Dense Compounded Features and Data Association Based Multiple People Tracking

Abstract

People tracking is a wide area of research in computer vision and machine learning. The challenging problem of multiple objects tracking (MOT) is further complicated by factors such as occlusion, varying number of targets, illumination variations and objects' appearances which may be similar.In this paper, a significant approach is proposed for MOT with a single static camera based on dense valuable global and local features. This includes cascaded HOG-III features, texture, and motion information of objects in order to build a robust tracking system. To speed up the system, a simple data association method is employed using Hungarian algorithm to associate candidate response to the target objects. Genetic algorithm is also used to provide a heuristic data association based multiple objects tracking. A comparison between two association algorithms is made based on the tracking results.TUD-crossing and TUD-campus datasets are used for validation purposes. A system's performance can be evaluated using a wide metrics of MOT performance indicators (MOTA, MOTP). The results reach (82.89 and 81.96) in terms of TUDcrossing dataset and (79.30 and 73.06) in terms of TUD-campus dataset respectively. The experiments show that the proposed method can be suitably employed during scale changes or in the presence of a cluttered background environment, in addition, our method achieves competitive results in comparison with state of the art approaches.