two transport modes. While the precision and recall values for the bus are 57.14% and 9.69%, respectively.
The confusion matrices show that the walk mode is detected more accurately since it has a larger number of point data in the training dataset. Moreover, Walk-bike and also bus-car are easily misclassified since these modes have some similarities in the motion, speed, and road network. Moreover, the model identifies car mode with higher accuracy compared to other
21 motorized modes due to more labeled data than bus and train in the dataset and different trip characteristics with walk and bike. The bus is mostly misclassified as a car due to similarities in the similar movement patterns. The bus is also misclassified as the bike. This happens due to the low bus speed in the city streets and the acceleration and turning rate when the bus is close to traffic lights. The train is also misclassified as a car due to some similarities in acceleration and deceleration.
4.4. Model evaluation based on the validation dataset
The trained model is applied to the validation dataset to predict transportation modes to assess how the model performed and how the labeling part is accurate. Table 6 lists the confusion matrix of predicted transportation modes.
Table 6. Confusion matrix of the validation dataset
Validation Dataset Predicted Mode Recall %
Real Mode Bike Bus Car Train Walk
Bike 19,375 0 39 78 76 99.01%
Bus 33 8 213 2,087 1,774 0.19%
Car 0 0 54,964 0 0 100%
Train 0 0 1,250 5,533 14,651 25.81%
Walk 0 0 0 0 168,438 100%
Precision % 99.82% 100% 97.33% 71.87% 91.07%
F1-Score 99.42% 0.38% 98.65% 37.98% 95.33% 92.47%
The F1 score for car and walk is 98.65% and 95.33%, respectively that are higher in comparison with other transportation modes. Consequently, higher precision and recall values (as presented in Table 5) result in a higher F1 score, as shown in Table 6. The proposed model results in a low F1 score value of 0.38% for the bus due to the low number of point data related to the bus in the labeled dataset. The train also is misclassified as walk and results in a low recall value of 25.81%. This can be due to the low number of point data for the train and the low speed of train before and after each train station. Another reason can be traffic congestion and the proximity of railways to roads and street sidewalks in the city. The model achieves an F1-score of 92.47%.
The proposed model develops a satisfactory result for the bike, car, and walking. However, the model fails to predict well the bus and train.
4.5.
Comparison with similar studies
Along with examining the proposed model's performance, it is also interesting to compare other studies' results. This study compares the identification accuracy of semi-supervised deep ensemble learning (James, 2020), semi-supervised convolutional autoencoder (SECA) (Dabiri et al., 2019), and LSTM (Asci & Guvensan, 2019) with our study. These three studies are
22 selected due to their similarity in the applied machine learning algorithm. Each of these three studies either used LSTM or an Autoencoder as part of their proposed model. The comparison result is summarized in Figure 8.
As shown in Figure 8, with only 10% of the labeled dataset, the proposed model results in an accuracy of 93.21%, while the proposed model by James, (2020), which is the most similar study to this thesis, obtained 89% accuracy. However, the results for the 100% labeled dataset shows 93.60% (This study) and 91.50% (James, (2020)) accuracy. Consequently, the proposed model by this study has the potential to predict transportation modes with higher accuracy.
Dabiri et al. (2019) used a Convolutional Autoencoder in their model, while this study uses LSTM Autoencoder. Their result showed 77.20% accuracy in 100% labeled data however, this work obtains 93.60% accuracy.
The proposed model performs better than the other three models (Figure 8). Moreover, it is evident from the result that the proposed model gets a higher accuracy than others, even with training on 10% of labeled data. However, the proposed model by James, (2020) performed better in identifying bus and train.
James, (2020) used six stacked LSTM to get the latent information and four different deep neural networks to predict the transportation mode, while this study uses a simple LSTM Autoencoder including four layers of LSTM (2 for encoding, 2 for decoding), combined with a
93.21% 93.31% 93.45% 93.60%
89.00% 90.00% 90.80% 91.50%62.30% 71.60% 72.90% 77.20%60.90% 70.10% 73.50% 81.70%
10% 20% 50% 100%
Proposed Model (James, 2020) (Dabiri et al., 2019) (Asci & Guvensan, 2019)
Figure 8. Comparison of transportation mode identification accuracy
23 deep neural network with three hidden layers. This study shows that the proposed model outperforms the other studies in accuracy with less complexity.
