Detection and Tracking of Moving Vehicles in Crowded Scenes

Abstract

Vehicle inter-occlusion is a significant problem for multiplevehicle tracking even with a static camera. The difficulty is that the one-to-one correspondence between foreground blobs and vehicles does not hold when multiple vehicle blobs are merged in the scene. Making use of camera and vehicle model constraints, we propose a MCMCbased method to segment multiple merged vehicles into individual vehicles with their respective orientation. Then a Viterbi algorithm is applied to search through the sequence for the optimal tracks. Our method automatically detects and tracks multiple vehicles with orientation changes and prevalent occlusion, without requiring a special region to initialize each vehicle individually. Tests are performed on video sequences from busy street intersections and show very promising results.