Abstract:

A novel frame work of multi-target tracking (MTT) based on block division, saliency centroid modeling and multi-level association is presented to enhance the robustness and accuracy of tracker under occlusion among targets and disturbance caused by similar target．Selfadaptive threshold value based background difference is employed to detect motion regions. Based on blockdivision of target region, the blocklevel color saliency computed from background difference at each moving pixel, is utilized to model the centroid being with motion and color saliency. To discriminate the targets and blocks being in occlusion, the global and bolck associations are established among tracked targets and motion regions, and the block occlusion matrix is built to update the target model. The valid color and motion pixels are utilized to calculate the shifting vetor and fusion weight of each block centroid, then the global centroid shifting vetor is gained and used to locate the target. Experiments demonstrate that the proposed method is robust enough for tracking the multi-target in scenarios of occlusion, similar target disturbance and appearance change.