Research on Enhanced YOLOv8 Gesture Recognition Method for Complex Environments*

This study proposes a gesture recognition method, named DF-YOLOv8s to address the issue of low recognition rate under complex environments. Our approach firstly replaces the SPPF module with the AIFI module to enhance feature extraction effectiveness by using attention-based internal scale feature interaction. Then we design a novel ZD feature fusion network according to the ASF-YOLO network structure, which improves our capability to extract and fuse features in images, thereby enhancing gesture recognition accuracy under complex environments. This improvement comprehensively considers features of different sizes, occlusion and uneven illumination that may result in small target loss. Finally, experiments are performed on our self-built dataset from publicly available datasets NUS-II and HaGRID, which riches lighting contrast, skin-colored background, and foreground occlusion characteristics. Our experimental results represent mAP50 score of 95.4% with a 2.8% enhancement over YOLOv8s. And also we have performed contrasted tests with other algorithms, the results illustrated the validity of the proposed method.