Human Detection Based On The Omega Shape Feature

Recent years, in order to improve the quality of contemporary social life, the human detection technology has been widely used in various technical fields. Researchers raised many many excellent feature extraction algorithms and classification algorithms to the detection based on human classification framework. These algorithms have achieved some results, but there are still many critical issues remain to be solved, such as the characteristics of non-rigid body, the changing model of human body, the background environment of diversity and light detection result interference problems. These problems interfere with the accurate modelling algorithm to human targets, reducing the accuracy of the human detection.To solve the problem above, this paper proposed detection method using body shape features of Omega based on the discriminant human target detection framework and the analysis of human academia more popular detection method. Specific methods include three parts: 1.This method detect the human based on the human head- shoulder Omega shape feature.Compared with the human detection methods based on the overall characteristics of the currently widely used,our method can reduce miss rate and the false rate caused by the body non-rigid plastic deformation and serious false positives and missed part of the block in the practical application. 2. The combination of gradient histogram(Histograms of Oriented Gradients, HOG) feature extraction algorithm and orthogonal non-negative matrix factorization(Orthogonal Non-negative Matrix Factorization, ONMF) algorithm. In this paper, we select the HOG features to characterize the shape of Omega and describe the contour, while effectively suppress the interference of light and tiny changes caused by deformation detection, thus satisfy the need to detect Omega shape feature extraction for analysis. Meanwhile, we deal with several key issues for human detection by using the characteristics of the HOG-ONMF to reconstruct and present the most essential features of the model of the sample library. thus reduce the partial occlusion and complex background on the detection from the algorithm 3. Use Large margin nearest neighbour methods to redefine the distance of the feature of samples. The method is effective to establish of multi-modal discriminative appearance model, more over enhance the identification of the characteristic features of the model from the perspective of the analysis.In this paper, the method described above is made a contrast with the traditional method. Experimental results show that the proposed method can successfully overcome aerial camera perspective problems caused by body parts mutilated limbs, and capable of complex background and human objects under partial occlusion detection. As can be seen from the whole sequence of results in public databases, compared with other methods, this paper presents a method to improve the accuracy of detection of human targets, reducing false positives and miss rate.And this method can get more robust detection results in real human detection system.