travis an efficient traffic visual monitoring system

TRAVIS AN EFFICIENT TRAFFIC VISUAL MONITORING SYSTEM Theodoros Semertzidis1, Alexandra Koutsia1, Kosmas Dimitropoulos1, Nikos Grammalidis1, Antonis Kantidakis2, Kyriakos Georgouleas2 and Petros Violakis2

ABSTRACT. The paper presents a real-time vision system for automatic traffic monitoring based on a network of autonomous tracking units that capture and process images from one or more pre-calibrated cameras. The proposed system, which has been developed within the framework of TRAVIS (TRAffic VISual monitoring) project, is flexible, scalable and can be applied in a broad field of applications in the future, such as the traffic control of tunnels at highways and aircraft parking areas at airports. Different image processing and data fusion techniques are evaluated, while a novel algorithm for the synchronization of all autonomous tracking units over a TCP/IP network is comprehensively demonstrated. Experimental results obtained from two pilot installations at Macedonia airport of Thessaloniki and at a road tunnel near Piraeus harbour have shown the great potential of the proposed system. INTRODUCTION Traffic control and monitoring using video sensors has drawn increasing attention recently due to the significant advances in the field of computer vision. Many commercial and research systems use video processing, aiming to solve specific problems in road traffic monitoring, Kastrinaki et al. (2003). An efficient application for monitoring and surveillance from multiple cameras is the Reading People Tracker (RPT), Le Bouffant et al. (2002), which was later used as a base for the development of a system called AVITRACK, which monitors airplane servicing operations, Thirde et al. (2005). Furthermore, in the FP5 INTERVUSE project, an artificial vision network-based system was developed to monitor the ground traffic at airports, Pavlidou et al. (2005). The system uses the Autoscope® Solo Wide Area Video Vehicle Detection System which has been successfully deployed worldwide for monitoring and controlling road traffic, Michalopoulos (1991). However, robust and accurate detection and tracking of moving objects still remains a difficult problem for the majority of computer vision applications. Especially in the case of outdoor video surveillance systems, the visual tracking problem is particularly challenging due to illumination or background changes, occlusions problems etc. In this paper, our aim is to present a novel, extensible multi-camera video surveillance system, which supports multiple functionalities such as detection, tracking 1 2

Informatics and Telematics Institute, 1st km Thermi-Panorama road, 57001 Thessaloniki, Greece MARAC Electronics, 165 Marias Kiouri & Tripoleos Str, 188 63, Piraeus, Greece

1

and classification of objects moving within the supervised area and can be used to address various applications. The proposed system, which has been developed within the framework of TRAVIS (TRAffic VISual monitoring) project, is based on a network of intelligent autonomous tracking units, which capture and process images from a network of pre-calibrated visual sensors. The fundamental goal of TRAVIS is the development of a moving target tracking system which is fully scalable and parameterized and can be applied in a broad field of applications in the future. Reaching the end of the project, two prototypes were developed, each one focusing on a different aspect of traffic monitoring: Traffic control of tunnels at highways: This application aims to investigate the ability of the proposed system to monitor and control highway tunnels traffic in order to trace events that can lead to accidents. The tracing of such events leads to instant warning of drivers through special traffic lights placed at the entrance of the tunnel. The system is also able to provide the traffic control centre with statistical information about the traffic inside the tunnel (traffic volume, average vehicle speed etc) as well as warnings in case of accidents using a user friendly graphical interface. Traffic control of the aircraft parking area (APRON) at airports: This prototype is addressed to the APRON controllers, who are responsible for the control of movements (airplanes, cars, buses, humans, etc.) occurring at the aircraft parking area. The objective of this application is to provide airport supervision authorities with a ground situation display in order to facilitate the traffic management at the APRON. The system also provides alarms for the avoidance of accidents, thus increasing the safety levels at airports. TRAVIS SYSTEM ARCHITECTURE The proposed system consists of a scalable network of autonomous tracking units (ATU) that use video sensors to capture images, detect foreground objects within their field of view and provide results to a central sensor data fusion server (SDF). The SDF server is responsible for tracking and visualizing moving objects in the scene as well as collecting statistics and providing alerts when specific situations are detected. In addition, depending on the available network bandwidth, images captured from specific video sensors may also be coded and transmitted to the SDF server, to allow inspection by a human observer (e.g. traffic controller). The system requires frame synchronization, i.e., simultaneous frame capturing from all cameras and constant frame rate from all ATUs, which are achieved with the use of the Network Time Protocol (NTP) and a synchronization procedure that is presented in following section. The topology of the ATUs in the network as well as their number varies in each application depending on the existing infrastructure, geomorphologic facts and the bandwidth and cost limitations. The network architecture is based on a wired or wireless TCP/IP connection as illustrated in Figure 1. These topologies can be combined to produce a hybrid network of ATUs, meaning a network combined using different network technologies and protocols. The video sensors are standard CCTV cameras, not necessarily of high resolution, equipped with a casing appropriate for outdoor use and telephoto lenses for distant observation. They are also considered as static (fixed field of view) and are pre-calibrated. Since the calibration technique is based on plane homographies, the observed targets have to be relatively small, in other words their distance from the video sensors has to be large enough, so that each target

