Deformable Model-Based Vehicle Tracking and Recognition Using 3-D Constrained Multiple-Kernels and Kalman Filter

Video-based vehicle tracking and recognition is an important application in the Intelligent Transportation System. High similarity among vehicle types, frequent occlusion and low video resolution in traffic surveillance are the major problems in this research area. In this paper, we proposed a vehicle tracking system by using 3-D constrained multiple-kernels, facilitated with Kalman filtering, to continuously update the location of the moving vehicles. To further robustly and efficiently track vehicles that are partially or even fully occluded, evolutionary optimization is applied to camera calibration for systematically building 3-D vehicle model, from which we can extract the vehicle's features such as the vehicle type, color and license plate. Then, a self-similarity descriptor is further introduced for vehicle re-identification. The proposed system is evaluated on the NVIDIA AI City Datasets and one self-recorded high-resolution video. The experimental results have shown the favorable performance, which not only can successfully track vehicles under occlusion, but also can maintain the knowledge of 3-D vehicle geometry.