The Study Of Pedestrian Detection In Traffic Environment

The safety of pedestrian is always a focus problem in traffic safety. As an important branch of object detection, pedestrian detection has a broad practical prospect in many fields. In traffic environment, drivers need to know the information of pedestrians surrounding the vehicles to give way to pedestrians and avoid causing casualties. In densely populated venues, pedestrian detection could provide the number and distribution of pedestrians at a specific time and place, which helps to guide pedestrians to the right emergency exit and minimize casualties.However, there are several problems in the research of pedestrian detection on computer vision which restrict the development of pedestrian detection. A single feature of the pedestrians only demonstrates partial information of them, and cannot represent the global information sufficiently. In traffic environment, pedestrians in surveillance videos have variable sizes, postures and clothing. Some pedestrians may be occluded partially by other things which results in the loss of the global structures of pedestrians. These problems have brought great challenges to the research of pedestrian detection.To solve the above problems, an integrated feature and the combination of local detections were utilized to propose a new pedestrian detection algorithm in this paper. Some improvements and innovations are listed as follows:(1)The Gabor feature and HOG feature were integrated to build pedestrians’ feature database. HOG feature can extract the outline of pedestrians sufficiently. Wavelet transform can represent both the high frequency and the low frequency information of the pedestrian images. 2D Gabor filter can represent the information of pedestrians in different scales and directions. The integration of the above features can implement a more comprehensive and sufficient representation of pedestrians.(2)For the diversity of the pedestrians’ features, the pyramid transform was applied to the pedestrian images before feature extraction. The features extracted from pedestrian images in different scales can increase multi-scale information of pedestrians.(3)The pedestrians’ features include a great deal of redundant information which leads to excessive dimensions of the feature vectors. PCA was applied to pedestrians’ features for dimension reduction, the main information was kept to adapt to Adaboost algorithm better and the speed of pedestrian detection was improved.(4)Both head and shoulder detection and leg detection were utilized. The traditional global detection was replaced by the combination of local detections to deal with the partial occlusion of pedestrians and increase the accuracy of pedestrian detection.The true positive rate and the false positive rate of the method in this paper were analyzed on the INRIA, MIT and Daimler data sets, and compared with other pedestrian detection methods. The experiment result demonstrated that the method proposed in this paper could represent the global information of pedestrians more sufficiently, increase the accuracy of pedestrian detection and satisfy the requirements of real-time detection.