A Gait Recognition Algorithm Based on Global-Local Fusion and Adaptive Convolution

Gait recognition is a biometric technology that identifies or verifies an individual’s identity by analyzing their walking patterns. In modern deep learning models, effectively integrating global and local features is crucial for enhancing the model’s perceptual and representational capabilities. To address this, we propose a gait recognition algorithm based on Global-Local Fusion and Adaptive Convolution. By integrating traditional 2D convolutions with adaptive convolutions, the algorithm leverages both global contextual information and local details. Additionally, incorporating a long-short temporal feature extraction module enables the model to simultaneously capture global dynamic information and local sequential details, significantly improving its ability to perceive complex temporal patterns. Extensive experiments on the CASIA-B dataset validate the superior performance of our model, achieving rank-1 accuracy rates of $\mathbf{9 7. 8 \%}, \mathbf{9 4. 3} \%$, and $\mathbf{8 3. 6 \%}$ under normal walking, carrying a bag, and wearing a coat conditions, respectively.