Research On Human Action Recognition Based On Motion History Image

With the progress of society and the development of science and technology,people's demands for public security and quality of life continue to increase.The living environment is expected to be more intelligent,comfortable and convenient,which makes human-computer interaction a hot topic in academia.As the basis of human-computer interaction,human action recognition has important theoretical research value and practical application value.In this thesis,we analyze the current research status of human action recognition,and find that the action recognition based on vision is easily affected by external conditions such as background noise,light change,and clothing occlusion,etc.Moreover,some algorithms are computationally complex and have poor stability,which makes it difficult to adapt to real-time application scenarios.To overcome the above shortcomings,this thesis focuses research on object detection,feature extraction and action classification.The main contents of the thesis can be summarized as follows:1.Motion segmentation method based on the improved motion history image(MHI).It is to use uniformly-spaced sampling to extract key frames to improve the computational efficiency,and obtain the object contour by frame difference method,and update the gray values of the target contour to further synthesize MHI.The proposed method significantly reduces the complexity of the parameter estimation and can capture the key motion information in videos more effectively compared with the original methods.2.Human action recognition based on multi-feature extraction.This thesis use histogram of oriented gradient(HOG)algorithm,local binary pattern(LBP)algorithm and Gabor wavelet to extract the gradient features,grayscale statistical features,and multi-scale multi-directional texture features of the improved MHI,and use Adaboost algorithm for classification.The experiment shows that the improved MHI has good separability from its gray,gradient and texture.3.Dimension reduction algorithm based on energy features.The high-dimensional LBP feature and Gabor feature are re-encoded by the proposed method to obtain a new low-dimensional energy block-local binary pattern(EB-LBP)feature and energy block-gabor(EB-Gabor)feature.Compared with principal component analysis(PCA)algorithm,the proposed method can retain more effective information.4.Decision-level fusion algorithm based on belief assignment.According to the recognition rate to allocate different belief,the proposed algorithm aims to obtain the final decision results by decision-level fusion to improve the system fault tolerance.The experiment shows that the recognition performance is good enough.This thesis analyzes the proposed human action recognition method based on KTH video database.Through experimental analysis,we can see that the improved MHI can obtain good motion segmentation and improve recognition accuracy.The energy feature extaction algorithm can effectively reduce the original feature dimension.The decision-level fusion algorithm based on belief assignment can further improve the system performance.The above research can be further extended to practical applications such as virtual reality,intelligent video surveillance and public safety,etc.Therefore,it has tremendous economic value and social value.