Object Tracking Algorithm Based on Multi-Time-Space Perception and Instance-Specific Proposals

Abstract

Aiming at the problem that a single correlation filter model is sensitive to complex scenes such as background interference and occlusion, a tracking algorithm based on multi-time-space perception and instancespecific proposals is proposed to optimize the mathematical model of the correlation filter (CF). Firstly, according to the consistency of the changes between the object frames and the filter frames, the mask matrix is introduced into the objective function of the filter, so as to extract the spatio-temporal information of the object with background awareness. Secondly, the object function of multi-feature fusion is constructed for the object location, which is optimized by the Lagrange method and solved by closed iteration. In the process of filter optimization, the constraints term of time-space perception is designed to enhance the learning ability of the CF to optimize the final tracking results. Finally, when the tracking results fluctuate, the boundary suppression factor is introduced into the instance-specific proposals to reduce the risk of model drift effectively. The accuracy and success rate of the proposed algorithm are verified by simulation analysis on two popular benchmarks, the object tracking benchmark 2015 (OTB2015) and the temple color 128 (TC-128). Extensive experimental results illustrate that the optimized appearance model of the proposed algorithm is effective. The distance precision rate and overlap success rate of the proposed algorithm are 0.756 and 0.656 on the OTB2015 benchmark, which are better than the results of other competing algorithms. The results of this study can solve the problem of real-time object tracking in the real traffic environment and provide a specific reference for the detection of traffic abnormalities.