2

can be locally well approximated by a plane. In our implementation, we obtained satisfactory results when the maximum target size is smaller than 10% of image width. . Sensor 1

Sensor k

Autonomous Tracking Unit 1

Sensor 1

....... ..

Sensor m

Autonomous Tracking Unit n

Sensor 1

Sensor k

Autonomous Tracking Unit 1

Sensor 1

....... ..

Sensor m

Autonomous Tracking Unit n

Ethernet WiFi / WiMAX SDF Server

(a)

SDF Server

(b)

Figure 1: Basic architecture. (a) Wired topology. (b) Wireless topology. THE AUTONOMOUS TRACKING UNITS Each autonomous tracking unit is a common processing unit (PC or embedded PC), which obtains frames from one or more video sensors, at fixed time intervals. Each autonomous tracking unit consists of the following modules: - Calibration module (off-line unit to calibrate each video sensor). To obtain the exact position of the targets in the real world, the calibration of each camera is required, so that any point can be converted from image coordinates (measured in pixels from the top left corner of the image) to ground coordinates and vice versa. A calibration technique, which is based on a 3x3 homographic transformation and uses both points and lines correspondences, was used, Dimitropoulos et al. (2005). - Background extraction and update module. This is the most critical part of the autonomous tracking unit as it is the most time consuming processing step. In order to make the system work effectively in real time, many methods were considered and implemented. For the first tests, the mixture of Gaussians algorithm was used KaewTraKulPong and Bowden (2001). Other methods that are supported are the Bayes algorithm, Li et al. (2003), the Lluis-MirallesBastidas method, Lluis et al. (2005), and non-parametric modelling, Elgammal et al. (2000). - Foreground segmentation module using connected component labelling to identify individual foreground objects (moving blob analysis). - An optional blob tracking module to track blobs. For this task the MHT tracker has been successfully used, although association and tracking of very fast moving objects may be problematic. - A blob classification module. This module classifies blobs identifying their probability to belong to each member of a pre-defined set of classes e.g. person , car etc. These probabilities are calculated using a formerly trained back propagation neural network. For each blob, 9 attributes, characteristic of its shape and size, are used as input to a neural network, whose number of outputs is equal to the pre-defined number of classes. The 9 attributes used for each blob are the two sizes of the major and minor axes of the blob s ellipse and the 7 Hu moments (Hu, 1962; Hu, 1961) of the blob, that are invariant to both rotations and translations of the blob.

