Energy Efficiency and Power Consumption Improvement of IR Illumination for Surveillance Cameras

(1)

Energy Efficiency and Power

Consumption Improvement of IR

Illumination for Surveillance

Cameras

CARLOS TORMO LLUCH

K T H R O Y A L I N S T I T U T E O F T E C H N O L O G Y

(2)

Master of Science Thesis

KTH School of Information and Communication Technology SE-100 44 STOCKHOLM, SWEDEN

Energy efficiency and power

consumption improvement of IR

illumination for surveillance cameras

(3)

-ii-

Master of Science Thesis

Energy efficiency and power consumption improvement of IR illumination for surveillance cameras.

Carlos Tormo

Approved Examiner Supervisor

(4)

-iii-

Abstract

The power and energy optimization of a device can lead to a reduced cost, smaller area, better temperature performance, and higher lifetime. Furthermore, in systems that have limited power budget, it allows running simultaneously more functionalities or using features that require higher power demand.

Therefore, both from the user and the company perspective, the value of a product increases as the energy optimization improves.

For nighttime surveillance video recording, it is common to use infrared illumination to light the target scene, which draws a significant portion of the total camera energy consumption. This master thesis examines and discusses how stroboscopic infrared illumination can enhance the energy efficiency in video recording cameras with rolling shutter image sensors. This report analyzes LED driver circuits,

recommends methodologies, and sorts the most relevant parameters to help to dimension and design the illumination system for a light-strobing system. A promising field of use for this technique has been found to be the license-plate recognition (LPR) scenario, for which this thesis dedicates a chapter in this

document.

This project has been developed at AXIS Communications, where a prototype has been built for one of their network security cameras. The prototype has been tested for LPR for both strobing light systems and conventional IR lighting systems. The results obtained prove that the energy efficiency of the illumination system can be improved more than 95% when stroboscopic illumination is used.

(5)

-iv-

Sammanfattning

Effektförbrukning och energioptimering av en produkt kan leda till lägre kostnad, mindre storlek, bättre temperaturprestanda och högre livslängd. I system med begränsad effektbudget möjliggör detta dessutom aktivering av fler funktioner samtidigt, eller användning av funktioner med högre strömförbrukning. Därmed gör energioptimering att produktens värde ökar både för användaren och för företaget som tillverkar den.

För videoinspelning med övervakningskamera nattetid är det vanligt att använda infraröd belysning för att belysa scenen, vilket ofta förbrukar en betydande del av kamerans totala effektbudget. Detta examensarbete undersöker och diskuterar hur blixtrande (Eng. strobed) infraröd belysning kan förbättra energieffektiviteten vid videoinspelning med bildsensorer med rullande slutare. I denna rapport analyseras LED-drivkretsarna, metodik rekommenderas samt att de mest relevanta parametrarna för att dimensionera och designa ett belysningssystem baserad på strobed IR-belysning sorteras ut. Ett lovande användningsområde för denna teknik har visat sig vara LPR-scenariot (License Plate Recognition), vilket diskuteras i ett eget kapitel i denna rapport.

Projektet har genomförts på AXIS Communications, där en prototyp har byggts baserat på en av dess nätverkskameror. Prototypen har utvärderats LPR-sammanhang med både strobed och konventionellt IR belysningssystem. De erhållna resultaten visar att energieffektiviteten hos belysningssystemet kan förbättras med mer än 95% när blixtrande belysning används.

(6)

-v-

Acknowledgments

I feel in debt with all the people that accompanied me along the duration of this project. First, to Steve, with whom I have worked and discussed for many hours. Also, to Johan, Christian, and Anders. They have all helped to turn problems into solutions, work into fun, and time into an outstanding experience. As this project has been a small piece of a vast puzzle, I would like to sincerely thank all the other more experienced puzzle builders that helped me: Ola, Jenny, and Andreas. Also, special thanks to Mark T. Smith, my examiner, for his willingness to help and his excellent advice.

Sweet thanks to Malte, Fei, Karolis, Hang, and Martin, for sharing a few well-deserved fikas.

(7)

-vi-

Abbreviations

CMOS Complementary Metal-Oxide-Semiconductor

DUT Device Under Test

EMI Electromagnetic Interference

EIT Extended Integration Time

LED Light-Emitting Diode

LET Long-Exposure Mode

LPR License-Plate Recognition

OF Overpower Factor

PAPR Peak-to-Average Power Ratio

PD Powered Device (Context: PoE)

PLD Programmable Logic Device

PoE Power over Ethernet

PSE Power Sourcing Equipment (Context: PoE)

IC Integrated Circuit

IR Infrared

IS Illumination System. The system composed of the LED driver circuit, LEDs, and additional optics.

STS Shared-time Strobing

(8)

-vii-

1. Introduction

1.1 Introduction: Surveillance cameras and IR illumination

AXIS Communications is a Swedish company whose core-business is video-surveillance network cameras. Video-surveillance applications often imply nighttime video recording, which usually uses IR illumination to light the target scene without disturbing users and reducing camera’s intrusiveness. Since AXIS cameras are connected to the network through ethernet cables, AXIS Communications has chosen to power them via PoE, thus easing the installation of the security system. However, PoE implies a limited available power for the cameras, which depends on their PoE class. While upgrading the PoE class is a possible solution to high power demands, it increases the production cost of the camera from the company’s perspective, as well as the installation cost from the user’s perspective, since the PSE´s higher power capabilities will be translated into a higher cost.

According to AXIS Communications (personal communication, May 2018) IR illumination can represent almost a half of the camera’s total consumption. Because of this, applications that require surveillance over large areas can often require the installation of separate IR LED lamps for those cases where the camera and the illumination system consumption exceed the power supply capabilities. Thus, optimizing the IR illumination system of the cameras is desirable, since it can ease the surveillance system’s installation and leave more room for additional image processing (more and/or more power demanding features in the cameras). Furthermore, systems with optimized energy consumption dissipate less heat, which might mean longer operation time when compared to systems that tend to overheat and need to reduce performance to cool themselves.

1.2 Background: The strobing light technique concept

Video-surveillance cameras that use IR illumination for dark environments recording emit a constant beam of light onto the area under surveillance. Hereafter, we will refer to this technique as constant-light illumination technique.

However, cameras are only sensitive to light when their shutter is open, that is, when light flows into the lens and excites the image sensor. The shutter time, which is the time that the shutter is open, is adjusted so that the output stream of frames has the correct exposure and no blur. In very bright scenes or when fast moving objects want to be precisely captured, the shutter time tends to be short. As can be seen in Figure 1, there is a fraction of the emitted light, that depends on the frame period and the shutter time, which is not utilized.

The exposure of the resulting image is also affected by the ISO and the aperture of the lens, but the selection of these parameters has multiple implications on the image quality other than the brightness of the image and thus is out of the scope of this thesis.

(11)

-10-

Therefore, in that situation, if the constant-light technique were being utilized, 95% of the emitted light would be wasted, since it would not be captured by the sensor. That is, all the energy required to emit light during that closed-shutter time, is energy wasted.

This reasoning brings us to a conclusion, instead of emitting light constantly, it would be more efficient to emit light only when the shutter is open. Hereafter, we will call this technique, strobing-light technique. The implementation of the strobing-light technique will be affected by the type of image sensor used. There are two types of shutter mechanisms for image sensors. The more intuitive one is the global shutter sensor, in which all the pixels are exposed to light at the same time. The less intuitive one is called rolling shutter sensor, in which each row of pixels exposure-time starts sequentially one after the other, delayed by 1H. ‘H’ units measure time as a function of the configured output data-rate of the sensor (more information about this can be read in “Appendix A: The ‘H’ unit”). In Figure 2, both shutter techniques are shown.

