Digital Assistant to Operator

AI technologies gain value when they become an essential tool for the industry. In this article, we show the result of turning image recognition technology into an industrial tool for vehicle fleet management. The relevance of the drivers' actions monitoring task is substantiated. The choice of a object detection model was carried out among two-stage and single-stage models. R-CNN, Fast R-CNN, Faster R-CNN, R-FCN, SSD, RetinaNet, CenterNet and YOLOv4 were considered. Comparison of the different object detectors specifications is convinced us to choose YOLOv4. This choice is due to a good balance between performance, the ability to control the probability of false alarms, high recognition accuracy, and the presence of a special simplified version. In the paper we presented the results of YOLOv4 and YOLOv4-tiny research. The result is applicable to solving the problem of real-time recognition of driver actions, including on low-power platforms.