3

- A 3-D observation extraction module that uses the available camera calibration information to estimate the accurate position of targets in the scene. Since the camera calibration is based on homographies, an estimate for the position of a target in the world coordinates can be directly determined from the centre of each blob in each camera. Each observation is also associated with a reliability matrix R , depending on the camera geometry and its position at the camera plane. The final output of each ATU is a small set of parameters (coordinates, classification information, reliability matrix) describing the moving objects in ground coordinates, which is provided to the central Sensor Data Fusion (SDF) server through wired or wireless transmission. In addition to these text observations, a properly compressed greyscale image is transmitted at each polling cycle only if SDF operates in the foreground map mode that is described at the following section. The data packet transmitted from each ATU is presented in the following figure:

Figure 2: TRAVIS Data packet THE SENSOR DATA FUSION SERVER The Sensor Data Fusion Server collects information from all ATUs using a constant polling cycle, produces fused estimates of the position and velocity of each moving target, and tracks these targets using a multi-target tracking algorithm. In addition, the SDF server produces a synthetic ground situation display (Figure 3), collects statistical information about the moving targets in the supervised area and provides alerts when specific situations (e.g. accidents) are detected. Moreover, important information, such as velocity, direction, class of object or lane are also obtained for each moving object. Using these data, various traffic statistics can be calculated, such as traffic density (in vehicles per Km), (both total and per lane) traffic flow (in vehicles per second), vehicle counts (both total, per class and/or per lane) and finally minimum, maximum and average velocity per lane.

Figure 3: SDF visualization window showing 3 targets on the airport APRON. Observation fusion A target that is present in the field of view of multiple cameras simultaneously will generally result in multiple observations due to the fact that the blob centres of the same object in two 4

different cameras correspond to close but different 3-D points. In order to group together all the observations that correspond to the same target, two techniques are proposed: One of the following techniques is selected for areas visible by two or more cameras depending on the application requirements, the computational power of the SDF server and the available bandwidth for the connection between the autonomous tracking units and the SDF. Observations belonging in areas visible by only one camera are simply added to the final result. The two proposed data fusion techniques were evaluated using results from two pilot applications, each using two cameras. The grid-based technique is simpler, faster and requires less bandwidth since it transmits less information, while the foreground map fusion technique was seen to robustly resolve occlusions and provide more accurate results, at the extra cost of transmitting a foreground probability map at each cycle. Grid-based fusion A grid that separates the overlap area (in world coordinates) in cells is defined. Optimal values for the cell size are determined considering the application requirements (e.g. maximum allowed separation between vehicles). Observations belonging to the same cell or to neighbouring cells are grouped together to form a single fused observation. The method used to implement this technique is based on expressing the grid as a binary image: cells that have at least one assigned observation are represented by a white pixel, while those with no observations are represented by a black pixel. A connected component labelling algorithm is then used to identify blobs in this image, each corresponding to a single moving target. Fused observations are produced by averaging the parameters of the observations that belong to each group. More specifically, each fused observation consists of an estimated position of the world coordinates, an uncertainty matrix as well as a classification probability matrix. An example of the grid-based fusion technique is presented in Figure 4. The white areas of this map are the areas covered by two cameras, while the grey areas are only covered by one. The black areas are not visible by the system. The dark-grey dots denote the initial observations while the light-grey ones are the ones derived from the fusion step, after the weighted averaging of all observations belonging to the same cell or neighbouring cells.

Figure 4: Example of grid-based fusion.