Figure 2 Rolling shutter and global shutter for a 6-row image (6-pixel height image).

As we can see in Figure 2, rolling shutter time extends the period in which one frame is being captured. This does not affect the number of frames per second of the camera, since different rows’ exposure can be overlapped in time, but affects the ratio of open-shutter time to frame period, and thus, the energy improvement between the two techniques. For the example in Figure 2 and a frame period of 20H, the strobing-light technique would be able to reduce the energy consumption by 20% for the rolling shutter and by 45% for the global shutter when compared to the constant-light technique. Although the shutter is open for the same amount of time per each row in both types of shutter techniques, the rolling shutter sensor requires a longer pulse of light (16H) than the global shutter sensor (11H) to achieve the same image brightness.

(12)

-11-

Figure 3 Effects of an unsynchronized light pulse in rolling shutter and global shutter sensors. The resulting image (car) in the rolling shutter sensor looks darker at the bottom since those pixels have been exposed to light for a shorter time than those above them. In the global shutter sensor, an

unsynchronized pulse only means a darker image when compared to a properly synchronized one.

Stroboscopic illumination effects on video-recording cameras are discussed in [1] for the first time. The goal of the report is to use synchronized pulses of light to remove the rolling shutter distortion, i.e., the distortion caused by each row capturing the image at different intervals of time. Furthermore, the report aims to synchronize the emitted light with several cameras. The method proposed in this report is used by companies like “Photometrics” and “Qimaging” to avoid the rolling shutter distortion.

For different purposes, J. C. Lee in [2] claims a similar method to the one presented by D. Bradley, B. Atcheson, I. Ihrke, and W. Heidrich in [1]. However, the method is intended to allow using modulated light for depth sensing in rolling shutter sensors.

Finally, US9854193[3] describes a method to improve the power efficiency and extend the useful life of products based on strobing light. Additionally, the patent claims the use of high-speed CMOS sensors to reduce the rolling shutter distortion.

The method in [3] consists of one or more sources of stroboscopic light, which are synchronized sequentially to yield a wider light pulse. Furthermore, the patent proposes to light the scene temporarily and discard the dark frames. To end with, the patent defines a region of interest as the part of the captured frame that contains useful information. By knowing this, they propose to exclusively light this part, so that no energy is wasted lighting useless parts.

This thesis focuses on the same purpose as [3] but with methods inspired by the idea proposed in [1]. In addition to the power and energy analysis, this thesis also studies the design of strobing-light illumination systems for rolling shutter sensors.

1.3 Goals and scope of the project

This master thesis is the result of AXIS Communications’ will to find how to increase the efficiency of the IR illumination system. Of foremost priority was to investigate how strobing the IR light could be beneficial when compared to constant-light illumination, i.e., how pulsing the light in a proper way could lead to better quality image or better energy/power performance. While image quality is not the main concern of this thesis, it is certainly a determining factor to guide the energy and power efficiency study.

The goals for this project as perceived by AXIS were:

- Determining the strobing-light technique advantages when compared to constant-light technique. - Developing a prototype that could be used as a proof of concept.

(13)

-12-

reasons, the model of the camera is not mentioned in this report and is referred to as the device under test (DUT). Also, the LED driver design will be mainly oriented to fulfill its characteristics. The circuit was chosen to operate at 12V, since it is a common available internal voltage in most of the cameras.

The image sensor on the DUT cannot either be disclosed. However, it is just necessary to know that it is a rolling shutter sensor with different available frame rates and resolution modes.

The completion of this project should provide the necessary data to determine whether the strobing-light technique could be useful or not, propose use cases, and provide a prototype that can be used as a proof of concept.

1.4 Methodology

The accomplishment of the goals requires an initial research, techniques analysis, testing, and prototyping. Since the most important goal is the prototype implementation, an iterative process will be followed, where the first iteration will aim to deliver a basic but fully functioning prototype. This results in a short research period at the beginning of project that sets the basic requirements for the prototype. After this short period, an initial prototype can be implemented and the concept can be tested. The main reason to follow this approach is to prioritize to have a working prototype by the end of this thesis. Since the development platform is complex, diverse, and unknown at the beginning of the project, prioritizing to have an early prototype reduces the risk of failing to achieve the main goal.

The initial research will include the state of the art of the technique and theory about the current image sensor used in the DUT. Developing the prototype will require microcontroller programming in C language, and PLD programming and testing with System Verilog. Also, it will be necessary to select or design an electronic circuit that fulfills the requirements set during the initial research. Evidently, the DUT will need to be reworked to include this new circuitry.

For testing purposes, the System Verilog hardware description code will be evaluated with testbenches. The microcontroller code will be empirically tested since the functionalities required by it will be simple. The proper functionality of the prototype will be tested at two levels: in the laboratory, by checking all the signals with an oscilloscope; and in a real-case scenario, by evaluating the recorded video.

The discussion about the circuitry of the LED driver will be supported by electronic circuit design theory, measurements on the prototype, and LTspice simulations, an electronic simulation tool from Analog Devices, which is offered for free at its website (Windows and Mac operative systems versions).

(14)

-13-

2. Strobing-light technique and constant-light technique

comparison

2.1 Constant-light analysis

In this chapter, we will compare the constant-light technique, the WFS technique, and the STS technique. We will evaluate and extract equations to describe energy efficiency. For this purpose, we need to start by evaluating the constant-light technique, so that we can later compare it to the strobing-light techniques. All the analysis targets rolling shutter sensors, since this is the type which is used at AXIS. Some of the equations might include specific sensor limitations of the image sensor of the DUT since they will be used later for our scenario.

In Figure 4, time has been discretized in slots of 1H unit each. The shutter timing (red colored) is the time in which the row exposure is being activated. The readout time (blue colored) is the time in which the exposure has finished, and the data of all the pixels in the row is sent to the processing unit. The time in between the shutter time and the readout time is the integration time and is set according to the image captured. The longer the integration time is, the brighter the output image will become. On the other hand, long integration times can produce blurry images if the recorded objects move too fast.

Figure 4 Constant-light technique example for two consecutive frames. For this example, the period (T) is 14H, and the integration time (Ti) is 5H.

From the example in Figure 4, we can see that light is emitted during the whole period (T=14H), while it is only being used for 5H units of time per row (Ti=5H). Hence, the light utilization is 35%, which is calculated

by using Equation 1. The time unit used in this section is the ‘H’ unless specifically stated. 𝑈𝐶𝐿𝑇 =

𝑇𝑖

𝑇 [1]

Equation 1 shows that the utilization of light is directly proportional to the integration time and that, when 𝑇𝑖 ≅ 𝑇, the constant-light technique utilization is almost 100%. We can calculate the maximum utilization

for the constant-light technique (𝑈𝐶𝐿𝑇 𝑚𝑎𝑥) with Equation 2. The 2H units of time that cannot be utilized

are due to the time required for the shutter and readout (limitations of the DUT image sensor). 𝑈𝐶𝐿𝑇 𝑚𝑎𝑥 =

(15)

-14-

By using Equation 2, we can calculate the maximum utilization in a constant-light scenario when the integration time is maximum (ti MAX). For this example, we will assume the sensor is working in Full HD

1080p operating mode, i.e., 1125 total number of vertical lines. 𝑈𝐶𝐿𝑇 = 𝑡𝑖 𝑀𝐴𝑋 1125𝐻= (1125 − 2)𝐻 1125𝐻 = 1123 1125= 99,8%

2.2 Whole-frame strobing and Shared-time strobing

In the following sections, the strobing-light technique will be discussed and studied, so that we can expose its advantages and disadvantages against the constant-light technique. Again, the analysis targets rolling shutter sensors. Note that, for the analysis in this chapter, we refer to the shutter state as either “open” or “closed” as an analogy to mechanical shutter systems. However, the DUT’s image sensor shutter is not controlled mechanically, but electronically.

