A Comparative Study of Object Detection and Keypoint-Based Methods for Static Chinese Sign Language Recognition

This study addresses the challenges of data scarcity and gesture complexity in Chinese Sign Language (CSL) recognition by constructing a high-quality image dataset that covers 19 common vocabulary categories across various real-world scenarios. Two static CSL recognition methods are systematically compared: YOLO-based object detection and a keypoint-based approach that utilizes MediaPipe and a 1D CNN for gesture classification. Experimental results show that both methods achieve overall accuracy close to 90%. The keypoint-based method demonstrates superior performance in distinguishing subtle gestures, while object detection performs better in complex backgrounds and real-time scenarios. The results provide a systematic performance evaluation and practical guidance for future CSL recognition system design.