Aiming at the problems of fast movement，occlusion，non-rigid deformation and illumination change of objects in complex scenes，an object tracking algorithm based on 3D attention and pyramid decoder is proposed. Firstly，VGG-16 neural network is introduced and its structure is optimized to improve the efficiency and quality of feature extraction. Secondly，by introducing 3D attention，the extraction ability of key features is enhanced. Then，the deep semantic fusion module is used to fuse feature information through upsampling to achieve accurate expression of features. Finally，a pyramid decoder is designed to improve the robustness of the model in complex scenes. Experimental results show that the success rate and tracking accuracy on OTB100 data set are improved by 15.8% and 16.2% respectively compared with those of the baseline algorithms.