Research On Video-based Pedestrian Targets Detection Technology In Tunnel Scene

Pedestrian detection is a necessary supporting means in modern traffic management.In tunnel scene, the traditional manual detection method is time-consuming and withlow degree of accuracy. Therefore, it is urgent to solve the problem that detecting thetunnel pedestrian automatically. Video detection technology has a good applicationprospect in pedestrian detection due to its advantages, such as large amount ofinformation, easy installation and maintenance, and so on. However, it is more difficultto detect the pedestrians in tunnel since the tunnel environment is highly sensitive to theeffects of illumination change, and most of the pedestrians are with low pixel and smallsize. Thus, the further research on video-based pedestrian targets detection technologyin tunnel scene has important theoretical and practical significance.Existing pedestrian detection algorithms based on video mostly focus on thedetection and identification of pedestrians targets in the near or medium view, but thereis relatively less researches about the pedestrian objects in the far view. In view of this,this thesis further researches a video-based detection method of the pedestrian targetsin tunnel scene. After deeply analyzing the problems and difficulties of the video-basedpedestrian targets detection technology in tunnel scene, the paper focuses on thefollowing two parts, video-based extraction of pedestrian targets in tunnel and therecognition of pedestrian targets in tunnel.In the part about the extraction of video-based pedestrian targets, a backgroundmodeling method based on nonparametric kernel density is used in the complex tunnelscene where illuminations change frequently, and a background updating method whichis suitable for the tunnel scene is proposed. And then, after the background subtraction,suspected pedestrian targets are extracted by a small area pedestrian targets denoisingmethod based on the probabilities analysis which is proposed in this paper. The resultsindicate that background modeling method based on nonparametric kernel density notonly has better effects, but also is time-saving. What’s more, the proposed backgroundupdating method can effectively inhibit the tunnel illumination abrupt changes andgradations. In addition, the presented small area pedestrian targets denoising methodcan effectively remove light noises and achieve extracting the suspected pedestriantargets more effectively.In the stage of pedestrian targets recognition, a recognition method based on multi-features is presented. Firstly, considering the shapes, areas and other features ofsuspected pedestrian targets, then to construct a description operator of contour featurefor suspected pedestrian targets. After that, to establish a contour feature recognizerbased on fuzzy c-means clustering algorithm to classify pedestrian targets andnon-pedestrian objects roughly. In addition, to build a motion feature recognizeraccording to the motion feature of pedestrian targets. At last, the pedestrian targets areidentified through the two-stage cascaded recognizer, and the automatic alarm isachieved.Finally, a pedestrian targets video detection system is established throughcombining the algorithm proposed in this thesis. Then, making experimental verificationin VC environment by using traffic surveillance videos data of Chongqing highwaytunnel scene. The results indicate that the method presented in this article can extract thesuspected pedestrian targets accurately, and the multi-features-based pedestrian targetsrecognition method proposed can identify the pedestrian targets effectively with highaccuracy and feasibility.