Target Tracking Algorithm Combining Twin Network and Semantic Segmentation

Abstract

Target tracking is widely used in automatic monitoring, vehicle navigation, video marking, humancomputer interaction, and automatic driving scenes. However, the accuracy of target tracking is limited by the influence of object deformation, lighting, occlusion, background interference, and other factors. In this study, the tracking accuracy of the twin network target tracking algorithm is reduced under the conditions of object deformation, illumination, occlusion, background interference, and so on. SiamUNet is a three-twin network target tracking framework proposed in this study. The basic network structure of SiamUNet is U-NET. The algorithm uses four-layer feature extraction, and after feature fusion and a four-step up-sampling process, the semantic information and background information of the target is fully utilized. At the same time, SiamUNetjudges whether each pixel belongs to the target by predicting the binary segmentation mask of the target in each frame of the video to obtain more accurate target information. SiamUNet was evaluated on VOT-2021, OTBIOO, and OTB50 datasets and compared with five popular trackers. Experimental results showed that SiamUNet had better tracking.