Work Safety Assessment through Contextual Analysis with Computer Vision

Work accidents are a reality for which companies must establish constant efforts to reduce both their likelihood of occuring and their impact should they occur and, where possible, eliminate. Accidents are caused by hazards which depend on the type of occupation and the activity carried out in the company. In Honduras, the third largest occupational group is the manufacturing industry. With the aim of reducing the impact of accidents, this research seeks to implement computer vision as a method of monitoring compliance of established safety measures. Firstly, the safety measures used in the industry are identified for the purpose of defining the measure to monitor with computer vision. Data is collected qualitatively in the form of images to perform the training and testing of the object recognition algorithm, YOLOv3-Tiny. Consequently, by applying a quantitative methodology, the performance of the prototype program is measured with the accuracy of the system compared to the ground truth established through the manual annotation of the images in the test data set. Finally, an evaluation of compliance with safety measures is incorporated through the qualitative labeling of the risk zone and the results delivered by the object recognition algorithm.