Study on the edge computing method for skeleton-based human action feature recognition

Human action recognition is an important task in intelligent security monitoring field. Traditional action recognition method uploads the original image signal to the cloud server in a centralized manner, which brings problems such as network congestion, server computing pressure and computing delay. This paper proposes an edge computing-based multi-time scale skeleton feature fusion method for action recognition. Firstly, the spatial positions of human key points are extracted from the original images of human action and the skeleton features of human action are constructed. Then, the skeleton feature extraction and recognition task are deployed to multiple edge nodes. In the edge nodes the skeleton features on different time scales are extracted and recognized respectively. The results in all the edge nodes are uploaded to the cloud server and fused to make a decision. The proposed method not only achieves the dynamic scheduling of computing resources through adjusting the number of edge nodes according to the accuracy requirements, which relieves network congestion and lightens server computing pressure, but also significantly improves the recognition accuracy, which has important practical application value for the human action recognition in intelligent security field. © 2020, Science Press. All right reserved.