5

Foreground map fusion In this technique, each autonomous tracking unit initially determines the pixels in each video sensor that are also visible by other video sensors. For these pixels instead of observations, foreground probability maps are generated, i.e. greyscale images describing the probability for each pixel to belong to a foreground object. These maps are then transmitted to the SDF, where they are fused together (warped to the ground plane and multiplied together), using a technique similar to Khan and Shah, (2006). The fused observations are then generated from these fused maps using connected component analysis and classification information is computed as in the ATU s blob classification module. Although this technique has increased computational and network bandwidth requirements, when compared to grid-based fusion, it can very robustly resolve occlusions between multiple views. For example, if two blobs are erroneously joined together in one view, the algorithm can still correctly separate the two targets, provided that they are identified as two different blobs in another view. Multiple Target Tracking The tracking unit is based on the Multiple Hypothesis Tracking (MHT) algorithm, which starts tentative tracks on all observations and uses subsequent data to determine which of these newly initiated tracks are valid. Specifically, MHT, Blackman and Popoli, (1999), is a deferred decision logic algorithm in which alternative data association hypotheses are formed whenever there are observation-to-track conflict situations. Then, rather than combining these hypotheses, the hypotheses are propagated in anticipation that subsequent data will resolve the uncertainty. Generally, hypotheses are collections of compatible tracks. Tracks are defined to be incompatible if they share one or more common observation. MHT is a statistical data association algorithm that integrates the capabilities of: Track Initiation: The automatic creation of new tracks as new targets are detected. Track Termination: The automatic termination of a track when the target is no longer visible for an extended period of time. Track Continuation: The continuation of a track over several frames in the absence of observations. Explicit Modelling of Spurious Observations Explicit Modelling of Uniqueness Constraints: An observation may only be assigned to a single track at each polling cycle and vice-versa. Specifically, the tracking unit was based on a fast implementation of the Multiple Hypothesis Tracking algorithm by Cox and Hingorani,( 1996). A 2-D Kalman filter was used to track each target and additional gating computations are performed to discard observation track pairs. More specifically, a gate region is defined around each target at each frame and only observations falling within this region are possible candidates to update the specific track. The accurate modelling of the target motion is very difficult, since a target may stop, move, accelerate, etc. Since only position measurements are available, a simple four-state (position and velocity along each axis) CV (constant velocity) target motion model in which the target acceleration is modelled as white noise provides satisfactory results. DATA AND CONTROL IN TRAVIS NETWORK The TRAVIS system is based on client-server network model where each ATU acts as a client to the SDF listening server. The clients are connected to the server via a wired 6

connection or a wireless point to point or point to multipoint wifi/wimax connection. The ATUs are remotely controlled from the SDF server through a signalling channel that enables the SDF server user to start/stop capture of frames, change operational mode, capture cycle and other parameters. The signalling channel has been implemented in order to separate data transmission of the ATUs from control signals sent by SDF. Data fusion and multiple hypotheses tracking algorithms, run by the SDF server, require synchronised capture of frames of the scene using a constant sampling period. To fulfil this requirement, the Network Time Protocol is used in order to synchronise system clocks of each computer in the TRAVIS network with reference to server s system clock. For this configuration the SDF server was set as stratum 1 clock level in the NTP hierarchy, whereas every ATU set at stratum 2 clock level. Even with synchronised system clocks, a capture command from the server will not arrive at every ATU on the same moment of time due to network latency, computer process delays etc. which makes the coinstantaneous frame grabbing from all ATUs with millisecond accuracy impossible. For this reason, a synchronized capture procedure is used, which was first proposed in Litos et al. (2006). According to this algorithm after the synchronisation of the system clocks, the SDF server sends to each ATU a specific timestamp, indicating the time in the future when the ATUs will start capturing frames as well as the value of the (constant) frame sampling period. Handshake procedure After installation all ATUs in the TRAVIS network are set in standby mode. At this phase each ATU is listening in a UDP port, defined during the installation procedure, for command by the server. When SDF server user clicks the start capturing button the handshake procedure starts and a command is sent through the signalling channel notifying the ATUs for SDF s IP address and port number. The next step is for ATUs to request TCP connection from the listening SDF server. After all connections have been established and the meeting time has been reached, the system enters normal operation. In the normal operation phase, SDF and ATU software maintain timers to control frame capture and synchronisation. Figure 5 presents an abstract state machine diagram for ATU standby, handshake and operational state, while Figure 6 presents a similar diagram for the server.

