Online Multi-Object Tracking based on Improved Kernel Correlation Filter

Abstract

In multi-target tracking, the target's scale change and occlusion problems easily lead to target loss. At the same time, multi-target tracking algorithm of data association based on deep learning for target detection requires a high hardware platform and cannot run on a multi-core DSP. Aiming at the above problems, this paper proposes an online multi-target tracking algorithm based on improved kernel correlation filter, using a multi-core parallel mechanism to establish a tracker for each target for tracking. The improved algorithm integrates multiple features of Hog, CN and LBP at the feature level to enhance the ability to represent the target; for the problem of target scale change, a 1D scale filter is introduced to estimate the new scale of the target; In terms of model updating, the APCE indicator is used to measure the confidence of the response map in order to detect whether the target is occluded or lost in time. By adopting a high-confidence model update method, the correlation filter model is prevented from being polluted by the wrong samples when the target is occluded. When the target tracking fails, the detection module is triggered to detect the target: firstly, the correlation matching algorithm based on the hog feature is used to preliminarily screen, and the relevant surface is analyzed, and then the kernel correlation filter based on the ridge regression classifier is used to detect the multi-peak of the correlation surface. Experimental results show that the improved algorithm is more robust to target occlusion, scale changes, background interference, target loss re-tracking, etc. compared to standard algorithms. This paper effectively extends the improved kernel correlation filter algorithm to multi-target tracking, which can be further transplanted to multi-core DSP platforms.