Research And Application Of Pedestrian Detection Method Based On Human Head And Shoulders

Human Detection is the basic research in a series of security field such as the intelligent video surveillance. The current human detection methods, on the one hand, are usually hard to get high detection rate as the result of occlusion, and on the other hand, are difficult to be real-time. Compared with the whole or other part of human body, head-shoulder is not susceptible to occlusion, so the head-shoulder is chosen as the detection object in this paper, and the human detection method based on the head-shoulder counter characteristics and the human detection method based on the head-shoulder HOG feature have been studied, improved and verified by experiments respectively. After that, the human detection method based on the secondary structure is proposed in the paper.First of all, the detection area is extracted from the video by the improved frame difference method, and then the target is detected by the method based on the head-shoulder counter feature as the first level detection. In this part, the integrity of the head-shoulder contour is improved by edge detection combined with the counter filter and regional clustering. Then the dimension of the counter feature vector is adjusted by experiments. Then training and optimizing the BP neural network to classify the feature vectors. Interference like occlusion can decrease the performance of the first level detection, so the object which has been detected as the non-person by the first level detection method should be continued to deal with by the method based on the HOG features as the secondary detection. In this part, the traditional HOG feature has been improved and a new feature called SDHOG has been proposed which considered the difference of opposite gradient direction. The parameters of SVM have been chosen based on the cross-validation, and then the feature vector is detected by SVM. In order to avoid the time-consuming when the occlusion occurring obviously and is still detected by the first level, the target duty ratio should be calculate, and as a rough basis to decide whether to detect the object on the secondary level directly.Experiments show that the method based on the secondary structure combined the advantages in detection speed and detection actuary rate of the method based on the head-shoulder counter characters and method based on the head-shoulder HOG features. It can guarantee the real-time speed basically; meanwhile keep high detection accuracy up to 93%.