In order to balance the tracking performance and inference speed, a lightweight Siamese-based tracker named SiamAHG is proposed in this paper. It employs the lightweight network ShuffleNet V2 for feature extraction and a novel adaptive hierarchical graph attention for feature fusion. Specifically, the feature fusion sub-network consists of an adaptive feature map refinement module (AFMRM) and a hierarchical graph attention module (HGAM). AFMRM refines feature maps of different stages and strengthens the model expression by combining multiple convolutional kernels dynamically based upon diverse attentions. HGAM realizes node-level information interaction of the same stage between the template and search branches, and integrates the features of different stages to get complementary information for tracking. Experiments on LaSOT, OTB100, GOT-10k, and UAV123 validate the effectiveness and efficiency of our proposed tracker. It has significant advantages in model complexity and inference speed, while achieving competitive results with state-of-the-art trackers.
10 September 2025 in Signal, Image and Video Processing
Object tracking is now widely used in sports event broadcasting, security surveillance, and human-computer interaction. It is a challenging task for tracking on unmanned aerial vehicle (UAV) datasets due to many factors such as illumination change, appearance modification, occlusion, motion blur and so on. To solve the problem, a visual correlation filter tracking algorithm based on temporal and spatial regularization is proposed. It employs boolean maps to obtain visual attention, and fuses different features such as color names (CN), histogram of oriented gradient (HOG) and Gray features to enhance the visual representation. New object occlusion judgment method and model update strategy are put forward to make the tracker more robust. The proposed algorithm is compared with other six trackers in terms of distant precision and success rate on UAV123. And the experimental results show that it achieves more stable and robust tracking performance.
Recent years, attention mechanism has been widely used in computer vision, such as object detection, tracking and recognition. Many studies reveal that attention has better performance than those CNN-based or RNN-based networks. However, the cost of better performance, such as complicated structure and complex algorithm, cannot be neglected. In this paper, a lightweight transformer tracker named LTT is proposed. Different from Transformer Tracking network (TransT), three aspects of lightweight operation are adopted: firstly, YOLO-nano Darknet is utilized as the feature extraction network; secondly, the size of template image is scaled to 1/4 of the original and removing the self-attention layer; finally, the combination layer of convolution and cross-attention (long-short range attention) is adopted for the sake of feature fusion. Experiments show that our tracker runs at roughly 70 fps on GPU while there is no significant performance loss compared with other networks like SiamFc or SiamRPN++. Moreover, the model size of our tracker is only 4.58M.