LCTrack: Leveraging Convolutional Neural Network and Transformer for Visual Tracking

In recent years, with the support of Transformers' ability to integrate information from the target region and search region, the performance of relevant trackers has dynamically improved. However, these Transformers methods that capture global information about the tracking target often overlook local information, which significantly affects the performance of the corresponding trackers. To address this issue, we propose a novel tracking framework that uses both convolutional neural networks (CNN) and Transformers for visual tracking, called LCTracker. Our main contributions are as follows: 1) We integrate a new tracking framework that combines local information extraction from CNNs and global information extraction from Transformers; 2) We incorporate a multi-layer Transformer aggregation module; 3) We implement cross-attention between CNN feature maps and Transformer attention maps. Additionally, extensive experiments demonstrate that our LCTracker can effectively handle challenges such as fast motion, background clutter, and low resolution, showing significant advantages in these areas. Impressively, LCTracker achieved leading performance in three tracking benchmark tests and significantly outperformed SOTA results.