EDGE AI-DRIVEN DUAL-MODE WORKPLACE SAFETY AND INDUSTRIAL INSPECTION SYSTEM

This paper has presented a complete Edge AI system for dual-mode workplace safety monitoring and industrial product inspection, deployed on the NVIDIA Jetson Orin Nano 4 GB Developer Kit with TP-Link Tapo C200 RTSP cameras. The system integrates five AI subsystems—YOLOv8n TensorRT FP16 object detection, InsightFace ArcFace face recognition, a convolutional autoencoder for unsupervised anomaly detec- tion, a DeepStream GStreamer multi-stream pipeline, and a FastAPI real-time streaming and alerting backend—within a unified architecture constrained by a 4 GB unified memory budget. A process controller enforces mutually exclusive GPU allocation between the two operating modes, ensuring hard- ware memory limits are never violated. Experimental results demonstrate 38.5 FPS batch inference throughput for PPE detection, 14.7 ms face recognition latency,4.9 ms autoencoder inference latency, and end-to-end alert latency of approximately 130 ms. The system achieves an overall PPE detection [email protected] of 89.3% and an anomaly detection AUROC of 0.913, validating practical utility for real deployment. The complete implementation is structured as a modular, documented reference that can be adapted to other edge AI monitoring applications on Jetson hardware. Future work will investigate INT8 quantisation for the TRT engine, DLA offloading for the autoencoder, online threshold adaptation, and multi-device federation to lift the single-device mode exclusion constraint while maintaining the local data sovereignty and low-latency guarantees of edge deployment.

[1]                          T. H. Vo and G. S. Lee, “A vision-based method for automatic detection of personal protective equipment usage on construction sites,” Automa- tion in Construction, vol. 69, pp. 97–107, 2016. [2]                          J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” in Proc. IEEE CVPR, 2016, pp. 779–788. [3]                          W. Liu, D. Anguelov, D. Erhan et al., “SSD: Single shot multibox detector,” in Proc. ECCV, 2016, pp. 21–37. [4]                          Y. Li, H. Wei, Z. Han, J. Huang, and W. Wang, “Deep learning-based safety helmet detection in engineering management,” J. Construction Engineering and Management, vol. 146, no. 8, 2020. [5]                          N. D. Nath, A. H. Behzadan, and S. G. Paal, “Deep learning for site safety: Real-time detection of personal protective equipment,” Automa- tion in Construction, vol. 112, p. 103085, 2020. [6]                          J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “ArcFace: Additive angular margin loss for deep face recognition,” in Proc. IEEE CVPR, 2019, pp. 4690–4699. [7]                          J. Guo and X. Zhu, “InsightFace: An open-source 2D and 3D deep face analysis library,” IEEE Trans. Biometrics, Behavior, and Identity Science, vol. 3, no. 3, pp. 380–391, 2021. [8]                          E. Principi, F. Vesperini, S. Squartini, and F. Piazza, “Acoustic novelty detection with adversarial autoencoders,” in Proc. IJCNN, 2017, pp. 3324–3330. [9]                          P. Bergmann, M. Fauser, D. Sattlegger, and C. Steger, “MVTec AD — A comprehensive real-world dataset for unsupervised anomaly detection,” in Proc. IEEE CVPR, 2019, pp. 9592–9600. [10]                      K. Roth, L. Pemula, J. Zepeda et al., “Towards total recall in industrial anomaly detection,” in Proc. IEEE CVPR, 2022, pp. 14318–14328.