Lightweight Siamese Tracking based on Cross-correlation Attention Mechanism

When performing object tracking tasks, precisely tracking the desired target is the optimization goal of the tracking algorithm. Generally, deep learning models with larger model sizes and higher computational costs tend to have stronger computational capabilities. However, uncontrolled increases in model size and computational cost can make the algorithm difficult to use in practice. Therefore, to find a suitable balance between increasing computational cost and improving computational performance, this paper introduces a lightweight object tracking method that introduces a CNN attention mechanism. Based on SiamRPN++, this algorithm employs a more concise corner prediction head and uses depthwise separable convolution to replace general convolution, significantly reducing the computational cost of ResNet-50. Moreover, symmetric point-wise expansion (PE) and point-wise linear projection (PLP) are utilized to maximize computational performance while minimizing computational cost. Additionally, the Squeeze-and-Excitation (SE) network is employed to introduce an attention mechanism to each convolutional module. These strategies effectively make the algorithm lightweight and improve its computational efficiency.