Research And Implementation Of Pedestrian Detection Based On Monocular Vision In The Scene Of Car Backing

With the continuous development of social economy, car as the basic transport vehicle ownership, its number is increasing, leading to urban traffic safety problems. Pedestrians as the major players of traffic behavior, often become the direct victims of traffic accidents, so how to protect the safety of pedestrians becomes a hot research field. At the same time, as the global science and technology continues to develop, applications based on computer vision techniques are becoming more and more widespread. This article used monocular vision sensor to get the vehicle surroundings information and completed the design of pedestrian detection and warning system.This thesis mainly completed the following four parts of research and implementation. First, the thesis used technologies of image preprocessing, edge enhancement, contour extraction, to segment pedestrian candidate regions based on effective marginal. Candidate region segmentation can be effective in reducing pedestrian recognition detecting range, thereby increasing pedestrian detection in real time. Secondly, Haar features were used to describe pedestrians, by learning and training the collected positive and negative pedestrian samples to complete the Adaboost classifier design which was based on cascade algorithm, the cascaded classifier can effectively detect the range of the candidate region identify pedestrian goals.Thirdly, under the HSV color environment, according to extraction of the color column diagram of pedestrians detected, this thesis implemented the tracking of pedestrian objects based on CamShift algorithm. While improvement on the convergence conditions of CamShift algorithm was proposed, and this reduced the ambient color interference effect while tracking objects. Lastly, the parameters of the camera were obtained by calibration of the experiment camera using Zhang Zhengyou calibration method. On the given constraints, the thesis established a geometric model of the vehicle rear. Then the geometry model and ultrasonic backing radar model were combined to build an early-warning model on the rear of vehicles. The model has a wide range of early warning, and can make different behaviores according to the area where the pedestrian object is in.Finally, through experiments in different environments, the results show that the pedestrian detection algorithm has better robustness, and recognition algorithm has higher detection rate, lower false detection rate. At the meanwhile, tracking algorithms and warning model is accuracy.