Based On Infrared Moving Human Target Detection And Tracking

Human detection and tracking in infrared images are widely studied in computer vision. As a result of thermal based imaging, infrared imaging system has many virtues, e.g. it has visual capacity in darkness and frog, can almost work in any environment and all-weather conditions. So it is a kind of useful techniques for intelligent video surveillance, automatic vehicle driver assistance.However, infrared image has its inherent defects, such as low signal-to-noise ratio and unitary information. So the human detection and tracking algorithms in infrared images is a complex challenge. It is involved with infrared image segmentation, human body features extraction, description and classification, no-rigid objects tracking, etc. We have to confront many difficulties, such as poor image quality including low resolution and narrow dynamic range;There are little color and texture information in infrared image;human are non-rigid object and their postures and appearances are complicated as well as changeable. In this paper, based on analyzing on infrared imaging properties, an integrated solutions is presented after studying on current human detection and tracking techniques in infrared images and visible images.Take overall consideration of the detection and tracking techniques of human in infrared images, this paper not only discriminate human and non human body, but also divide human targets into two kinds to provide basis for human body tracking. In image segmentation,2-D Maximum Entropy Method as well as Mean-Shift algorithm are studied. Then, based on segmentation results, candidate areas are determined. Finally, Two-grade SVM classifier is adopted to detect candidate targets. Intensity-distance histogram character is firstly used to distinguish human body from other targets. Then invariable square features are used to divide all human body targets into two class and identify them with different flag.In infrared human body tracking, a kind of object tracking algorithm based on simple voting is proposed. The centroid character, area and eccentricity ratio are extracted and used together with class information obtained in the last chapter as feature set for target tracking. Then, voting algorithm is used to determine tracking target. Experimental results show that the algorithm provided in this paper is robust when there is adhesion phenomenon.