Expandable residual attention-based high-performance embedded gesture recognition
Shuyu Chen · 2024
In embedded human-computer interaction systems, the development of high-performance gesture recognition technology is crucial due to its demand for low power consumption and efficient processing. Addressing the challenge of highprecision gesture recognition in complex backgrounds, a high-performance embedded gesture recognition method based on the Expandable Residual Attention mechanism is proposed. This method enhances the capability of extracting differentscale gesture features by introducing the Expandable Residual Attention mechanism into YOLOv7. Additionally, to address the characteristics of high degrees of freedom and self-occlusion in hand gestures, SoftNMS is introduced with a penalty term to effectively reduce the probability of target omissions. Finally, the gesture recognition model is compressed and accelerated with TensorRT. Experimental results on the Jochen Triesch Static Hand Posture Database demonstrate that the proposed method significantly improves gesture recognition accuracy while maintaining high inference efficiency.