A fast multi-scale correlation filter tracking algorithm for anti-occlusion

Abstract

Aiming at the problem of scale transformation or occlusion in the process of visual object tracking, an anti-occlusion fast correlation filtering object tracking algorithm is proposed in this paper, which is mainly composed of three parts: position filter, fast scale filter and Kalman filter. Firstly, the position filter and fast scale filter were initialized. Secondly, during tracking, the average-peak to correlation energy (APCE) occlusion criterion was used to judge whether the target was occluded. If there was no occlusion, the target position was estimated by position, scale filter and feature. When the target was occluded, the current position of target occlusion was estimated by Kalman filter instead of the estimated value of position filter. Finally, the correlation tracking results were calculated. The experimental results show that when tested on the CVPR 2013 benchmark public data set, the algorithm is compared with the mainstream algorithm, the accuracy of video sequences dealing with occlusion and scale transformation attributes is improved by 0.047, the success rate is improved by 0.067, the center position error is improved by 21.128 pixels, and the distance accuracy is improved by 0.153. It can better solve the problems of occlusion and scale change, and has good tracking speed.

Keywords

• object tracking
• average peck-to correlation energy
• occlusion criterion
• fast scale filter
• kalman filter