CG-YOLO: A Lightweight, High-Accuracy Gesture Recognition Method for Human-Computer Interaction

Although gesture recognition technology has made significant progress, its accuracy and robustness in practical applications are still insufficient due to factors such as complex environments, lighting variations, and background interference. This paper proposes a lightweight and high-accuracy humancomputer interaction gesture recognition method called CGYOLO to address these issues. We reconstruct the backbone network of YOLOv8s by introducing a Contextual Feature Aggregation Backbone Network (ContextRepViT) to enhance the model's accuracy in recognizing small-sized hand targets, thus solving the problems of missed and false detections in complex backgrounds. Additionally, we propose a lightweight feature extraction module (GSRepNCSP) to redesign the Contextual 2D Features (C2f) modules in the Head network of YOLOv8s, improving parameters utilization and reducing computational load. Experimental results on the public HaGRID dataset show that, compared to YOLOv8s, the CG-YOLO reduces the number of parameters by 1.3M, and improves Precision, Recall, [email protected], and [email protected] by 3.2%, 1.5%, 2%, and 3.9%, respectively. Moreover, the human-robot interaction experiments with the quadruped robot also verified the effectiveness of the CG-YOLO.