Athlete Training Sensor Data Detection Research Based on Optimized Convolutional Neural Network

Abstract

The traditional convolutional neural network has matured in image detection and recognition, but maintaining its high accuracy requires a large memory for computation, which increases the volume of sensors and puts additional strain on athletes. Therefore, the research uses local error calculation to replace global error, and uses layer-by-layer residual unit structure to replace gradient calculation, to apply it to athlete training sensor data detection. Through the simulation experiment and the example experiment of the data set, the optimal residual unit structure and the local error joint weight of the residual network human motion detection are obtained. At the same time, in the simulation experiment and example experiment analysis of the school physical education curriculum, the accuracy of the algorithm model proposed in the study is higher than the first five CNN network structures, and its training time is the shortest. The experiment demonstrated that the local error method combined with the residual network structure had improved precise performance in motion detection, and the computational load was low, making it appropriate for wearable sensors during training.