Design And Implementation Of Pedestrian Warning System Based On On-board Videos

With the development of economy, car ownership in our country is increasing. Meanwhile, the situation of traffic safety is getting worse. Thus, more and more attention has been paid to driving safety, and the most important object of the problem is pedestrian. How to effectively protect pedestrian in the increasingly frequent traffic environment has been always one of the hot issues in society.Pedestrian detection warning system based on computer vision is of great significance in driver assistance, which has been one of the hottest research subjects in the field of intelligent video processing and intelligent vehicle. The on-board camera is used to shoot the scenes in front of the car to recognize the pedestrian and estimate the degree of danger and give the response to protect pedestrian in this system. Pedestrian detection is not only the core of the system but also the difficulty in the field. Due to the variability of pedestrian states and complexity of traffic environment, there is not any common, robust and high real-time algorithm currently. There is a variety of methods based on using the different features of pedestrian and background to treat the detection problem as recognition issue.Pedestrian detection technology based on monocular camera is deeply studied in this dissertation. On this basis, a real-time pedestrian warning system based on aggregate channel features is designed and realized. The main work is as follows:(1) The evaluation criterions and methods of pedestrian detection algorithms are studied in detail. Two criterions and three methods are introduced in this paper. A variety of pedestrian databases are systematically studied and two databases, including INRIA and Caltech, are detailed introduced. Some concepts about classifier are learned and a brief introduction is given. The principles and experimental results of three classical pedestrian detection algorithms, including BING algorithm, HOG algorithm and DPM algorithm, are studied and described.(2) Common pedestrian features are learned systematically. The fusion feature, consists of three color channels, one gradient magnitude channel and six quantized gradient direction channels, is decided to be extracted as the input of classifier. Estimating method is used to compute the feature instead of traditional method. The principle of adaboost algorithm and the advantage of cascade classifier are introduced in detail. Experiment results show that both moving and still pedestrian in front of the car can be effectively detected.(3) A tracking algorithm based on context information is studied. It runs fast, but performs poorly when the features of target and background are similar, and it is easy to track failure when the occlusion area is large. In this thesis, we propose a novel Spatio-Temporal Context algorithm by the fusion of infrared information and visual information. Experimental results on public databases and our own surveillance videos show that proposed algorithm outperforms the original STC algorithm. The content above lays the foundation of the system’s future work.(4) A pedestrian detection warning system based on on-board videos is designed and implemented. The system consists of pedestrian detection module, regional division module, monocular vision distance measurement module and warning module. Three methods are used and manual division is selected as the method of regional division module. The pedestrian detection module uses the method of combing aggregate channel features and cascade adaboost classifier to construct channel pyramids to detect the pedestrian quickly in video and obtains the key information and motion properties. The monocular vision distance measurement module estimates the distance of pedestrian and vehicle. The warning module judges the degree of danger based on the motion information and presents the response type. This paper experiments with the record videos in urban environment, and validates the real-time capability and accuracy of system.