Two different types of strobing techniques have been identified as possible candidates: the whole-frame strobing technique (WFS), and the shared-time strobing technique (STS). WFS consists of illuminating the scene during the whole time that the shutter is open, while STS illuminates only during the fraction of the time that all the pixels are sensitive to light (Figure 5).

Figure 5 Two different types of strobing are shown: whole-frame strobing (WFS), and shared-time strobing(STS).

The time slots which are exposed to light are shown in green, while those which are not, are white. The addition of all the green and white time slots for one row is called integration time. To ease the analysis and discussion, we have defined the shared time interval, which is a new term that gives its name to the STS technique. The shared time is the interval for which all the rows (and all the pixels) are exposed simultaneously to light. These terms can be seen in Figure 6.

(16)

-15-

From Figure 5, we can see that the STS technique yields a higher utilization of the emitted light (more efficient lighting) than the WFS technique. This is because all the light emitted in STS is being captured by all the pixels of the sensor, whereas in WFS, this does not happen. We can also observe that STS might require high peaks of power since the light must be squeezed into the shared time.

In the following sections, we will analyze the power and energy improvement for both WFS and STS when compared to the constant-light technique.

2.3 WFS analysis

As we have already mentioned, WFS does not offer a better light utilization compared to STS, but this can be the case when compared to the constant-light technique. This will happen in those configurations in which there are time windows when no row is being exposed to light. Hereafter, we will call that time dead

time.

In Figure 7, we can see an example of WFS. In this example, the integration time is 5H (𝑇𝑖), the shared time

is 1H (𝑇𝑆), the dead time is 5H (𝑇𝐷), and the time period is 14H (𝑇). For this situation, the IR light is on for

9H and off for 5H units. The light utilization of the WFS technique can be calculated as in Equation 3, which is 55%, which is greater than 𝑈𝐶𝐿𝑇 = 35%. Thus, the efficiency improvement compared to the

constant-light technique (𝜂𝑊𝐹𝑆) is 35%, which can be calculated as in Equation 4.

𝑈𝑊𝐹𝑆= 𝑇𝑖 𝑇𝑖+ (𝑁𝑅− 1) [3] 𝜂𝑊𝐹𝑆 = 𝑇 − (𝑇𝑖+ (𝑁𝑅− 1)) 𝑇 = 𝑈𝑊𝐹𝑆− 𝑈𝐶𝐿𝑇 𝑈𝑊𝐹𝑆 [4] 𝑁𝑅: Number of rows.

Figure 7 Two consecutive frames illuminated with the WFS technique.

Note that for this technique, 𝑈𝑊𝐹𝑆 𝑚𝑎𝑥= 𝑈𝐶𝐿𝑇 𝑚𝑎𝑥when TD is null. Also, 𝑈𝑊𝐹𝑆> 𝑈𝐶𝐿𝑇 when the frames do

not overlap in time with each other, while 𝑈𝑊𝐹𝑆= 𝑈𝐶𝐿𝑇 = 𝑈𝐶𝐿𝑇 max when they do, since the WFS technique

will behave as the constant light technique (no deadtime).

2.4 STS analysis

Energy efficiency

(17)

-16-

While the constant-light technique and the WFS technique emit light outside the shared time, the STS does not. Because of this, the light utilization of STS is always 100% (𝑈𝑆𝑇𝑆 = 100%) and the energy

improvement can be written as in Equation 5.

𝜂𝑆𝑇𝑆= 1 − 𝑈𝐶𝐿𝑇 [5]

However, STS technique involves an important limitation: light can only be emitted during the shared time. Because of this, to maintain a certain level of brightness, an IS using the STS technique might need to increase the power, i.e., the exposure time reduction is compensated by a power increment.

Power demand

In this section, we will discuss how the power demand is affected when applying the STS technique. In Figure 8, a simple 5-line frame illuminated by STS is shown. It can be concluded that, when the integration time is equal to the number of rows, a 1H interval of shared time will be available.

Figure 8 Number of lines to achieve 1H interval of shared time.

Equation 6 is the general formula to calculate the available shared time (TS) in a scenario with ‘NR’ number of

lines and ‘Ti’ H units of integration time.

𝑇𝑆= 𝑇𝑖− 𝑁𝑅+ 1 [6]

For a 1110-row frame (1125 total lines1_{), the maximum allowed integration time is 1123H. Because of this,}

the maximum available shared time (𝑇𝑆 𝑀𝐴𝑋@1125) will be of 14H units. This number does not depend on the

frame rate since in one frame there are always the same number of ‘H’ units. Also, we can conclude that when 𝑇𝑖≤ 𝑁𝑅− 1 = 1110 − 1 = 1109𝐻, there will be no shared time available and thus, the STS technique

cannot be used. For the power demand analysis, we will not consider those cases in which 𝑇𝑖 ≤ 1109𝐻,

since as we demonstrated, there’s no shared time available.

For each row, the STS technique must fit the same energy (E’) as the constant light technique (E) during TS

to achieve the same brightness. The power relation between the STS scenario and the constant-light one, hereafter called overpower factor (OF), can be calculated as it is shown in Equation 7. In Graph 1, OF equation

is plotted over the valid Ti values.

𝑂𝐹 =𝑃 ′ 𝑃 = 𝐸′ 𝑇𝑆 𝐸 𝑇𝑖 𝐸=𝐸′ → 𝑃 ′ 𝑃 = 𝑇𝑖 𝑇𝑆 [7]

Where P’ and P are the power demand of the STS technique and constant-light technique, respectively.

1_{For the image sensor of the DUT, 1110 pixel-rows contain useful/valid image information (valid lines), while the}

(18)

-17-

Graph 1 Overpower factor (OF) for STS against constant-light technique as a function of the integration time. The function is plotted for an image sensor configuration of 1125 rows.

For the full-HD capture mode (1125 total rows and 1110 valid ones), the worst-case scenario occurs when TS is minimum (1H) at Ti=1110H, in which OF=1110. The best-case scenario occurs when the TS is

maximum at Ti=1123H, in which OF=80. From this graph and the previous calculations, we can see that

STS presents significant issues regarding the power demand.

2.5 Extended Integration Time (EIT) technique

Introduction

Although the STS technique looked promising from the energy efficiency perspective, there are severe issues regarding the OF. Furthermore, we found that it can only be used under a very restrictive condition: 𝑇𝑖>

𝑁𝑅− 1. In this section, we propose and explore the Extended Integration Time (EIT) technique, which

aims to improve the performance of STS.

As it names indicates, the EIT technique extends the integration time to create a shared time interval long enough to fit the strobing light pulse. In Figure 9, an example of the application of the EIT is shown. As it can be seen, the new integration time (Ti’=5H) is longer than the initial one (Ti=1H). As we previously

calculated, we can only have 14H of shared time. Because of this, we can only apply EIT for initial integration times of 1H to 14H, when Ti>14H, the integration time will already be at its maximum value.

Figure 9 EIT technique. Initially, the integration time is 1H and thus, every line is exposed to light during 1H. After applying the EIT technique, the integration time is increased to 5H and light is only emitted for 1H time interval, which yields the same image brightness but 100% of light utilization.

Overpowering the LED driving circuit

(19)

-18-

Hereafter, we will refer to this technique as the overpowering technique. For example, if the driving circuit can provide four times more power (overpower factor of 4, 𝑂𝐹 = 4), the EIT can be used until the integration time is 56H, i.e., 14H x 4. That is, the EIT applicability threshold has been increased to 56H.

This poses a technical challenge not only on the capability of providing the power but also on the capability of storing that additional power so that the peak power consumption is not increased.

