Research Of Pedestrian Detection Based On Combined Feature Of Part

With the development of national economy and people’s living standards, China’s car ownership continues to grow, but at the same time causing serious driving safety issues. Vehicle and pedestrian’s collision is the main type of road traffic accidents and pedestrian is always in weak position. Carrying out detection of pedestrian in critical condition and developing a timely warning system can effectively reduce the number of pedestrian casualties in road traffic, and will make great contributions on promoting the vehicle initiative safety.How to achieve a rapid and accurate pedestrian detection is the key question in vision-based pedestrian detection methods. As methods based on statistical classification are capable to overcome the impact of unfavorable conditions such as various pedestrians, scenes and lighting conditions, they are more prevalent. But extracting and taking which kind of features and classifiers on human bodies or their typical parts needs a further research. In the classifier design, mechanism with combination of multiple classifications can take advantage of the complementary information between classifiers, and gives full play to the advantages of each classifier to improve the accuracy of supervised classification. Our pedestrian detection research is carried out on the basis of in-depth discussion about these issues, and a pedestrian detection method based on combined feature of part is proposed in this paper. The specific contents are as follows.(1) As the detection speed of traditional pedestrian detection method based on histogram of oriented gradient (HOG) features is quite time consuming, the paper presents a pedestrian detection approach based on optimized HOG features of the legs. The week classifier designed by weighted liner fisher discriminant is employed to select high discriminative feature, which can significantly decrease the training time and space while maintaining the comparable classification rate. Moreover, the look up table Gentle Adaboost algorithm can optimize the weighted combined HOG features and form a strong classifier to identify the pedestrian.(2) The contours of human’s head have a good invariance. In this case, the template matching method can describe the target better, so it is adopted to detect the head on the basis of leg detection. Then a comprehensive method using parts constraint, feature transformation and classifier threshold adjustment is employed to integrate the test results of legs and head, which can significantly enhance the ability to identify the pedestrian. (3) For pedestrian tracking, Camshift algorithm is a good choice as its better real-time. But in tracking process, the Camshift algorithm is susceptible to color changes. Through embedding the Kalman filter prediction mechanism, color and motion information are combined to promote the robustness of the pedestrian tracking.(4) Finally, according to the foregoing findings, the pedestrian detection and tracking algorithm are organically combined to form a complete system through the matching technique. Experimental results indicate that the algorithm introduced could achieve effective recognition of pedestrians on the front of the vehicle within a certain distance. At the same time, the algorithm has good detection effect both to the still and moving pedestrians. The pedestrian detection system can accurately depict the trajectory of pedestrian movement, and provide the basis for the analysis of its behavior.