Object tracking algorithm with multi-scale channel attention

Abstract

In recent years, the application of Siamese networks for visual target tracking has brought a great improvement in tracker performance, allowing both accuracy and real-time performance. However, the accuracy of Siamese network trackers is largely limited. Object tracking algorithm with multi-scale channel attention is proposed based on the Siamese object tracking algorithm of fully convolutional classification and regression. Firstly, the backbone network ResNet50 is improved, which combined with multi-scale channel attention for feature extraction and enhancement. Then, adding a spatial attention mechanism to focus on information about the location features of the target image within each channel after the features are extracted from the template branch. At last, it performs fusion, classification and regression successfully. Compared with other advanced trackers, our approach achieves higher accuracy and success rate, especially in complex scenarios such as fast motion, occlusion, similarity interference, and scale changes.