Research On Key Algorithm For Human Gait And Activities Recognition

In the future of human-computer interaction, that computer perceives the purpose and intention of human may greatly rely on the vision system. With small, passive and non-contact characteristics, the sensor of vision and vision information system are usable everywhere. Research on human gait and activities based on vision has got more and more attention recently, which contributes to the theoretical development of artificial intelligence and pattern recognition, and has important meaning. Gait and activities recognition technology plays an essential role in intelligent security surveillance, elderly care, and human identification, etc.But gait and activities recognition have been confronted with many difficulties so far for the complexity of environment and human motion. This paper aims to improve the performance of gait and activities recognition algorithms, to reduce the influence of the objective environment, and to prompt the gait and activities recognition technology to more practical and real-time development. The research results and major innovations of this paper are shown as follows:(1) To propose the moving human segmentation based on texture analysisThere are usually shadows in the foreground image extracted from video image sequence. In order to remove the shadows, texture analysis method is proposed which try to match the texture of foreground and background. If the textures of images are similar, then we can judge that they are shadows. When we match the textures, blending eigenvector based on the color and texture of pixel is adopted, which support vector machine to perform vector classification. The experiments show that the method can get better segmentation result.(2) To study the gait recognition method that based on subspaceAccording to the gait silhouette sequence, the feature image of gait is worked out, the feature image gait energy image (GEI) and active energy image (AEI). The study adopts the kernel principal component analysis method and kernel discriminant analysis method to extract the feature of GEI and AEI, and then to reduce the dimensions, which project the sample data to the kernel space and calculate there; and classify the eigenvector with support vector machine. The experiments show that the method can get better recognition rate in the stable conditions.(3) To propose gait recognition method based on sparse representationIn order to deal with the problems that high recognition rate algorithm always cost more time and can’t be used in real-time condition, hidden markov model is proposed that used part frame difference energy image’s sparse representation coefficient as it’s status feature, then represent the image in sparse way, finally to form image dictionary and achieve the rapid recognition of gait with fast algorithm for sparse decomposition. Moreover, the gait recognition methods based on reconstruction errors and discriminant dictionary are studied. The experiments show that the method gets better recognition rate and cost less time for recognition that can be used in real-time condition.(4) To proposed gait recognition method that based on dynamic Bayesian networkIn order to deal with the problems that gait recognition result is affected by dress changes, such as wearing coat or carring bag, the dynamic Bayesian network gait recognition model based on the fusion of dynamic information and static information is proposed. It tries to reason and study the model in probability, and analyzes model in robust way. Furthermore, according to the multi-dimension information of the gait, the two-dimensional dynamic Bayesian network&multiple informations fusion gait recognition method is proposed, and the recognition model is established on global and local feature of gait image. The experiments show that the method can greatly reduce affect of dress changes in gait recognition.(5) To proposed the segmentation and recognition method for continous activitesIn order to deal with the problems of being continuous, longtime and inseparable for multiple behaviors, human continuous activities segmentation and recognition method is proposed. The method constructs dynamic Bayesian network model to recognize the classified behavior sequence. And to form continuous activities classification network (CACN) based on time sequence, which segments twice to achieve the best action segmentation. The first segmentation adopts input video to make the coarse segmentation, and the second segmentation uses the viterbi algorithm to achieve the best segmentation of activities after the coarse segmentation. The experiments show that the method can get better activities segmentation rate and better activities recognition rate.