A Visual Tracking Algorithm Based on Estimation of Regression Probability Distribution

Abstract

During target tracking, the target is often interfered by uncertainties like occlusion and motion blur. The interference leads to inaccurate tracking and even loss of the target. To solve the problem, this paper designs a target tracking algorithm based on the estimation of regression probability distribution (RPDE). Specifically, the uncertainty degree of the tracking frame was estimated by learning the statistical properties of regression parameters, and the quality of that frame was evaluated by fusing the predicted regression probability scores with classification scores. Next, an anchor-free regression mechanism was introduced to improve the computing speed. During network training, a simple and efficient strategy was presented for joint prediction, which jointly expresses classification scores and regression scores to eliminate the extra quality estimation branches in training and prediction. After that, the performance of our algorithm was tested on several public benchmarks, namely, OTB2015, VOT2016, GOT10k, and UAV123, and contrasted with several state-of-the-art algorithms. The results show that the proposed algorithm, named SiamRPDE for short, performed excellently on several benchmarks, and achieved the speed of 125 frames per second (FPS).