Yolo algorithms to improve the detection of personal protective equipment thought industrial computer vision

Velasquez R.A.; Lara J.V.M.; Velasquez R.A. · Proceedings of the 2023 IEEE 30th International Conference on Electronics, Electrical Engineering and Computing, INTERCON 2023 (2023) · DOI: 10.1109/intercon59652.2023.10326066

Resumen

During 2023, accidents at work are unexpected situations that cause losses in both human and material resources. In Peru there is current legislation that encourages a safe work environment, however, according to 2019 reports from the Ministry of Labor and Employment Promotion, 31.92% of accidents caused this year directly affected the cranial area. The use of personal protective equipment requires finding an optical way to identify the required safety equipment in real time. In response to this problem, the use of artificial vision or computer vision for helmet recognition. The trained model for online personal protective equipment classifier algorithms with computer vision in industrial applications allows to detect helmets, head, safety glasses, industrial vests with precision metrics greater than 88% according to light intensity. Additionally, it is possible to improve the accuracy of the other clusters by training with higher epochs, increasing batch processing from eight to sixteen, and using parallel processing. Despite the fact that the configuration profile for the training process carried out was basic, the database used, due to the varied labeling and number of images, acceptable results were obtained. Also, due to the selection of the YOLOV5m model, transfer learning training contributed to the training time without reducing the accuracy rates. Our findings is the recognition of a model with 5000 pictures with an accuracy of 88% from 400 to 800 lux and limited for the light intensity lower than 400 lux, with 78.63%. With light intensity higher than 800 lux, the recognition time of less than 0.3 seconds/image together with a rate of 98% with a low-cost camera and parallel processing with a Raspberry Pi version 4.

Instituciones: Universidad Nacional de San Agustín de Arequipa; Universidad Tecnológica del Perú