Performance Evaluation of Lightweight Object Detection Models for Real-Time Personal Protective Equipment Detection in the Construction Sites

The environment of construction industry was known to have a high risk and high number of occupational accidents and injuries. One of the main causes of the occurrences was the construction workers' negligence in wearing personal protection equipment. Computer vision-based approaches were developed to assist in personal protective equipment adherence to address this issue. Using lightweight machine learning algorithms, object recognition can help to detect if the PPEs are worn correctly. We evaluated performance of YOLOv8-Nano and YOLOv9-Tiny (state of the art lightweight object detection models). Custom dataset was used for training the models and then metrics like F1 score, precision, recall mAP50 and mAP50-95 were used to evaluate both models’ performance. Results found that both models were able to show promising real time detections, but the YOLOv9-Tiny model was able to outperform the YOLOv8-Nano model on many evaluation metrics. Specifically, in terms of mAP, YOLOv8-Nano achieved an mAP50 of 81.48, while YOLOv9-Tiny attained a slightly higher mAP50 of 82.70. Higher efficiency in these parameters will help small industry to enforce PPE adherence monitoring using edge device at a relatively low cost. Lastly, enhanced enforcement of PPE regulations through automated detection system can contribute to improve workplace safety which in turns will lead to less injuries. Keywords: Object Recognition, Computer Vision, Machine Learning, Lightweight, Personal Protective Equipment, YOLO