Computer Vision-Based Personnel Management in Complex Scenarios of Industrial Enterprises

Some companies have extremely strict requirements for the safety management of personnel and equipment in the production plant, including real-time counting of the number of personnel. However, in practice, it may not be convenient to use traditional access control for real-time counting due to production efficiency. This study aims to design a video streaming-based visual counting method for personnel at critical entrances and exits. To achieve this goal, we extensively evaluated and compared the performance of traditional target detection algorithms to find the one that is suitable for our task scenario. Subsequently, we delved into the RCNN [7] family of algorithms and the YOLO [8], [9]family of algorithms for both two-phase target detection algorithms and single-phase target detection algorithms. We used manually annotated datasets to train the relevant models and tested these algorithms for performance evaluation in task scenarios. In the end, we chose the YOLOv7 and StrongSORT [9], [13] based algorithms to successfully implement the personnel counting task and achieved a small error through model optimization to make it suitable for complex scenarios.