Research On Human Action And Gesture Recognition Based On Vision

Human motion analysis based on vision has been one of the active topics in computer vision research and has broad application scenarios,such as human-computer interface,automatic video indexing and surveillance.In recent years,human motion analysis based on vision has been extensively researched.The procedure of human motion analysis consists of detecting and analyzing the human posture from the images or videos,understanding and describing its motion.However,the study of human motion analysis also remains a challenging problem due to various factors such as the complexity of gesture,occlusion and clutter background,perspective and scale changes,etc.For the key problems of human motion analysis,action and gesture recognition are studied in this thesis.The previous research of human motion analysis is introduced.And some basic methods of action and gesture recognition are summarized.Specifically,the related theory of the research in action and gesture recognition is introduced.We propose an algorithm of human action recognition based on the feature level fusion and random projection for the action recognition,in which the spatial-temporal interest points are detected from the video,and then integrate the spatial-temporal gradient descriptor with the Gabor descriptor,which can improve the discriminant ability of the final feature.Moreover,random projection is employed to reduce the dimensionality of the fused features and the computation complexity.Then,the bag of word model is used to describe the action,and form the feature words.Besides,in order to handle the complex iteration issue in the parameter estimation,the Bayesian parameter estimation is applied to the parameter estimation of LDA topic model.In the stage of training,it can provide more priori information and improve the recognition performance.We propose an algorithm of human gesture recognition based on random projection and multi kernel learning for the gesture recognition,in which some preprocessing steps are employed in the training images,including Grayworld light compensation algorithm,gesture segmentation and morphological operations.Then,the SIFT features are extracted and the dictionary is learned in two stages.In the first stage,K-means algorithm is used to train the dictionary.In the second stage,the incremental codebook optimization is used to update the dictionary learned previously.Afterwards,we use the spatial pyramid machine in the gesture image after segmentation and locality-constrained linear coding to encode the feature.Then,the random projection is employed to reduce the high dimension of space pyramid feature.Finally,we use the multikernel learning to classify the gesture by the optimal combination of the kernel matrix coefficient.