The Geometric Algebra Methods For Body Recognition In Color Video Image

The human body recognition of massive video images is one of the basis of video surveillance analysis which face to the public safety. As the Geographical scene of video images can be very complex with a huge data capacity, the human body detecting is always disturbed by the change of background, the various posture of humans, the image distortion and so on. With the result that the traditional video image processing methods has some weakness in processing human body recognition of color video image, such as lower-efficiency algorithm, unstability and low applicability. Due to color image providing more information than monochrome image, it can be a radical boost for the accuracy and efficiency of human body recognition of color video image to construct a image recognition method based on the mathematical basis which integrally support multidimensional operation.Geometric algebra is a new kind of mathematics tool based on dimension computation and integrally-support multidimensional computation. With the autonomously defined dimension space of geometric algebra the multi-dimension object can be expressed and computed unifiedly. So geometric algebra provides the theoretical and practical basis for representing high-dimensional video image features. The paper introduced geometric algebra to the color video image processing, and the customizable color dimension space was constructed to build the new methods for the edge detection, human body recognition and matching. The paper try to provide a new theory and method for color video image analysis, the following main parts were included in this thesis:(1) Based on geometric algebra, a color-space representation method is created. With the support of customizable dimension space, paper constructed the RGB and HSV color space, and the color information was coded with geometric algebra in order to embed the multi-dimension vectors of color video image into the representation. After uniformly representing color video image, the paper designed some matching operators and algorithms. Then the commonly used color features were extracted and the mixed color feature operators were constructed and calculated.(2) The imaging model of the color sensor was modeled to build a constant color invariant system. The paper also constructed a set of color feature operators, which were combined with the texture and shape features to form the more stable feature identification methods for color image. In the case study and algorithm validation, we extracted edge structure by the gradient of saturation constructed with geometric algebra. Compared with the traditional algorithms, our method extracted more continuous edge structure and higher degree of differentiation between foreground and background.(3) A template matching method is designed in color space by Clifford algebra basic operators. We optimized template matching and correlation by Clifford FFT. Then, the scale adaptive and angle adaptive template matching method was designed, with the correlation coefficient calculation between feature template and video image, the paper matches and evaluates location of human body in video scene were worked out. The case study shows that our method has the features that templates stability, human body postures self-adaption and scale self-adaption.Geometric algebra is introduced into color video image processing. The color vector combined the each component of colors, and we expressed color vector and its transformation in a more intuitive and more compact method. The paper designed an image feature extracting and identifying method, which could be a new idea and solution for video image analysis and human action analysis in complex geographical scene.