Region Restraint HOG-LBP Feature Based Human Target Detection Algorithm

As the very basis upon which human tracking, identity recognition and behavior analysis are implemented in video surveillance, human target detection has been widely studied in the field of computer vision. In terms of robustness and efficiency, although many approaches with great progresses have been introduced by researchers, human detection remains a challenging task owning to targets’variable appearances and gestures, complicated and cluttered background, the occlusion between different targets, etc..Currently, binary sliding window classifier serves as the predominant method when it comes to human detection. To adopt this method, an image is scanned with very high resolution from the top right to the bottom left with multi-scale rectangular windows. And for each window, feature vectors representing targets should be extracted for classifier’s offline labeled data training in order to establish an effective classifier. In2005, Dalal and Triggs introduced Histogram of Oriented Gradient (HOG) descriptor for human detection based on the basic idea that "local object appearance and shape can be characterized rather well by the distribution of local intensity of gradient or edge directions". They also established an image database named INRIA data set which is very challenging and generally acknowledged in the field of pedestrian detection.Based on Dalal and many other researchers’work, in this thesis, we will introduce an improved feature extraction method in which, with added Local Binary Pattern (LBP) feature, HOG descriptor is extracted in the Region of Interest (ROI) of human body edge for the purpose of reducing the vector dimensions. For each scanning window, it is doubtless that people’s shape and edge information are not evenly distributed, which means the possibility of setting ROI for the extraction of HOG features. By using Sobel operator to extract and display the shape and edge information for grayscaling positive sample images in INRIA data set, we set four ROIs which are respectively head-torso, right arm, left arm and legs. Utilizing the concept of uniform pattern, LBP operator is a texture descriptor that can filter out noisy edges. Because HOG descriptor, focusing purely on capturing edge and shape information, might perform poorly when the background is cluttered with noisy edges, the combination of LBP and HOG descriptor makes up the deficiency and contributes significantly to calibrate human detection.Using the Part-template Matching Theory introduced by Zhe Lin for reference, in this thesis, we also make human shape part templates for further analysis of the distribution of HOG feature in each scanning window. The whole shape templates tree includes, three layers, which are head-torso layer, upper-legs layer and lower legs layer from the top to the bottom respectively. In each layer, we compute and compare the matching score for every part template candidate using an approach similar to the extraction of HOG feature. We adopt top-down sequence when doing the matching and the matching result of the upper layer will influence the matching choices in the lower layer directly.In order to evaluate our ROI-HOG and LBP human detector, we conduct experiments on INRIA person data set. Using the same linear SVM, our approach achieves better performance when comparing it with Dalal’s via Detection Error Trade off (DET) curve experiment. We also evaluate our part templates matching human edge analyzing approach by applying it on both INRIA person data set and the practical campus video surveillance. We get very distinct human detection results. Generally, the edge segment for testing images performs also very well.