Study On Human Detection In A Static Image

The algorithm for human detection in a static image can be applied in many areas, such as driver assistance system, motion capture system, adult image filtering system and virtual video. However, as a human object may have various shapes and can be dressed in many different colors and styles, it is a challenging task to detect a human in a static image. In this paper, we outline some possible remedies in solving this problem, and the following contributions distinguish this dissertation from previous works: Although the shapes of humans can vary dramatically, the shape of limb is relatively simple. Based on the observation, we present a human model that emphasizes the shape of limbs, and we introduce a probability model for the human body to evaluate the variations of human shapes in chapter 2. Edge is an important image feature for object detection system, and we present a color edge detector based on the intensity variations and colors' spatial distribution in chapter 3. Firstly, the algorithm detects gray edges and then further makes effort to find the color-texture edges. Finally, the detector removes the false edge pixels inside texture regions. The set of edge pixels detected by this algorithm is a useful cue for object's boundary, so it can be taken as an important feature for the human detection system. An algorithm of human detection is presented based on rectangle fitting in chapter 4. According to the principles of perceptual organization, the human detector takes edges as input, and then makes effort to obtain higher feature details step by step. The higher detail features include line segments, rectangles, limbs, OSBC (one-side body contour). Finally, we try to find the reasonablecombinations among all the OSBCs to detect possible human objects. The algorithm can deal with the variations of human shapes effectively. An algorithm is presented to detect the combination of face and arms in chapter 5. The skin pixels are detected in the YCbCr space, and then the skin regions are determined according to the connectivity analysis. After that, face candidates are detected by the aspect ratio of the smallest bounding rectangles corresponding to the skin regions. The concept of connectivity distance is defined based on the edges of the original image. For each face candidate, its corresponding arms are detected in terms of the connectivity distance between the skin regions. If a combination of face and arms satisfies some geometric and topological constrains, it can be taken as a human object. As this algorithm makes use of skin colors, the detection results are more reliable. In chapter 6, we discuss another work that is an algorithm for rendering the same scene from multiple viewpoints. In the pre-process, we define a splat for each polygon, and each splat is composed of a contour trace and a set of strips. After that, we can utilize the re-usable information contained in the splat to render each polygon. Compared with the traditional polygon scan-line algorithm, our new method can accelerate the rendering process effectively.In this dissertation, we mainly study the algorithms for the human detection in a static image. A detection system is implemented in C++ under windows 2000 operating system, and some experiment results are presented. The algorithms may be applied in the driver assistance system, motion capture system, image filtering system and virtual video system in the future.