The Research Of Human Detection Based On Color Image And Depth Information From Kinect

Human identification and detection technology is becoming more and more important in the daily life of people with the continuous advances in computer science and technique. In the early stage of human body detection, color image is the medium of the detection method,but it is very easy to be affected by human body posture change, clothing difference, light intensity and complex background environment. This study is to detect the human target based on the hardware of Microsoft Kinect camera by using the synchronous generated color and depth data, in order to obtain better robustness in complex environment. The specific work includes the following parts:1. Image preprocessing is carried out on the depth image, and collected the sample database of the human body detection system. Firstly, using the HSL color space for color rendering of Kinect depth data, the traditional pseudo gray depth image were replaced by a more intuitive color image. Then based on the color and depth camera of Kinect, A set of color and depth training samples is set up, and the experimental color and depth test samples were collected in general environment, complex environment and weak light environment.2. Based on the image features of Local Binary Pattern model(LBP) and orientation gradient histogram(HOG), the hierarchical feature fusion algorithm PHOG-LBP is proposed by using the layered Pyramid space theory. Firstly, extract the overall contour feature and the region feature of the image by defining the different sizes of the test blocks. Secondly, the texture features of the image are extracted by using the uniform LBP feature. Finally, the feature vectors of the multi-layer contour feature and the texture feature are normalized by L2-Hys method, and the final feature vector is connected form these feature vectors. In the experiment, all the feature descriptors trained by the same samples will be evaluated in a set of independent sample, and the Receiver Operating Characteristic(ROC) curves and the AUC(Area Under ROC Curve) values were plotted and calculated. After the comprehensive comparative experiments show that this method can get more recognition in the human body for excellent results.3. Based on decision template(DT) algorithm, achieved the effective integration by Kinect color and depth image detection results. Firstly, The HOG features of depth and color are extracted based on training sample set, and complete the training of color and depth classifier to get the classification models. Then, put the positive and negative training samples to the corresponding classifier to get the DT template, and the test samples are put into the classifier to get the classification support to build the DP profile template. Finally, taking the Euclidean distance as a measure, the DP template matrix is compared with the DT template matrix to make the final class label. In order to verificate the robustness of the classification,color image detection and depth image detection and fusion detection were compared and validated in general environment, complex environment and weak light environment. The performance of human detection system based on DT fusion algorithm has the most outstanding performance.