Furthermore, the result shows that only a small proportion of labeled data (10%) can maintain a high level of accuracy. Therefore, having a fully labeled dataset is not mandatory to have a high-performance level of accuracy.
24
5. Conclusion
This study proposes a semi-supervised model to identify transportation modes. This model can handle unlimited unlabeled GPS trajectories. The model employs an LSTM Autoencoder and a simple deep neural network to detect modes of transportation. The deep neural network is proposed to get latent information from the LSTM Autoencoder to generate the label for unlabeled data based on the knowledge of exiting labeled data in the dataset.
To evaluate the performance of the proposed model, this work first assesses the accuracy of the model with the different proportions of the dataset. Second, it compares the predicted result from the validation dataset with the truth table. The result shows that the proposed model can perform notably well in predicting transportation modes for unlabeled data.
The obtained result from model accuracy reveals that the proposed model can predict the transportation mode accurately for car and walk with just 10% of the labeled dataset. However, the bus detection result is not as accurate as other transportation modes due to fewer point data.
Evaluation of the proposed model shows high accuracy for bike mode, besides the car and walk.
Comparing the result of current work with previous studies indicates that the proposed model in this study could successfully predict transportation modes with higher accuracy and less complexity, with just 10% of the dataset.
The findings of this study are useful for the government to develop the city structure based on human behavior accordingly. Moreover, if the accurate mode of transportation of individual users is identified, it is possible to have a more realistic picture of travel demand. Another application of detecting transport mode is the detection of real-time traffic state because companies collect data from GPS datasets in order to estimate the traffic speed on roads.
As future work, the author suggests evaluating the performance of the proposed model with a larger dataset. Moreover, another suggestion for future research can be applying the proposed model in a balanced dataset. Modification of the current model can help obtain a more accurate identification between modes, especially bike, bus, and car. Furthermore, the time complexity of the proposed model can be defined as the required calculation time for the performance of the model with respect to the input data.
25
References
Adler, J. L., & Blue, V. J. (1998). Toward the design of intelligent traveler information systems.
Transportation Research Part C: Emerging Technologies, 6(3), 157–172.
Asci, G., & Guvensan, M. A. (2019). A novel input set for LSTM-based transport mode detection. 2019 IEEE International Conference on Pervasive Computing and Communications Workshops (PerCom Workshops), 107–112.
Biljecki, F., Ledoux, H., & Van Oosterom, P. (2013). Transportation mode-based segmentation and classification of movement trajectories. International Journal of Geographical Information Science, 27(2), 385–407.
Chen, C., Gong, H., Lawson, C., & Bialostozky, E. (2010). Evaluating the feasibility of a passive travel survey collection in a complex urban environment: Lessons learned from the New York City case study. Transportation Research Part A: Policy and Practice, 44(10), 830–840.
Dabiri, S., & Heaslip, K. (2018). Inferring transportation modes from GPS trajectories using a convolutional neural network. Transportation Research Part C: Emerging Technologies, 86, 360–371.
Dabiri, S., Lu, C.-T., Heaslip, K., & Reddy, C. K. (2019). Semi-supervised deep learning approach for transportation mode identification using GPS trajectory data. IEEE Transactions on Knowledge and Data Engineering, 32(5), 1010–1023.
Ebrahimpour, Z., Wan, W., Velázquez García, J. L., Cervantes, O., & Hou, L. (2020).
Analyzing Social-Geographic Human Mobility Patterns Using Large-Scale Social Media Data. ISPRS International Journal of Geo-Information, 9(2), 125.
Fang, S.-H., Fei, Y.-X., Xu, Z., & Tsao, Y. (2017). Learning transportation modes from smartphone sensors based on deep neural network. IEEE Sensors Journal, 17(18), 6111–
6118.
Feng, T., & Timmermans, H. J. P. (2016). Comparison of advanced imputation algorithms for detection of transportation mode and activity episode using GPS data. Transportation Planning and Technology, 39(2), 180–194.
Frappart, F., & Bourrel, L. (2018). The Use of Remote Sensing in Hydrology. MDPI- Multidisciplinary Digital Publishing Institute.
