Research On Human Action Recognition Method Based On 3D CNN

Recognizing human actions in the real-world environment finds applications in a variety of domains including intelligent video surveillance,customer attributes,and shopping behavior analysis.However,accurate recognition of actions is a highly challenging task due to cluttered backgrounds,occlusions,and viewpoint variations,etc.Therefore,it is very important to develop advanced actions recognition algorithms.The traditional behavior identification method mainly consists of two steps.The first step is to extract the features of the video image.The second step uses the learned classifier to classify the features.In real-world scenarios,different action classes may appear dramatically different in terms of their appearances and motion patterns.Therefore,it is difficult to select suitable features,and the deep learning model can learn characteristics through the sample,and thus has a better advantage than the traditional behavior identification method.This article is based on deep learning theory,proposes a human behavior recognition method based on 3D Convolution Neural Network.Mainly consists of the following steps:First step,Constructing 3D Convolution Neural Network models(C3D),training and extracting features of C3D Convolution Neural Network.Second step,a 3D Convolution Neural Network based on motion trajectory(TC3D)is constructed based on the C3D Convolution Neural Network,and the features of the TC3D Convolution Neural Network are trained and extracted.Third step,the optical flow image of the video is sent to the VGG16 network model to extract the time domain features.Fourth step,the features of the C3D Convolution Neural Network and the time domain features are merged and sent to the SVM for category classification.Finally,the experimental results are tested,analyzed,and prospected.This paper tests on the UCF101 dataset.When the pre-training model is not used,the accurate recognition rate of C3D network model is 43.171%,the accurate recognition rate of TC3D network model is 38.92%,4.52%reduction compared to C3D.and the accurate recognition rate of pre-training model C3D network model is 79.1551%.The accurate recognition rate of the VGG16 network model is 68.5961%,and the accurate recognition rate of the fusion of the C3D feature and the time domain serial feature is 88.5547%,9.3996%increase compared to C3D.and the accurate recognition rate of the weighted fusion is 87.1545%,7.9994%increase compared to C3D.The results show that TC3D can't learn behavior characteristics better,but the fusion of C3D features and time domain features can improve the accuracy of recognition and have better robustness.