Long-Exposure technique (LET)

In the previous section, we have seen how to increase the EIT applicability threshold by overpowering the LED driving circuit. In this section, we will look at how to increase even more the EIT threshold by using the long-exposure mode of the image sensor of the DUT. This mode allows to have more integration time at the cost of reducing the fps. However, the readout time is not increased. This results in more shared time than if the frames per second were directly reduced.

When the target integration time is above our EIT applicability threshold, the longer-exposure mode can be used. The new EIT applicability threshold (EITt) will be significantly increased and can be calculated

with Equation 8.

𝐸𝐼𝑇𝑡= 𝑡𝑖 𝑀𝐴𝑋· (

𝑓𝑝𝑠

𝑓𝑝𝑠′− 1) + 𝑡𝑆 𝑀𝐴𝑋 [8]

𝑡𝑖 𝑀𝐴𝑋 [H]: maximum integration time that it is allowed in the new mode.

𝑓𝑝𝑠: frames per second of the previous mode. 𝑓𝑝𝑠′: frames per second of the new mode.

𝑓𝑝𝑠

𝑓𝑝𝑠′: frames per second ratio, which indicates the fps reduction after applying the long-exposure

technique.

𝑡𝑆 𝑀𝐴𝑋 [H]: maximum shared time of the previous mode (14H in the 1125-line mode).

We refer to this technique as the long-exposure technique (LET). For example, in a 1125-line scenario running at 120fps, the EIT threshold would be 14H. By using EIT and LET at half the fps (𝑓𝑝𝑠′ _{= 60𝑓𝑝𝑠), the}

EIT applicability threshold could be increased to:

𝐸𝐼𝑇𝑡= 1123 · (

120

60 − 1) + 14 = 1137𝐻 = 1137 · 7,41𝜇𝑠 = 8,425𝑚𝑠

This new obtained EITt is much longer in time than the EITt at 60fps, which would be:

𝐸𝐼𝑇𝑡 60𝑓𝑝𝑠= 14𝐻 = 14 · 14,82𝜇𝑠 = 0,207𝑚𝑠

2.6 Constant-light dimming at maximum integration times

Outside the strobing-light techniques repertoire, dimming a constant light source can be a very efficient way of illumination. We have already seen that the efficiency of constant-light illumination is nearly 100% when the integration time is close to the maximum value for its frame rate.

Using this fact to our advantage, if the integration time is kept maximum and the image brightness is controlled by dimming the light source, the efficiency will remain almost 100% always. Also, the energy consumed is spread over a longer time, which results in lower peak power.

This technique can be combined with the strobing-light techniques: low integration times can be tackled with STS and WFS, while long ones with analog dimming and constant-light illumination.

(20)

-19-

2.7 Light techniques comparison

Figure 10 shows a summary of all the techniques placed in the initial integration time areas that they can be applied. Also, the efficiency of the constant light technique is displayed. Every block represents a technique that can be used to improve the efficiency in its range of initial integration times. Since the blocks represent the strobing-light techniques, the efficiencies for the blocks are always 100%. Some concerns when using each technique are shown inside brackets (e.g., increase of power and frame rate drop).

When two blocks overlap with each other, it means that both techniques can be used in that region. This is the case of a the EIT+overpower region, where the EIT+LET can also be used. If the former technique is used, an increase of power by a 𝑃𝑓 factor will be required, whereas, if the latter is used, a frame drop will

result.

Figure 10 Available techniques depending on the target integration time for Full-HD operating mode.

By properly using the techniques in Figure 10, we can translate any initial integration time into an equivalent integration time that yields the same brightness but achieves virtually a 100% of light utilization. Note that, as we highlighted in the previous sections, for integration times close or higher than the frame period, we might decide to use constant-light illumination since the light utilization is almost 100%. For example, in the example above, we could decide that above Ti=1080H, the constant-light technique should be used,

because the efficiency would always be higher than 96%.

Sometimes, applying the previously described techniques will impact the image quality. This is the case, for example, when any source of light other than the one emitted by the camera’s illumination system is present in the scene. In this situation, extending the integration time with the EIT technique will yield a brighter image. In [10], the effects on the image quality when using this kind of techniques are evaluated, therefore, we will not dwell on the details.

2.8 LED Driver general specifications and design

Introduction

From the previous sections analysis, we conclude that the strobing-light technique imposes additional requirements on the LED driver when compared to the constant-light technique: fast switching capability of the LEDs, and more peak-power demand if overpowering is used.

The edge times are limited by the minimum pulse of light that needs to be emitted, which depends on the camera’s application. During our research, we have found that the shortest integration times considered at AXIS are around 500µs, which are used for license plate recognition. Because of this, we set the rising and falling time requirements one or two orders of magnitude smaller, from 5µs to 50µs.

(21)

-20-

On the one hand, the image quality can be enhanced in those cases in which temperature is considerably reduced, as so will be the thermal noise in the CMOS sensor [4]. On the other hand, the switching of the current of the LED driver will introduce additional noise into the circuit.

For those cases in which overpowering is used, we must make a proper selection of the components so that they can withstand the maximum peak power during the longer output current pulse.

Buck-converters overview

Even though there might be multiple ways of designing an LED driver, one of the most versatile is to use a switching converter LED driver IC. This solution is currently used at AXIS, and thus, it is interesting to evaluate if it would still be appropriate for the strobing-light technique. Switching converter LED drivers can offer output current control, stability control at the output, safety current limiting, overvoltage protection, analog dimming, amongst other functions. Because of this, they are valuable components and often it will not be worth the effort of designing an LED driver with discrete components.

In our case, where we have a 12V supply voltage (𝑉𝑆) available and a single LED, the LED driver IC must

be a buck converter, since the forward voltage of the IR LED (𝑉𝐹) is smaller than the supply voltage. Current ripple in buck-converter LED drivers for video-recording illumination

Current ripple is one of the most common constraints to consider when starting the design of a buck-converter LED driver. For our scenario, we would like to evaluate the effects of current ripple regarding image quality, LEDs forward current, component cost, and electrical noise.

To begin with, it is important to realize which are the consequences of ripple current on image quality. As light hits the camera’s sensor, it is accumulated in every pixel of the sensor, yielding an image in which the light has been integrated over the whole shutter time interval. Buck-converters usually work at switching frequencies in the order of hundreds of kilohertz, while video-recording cameras frame-rate ranges from 25fps to 120fps. That is, for shutter times close to the camera´s frame-rate, the irradiance variation due to the current ripple cannot be seen in the output image stream since the irradiance ripple will be integrated over the whole shutter time. Fast shutter speeds are required when fast objects need to be captured, to avoid blurry images. The fastest shutter speed used at AXIS is 500µs, for LPR purposes, which compared to a slow switching period of 10µs (100KHz), is still 50 orders of magnitude greater. This means that 25 pulses of light will be integrated per frame2_{, yielding no noticeable light exposure difference between frames. Below}

200µs shutter time (10 pulses of light integrated per frame), the buck-converter design should take into consideration the ripple regarding image quality, since the light exposure could significantly differ from frame to frame. The value for the switching frequency (𝑓𝑆𝑊) that allows ‘𝑁’ waves of ripple per frame for a

minimum light pulse of ‘𝑇𝑝 𝑚𝑖𝑛’ seconds can be calculated with Equation 9.

𝑓𝑆𝑊 > 2 ·

𝑁 𝑇𝑝 𝑚𝑖𝑛

[9]

The LED forward current that the LED driver can supply also depends on the ripple. This is because the components have maximum current ratings that cannot be exceeded, thus limiting the peak current that the driver can withstand. For a constant maximum peak-current, the average forward-current decreases as the ripple is increased. This can be tackled by selecting components that can tolerate higher peak currents, but this generally means increasing the cost and area of the circuit.

