Research on object tracking algorithm based on Siamese network

Object tracking has long been a critical research direction in the field of computer vision, playing a vital role in practical applications such as human-computer interaction, autonomous driving, and video surveillance. Siamese network-based algorithms have demonstrated excellent performance in testing, balancing tracking accuracy and computational speed, and show promising development prospects. This paper proposes a Siamese network-based object tracking algorithm with the following key innovations:(1) Feature Extraction Enhancement: The original Siamese algorithm employs AlexNet for feature extraction. To leverage deeper networks, this work replaces AlexNet with EfficientNetV2. Further improvements are made to adapt EfficientNetV2 for tracking tasks, including integrating dilated convolutions to expand the receptive field and adjusting network structure parameters such as feature map resolution and stride, thereby enhancing feature extraction capabilities.(2) Anchor-Free Design: An anchor-free approach is proposed to reduce network parameters and training memory consumption. By eliminating predefined anchor box dimensions, the algorithm gains greater flexibility in generating bounding boxes. Additionally, a channel attention module is embedded into the classification branch, enabling the network to learn channel-wise weights and prioritize channels containing critical information, thereby improving discriminative power.(3) Multi-Level Feature Fusion: Unlike other computer vision tasks, object tracking requires both high-level semantic features to distinguish targets from backgrounds and low-level positional features for precise localization. This work fuses detection results from high-level and low-level features to achieve more accurate target tracking.To validate the algorithm ’ s effectiveness, evaluations are conducted on public datasets OTB50. Comparisons with state-of-the-art real-time trackers such as SiamFC++, DaSiamRPN, GradNet, and ATOM demonstrate the superior performance of the proposed method.