Linear-Complexity Gate Approximation Tracker

In response to the complex challenges of object tracking tasks, such as occlusion, fast motion, and shape deformation, existing Siamese networks have integrated Transformer structures based on attention mechanisms. Through their remarkable ability for global modeling and modal adaptation, these structures have significantly improved tracking precision and robustness. However, the global attention mechanism in Transformers introduces computational complexity quadratic to the length of the input sequence, thereby limiting its inference speed and affecting its widespread applicability in practical scenarios. To address these challenges, this paper innovatively proposes a linear complexity tracking algorithm based on gate approximation, termed the Gate Approximation Tracker (GAT). Specifically, we introduce a dual-branch attention architecture into the existing Siamese network framework and enhance the feature extraction capability of the tracker through the design of a multi-branch feature fusion strategy. Building upon this, we construct a linear-complexity approximation module, which based on gated attention and a linearized multi-layer perceptron architecture, significantly improves computational efficiency while retaining excellent adaptability to target features. Experimental results demonstrate that on the OTB2015 dataset, the GAT algorithm achieves state-of-the-art performance with a success rate of 70.4% and can achieve real-time tracking at 40FPS, showcasing its robust tracking capability.