Human Visual Analysis Of Feature Extraction And Classification Algorithm

Visual analysis of human is currently one of the most active research topics in the domain of computer vision. This strong interest is driven by a wide spectrum of promising applications in many areas such as virtual reality, smart surveillance, perceptual user interface, content-based image storage and retrieval, athletic performance analysis, etc. An angle-based feature extraction algorithm (AFE) of human shape and a Similarity of Larger Scatter between Sorts (SLSS) for classification are proposed in this paper. The purpose is to recognize and analyze normal human postures by pattern recognition from single video image. As far as human feature extraction is concerned, traditional algorithm is of much computation complexity because of Eigenspace Transformation. Another feature extraction method based on extreme points proposed by Robert T. Collins results of feature vectors with unequal divisions. Considering theses shortcomings, we designed the AFE algorithm which can reduce the dimensionality of the original feature vectors based on fiducial joint angles. The AFE can extract features without calculating all edge pixels, which decreases the computing complexity of Normalizing and Eigenspace Transformation dramatically. Moreover, it can analyze human posture and motions without calculating geometry characters such as the curve of the contour. But it is sensitive of noise and Occlusion Handling, which is also the problem of most Visual Analysis System. In recognition part, we modified Euclidean Distance and proposed the Similarity of Larger Scatter between Sorts (SLSS). The main thought is to use weighted vectors on original feature vectors of corresponding classes and to make the scatter between sorts largest in the eigenspace. Differently, Linear Discriminance Function uses the best weighted vector, but SLSS uses n weighted vectors which not only leaves out the computation of sample training, but also ensures robust recognition of human posture. SLSS is simple and wieldy, but it requests samples with interlaced peak value which is fit for recognition of human posture.