To address the problem of insufficient representational capability and lack of online update of the fully-convolutional Siamese network (SiamFC) tracker in complex scenes, this paper proposed a visual tracking algorithm based on multi-attention and dual-template update. First, the feature extraction network was constructed by replacing AlexNet with the VGG16 network, and SoftPool was used to replace the maximum pooling layer. Secondly, the multi-attention module (MAM) was added after the backbone network to enhance the network’s ability to extract object features. Finally, a dual template was designed for feature fusion and response map fusion, and average peak-to-correlation energy (APCE) was used to determine whether to update the dynamic templates, which effectively improved the tracking robustness. The proposed algorithm was trained on the GOT-10k dataset and tested on the OTB2015, VOT2018, and UAV123 datasets. The experimental results show that, compared with the benchmark algorithm SiamFC, the tracking success rate of the proposed algorithm is increased by 0.085 and 0.037, and the accuracy is increased by 0.118 and 0.058 on the OTB2015 and UAV123 datasets, respectively. On the VOT2018 dataset, the tracking accuracy, robustness, and expected average overlap (EAO) are improved respectively by 0.030, 0.295 and 0.139. The proposed algorithm achieves high tracking accuracy in complex scenes, and the tracking speed reaches 33.9 frame/s, which meets the real-time tracking requirements.

The bounding box regression loss effect is limited, and the multi-scale feature representation ability is insufficient in the YOLOX algorithm, which leads to inaccurate detection results. To address this issue, an object detection algorithm based on distance shape of generalized intersection over union (DSGIoU) loss and dual branch coordinate attention was proposed. Based on the intersection over union (IoU) loss term, the regression convergence effect of the bounding box was optimized by adding three penalty terms: non-overlapping area, distance from the center, and aspect ratio between the true box and the predicted box. Meanwhile, the feature was encoded in two directions by using average pooling and max pooling to obtain directional perception information and position information, so as to enhance the feature. To demonstrate the detection performance of the proposed algorithm, YOLOX with network sizes of Tiny, S, and M was used as the benchmark to carry out tests on PASCAL VOC and KITTI datasets. The experimental results show that the detection accuracy of the proposed algorithm on the PASCAL VOC dataset reaches 80.0%, 82.6%, and 85.8%, respectively, which is 1.5%, 1.6%, and 2.0% higher than the YOLOX as the benchmark. On the KITTI dataset, the detection accuracy reaches 87.7%, 89.7%, and 90.7%, which is increased by 1.7%, 2.9%, and 1.3%, respectively. The proposed algorithm can optimize the network convergence, improve the representation ability of multi-scale features, and significantly boost the detection accuracy.