Real-time detection and tracing of vehicles via camera systems

Traffic surveillance and analysis is an important matter to increase the safety on the roads. By monitoring the traffic appropriate measures be taken when dangerous situations appear or to keep the traffic flow steady. Therefore it is necessary to use a system that can detect and track vehicles so that details of the traffic situation can be evaluated. A distinction of the vehicle types is also useful to get more detailed information about the traffic. In this diploma thesis different approaches for vehicle detection and tracing via camera systems are implemented and evaluated. The implementation is a plug-in for the program FloDiEdi done in C++. Three different approaches are implemented and evaluated. The first and second approach use the Shi-Tomasi feature detector and an modified Lucas-Kanade feature tracker. Perspective transformation is used in the second and third approach too. In the third approach a combination of position estimation and template matching is used to track vehicles. For speed calculation the first approach uses cross ratio. The second and third, due to perspective transformation, use a pixel to meter conversion to get the distance in meters. The extracted information are visually presented such as the ID, the type and the speed of the vehicle is displayed and additionally stored to an xml file. Furthermore in the xml file the lane and the size of the vehicle is stored. The camera system is a stationary system which captures the images of the street. The first approach achieves a high detection rate but the tracking does not work well, so just around 10% of the vehicles were successfully tracked. In the second approach the detection rate is lower than in the first but tracking works more reliable with an tracking rate of around 30%. The third approach has the same detection rate as the second approach but the amount of successful tracked vehicles is over 70%. Concerning the computational time, the third approach is seven times more faster than the first and second, which are similarly slow. After evaluating the three methods the fastest of the three approaches and most accurate approach is the third one. There is still room for more improvements such as achieving a speed up by using the GPU for matrix operations or parallel programming.