Abstract
Detection of abnormal human postures holds significant application value in fields such as smart healthcare, public safety, and elderly care monitoring. Timely and accurate detection of abnormal postures provides reliable support for emergency alerts and intervention. However, challenges such as occlusion interference in complex scenarios, pose variations, data imbalance, and excessive model parameters often hinder detection accuracy and real-time performance. To address these issues, this paper proposes PSD-YOLOv8n (Pose-Spatial-Dynamic-YOLOv8n), a lightweight human abnormal posture detection algorithm based on YOLOv8n. This approach incorporates a PoseMSA module integrating spatial and channel attention, combined with separable convolutions and residual bottleneck structures to enhance feature extraction while reducing model parameters. It introduces a keypoint-aware KA-Sample upsampling module for high-fidelity spatial reconstruction of abnormal posture regions. Furthermore, a Detect-PSA detection head integrating multi-scale features and distributed regression is constructed to effectively strengthen the model's ability to model global spatial relationships. Experimental results on the self-built SSHDataset demonstrate that PSD-YOLOv8n achieves 97.8% and 75.8% on the mAP@0.5 and mAP@0.5:0.95, respectively. These results not only surpass the baseline YOLOv8n model but also outperform advanced models such as YOLOv9-T, YOLOv10n, and YOLOv11n. Concurrently, the model features a compact parameter count of 2.07 M and a weight size of merely 4.5 MB, fully demonstrating its lightweight yet high-precision advantages. Visualization experiments further validate the method's robustness in complex environments and diverse postures. PSD-YOLOv8n provides an efficient and reliable solution for detecting abnormal human postures.