2_{Note that the period of the ripple is twice the switching period, since one ripple period consists of a rising edge (first}

(22)

-21-

Finally, the ripple current is related to electrical noise generated by the driver circuit. This is due to the change of current demand through the wires, that can generate conducted and radiated interference. As the ripple increases, so will the noise, making EMI problems more likely to appear.

For the above reasons, we can conclude that, even though current ripple does not pose a problem regarding image quality, it will determine the maximum peak current and thus, the cost and area of the driver. Also, because of EMI, it is a smart choice to keep it low.

Edge times in buck-converter LED drivers

As we have seen in the previous sections, the LED driver has a strict requirement on the switching time of the IR LEDs. In this section, we will study the switching time in buck-converter LED drivers. We will assume that the current through an IR LED is proportional to the IR light emitted [5], [6], i.e., evaluating the edge times of the current is equivalent to evaluating the edge times of the emitted light.

In Figure 11, a simplified circuit to evaluate the edge times of a buck-converter LED driver is presented. This model consists of a voltage source (𝑉𝑆𝑊) that is not the supply power source of 12V, but the voltage

that simulates the initial state before the rise and fall edges. It also includes a voltage-controlled switch (𝑆1) that simulates the switching transistor in the buck-converter, an 𝐿1 and 𝐶1 low pass filter (with ideal components), a recirculating diode (𝐷2), and a light-emitting diode (𝐷1). As we said in the Methodology section, we use the simulation tool LTspice to support the discussion presented in this chapter.

Figure 11 Simplified circuit of a buck-converter LED driver to evaluate rise and fall times.

Buck-converter LED drivers act as a controlled constant-current source. For this matter, since there are supposed to be no current load transients, they can be designed without any output capacitor. Thus, the design flow that we propose for the LED driver design is based on this fact: first, the LED driver will be designed as if there was no output capacitance; later, the output capacitor will be added to improve the performance.

For the falling-edge time, the initial voltage will be the one that makes the desired on-state current flow through 𝐷1. For example, for the diode in the simulation with 550mA as forward current, the voltage drop (𝑉𝐹1) is equal to 4V. Assuming no losses in the other components, 𝑉𝑆𝑊 must be of 4V to simulate properly

the starting conditions for the falling-edge time evaluation. The simulation will reach the steady state with 𝑆1 closed, that is, 550mA flowing through 𝐷1 (light-emitting state). When the switch is opened, the current will stop flowing from 𝑉𝑆𝑊 and the storage elements will discharge through 𝐷1 and 𝐷2.

(23)

-22-

This means that the current in the inductor decays at the rate indicated in Equation 10. 𝛾𝐿1= 𝑣𝐿1 𝐿 = 𝑉𝐹1− 𝑉𝐹2 𝐿 ≈ 𝑉𝐹1 𝐿 [10]

𝑣𝐿1 at 𝑡1 can be calculated as 𝑉𝐹1− 𝑉𝐹2= 4 − 0,3 = 3.7𝑉. Because of this, in our example the decay rate is

𝛾𝐿1= 3,7

33µ= 0.1121 𝐴

µ𝑠 . This value is very close to the simulated value of the circuit in Figure 12, which is

0.1122 A/µs.

The approximation holds properly when 𝑣𝐿1 variation is small. For driving voltages (VSW) significantly

greater than the load’s conduction voltage, i.e., when the inductor voltage difference between the initial and final states is high, the linear approximation will not hold as good anymore, and will introduce a considerable error in the equation. For example, when 𝑉𝑆𝑊= 8𝑉, Equation 10 predicts a decay rate of 𝛾𝐿1=

8−0,8

33µ =

0,218𝐴

µ𝑠, while the simulated one is 0,173 𝐴

𝜇𝑠. This situation though is not very common since the LED will

usually operate close to its conduction voltage, allowing us to consider the voltage drop across the inductor to be constant.

Note that, this approximation does not include any effects from the output capacitor (𝐶1). This is because

of the LED driver design flow that we mentioned at the beginning of this subsection: the capacitor at the output is neglected as it is a component that is added only to improve the performance after the driver is designed. Therefore, its capacitance value is chosen so that the energy stored is neglectable when compared to the inductor one.

From this analysis, we can conclude that the fall time will improve (shorten) if: - The recirculating diode (𝐷2) forward voltage is lowered.

- The forward voltage of the load LEDs is increased. - The inductor value is reduced.

For the rise time, the analysis can be performed the same way. In Figure 12 a simplified circuit for the buck-converter rise time analysis is shown. Note that in the circuit, the output capacitor has not been considered yet. In this situation, ‘S2’ has been added to ensure that ‘D1’ was completely discharged at the beginning of the rising edge and helps no purpose other than setting the initial conditions for the simulation.

Figure 12 Buck-converter LED driver simplified circuit for the rise time analysis. ‘S1’ is the switch that controls the rise transient, while ‘S2’ is the switch that ensures that the LED is completely discharged at the beginning of it.

By considering the voltage across the inductor constant, we can approximate the increasing rate (𝛿) as in Equation 11.

𝛿 =𝑣𝐿1 𝐿 =

𝑉𝑆𝑊− 𝑣𝐷1

(24)

-23-

𝑣𝑆𝑊: supply voltage seen by the inductor during the rising edge.

𝑣𝐷1: forward voltage-drop on the LED load (approximately constant).

For 𝐿 = 66𝜇𝐻, 𝑉𝑆𝑊= 12𝑉, 𝑉𝐹 𝐷1= 2,9𝑉, Equation 11 yields a 𝛿 = 0,137 𝐴

𝜇𝑠, which is very close to the value

obtained by simulation: 0,133𝐴

𝜇𝑠 over the operating currents of the LED.

However, for the rise time analysis, it is important to consider the effects of the output capacitor. Figure 13 shows the simplified circuit of the rising edge scenario including the output capacitor (C1).

Figure 13 Buck-converter LED driver simplified circuit for the rise time analysis with output capacitor (C1).

For this circuit, the evolution of the most important signals are plotted in Figure 14 (𝐿 = 66𝜇𝐻, 𝐶 = 1𝜇𝐹, 𝑉𝑆𝑊 = 12𝑉, 𝑉𝐹 𝐷1= 2,9𝑉). As it can be seen, the LED starts conducting after approximately 6µs

(point A in Figure 14), since the capacitor is storing all the energy. At point A, the capacitor’s voltage reaches the LED’s forward voltage (𝑉𝐹 = 2,9𝑉) and thus, the LED starts to conduct. During the interval A→B, the

LED’s current increases rapidly (faster than the increase rate of the inductor). At point B, the current of the LED and the one provided by the voltage source (the same as the one flowing through the inductor) are virtually the same.

Figure 14 Most relevant signals in the buck-converter LED simplified driver circuit for the rise time analysis. Point ‘A’ in time indicates when the LED starts to conduct current, i.e., its voltage exceeds the conduction voltage. Point ‘B’ indicates when the LED’s current is virtually the same as the

inductor’s one.

(25)

-24-

introduced by the capacitor (𝑡𝑑= 𝐴). However, in terms of current rise time, i.e., without including 𝑡𝑑, the

capacitor boosts the increase rate of the LED current, since it allows the current in the inductor to increase steadily until its accumulated charge builds a voltage greater than the LED’s conduction voltage (point A). In the 0→B time interval, the LED starts conducting, reducing its impedance as the voltage gets higher. At point B, the diode will virtually behave like a very small resistor, whereas the capacitor will behave like an open-circuit, since the ‘dv/dt’ will be almost null.

In the following paragraphs, we will find the parameters to which point ‘A’ is dependent on, this way we will be able to predict the capacitor delay. We will also find a way to estimate the rise time and point ‘B’. Hence we will be able to extract the current value for which the step appears (𝐼𝑠𝑡𝑒𝑝 = 1,3𝐴 in Figure 14).

