Research on Human Action Recognition Based on Convolutional Neural Network

Abstract

This paper proposes a human action recognition (HAR) algorithm based on convolutional neural network, which is used for human semaphore motion recognition. First, collecting datas in three scenarios and Deep Convolution Generative Adversarial Networks(DCGAN) is used to implement data enhancement to generate the dataset (DataSR). Then, the 1*1 and 3*3 convolution kernels are used to design the full convolution network and the model is further compressed using the group convolution to obtain the new model HARNET. Experiments show that the mAP of HARNET on the DataSR dataset is 94.36%, and the model size is 76M, which is 30% of the size of the YOLOv3 model.