Grid-based multi-object tracking with Siamese CNN based appearance edge and access region mechanism

Receiving growing attention for its various applications during the last few years, multi-object tracking remains a complex and challenging problem. Conventional grid-based tracking method is an efficient and effective method to tackle multi-object tracking, whose performance can be further boosted by intuitively taking into account the appearance similarity information yet. Therefore, we introduce appearance similarity edge into the grid-based method, where a Siamese network is utilized to produce the proposed similarity edge. In addition, we build a grid model with hexagonal cells and propose an access region mechanism including accessible area definition and an automatic-generation approach for entrance/exit grids. Since our tracking framework follows 'tracking-by-detection' paradigm, the corresponding detection information is available to be integrated into access region mechanism, which will facilitate appropriate grid modeling. We verify the proposed Siamese network based appearance edge and access region mechanism through the experiments on some popular datasets like PETS-09, KITTI.