Real-time vehicle tracking using CUDA

Abstract

Computer vision is one of the most actively researched topics in computer science. This interest is sparkled not only by the idea of making computers understand and reason about events happening in a video sequence, but also by the successful implementation of computers in various fields, including surveillance applications. Traditionally, monitoring behaviour of traffic involved installing cameras on junctions and highways, and having personnel looking at monitors for hours on end. A computer capable of understanding its surroundings would be able to partially or totally replace humans, resulting in reduced costs and greater security. This thesis investigates the techniques involved to track vehicles over a sequence of consecutive frames, given a fixed monocular setup. To increase robustness, tracking is carried out in 3D using a priori knowledge incorporated into the system, in the form of wireframe vehicle models, the ground plane constraint, and vehicle dynamics equations. The Mixture of Gaussians algorithm is used to detect areas of movement in the scene. A shadow removal algorithm is employed to remove shadow pixels and the object contours extracted using connected component analysis. A number of vehicle pose recovery algorithms were implemented and compared, specifically two orientation angle recovery algorithms and two pose evaluation algorithms. An improved method for vehicle localisation based on existing methods was also proposed and tested in this dissertation. The possibility of using a genetic algorithm to fit a single deformable vehicle model to image data was also investigated. Finally, an extended Kalman filter was implemented to filter noisy measurements and calculate a better estimate of the true state of the system. The main output of the application is a 3D wireframe model superimposed over each moving vehicle in the scene. A top-view trajectory of the vehicles is also presented in graphical format. The Mixture of Gaussians and shadow detection algorithms were implemented in parallel on the NVIDIA CUDA architecture, and compared against their sequential counterparts. The final artefact was tested for accuracy using standard benchmark videos available for evaluation of tracking and surveillance systems. The different methods that were implemented were compared and the combination of algorithms that gives the best overall result was outlined.