<p>As a traditional Chinese health-preserving exercise, Baduanjin has gained widespread popularity among the general public due to its gentle movements and ease of learning and practice. The head, serving as the pivotal point for the ascending and descending of vital energy within the human body, directly influences the health benefits of practicing Baduanjin. However, existing head motion detection methods still face challenges from complex background interference, partial occlusion, and variable movement poses, hindering the achievement of satisfactory results. This study proposes a multi-branch direction perception fusion network for detecting head movements in Baduanjin (MDF-YOLO) based on the YOLOv11n framework. Specifically, a multi-branch frequency-domain spatial fusion module (MBFS) is designed to effectively suppress background noise by leveraging enhanced frequency-domain features, thereby improving the representation of head-region characteristics. A multi-scale feature fusion module (MSF) is constructed that employs a hierarchical feature aggregation strategy to enhance robust feature representation under occlusion conditions. A multi-scale direction-perception module (MSDP) and an upsampling module (UDCS) are designed, which significantly improve the model’s ability to distinguish multi-pose head movements by enhancing directional gradient features. Experimental results on the self-built BDJH dataset demonstrate that MDF-YOLO achieves 94.5% mAP@0.5 and 63.6% mAP@0.5:0.95, substantially outperforming baseline models. This work provides a viable technical pathway for the digital analysis and intelligent inheritance of traditional health-preserving exercises.</p>