Vehicles detection and tracking in videos for very crowded scenes

Counting and tracking vehicles in very crowded scenes is a very challenging problem, where many products require discipline conditions to deal with. In this paper, a novel algorithm for automatically counting the number of moving vehicles and estimating their velocities and paths in regular and very crowded scenes, under different conditions, is presented. In this method, interest points are detected and trajectories are calculated independently where confusing trajectories are removed. Initial clustering of the interest points based on proposed mathematical relations is performed. The number of moving vehicles is estimated by grouping the initial clusters based on a new adaptive background construction method, maximum sub-rectangle sum algorithm and disjoint set data structure. Our algorithm has been applied to a collected dataset representing very crowded traffic scenes, where it showed an excellent high accuracy. In addition, it has low storage and computational requirements, which promotes it for real time applications.