Gong, H., Chen, C., Bialostozky, E., & Lawson, C. T. (2012). A GPS/GIS method for travel mode detection in New York City. Computers, Environment and Urban Systems, 36(2), 131–139.
Gong, L., Morikawa, T., Yamamoto, T., & Sato, H. (2014). Deriving personal trip data from GPS data: A literature review on the existing methodologies. Procedia-Social and Behavioral Sciences, 138(0), 557–565.
Jackson, B., Scargle, J. D., Barnes, D., Arabhi, S., Alt, A., Gioumousis, P., Gwin, E., Sangtrakulcharoen, P., Tan, L., & Tsai, T. T. (2005). An algorithm for optimal partitioning of data on an interval. IEEE Signal Processing Letters, 12(2), 105–108.
James, J. Q. (2018). Two-stage request scheduling for autonomous vehicle logistic system.
IEEE Transactions on Intelligent Transportation Systems, 20(5), 1917–1929.
26 James, J. Q. (2020). Semi-supervised deep ensemble learning for travel mode identification.
Transportation Research Part C: Emerging Technologies, 112, 120–135.
James, J. Q., & Lam, A. Y. S. (2017). Autonomous vehicle logistic system: Joint routing and charging strategy. IEEE Transactions on Intelligent Transportation Systems, 19(7), 2175–
2187.
Killick, R., Fearnhead, P., & Eckley, I. A. (2012). Optimal detection of changepoints with a linear computational cost. Journal of the American Statistical Association, 107(500), 1590–1598.
Lari, Z. A., & Golroo, A. (2015). Automated transportation mode detection using smart phone applications via machine learning: Case study mega city of Tehran. Proceedings of the Transportation Research Board 94th Annual Meeting, Washington, DC, USA, 11–15.
Li, Z., Yuan, C., Lu, J., & Kerre, E. E. (2020). Developments Of Artificial Intelligence Technologies In Computation And Robotics-Proceedings Of The 14th International Flins Conference (Flins 2020) (Vol. 12). World Scientific.
Lin, M., & Hsu, W.-J. (2014). Mining GPS data for mobility patterns: A survey. Pervasive and Mobile Computing, 12, 1–16.
Manzoni, V., Maniloff, D., Kloeckl, K., & Ratti, C. (2010). Transportation mode identification and real-time CO2 emission estimation using smartphones. SENSEable City Lab, Massachusetts Institute of Technology, Nd.
McGowen, P., & McNally, M. (2007). Evaluating the potential to predict activity types from GPS and GIS data. Transportation Research Board 86th Annual Meeting, Washington.
Nguyen, H. D., Tran, K. P., Thomassey, S., & Hamad, M. (2020). Forecasting and Anomaly Detection approaches using LSTM and LSTM Autoencoder techniques with the applications in supply chain management. International Journal of Information Management, 102282.
Nikolic, M., & Bierlaire, M. (2017). Review of transportation mode detection approaches based on smartphone data. 17th Swiss Transport Research Conference, CONF.
Noulas, A., Scellato, S., Lambiotte, R., Pontil, M., & Mascolo, C. (2012). A tale of many cities:
universal patterns in human urban mobility. PloS One, 7(5), e37027.
Rasmussen, T. K., Ingvardson, J. B., Halldórsdóttir, K., & Nielsen, O. A. (2015). Improved methods to deduct trip legs and mode from travel surveys using wearable GPS devices: A case study from the Greater Copenhagen area. Computers, Environment and Urban Systems, 54, 301–313.
Rezaie, M., Patterson, Z., Yu, J. Y., & Yazdizadeh, A. (2017). Semi-supervised travel mode detection from smartphone data. 2017 International Smart Cities Conference (ISC2), 1–8.
Sauerländer-Biebl, A., Brockfeld, E., Suske, D., & Melde, E. (2017). Evaluation of a transport mode detection using fuzzy rules. Transportation Research Procedia, 25, 591–602.
Scheiner, J., & Holz-Rau, C. (2007). Travel mode choice: affected by objective or subjective determinants? Transportation, 34(4), 487–511.
Shin, D., Aliaga, D., Tunçer, B., Arisona, S. M., Kim, S., Zünd, D., & Schmitt, G. (2015).
Urban sensing: Using smartphones for transportation mode classification. Computers, Environment and Urban Systems, 53, 76–86.