For LED currents greater than 𝐼𝑠𝑡𝑒𝑝 the current-step shape will be observed, whereas lower current will not

and will yield a faster rise edge. Our ultimate goal is not to find exact values, but the relation of the variables that influence those values, so that better design decisions can be taken when designing the driver.

To calculate the output-capacitor delay, we will assume that all the components are ideal and have no losses. This way we can conclude that the energy stored in the capacitor (𝐸𝐶1) and the inductor (𝐸𝐿1) at

point ‘A’ (𝑡𝐴) is the total energy provided from 𝑡0→𝐴 by the power supply. The following 5 steps are used

to find 𝑡𝐴 and the current at 𝑡𝐴 (𝐼𝑡_𝐴).

(I) Energy in the inductor (tA=A): 𝐸𝐿1= 1 2𝐿𝐼

2₌1

2· 𝐿 · (δL1 0→tA· 𝑡𝐴) 2

Where δL1 0→tA is the current increase rate through the inductor from t=0 to t=A. This

current rate has been approximated to the value shown in Equation 12. Care should be taken when 𝑉𝑆𝑊 ≈ 𝑉𝐹, since this linear approximation will be less accurate.

𝛿𝐿1 0→𝑡_𝐴 =

(𝑉𝑆𝑊−𝑉_{2 )}𝐹

𝐿 [12]

(II) Energy in the capacitor (tA=A):

𝐸𝐶1 = 1 2𝐶𝑉 2₌1 2· 𝐶 · 𝑉𝐹 2 (III) Energy provided by VSW (0→A):

𝐸𝑆𝑊 = ∫ 𝑃𝑆𝑊 𝑡𝐴 0 𝑑𝑡 = ∫ 𝑉𝑆𝑊· 𝐼 𝑡𝐴 0 𝑑𝑡 = 𝑉𝑆𝑊∫ 𝐼 𝑡𝐴 0 𝑑𝑡 = 𝑉𝑆𝑊∫ δL1 0→tA· 𝑡 𝑡𝐴 0 𝑑𝑡 = 𝑉𝑆𝑊· δL1 0→tA· (𝑡𝐴) 2 2 𝑃𝑆𝑊: Power of the voltage supply VSW.

(IV) Energy conservation:

𝐸𝑆𝑊 = 𝐸𝐿1+ 𝐸𝐶1 (V) In Equation 13 we use (I) to (IV) to find 𝑡𝐴.

(26)

-25-

In Figure 15 the circuit has been simulated for VSW=12V, C1=1µF, L1=66µH and 1 LED load (VF=2,9V).

Since VSW is considerably larger than VF, the linear approximation for the current through the inductor will

be accurate (Equation 12). δL1 0→t_A = (12 −2,9_{2 )} 66𝜇 = 159848 𝐴 𝑠 The simulation yields δL1 0→t_A= 165446

𝐴

𝑠, that is, our equation has a 3,5% of error for this case. By using

(I) to (IV), we can extract tA:

𝑡𝐴 = √

2 · 1𝜇 · 2.9

159848 = 6.02𝜇𝑠

The value found in the simulation is 5.7µs, which means that our approximation introduced a 5,3% of error. With both δL1 0→tA and 𝑡𝐴 we can find the current that flows through the inductor (and into the capacitor)

at point A. The approximated value is virtually the same as the simulated one (0,95A): 𝐼𝑡𝐴 = δL1 0→tA· 𝑡𝐴 = 159848 · 5,92µ = 0.94𝐴

Figure 15 Simulated ‘0’ to ‘A’ time interval for VSW=12V, C1=1µF, L1=66µH and 1 LED load (VF=2,9V). The increase rate of the current through the inductor is 165446 A/s.

At point A, the current will start flowing through the LED and will rapidly increase during the time interval A→B. In the following paragraphs, we will propose a method to estimate with a linear approximation the increase rate of the current for the time interval A→B. For our approximation, the starting conditions are well known. However we need to determine the final conditions, i.e., conditions at point B in time. At that point, we will consider that the current through the LED is much higher than the one through the capacitor (𝐼𝐷1≫ 𝐼𝐶1), or equivalently, since both form a current divider, that the capacitor’s resistance is much higher

than the LED’s resistance (𝑍𝐶 ≫ 𝑅𝐷).

(I) Capacitor and LED resistance relation at point B: We will assume that a good

approximation is when the resistance of the LED is 10 times the resistance of the capacitor.

10 · 𝑅𝐷1= 𝑍𝑐1

(II) Current divider:

𝐼𝐷1= 10 · 𝐼𝐶1 → 𝐼𝐿1=

11 10𝐼_𝐷1

(27)

-26- 𝑍𝑐1= 𝑉𝐶1 𝐼𝐶1 = 𝑉𝐷1 𝐶1 ·𝑑(𝑉𝐷1) 𝑑𝑡 = 𝑉𝐷1 𝐶 ·𝑑(𝐼𝐷1) 𝑑𝑡 · 𝑑(𝑉𝐷1) 𝑑𝐼𝐷1 = 𝑉𝐷1 𝐶 ·𝑑(𝐼𝐷1) 𝑑𝑡 · 𝑟𝐷1 𝑟𝐷1= 𝑑(𝑉𝐷1)

𝑑𝐼𝐷1 : Dynamic resistance of the LED D1. (IV) We use (I) to (III):

𝑑(𝐼𝐷1) 𝑑𝑡 = 𝑑 (10₁₁𝐼𝐿1) 𝑑𝑡 = 𝑑 (10₁₁(𝐼𝑡𝐴+ δL1 tA→tB· 𝑡)) 𝑑𝑡 = 10 11δL1 tA→tB 𝑉𝐷1 𝑅𝐷1· 𝑟𝐷1 =100 11 𝛿𝐿1 𝑡𝐴→𝑡𝐵· 𝐶1 [14]

We can rewrite Equation 14 to ease its use by defining a new parameter (K) and rearranging the variables of the LED:

𝐾 =100 11 δL1 tA→tB· 𝐶1 𝑉𝐷1 𝑅𝐷1· 𝑟𝐷1 =𝐼𝐷1 𝑟𝐷1 𝐼𝐷1 𝑟𝐷1 ≡ 𝐼𝐷1· 𝑔𝐷1= 𝐾 [15] 𝑔𝐷1= 1

𝑟𝐷1: Dynamic conductance of the LED D1.

Equation 15 contains the condition that fulfills the initial resistance relation (I). Thus, finding

the point in the ‘I/V’ curve of the LED (D1) that fulfills the equation yields the value of the current 𝐼𝐷1 at point B. From this current and the increase rate of the current through L1, 𝑡𝐵

can be found (Equation 16). Also, the increase rate of the current through the diode in the

interval A→B can be written as: δD1 t_A→t_B= 𝐼_{𝐷1 𝑡𝐵} 𝑡𝐵−𝑡𝐴. 𝑡𝐵= 𝐼𝐷1 𝑡_𝐵− 𝐼𝐿1 𝑡_𝐴 𝛿𝐿1 𝑡_𝐴→𝑡_𝐵 + 𝑡𝐴 [16]

𝐼𝐷1 𝑡𝐵: Current value of the diode (and inductor) previously found with Equation 15.

For the previous circuit example, 𝐼𝑡𝐵can be calculated as:

𝐾 =100 11 δL1 tA→tB· 𝐶1 = 100 11 · 12 − 2.9 66µ · 1µ = 1.25

The point in the LED’s ‘I/V’ curve in which 𝐼𝐷1· 𝑔𝐷1= 1.25 can be found by using its

(28)

-27-

Figure 16 'I/V' curve of the LED in the example circuit. The point in which ID·gD equals K=1.25 is shown (in red).

