Research On Human Body Action Recognition By Convolutional Neural Networks

In recent years,the launch of high-definition video equipment has promoted the rapid development of Artificial Intelligence based on action recognition technology in the area of Smart City,intelligent home,military security and so on.Due to extensive application prospects and economic values,action analysis and recognition technology comes to be hotspot in the computer vision field.Generally,traditional algorithm of action recognition consists of three procedures:motion foreground detection,feature extraction and recognition after training.Although above method can achieve acceptable accuracy,its robustness is not good and it has a huge workload.In addition,it may be difficult to recognize actions in actual scene,that is caused by many situations,such as overlaps between objects,complicated background and various shooting angles.This paper aims at improving above issues with Convolutional neural networks(CNN),that is,to increase the accuracy of recognition as well as the robustness of algorithm.Since the methods of background subtraction and frame difference can not get complete foreground under not too large range of motion,the paper proposed a kind of body silhouette extraction method based on the Difference of Gaussian(DoG)image.Firstly,this method establishes the difference image containing contour information of the human body by subtracting two images in nearby Gaussian scale spaces.And it takes binary strengthening and morphology processing operations on the difference image to produce rough image of body silhouette.The region of rough body silhouette is scanned with thresholds row by row,and close operation is also applied to obtain the exact image of body silhouette.To integrate temporal information of image sequences,this article accumulates periodic images of body silhouette,generating two-dimensional feature maps to be trained and recognized by CNN.At last,the average accuracy reached 85.3%on the KTH common database after adjusting parameters of the network and 5-fold cross validation,indicating that the recognition framework possesses certain feasibility.In order to deal with video data preferably,researchers have extended CNN to three-dimensional space.Based on it,this paper utilized 3D CNN to do some experiments,and the results showed that combination of "Optical flow,Frame difference,Three frame difference" could obtain the optimal recognition rate.And it reached 92.0%on the KTH common database after adjusting parameters of the network and 5-fold cross validation.After analyzing the samples' number and their accuracy of every class in the KTH database,the paper put forward three improved approaches,including training twice,oversampling and data expanding,to demonstrate that unbalanced data could make influence on the results.Ultimately,these ways increased the accuracy,and reached the average accuracy of 93.5%,92.8%and 94.7%respectively,which can provide solutions for classification of small database or unbalanced data.Furthermore,the algorithm of action recognition with 3D CNN solves certain problems in traditional methods,namely reducing the workload in feature extraction and improving the robustness of the algorithm.