Figure 5: ATU states from standby to normal operation

7

Figure 6: SDF handshake and event loop in normal operation. This handshake scheme simplifies both first installation and re-configuration of the nodes of TRAVIS network, since: The SDF or ATU IP addresses are not a-priori required Only the UDP port number where ATUs listen needs to be known a-priori. The SDF creates a list of the connected ATUs and waits for their data. Synchronization maintenance During normal operation, the ATUs calculate the timestamp for capturing every next frame in order to keep being synchronized to each other. After capturing the frame, image processing techniques, as presented at previous section, are applied to the captured frame, observations are extracted and finally a packet is sent to SDF. On the other side, the server receives one packet from every ATU, checks the timestamps to certify that frames are synchronized and determines that data are ready to be fused. If one of the ATUs does not complete frame processing within the available time cycle (abnormal function), the time window for sending data is lost and the data will arrive at the SDF server with delay. In this case this packet is discarded from the server, while the ATU will wait for the next valid window to capture a new frame and thus re-synchronize with the rest. The following figures show the above reasoning in a schematic way. Figure 7 presents the case where every ATU sends the extracted data in-time and waits until the next time window has been reached for capturing the next frame, whereas in Figure 8 a delay in processing occurred at ATU 2. In the latter case, ATU exceeds the given time cycle so it will send its data within the next time window and will then wait for the next sampling time instant in order to capture the next frame. The delayed data packet will be discarded from the SDF server after a simple timestamp check.

8

Figure 7: capture cycle and processing time window with no delays.

Figure 8: ATU 2 exceed the given time for processing and data transmission

9

Secondary video streaming system As a secondary backup system the ATUs support on demand media streaming directly to the control center, where SDF server is installed, in order to assist operators when an accident or other abnormal situation is reported. Depending on the available bandwidth this secondary backup system is able to stream compressed video or just a number of compressed images. The application installed at the control center may render this input as Motion JPEG, MPEG4 or any other available format and present it in a new window. This service allows the operator to select the ATU in subject, watch the scene and validate the alarm. EXPERIMENTAL RESULTS In order to decide on the most appropriate background extraction technique for the specific applications, tests have been conducted using sample sequences. The masks shown on Figure 9 are obtained from the pilot installation of the TRAVIS system at Macedonia airport of Thessaloniki, Greece. Figure 9 (a) shows the original image, while in Figure 9 (b) the three moving objects that need to be detected by the background extraction methods are marked with circles. As seen in Figure 9 (c) the objects are detected with the mixture of Gaussians method, although the shape of the masks is distorted due to shadows. The results of the Bayes algorithm are shown in Figure 9 (d). This method fails to detect slowly moving objects like the one on the left of the image. The Lluis et al method shown in Figure 9 (e) produces masks with low level of connectivity, which are not suitable for the following image processing steps. Finally the non-parametric model method (in Figure 9 (f)) yields very accurate results, while coping well with shadows.

(a)

(b)

(c)

(d)

(e)

(f)

Figure 9: a) Original image b) Moving objects c) Mixture of Gaussians mask d) Bayes algorithm mask e) Lluis-Miralles-Bastidas mask f) Non-parametric model mask.

10

Another critical issue when deciding on the most appropriate background extraction algorithm is the execution time of each algorithm. To evaluate the computation complexity, all four methods were applied on sequences of different sizes. The results obtained after running the methods for 50 frames are presented on Table 1. Standard Intel Pentium 4 3.2GHz PCs with 1GB of RAM running on Windows XP Pro were used. All algorithms were implemented in C++ using the open source library OpenCV. Table 1: Background extraction methods execution times and FPS Method

BAYES

GAUSS

LLUIS

NON- PARAMETRIC

Resolution

Time(s)

