Deep Learning-Based Human Pose Anomaly Recognition Using Stable Diffusion Generated Images and EfficientNetV2 Classification
Authors-S.Akhila, T. kavya
Keyword-Human Pose Anomaly Detection, Stable Diffusion, EfficientNetV2, Deep Learning, Image Classification, Industrial Safety, Artificial Intelligence, Worker Monitoring, Computer Vision, Synthetic Dataset.
Human pose anomaly recognition plays a significant role in improving workplace safety, intelli-gent surveillance, healthcare monitoring, and human–robot collaboration. Conventional ap-proaches generally rely on pose estimation algorithms to extract skeletal keypoints before per-forming anomaly classification, making the overall pipeline computationally complex and highly dependent on pose estimation accuracy. This research introduces a simplified image-based anomaly detection framework that directly classifies worker poses without employing any inter-mediate pose estimation module. A synthetic dataset containing both normal and abnormal in-dustrial worker poses is generated using the Stable Diffusion image generation model, allowing the creation of a large and consistent training dataset with well-defined anomaly categories. The generated images are subsequently used to train EfficientNetV2 deep convolutional neural net-work variants for binary pose classification. The proposed framework reduces processing com-plexity while maintaining excellent recognition capability. Experimental evaluation demonstrates that the EfficientNetV2-M architecture provides the highest classification performance, achieving an accuracy of 95.66%, outperforming the remaining EfficientNetV2 variants. The findings indicate that direct image classification is capable of learning discriminative pose representations without requiring explicit skeletal information. The proposed framework provides an efficient, scalable, and practical solution for industrial safety monitoring and intelligent worker surveil-lance applications.
Doi-[https://doi.org/10.5281/zenodo.21735776]