Study On Pedestrian Detection Ahead Of Vehicle Based On Monocular Vision

It goes without saying that the road traffic accidents could cause great harm to our society, so how to reduce the traffic accidents and the corresponding loss that they caused has become the topic to which people pay main attention. The research of vehicle safety driving assistant has been more and more concerned because it plays an important role in reducing the accidents and the corresponding loss. As pedestrian is the main participator of the road, the traffic accidents, especially those caused by vehicle and pedestrian's collisions, are the main reasons that cause people hurt or die. In this case, the accurate preceding pedestrian detection has become the main function of vehicle safety driving assistant system. It could effectively warn the driver of the incoming collide with pedestrian and help to prevent the collision. Now pedestrian detection has aroused much attention or interests of the government departments and some research institutions.In recent years, great efforts have been made by our country in the research of the protection to the drivers and passengers. However, little developments on the protection to the pedestrian have been got. In fact, the study on the safety of pedestrian is a new research field. In order to effectively guarantee the pedestrian safety and remind the driver of the possible collision with the pedestrian in time, the research of the pedestrian detection technology will provide powerful theory and technological support to the development of safety driving assistant technology in our country. At the same time, the pedestrian detection technology has potential economic value and application prospect. At present, the commonly used sensor to realize the pedestrian detection is machine vision for the reason that the digital image contains more abundant information than other types of sensors. Furthermore, if the machine vision is integrated with the knowledge of image processing, pattern recognition and other relevant methods, the objects can be detected successfully. In this paper, the preceding pedestrian detection technology based on monocular vision is researched systemically. The specific content is as following:1. Research on the segmentation of preceding pedestrianThe accurate and fast segmentation of pedestrian candidate regions is the major premise to realize the detection of pedestrian. Based on the different position of the pedestrian, two segmentation methods are put forward to extract the area of interest where pedestrian may exist in. The next pedestrian recognition method can be performed in the area of interest segmented from the image which helps to reduce the algorithm time.Because the collision between vehicle and pedestrian mostly happens in road region ahead of vehicle, so this paper uses the shape information of pedestrian to segment candidate region in the road area ahead of vehicle. Firstly the road area is extracted by using the existing lane recognition algorithm. Then the vertical edge in the area is enhanced and the noise is removed. Considering that the pedestrian's legs have high vertical edge symmetry, the vertical edge was extracted to determine the symmetry axis. Finally the width and height of the pedestrian was decided combining with pedestrian's morphological character.In order to segment the pedestrian who is possible to enter the road area ahead of vehicle, this paper puts forwards a fusion segmentation algorithm based on AdaBoost and pedestrian shape feature. This method makes full use of the characteristics that vision sensor could capture abundant information and that AdaBoost algorithm is simple and reliable, and has high accuracy. The AdaBoost algorithm is adopted to train the ideal pedestrian classifier with high recognition rate using Haar-like features. Then the trained cascade classifier is introduced to segment the initial pedestrian candidate region in the image. Finally, the candidate pedestrian in this region is segmented taking full advantage of the vertical symmetry of pedestrian legs, size and aspect ratio of pedestrian.2. Research on the recognition method of pedestrianAfter the candidate pedestrian areas segmented from the image by the algorithms described from above, a certain technology is needed to confirm that whether those areas is pedestrian or not. Through analyzing to the pedestrian sample images, we can know that the gray image of pedestrian has some texture and gray symmetry features. In addition, the continuous edge of pedestrian make the contour feature of the gratitude image obvious, so the extracted edge has certain boundary moment and gradient direction character. On the basis of these characters, every sample image is expressed by a 23-dimension characteristic vector, and the pedestrian classifier is formed adopting the supporting vector machine algorithm. The testing results show that the pedestrian detection rate of the classifier could achieve 83.46%.3. Research on the tracking method of pedestrianIn order to guarantee the good real-time performance of our system, this paper estimates the position of the pedestrian bounding box with Kalman filter based on the character that the pedestrian walks in low speed. A dynamic and variable area of interest is formed to narrow the seek range of pedestrian, which greatly reduces the image processing time to satisfy the real-time demand.4. The information acquisition to prevent from collisionIn this paper the distance between the vehicle and preceding pedestrian and his direction of motion is acquired. The distance measuring principle is analyzed based on the monocular vision to estimate the distance between current vehicle and preceding pedestrian. The pedestrian's motion direction is judged based on the changing of angle between the pedestrian and the vehicle's longitudinal middle line. At last, preliminary discussion about the collision avoidance rules is made according to the distance and motion direction.The experiment results indicate that the algorithms introduced could achieve effective recognition of the proceeding pedestrians with different sizes, colors and shapes. At the same time, the algorithms have good detection effect both to the still and moving pedestrians in good time.In summary, the paper has realized the detection and recognition of preceding pedestrian based on monocular camera. The pedestrian detection and track is performed with the technology of computer vision, image processing, and pattern recognition etc. The problems such as the acquisition of preceding image in the urban transportation, the real-time pattern recognition and track of the pedestrian, the pedestrian's motion analysis are settled, which would provide key technology support for developing advanced safety driving assistant system, so it could offer high and new technology support to reduce the traffic accidents caused by the collision between vehicle and pedestrian. The research results will make active contributions on promoting our country to catch up with the world's advanced level in this field.