FPS

320x240

0.13

7.68

640x480

0.59

1.68

768x576

0.73

1.38

320x240

0.07

14.38

640x480

0.29

3.39

768x576

0.39

2.59

320x240

0.05

22.19

640x480

0.19

5.16

768x576

0.28

3.63

320x240

0.054

18.57

640x480

0.35

2.83

768x576

0.38

2.65

Based on the above results, the non-parametric background extraction method was selected for both pilot applications. Also since the angle between the optical axis of the cameras is small due to installation constrains the grid method was selected for data fusion. Extensive tests were made during pilot installations both at Macedonia airport of Thessaloniki and at a tunnel near Piraeus harbour, in order to optimize TRAVIS system for each application. The system s maximum speed limitations depend on the installation, while the minimum speed the system can handle is 0 Km/h. Moreover, the parameters that affect the maximum speed limitation are the system s area of coverage and the FPS. For example, a system configured to run with 2 FPS and cover an area with diameter 40 meters can estimate target velocities upto 144Km/h. In the following subsections these two applications are described in detail. Airport installation Two cameras were installed at the 4th floor of the control tower, mounted at the balcony ledge left and right of the main tower. As shown in Figure 10, the two cameras cover a great percentage of the APRON area (approximately 25000 square meters) while focusing in parking place number 4 of the airport APRON. In airport installation two ATUs were installed, each consisted of one PC with a frame grabber and a camera. Both these PCs and SDF server were placed in the APRON control room of the tower and are interconnected through an Ethernet network.

11

Figure 11(a,b) presents synchronized frames from the two cameras, with the moving objects marked with a circle, while in Figure 11(c, d) the corresponding foreground maps are shown. Finally, in Figure 11(e) the final results of the SDF are presented in its visualization window. The same window is used to inform the operator for possible warnings or alarms in observation area. The available alarms are: person walking, speed limit violation, aircraft in movement. More alarms and warnings may be supported by the system in the future.

Apron

Camera 1

Taxiway

Camera 2

Control tower

Figure 10: Cameras Field Of View. Airport installation

(a)

(b)

(c)

(d)

(e) Figure 11: (a) frame from camera 1 (b) Frame from camera 2 (c) mask from cam1 frame (d) mask from cam2 frame (e) SDF visualization window showing the moving targets

12

Tunnel installation Taking into account existing installation constraints (limitations from tunnel authorities, slight left turn near the entrance, etc) two cameras were finally installed at the entrance of the tunnel placed in a way that the two fields of view overlap, with the first one being an extension of the second one as shown in Figure 12. The tunnel coverage resulting for this pilot installation is about 35 metres.

Camera 2

Camera 1

Figure 12: Cameras Field Of View. Tunnel installation Partial overlapping between the fields of view is recommended in order to make use of fusion techniques and overcome possible occlusion problems. Figure 13 show the positions of the two installed cameras. Each ATU consists of an embedded PC and a camera. The embedded PCs were installed on top of the tunnel roof and inside an IP-66 box, which is highly protective against environmental conditions. Using a WiMAX interface the ATUs were connected to the SDF server which was installed in a building about 1 Km away (Line of Sight). In Figure 14(a,b) two sample synchronized frames from the different cameras are shown while Figure 14(c,d) present the corresponding thresholded foreground maps. Figure 14 (e) is the visualization window of SDF which displays the fused observations as well as the calculated traffic statistics in real time. CONCLUSIONS AND FUTURE WORK A novel multi-camera video surveillance system for traffic monitoring applications was described. Different background extraction methods and data fusion techniques were evaluated. The non-parametric modelling method seems to provide improved background extraction results in terms of accuracy and time efficiency. Two data fusion techniques were examined, resulting to a trade-off between efficiency, bandwidth and computational complexity. Promising results were obtained from the two pilot applications for road and APRON traffic monitoring and surveillance, each using two cameras. Future research will be focused on a more detailed evaluation of the two fusion methods in terms of performance, speed and bandwidth requirements. A future goal is the distribution of SDF output results to remote clients (authorities, traffic information web services, etc.) using different profiles. Furthermore, an image stabilization algorithm can be used to compensate for slight camera movements caused by vibrations, wind or camera drift over time.

