Human Body Recognition Based On Color Sequent Images Under Complex Indoor Environment

Visual analysis of human motion has become the great significance forefront in the field of computer vision in recent years. It can detect,identify and track people , and can understand and descript their behavior from the sequence image which belongs to the scope of image analysis and understanding. Human motion analysis has great deal of contents to research from the view of technology. It has been related to many subjects, such as Pattern Recognition, image processing, Computer Vision and Artificial Intelligence etc. At the same time, the rapid segment of movement target under dynamic scene, the non-rigid move of human and the handling of target occlusion have become the challenge to human motion analysis. It can be used widely, such as man-machine interaction intelligent, security monitoring, video conferencing, medical diagnosis and content-based image storage and retrieval.Usually, a full process of movement analysis covers the following stages: object detection, object recognition, object tracking, understanding and analysis of the movement. All of the stages must not exist at the same time depending on the occasion of application. Generally, object detection and recognition is necessary commonly. The movement target recognition was focused in this paper.Firstly, the study background, the status quo at home and abroad, research tasks and goals were introduced. Next, the algorithm of moving target detection, moving target extraction and model building of movement target was described during the movement target detection stage. Object detection which task separates the moving object from the sequent images is the foundation of subsequent treatments. First, on the basis of analyzing three commonly algorithm of movement target extraction (background difference, inter-frame difference and optical flow method) at this stage, background difference method is been used with taking into account the application occasion. Second, the test results will been pretreated: getting rid of the wrong targets, and handling the movement targets (including removing noise and miscellaneous points, filling empty, and so on). The pretreatment is beneficial to reduce the amount of computation and to improve the efficiency of follow-up treatment. Finally the 2D model for recognition is built—head-shoulder model. In this paper, the author proposes the improved algorithm based on human anatomy against the shortcoming of paper [38] during the process of building the model. The algorithm can build a better head-shoulder model, and it is a good foundation for feature extraction and recognition.The feature of human model was described before recognize. In the first place, introduced some kinds of features which have the rotation, translation and scaling invariance. Secondly, seven moment invariants of the target's parts of outline are extracted as eigenvector for this application environment. In order to help identification, moment invariant was been compressed by computed its absolute value and then got its evolution because the range of moment numerical value is large and may be negative.During the phase of moving target recognition, target recognition is classification. The moving target has two kinds: human and inhuman. On the basis of introducing the principle, performance and learning rules of Back-propagation neural network, this chapter explores improvement from many ways against the shortcomings of BP neural network classifier, and depicts several BP algorithms based on numerical optimization. The best hidden node number is been found by training and testing of neural network with the different hidden node number. Then the pros and cons are analyzed by compare experiment which covers three improved BP algorithms, adaptive learning rate and momentum BP algorithm.Finally, the experimental verification has been done. Preliminary experiment show that the effective of the algorithms in this paper.