Abstract:A multi-scale Transformer online update workpiece tracking algorithm is proposed to address the issues of low accuracy and high failure rate in industrial scene target workpiece tracking tasks.Firstly,a Transformer feature pyramid structure is adopted to fuse multi-level feature information to achieve robust apparent modeling of the target;Secondly, using the Transformer module for feature fusion of advanced semantic information enables the network model to focus on the target artifact itself;Then,an loU Loss function optimization strategy based on sorting is proposed to effectively suppress the influence of the interference on the tracker;Finally,design an online update strategy to update the target template and enhance the robustness of the network.The experimental results show that the accuracy and failure rate on VOT-2018 are 3.8%and 4.1%higher than the benchmark tracker,respectively,and can maintain a real-time tracking speed of 53 fps;The accuracy and success rate on the LaSOT dataset are 0.578 and 0.573,both of which are better than the benchmark tracker.The algorithm proposed in this paper can accurately and robustly track the target workpiece by capturing video sequences using a CCD industrial camera.