13

Camera 2

Camera 1

Figure 13: cameras installation point

(a)

(b)

(c)

(d)

(e) Figure 14: a) frame from cam 1 b) frame from cam 2. c) Mask from cam 1 frame. d) Mask from cam 2 frame. e) Scene visualization window of SDF server.

14

ACKNOWLEDGEMENTS This work was supported by the General Secretariat of Research and Technology Hellas under the InfoSoc TRAVIS: Traffic VISual monitoring project and the EC under the FP6 IST Network of Excellence: 3DTV-Integrated Three-Dimensional Television - Capture, Transmission, and Display (contract FP6-511568). REFERENCES Blackman S, Popoli R., (1999) Design and analysis of modern tracking systems , Artech House, Boston, USA. Cox I. J, Hingorani S.L.,( 1996) An Efficient Implementation of Reid's Multiple Hypothesis Tracking Algorithm and its Evaluation for the Purpose of Visual Tracking , IEEE Transactions on pattern analysis and machine intelligence, Vol. 18, pp. 138-150. Dimitropoulos K., Grammalidis N., Simitopoulos D., Pavlidou N. and Strintzis M.,( 2005) Aircraft Detection and Tracking using Intelligent Cameras , IEEE International Conference on Image Processing, Genova, Italy, pp. 594-597. Elgammal Ahmed, Harwood David and Davis Larry,(2000) Non-parametric Model for Background Subtraction.Computer Vision , ECCV. Hu M-K, (1962) Visual pattern recognition by moment invariants , IRE Trans. on Information Theory, IT-8:pp, 179-187. Hu M-K.,(1961) Pattern recognition by moment variants , Proc. IRE, 49:pp, 1428. KaewTraKulPong P., Bowden R.,(2001) An Improved Adaptive Background Mixture Model for Real-time Tracking with Shadow Detection , in 2nd European Workshop on Advanced Video-based Surveillance Systems, Kingston, UK. Kastrinaki V., Zervakis M., Kalaitzakis K.,(2003) A survey of video processing techniques for traffic applications , Image Vision Computing, Volume: 21,Issue: 4,pp. 359 - 381 Khan S. and Shah M., (2006), A multiview approach to tracking people in crowded scenes using a planar homography constraint , 9th ECCV Conference, Graz, Austria. Le Bouffant T., Siebel N. T., Cook S. and Maybank S. (2002) , Reading People Tracker Reference Manual (Version 1.12) , Technical Report No. RUCS/2002/TR/11/001/A, Department of Computer Science, University of Reading. Li L., Huang W.,. Gu I. Y.H and Tian Q.,(2003) Foreground Object Detection from Videos Containing Complex Background , in International Multimedia Conference,. Litos G., Zabulis X. and Triantafyllidis G.A.( 2006): "Synchronous Image Acquisition based on Network Synchronization", IEEE Workshop on Three-Dimensional Cinematography (in conjunction with CVPR). Lluis J., Miralles X. and Bastidas O.,( 2005) Reliable Real-Time Foreground Detection for Video Surveillance Application , in VSSN'05. Michalopoulos G.1991), Vehicle Detection Video Through Image Processing: The Autoscope System, IEEE Transactions on Vehicular Technology, Vol. 40, No. 1. Pavlidou N., Grammalidis N., Dimitropoulos K., Simitopoulos D., Gilbert A., Piazza E., Herrlich C., Heidger R. and Strintzis M., (2005) Using Intelligent Digital Cameras to Monitor Aerodrome Surface Traffic , IEEE Intelligent Systems, Vol. 20, No. 3, pp.76-81. Thirde, D., Borg, M., Ferryman, J., Fusier, F., Valentin, V., Bremond, F., and Thonnat, M. (2006.) A Real-Time Scene Understanding System for Airport Apron Monitoring . In Proceedings of the Fourth IEEE international Conference on Computer Vision Systems.

15