Object Tracking Algorithm Based on Channel-interconnection-spatial Attention Mechanism and Siamese Region Proposal Network

Abstract

The target tracking algorithm based on the Siamese network has become one of the most mainstream and best tracking algorithms because of the balance of accuracy and speed. However, target tracking algorithms based on the Siamese network are affected by factors such as occlusion, illumination changes, motion changes, size changes and other factors in natural scenes, making designing a robust tracking algorithm a challenging task. In order to improve the feature extraction and discrimination capabilities of the algorithm in complex scenes, a tracking algorithm combining channel-interconnection-spatial attention mechanism was proposed. First a Siamese tracking framework with a deep convolutional network ResNet-50 as the backbone network was built to enhance feature extraction capabilities, then the channel-interconnection-spatial attention module was integrated to enhance the adaptability and discrimination capabilities of the model, then the multi-layer response maps were weighted and fused to make results more accurate, and finally the largescale datasets were used to train the network, and tracking tests on the benchmark OTB-2015 and VOT2016 and VOT2018 were completed. The experimental results show that the proposed algorithm is more robust and better adapt to complex scenes such as target appearance changes, similar distractors, and occlusion than the current mainstream.