Gesture Recognition in Complex Backgrounds Based on FCA-YOLOv8
Taiping Mo, Chang Ying, Peng Sun, Xiangwen Zhang · 2025
Gesture recognition is a widely used human-computer interaction method, particularly in smart home and smart industry applications. However, its accuracy is often compromised by factors such as skin-like objects, light intensity, and environmental clutter. To mitigate the impact of complex backgrounds, this paper proposes a recognition method, FCA-YOLOv8, based on edge enhancement and contextual guidance. First, we introduce the EdgeFusionNet module, which incorporates the Sobel operator to enhance the model's ability to extract object edge information. Second, in the neck part, the ContextAwareFusion module is applied, combining the SEAttention mechanism to improve feature representation by guiding contextual information and adaptively adjusting multi-scale feature fusion. This dynamic adjustment emphasizes important features, suppresses background noise, and enhances the model's detection accuracy and robustness. Finally, the Head-Attn module is proposed for the head section, which extracts multi-scale features using deeply separable convolution, fully connected layers, and exponential normalization. The attention mechanism strengthens the feature response of occluded regions, improving detection accuracy and robustness to occluded targets. The model demonstrates superior performance on the NUS-II dataset.