From the analysis of the buck-converter circuit and Equations 13 and 15, we can make some conclusions:

- The output-capacitor delay increases as the capacitance increases since more energy will be required to reach the LED´s conduction voltage.

- The output-capacitor delay increases as the conduction voltage for the LED(s) increases since the capacitor will need to charge up to a higher voltage for the LED(s) to start conducting.

- The output-capacitor delay also increases as the difference between the voltage source and the load conduction voltage decreases. This is because the current through the inductor will be lower. - The output-capacitor delay increases as the inductor’s value increases since a bigger inductor will

oppose more resistance to the change of current, thus providing less energy per unit of time to the rest of the circuit.

- The rise-time decreases as the inductor value decreases since the rate of current change in the circuit is higher, thus reaching faster the region in which the LED has less resistance.

- The rise-time decreases as the capacitor value decreases since it will oppose more resistance to the current, thus making it flow through the load diode instead.

- The rise-time decreases as the difference between the voltage source and the load conduction voltage increases, due to the increased current flowing through the inductor.

- The rise-time decreases the more abrupt the change of the ‘I/V’ curve is, since the low-resistance region of the LED will be reached sooner.

Switching techniques: Series and shunt switching

We have only discussed the switching performance of a circuit that has the switching mechanism in series with the LED (series switching). However, the switch can be placed in parallel to the LED (shunt switching). In this new situation, when the switch is activated, the current flows through it and thus, the current through the LED stops.

This technique allows faster falling edges since the current path for the LED and the inductor is not the same anymore. Instead, the switch allows the current from the inductor to flow across it, leading to a sudden drop of the current through the LED. This technique is described in the US11865695 patent[7]. For the shunt-technique, the fall time is neglectable when compared to the rising edge, only limited by the transistor switching capabilities and parasitic capacitances.

(29)

-28-

‘V_SHUNT’. When the shunting device opens, capacitor ‘C1’ charges until 𝑉𝐹1 and then, current starts to

flow through the diode. For the shunting technique, the output capacitor delay is lower since the initial current is higher than for the series shunting technique and thus, charges faster. For the same reason, the rise time is also faster in the shunting technique than in the series technique. The results of the simulation can be observed in Figure 18. We can see that the improvement of the rise time of this technique is not very significant when compared to the fall time one.

Figure 17 Simplified circuit to evaluate edge times for the shunting technique.

Figure 18 Simulation results for the rising edge of the shunting simplified circuit. Initially 1A flows through the inductor and the ‘SHUNT’ switch, which is closed. At t=0us, the switch opens, the capacitor ‘C1’ charges, and current starts flowing through ‘D1’ at 2.41us.

The same equations derived in Section 2.8.4 can be used to approximate the behavior of the circuit if the initial conditions are properly considered.

The edge time improvement in the shunting technique comes at the cost of a higher power consumption, due to the additional power demand of the switching device during the LED off time. To minimize this loss, the switching mechanism is chosen to have the minimum series resistance possible.

Buck-converter LED driver design

We have studied the parameters that affect the requirements of a buck-converter strobing-light LED drivers. There are many ways and approaches to the buck-converter design, and it is out of the scope of this thesis to create a guideline for it. In this section, we will discuss and propose when and how to consider the particularities of a buck-converter for the strobing-light technique.

(30)

-29-

size and cost within the desired limits. As we already mentioned, explaining how to find the right values for the buck-converter design is out of the scope of this thesis.

However, as we demonstrated in the previous section, the inductance value is not the only parameter that determines the switching speed, but also the 𝑉𝑆𝑊 and the total 𝑉𝐹 of the load. This gives rise to another

design decision: should light emitted by the IS be increased by adding more LEDs or by increasing the forward current in the LED driver?

From the formulas in the previous sections, we can conclude that to balance the rise and fall times, the conduction voltage of the load must be half the supply voltage. If we consider the number of LEDs that fulfill this condition as the starting point, adding more LEDs in series (higher forward voltage) will increase the rise time, while removing LEDs will increase the fall time. If the shunting technique is used, it is convenient to reduce the rising edge by having a low forward voltage since the falling edge is much faster. Regarding area and cost, increasing the number of LEDs implies the adding the cost and area of the additional LEDs. Also, the output capacitor must withstand the higher forward voltage. On the other hand, increasing the forward current might imply selecting a new LED driver IC, transistors, L1, and C1.

If we need to achieve faster edge times or reduce the cost of the driver, we can iterate over the above considerations after increasing the switching frequency of the buck-converter. Shorter switching periods allow smaller components but decrease the efficiency of the converter.

2.9 Health regulations

According to [8], devices can be categorized in four risk groups depending on their infrared radiation, and retina and near-infrared retinal thermal hazard duration. The hazard exposure limits for each group can be calculated with the equations provided in that technical report.

(31)

-30-

3. License Plate Recognition (LPR)

3.1 Introduction

In Chapter 2 we concluded that very significant energy savings could be achieved by the strobing-light technique for short shutter times. Generally, short shutter times are used to capture fast moving objects. Thus, a suitable scenario for this technique is one that occurs in darkness and involves fast moving objects, like LPR.

Since LPR is a perfect match for the strobing-light technique and a hot-topic in AXIS, special effort has been put into determining how much can it outperform the constant-light technique. Furthermore, a real scenario in which to apply this technique allows determining the specifications for a strobing-light LED driver, that is, going one step further into understanding the problem and its solutions.

3.2 Scenario overview

For LPR purposes, cameras can be placed on the side of the road or above it (Figure 19). As we will see later in this section, the position of the camera and the angle of its optical axis with the vehicle’s movement direction are the most relevant parameters to consider, as they will determine the power and width of the light pulses. This also means that the specifications of the LED driver have a direct connection with the LPR system installation.

Figure 19 Different positions of cameras for LPR.

For simplicity, during the study of the LPR scenario, we will consider only the case in which the camera is on top of the road. However, at the end of the explanation we will discuss how to derive the analogous process for a camera on the side of the road, and for a camera that is both above and on the side. In the last case, where the camera is both up and on the side of the road, we will also see how to consider both simultaneous effects.

(32)

-31-

A graphical representation of a vehicle´s velocity components for an LPR scenario is shown in Figure 20. The sensor-plane movement (𝑣𝑠), as it is called the perpendicular component in [10], can be easily calculated

by trigonometric analysis from the vehicle’s speed (𝑣𝑣) and the angle of the camera (𝛼) as in Equation 17.

𝑣𝑠 = 𝑣𝑣· 𝑐𝑜𝑠(𝛼) [17]

Figure 20 Vehicle speed components.

This means that, from the camera’s perspective, vehicles move faster the narrower the camera´s angle gets, thus having stricter requirements to yield non-blurry images. The parameter directly affected is the pulse width, which in the strobing-light technique is equivalent to the camera’s shutter time and will have an impact on the peak power required during the strobe. This is because the amount of energy per unit of time in the strobe increases (the peak power increases), to achieve the same brightness during a narrower strobe time.

On the other hand, a wider angle means longer distance camera-to-car. This implies that the power of the LED driver must be higher since the beam of light must reach the plate with the same intensity at a longer distance.

The last effect we will consider for the analysis is the distortion perceived in the license plate’s letters due to the angle of the camera. In this case, as opposed to the previous one, as α is reduced, the distortion is more significant. In Section 3.3 we will propose a method to quantify this distortion, at this point we just need to know that it is an effect that will require shorter strobes to reduce the blur as α gets smaller. Sensor-plane speed, distance, and angle distortion are thus the variables that determine the characteristics of the IR light pulses, width and peak power, which are needed to design and dimension the LED driver. By finding the relation between these variables, an optimal LED driver can be designed that satisfies the requirements for a specific scenario.

