This Master thesis explores car tracking and re-identification on videos with a single and stationnary camera. Several works have tackled this issue with appearance models based on CNNs but occluded cars make it quite difficult to track all the vehicles efficiently. The proposed solution is to improve the appearance model by skeletonizing the road and turning it into a graph to better model the trajectories of the cars and bring geometric information. The results show that this strategy can reduce the number of counted vehicles, making it closer to reality. As future work we suggest to adapt the hyperparameters to each video, to use the graph more efficiently and to get more data.
Keywords: Deep Learning, Computer Vision, car Re-identification, tracking, detection, mathematical morphology, skeleton, graph, traffic supervision.