Quantitative refueling action recognition algorithm

Abstract

An algorithm for identifying the action of quantitative refueling that can be deployed to edge equipment is proposed for the problem that refueling' action in production scenarios is not subject to real-time supervision. The algorithm firstly uses a YOLOv5s-improved object detection network for rapid human detection, then uses a tracking algorithm that combines IOU and histogram similarity to track the detected human. The traced sequence images are used to predict the skeletal key-point sequence of the human body through a quantitative pose estimation network, and finally, the skeletal key-point sequence data is input into the fully-connected network classifier on the sixth floor for action classification, to determine whether the refueling's actions are normally completed. Experimental data show that the algorithm greatly reduces the network weight and calculation amount. The human body detection speed on the BITMAIN Sophon SE5 terminal can reach 18 ms, and the action detection accuracy can reach 95.92% on the actual scene dataset.