In order to improve the tracking performance and tracking speed of the Siamese tracking algorithm in complex situations such as fast motion and similar objects tracking,a Siamese target tracking algorithm combining residual connection and depth separable convolution was proposed.The 5×5 convolution in the original feature extraction network was replaced with a normal 3×3 convolution,by which the amount of network calculations could be reduced and its ability to learn features could be improved.A smaller computational depth separable convolution was used to replace all the ordinary 3×3 convolutions in the original network,not only the network inference speed could be accelerated,but also the depth of the feature extraction network could be deepened,so a more characterizing ability for the target deep semantic information was obtained.A residual connection was added to the deep separable convolution module to form a residual block,to fuse the features of different layers extracted by using the network and improve the utilization of feature information.The results show that the proposed algorithm has improved tracking accuracy and success rate,and is superior to other algorithms in real-time and reliability.