Yolov5-Based Attention Mechanism for Gesture Recognition in Complex Environment

-Object detection is a fundamental task in gesture recognition, involving identifying and localising human hand or body gestures within images or videos amidst varying environmental conditions. To address the inadequate recognition rate of gesture detection algorithms in intricate surroundings caused by issues such as inconsistent illumination, background colors resembling skin tones, and diminutive gesture scales, a gesture recognition approach termed HD-YOLOv5s is presented. An adaptive Gamma image enhancement preprocessing technique grounded in Retinex theory is employed to mitigate the effects of lighting variations on gesture recognition efficacy. A feature extraction network incorporating an adaptive convolutional attention mechanism (SKNet) is developed to augment the network's feature extraction efficacy and mitigate background interference in intricate situations. A novel bidirectional feature pyramid architecture is implemented in the feature fusion network to fully leverage low-level features, thereby minimizing the loss of shallow semantic information and enhancing the detection accuracy of small-scale gestures. A cross-level connection strategy is employed to enhance the model's detection efficiency. To assess the efficacy of the suggested technique, experiments were performed on a custom dataset featuring diverse lighting intensity fluctuations and the publicly available NUS-II dataset with intricate backdrops. The recognition rates attained were 99.5% and 98.9%, respectively, with a detection time per frame of about 0.01 to 0.02 seconds.