Gesture Recognition Model Based on YOLOv8-AW
Yiqing Liu, Linxiao Zheng, Zijun Lu, Lin Wang, Lin Zhou · 2023
This paper proposes an improved deep model, YOLOv8-AW, based on the YOLOv8 algorithm for gesture detection. The paper enhances the feature pyramid network of the head by introducing the Asymptotic Feature Pyramid Network (AFPN). Feature fusion is conducted progressively, fostering better interaction between non-adjacent hierarchical levels to extract more valuable information and enhance the model’s perceptual capabilities for gesture features. To effectively reduce the genuine differences between predicted box dimensions and true box dimensions, modifications are made to the loss function. The Wise-IoU(WIoU) loss is introduced in the position regression loss function, focusing more on samples of ordinary quality, thereby improving the model’s generalization ability and overall performance. The WIoU loss has been introduced into the position regression loss function. Through experiments, the overall mAP of the improved model is increased by 0.2% over the basic YOLOV8, reaching 97.4%. The recall rate of gesture detection has increased by 8.3%, reaching 96.9%. The comprehensive performance shows that the improved model is more accurate in gesture recognition.