Handball Action Evaluation Method and System Based on Deep Learning

Abstract

Deep learning system aims to use hierarchical model to learn high-level functions from low-level functions. This study mainly discusses the handball action evaluation method and system based on deep learning. In the process of video data acquisition, the binocular RGB camera system is installed on a fixed iron frame 2.2m above the ground and 3M away from the end of the court in order to capture a wide range of field images including the whole plane of the court. In order to complete a variety of handball recognition in complex background environment, this paper builds a convolutional neural network to accept image input and output image categories, and selects Caffe as an open source, efficient and stable deep learning framework. In handball video processing, convolution neural network is used to detect the position of human body in each frame. Then, the pose of the human body in each bounding box is predicted, and the multi approximate pose of each frame is filtered, and a two-dimensional coordinate estimation of human joint is output. Finally, the 2D coordinates of human joints are used as input to fit the 3D coordinates of human joints. Then CNN is used to process the video data and optical flow data at the same time, and the motion evaluation entrance gives the results by calculating the angle at the joint point. In the experiment, the recognition rate is 88.89%. Among them, the recognition accuracy of other movements is 97% except for push and block. This study is helpful to provide positive guidance for handball training.