The Research Of Moving People Extracting And Action Identifying

In the era of big data,with the growing demand for high-quality video information,people increasingly pay more attention to the technology intelligent video analysis.As one of the key technologies in the intelligent video analysis,Human action analysis is an important hot topic in the field of computer vision,with great research value and significance,which has widely applied in the areas like smart surveillance,elderly guardianship,virtual reality,athletic performance analysis etc.Traditional human action recognition methods are mostly based on the image or RGB video,but there are still many difficulties in practice,due to the complexity of background disturbance and the interference of noise.With the availability of low-cost Kinect,action recognition based on RGB-D data launched new trends in the computer vision and the image processing.This paper focused on the method of human action recognition based on depth image and RGB image from by Kinect sensor.An automatic matting method was proposed to extract the moving people for the RGB video so as to reduce the influence of complex background in behavior recognition.Moving human detection is a fundamental research for high-level computer vision technology such as human action recognition.In this paper,we take the advantages of saliency detection as prior information replace manual annotated tri-map.After getting the saliency map,a cost function was constructed to obtain a foreground.The cost function comprehensively concerned that prior knowledge,color and distance how to the alpha value of pixels.Based on HOG absence of time information in Histogram of Oriented Gradients(HOG),we proposed a new feature method called histograms of temporal gradient(HTG)which can get both the space and time information.And we also proposed a novel HTGHOG feature representation method.First,we extract the HTG features from depth image or the extracted moving people images;Second,in order to improve our feature's discrimination ability,we applied the histogram of oriented gradients(HOG)on the HTG features to re-extract temporal information for entire video sequence;Last,for better distinguishing the similar actions with different time sequence temporal order,we introduced a simple temporal hierarchical construction.In the meanwhile,Kernel extreme learning machine(KELM)is utilized to provide the probability outputs of classification.Experimental results show that the moving object extraction method proposed in this paper can efficiently extract moving people.At the same time,experiment verified that HTGHOG has a strong discriminative ability in depth video sequences and RGB video sequences.The proposed method performs better than the state-of-the-art methods.