3.3 LED Driver specifications for LPR

To understand how the angle affects the LED driver pulse peak power, we analyze the situation represented in Figure 21. In all the situations, we will consider that the resolution is kept constant (equal resolution condition), i.e., if two situations with different camera-to-vehicle distances are compared, it will be assumed that proper zoom is used to get the license plate with the same resolution.

First, to obtain the same brightness in the recordings for both positions (𝛼1 and 𝛼2), the amount of

irradiance (𝐸) at the plate must be the same: 𝐸1= 𝐸2. For this to happen, the ratio between the power

radiated by the illumination system in position 1 (𝑃1) and position 2 (𝑃2) must be the one in Equation 18.

As it is shown in Figure 21, 𝑑1 and 𝑑2 are the distances from the camera to the license plates for 𝛼1 and

(33)

-32- 𝑃2 𝑃1 =𝑑2 2 𝑑₁2 = ( 𝑑2 𝑑1 ) 2 [18]

Figure 21 Representation of two recording situations with different angles. Note that the car speed and the camera's height is the same for both cases.

To obtain the same blur (𝐵), the displacement of the vehicle during one frame must be the same. We will consider both blur and displacement the same parameter, and it can be calculated as in Equation 19, as the product of the speed in the sensor plane [pxl/s] and virtual shutter time [s].

𝐵 = 𝑣𝑠· 𝑡𝑣𝑠 [19]

Considering 𝑡𝑣𝑠 as the virtual shutter time (strobe time), the relation between 𝑃1 and 𝑃2 can be found with

the following procedure:

𝐵1= 𝐵2 → 𝐵1 𝐵2 = 1 𝐵1= 𝑣𝑣· cos(𝛼1) · 𝑡𝑣𝑠1 ; 𝐵2= 𝑣𝑣· cos(𝛼2) · 𝑡𝑣𝑠2 𝑣𝑣· cos(𝛼1) · 𝑡𝑣𝑠1 𝑣𝑣· cos(𝛼2) · 𝑡𝑣𝑠2 = 1 𝑡𝑣𝑠1 𝑡𝑣𝑠2 =cos(𝛼2) cos(𝛼1)

Since the peak power is inversely proportional to the virtual shutter time, we can rewrite the last equation as:

𝑃2

𝑃1

=cos(𝛼2) cos(𝛼1)

By trigonometry, we can translate the equation above as a function of the distance from the camera to the car: ℎ 𝑑1 = cos(𝛼1) ; ℎ 𝑑2 = cos (𝛼2) 𝑃2 𝑃1 =𝑐𝑜𝑠(𝛼2) 𝑐𝑜𝑠(𝛼1) =𝑑1 𝑑2 [20]

As we can observe in Equations 18 and 20, the ratios between distance are inverse to each other. Also, the

(34)

-33- 𝑃2 𝑃1 =𝑑2 𝑑1 [21]

To exemplify graphically what this equation means we have added Figure 22. In it, it can be seen that a certain driver will need to provide four times more peak power when the camera-to-vehicle distance doubles (equal brightness condition). However, from the equal blur condition perspective, the requirements are laxer since the same blur will be generated from a twice longer virtual shutter time. Both considerations yield the result predicted by Equation 21: the resulting peak power ratio is the ratio between the distances.

Figure 22 Visual exemplification of the equal brightness and equal blur conditions.

This does not resemble reality good enough yet, because in that situation the plate would be 90º inclined with respect to the camera’s optical axis and the letter would not be visible. However, we still need to consider the angle distortion, which we previously introduced in this chapter.

To consider the effect of the angle distortion we need to rewrite the blur condition. Regarding image quality, the absolute amount of blur (B) is not relevant, but the amount of blur with respect to the letters size (Br). This means that bigger letters can tolerate more blur than smaller ones, as it is shown in Figure 23.

Figure 23 In this picture two different sized letters suffer the same amount of blur (B), 4 pixels in the vertical direction. In green, the position of the letter at the start of the shutter time is shown; in grey, the final position; and in black, the result captured by the camera. While the absolute blur does not

(35)

-34-

Angle distortion can be interpreted as a change of the size of the letters, i.e., the license plate and its letters height, when looked from any 𝛼 < 90º, will shrink. The shrinking factor (𝑠) that the plate suffers can be derived from Figure 24, as it is done in Equation 22.

ℎ𝑝′ = ℎ𝑝· 𝑠 = ℎ𝑝· 𝑠𝑖𝑛(𝛼) [22]

Figure 24 The height of the license plate (hp) shrinks to ‘hp'’ when looked from an angle ‘α ’.

If we consider 𝐵max @𝑟𝑒𝑠 as the maximum blur in pixels that we can tolerate at a certain resolution (𝑟𝑒𝑠),

we can describe the blur as a relative variable to it as in Equation 23. Since we are always assuming to have the same resolution, hereafter we will stop using ‘@res’.

𝐵𝑟 @𝑟𝑒𝑠=

𝐵 𝐵𝑚𝑎𝑥 @𝑟𝑒𝑠

[23]

𝐵max is a constant that could be extracted from image testing, qualitatively identifying the maximum blur

that a character can tolerate and still be readable, or by means of the analysis of the LPR algorithm results under different amounts of blur. We will not find a method to extract values for this parameter.

The same blur condition can be rewritten using the relative blur instead of the absolute one, as the same

relative blur condition, as shown below.

𝐵𝑟1= 𝐵𝑟2

𝐵1

𝐵max 1

= 𝐵2 𝐵max 2

However, the amount of permissible blur will depend on the angle of the camera (𝛼). For instance, if we

were looking frontally at the plate (𝛼 = 0º) we would tolerate the entire Bmax blur. In the most extreme case,

if we were looking the plate from the top (𝛼 = 90º), we would not see any character, thus the permissible

blur would be null. By trigonometric analysis, the perceived height of the plate can be found to be directly related to the sinus of the camera position angle (𝛼). We can find the peak power relation by assuming that

the amount of maximum blur that can be tolerated (Bmax) is proportional to the height of the plate perceived

by the camera. The procedure is the one that follows:

(36)

-35- 𝑃2 𝑃1 =𝑡𝑔(𝛼1) 𝑡𝑔(𝛼2) [24]

By basic trigonometry, Equation 24 can be rewritten as the power ratio for the absolute blur equation

(Equation 20) multiplied by the angle distortion ratio, as in Equation 25. Therefore, the angle distortion ratio is sin(𝛼1) sin(𝛼2). 𝑃2 𝑃1 =𝑠𝑖𝑛(𝛼1) 𝑠𝑖𝑛(𝛼2) ·𝑑1 𝑑2 [25]

We can extract the power ratio between two camera positions with different α by combining the same relative blur and the same brightness conditions. The result is shown in Equation 26.

𝑃2 𝑃1 =𝑠𝑖𝑛(𝛼1) 𝑠𝑖𝑛(𝛼2) ·𝑑2 𝑑1 [26]

For our comparison purposes, it is only relevant to us the case in which the cameras are at the same height, which yields Equation 27. Several curves for different 𝛼1 are plotted in Graph 2.

𝑃2 𝑃1 =𝑠𝑖𝑛(𝛼1) · 𝑐𝑜𝑠(𝛼1) 𝑠𝑖𝑛(𝛼2) · 𝑐𝑜𝑠(𝛼2) =𝑠𝑖𝑛(2 · 𝛼1) 𝑠𝑖𝑛(2 · 𝛼2) [27]

Graph 2 Power ratio between 2 cameras at different angle positions. Note that for α1 it is only plotted the 0° to 45° range since their respective complementary angles yield the same curve (e.g., the a1=80° curve is equal to the a1=10° one).

In Graph 2 we can observe how the power ratio decreases for any initial angle ‘𝛼1’ until ‘𝛼2= 45º’, thus