Multi-person Action Recognition Based On HARNet Network In The Park Environment

In China’s urbanization process,the increasing number of town parks meet people’s increasingly elevated living needs and promote the high-quality development of town construction,but also accompanied by social problems about park safety management,such as the elderly fall,children fall,sedentary and other safety incidents occur from time to time.Currently,park safety management is based on manual real-time surveillance video,which has problems such as low efficiency,high cost and poor stability.Therefore,how to solve the problem of identifying unsafe behaviors of visitors in the park environment is significant.Based on this,in order to ensure the safety of park visitors,this paper takes the park environment as the research object and aims to study a multi-person action recognition method serving the park environment.The main work of this paper is as follows:1.Analyzing the network structure and characteristics of the human pose estimation method,proposing a human pose estimation method based on the High-Resolution Graph Convolutional Network(HRGCN),which treats human keypoints as graph nodes and processes them through graph convolutional neural networks to improve the accuracy of keypoints positioning.The effectiveness of this method was verified by comparing it with multiple methods on the MPII dataset and the COCO 2017 dataset.2.Proposing a Heatmap-based Action Recognition Network(HARNet)to address the shortcomings of action recognition methods based on static images in human action recognition.The network takes human keypoint heatmaps as input and first pre-processes the confidence threshold of keypoints to avoid the influence of factors such as occlusion,image quality,and human appearance on keypoints.It then learns the effective relationship between human keypoints and further strengthens the probability of correct action classification through a classification fusion module.The effectiveness and applicability of HARNet are verified on a custom outdoor action dataset.3.Taking a park in Shenzhen as a practical application scenario,we apply HRGCN and HARNet network to human action recognition,and propose a multi-person action recognition method based on HARNet network to realize visitor action recognition in the park environment.Firstly,video frames are extracted from the park environment video captured by surveillance in real time;secondly,the YOLOv5 target detection model is used to detect all visitor target frames in the video frames and crop the target frames;again,the HRGCN network is used to predict the human key point heat map of the target frame image;finally,the HARNet network extracts the human key point heat map features and learns the relationship between human skeleton structure and action to achieve accurate action recognition.A custom park action dataset is constructed and experimented to verify the superiority of HARNet network in park scenarios.