Implementation of Object Tracking Using YOLOv7 Pose Estimation for Human Activity Recognition in Fabrication Yard via CCTV Videos

A new dataset is created in a fabrication yard environment for human activity classification. YOLOv7 object detection model is used to identify 5 classes which are welding, grinding, standing, sitting and walking. Five experiments were conducted, and Experiment 5 achieved the highest mean Average Precision (mAP_ 0.5) of 0.621. The mAP@0.5 value is the most critical metric as it reflects the model's ability to balance precision and recall across all confidence thresholds. This higher mAP@0.5 suggests that Experiment 5's model is better at accurately identifying. However, YOLOv7 object detection posed limitation to differentiate classes for sitting, standing and walking. The implementation of pretrained models YOLOv7 Object Tracking and YOLOv7 Pose Estimation is done respectively with YOLOv7 Object Detection to improve the detections. The YOLOv7 Object Tracking showed significant improvement in the detections of standing and walking but had limitation for sitting. The YOLOv7 Pose Estimation which uses the heuristic method showed significant improvement in the detections of standing, sitting and walking compared to YOLOv7 Object Tracking.