Research On Key Theories And Technologies For Detecting Pedestrian Traffic Information Based On Computer Vision

With the popularization of video surveillance system and the development of video image processing technology, the application of intelligent transportation system based on computer vision technology is paid more and more attention. It can comprehensively utilize image processing, pattern recognition, artificial intelligence and other technologies to process and analyze the video image sequence. Through intelligently understanding and processing the content of videos, the system can deal with various problems such as accident information judgment, pedestrian and vehicle classification, traffic flow parameter detection, moving target tracking and so on, which make the intelligent transportation system more intelligent, and provide a comprehensive and real-time traffic information for traffic management. Therefore, the research on the method of traffic information detection based on computer vision has important theoretical value and practical significance.Though the intelligent video surveillance technology has been studied for many years, intelligent traffic information detection system based on computer vision is still in developing stage and some key technical problems still need be further studied. At present, there is not a standard, sound, accurate and high-performance target detection and tracking algorithm, which couldn’t acquire real-time and effective pedestrian traffic data and is difficult to analyze and judge rules of pedestrian traffic. The effective management and control of traffic environment couldn’t be done up to now.In this context, the research on the field of traffic information detection based on computer vision has gradually expanded, and has shown a good application prospect.Based on The National High Technology Research and Development Program of China (863 Program) and Doctoral Research Fund Project, the fundamental theories and technologies of intelligent transportation system is researched. Combined with the theory of computer vision, based on learning to use MATLAB computer vision development platform, this paper selects the pedestrian in traffic video as the research object and researches moving target detection, extraction, recognition, tracking and calculation and traffic flow parameters, providing technical support for more intelligent ITS. The main research contents of this dissertation are as follows:(1) Based on the related theory of computer vision and intelligent traffic information detection, traffic semantic information are layered by the image semantic hierarchy method. The whole process include the underlying visual-layer, middle visual-layer, high visual-layer and application layer, and to define the function of each layer. The key technology structure of intelligent detecting traffic information are presented and the relevant theoretical knowledge applied in this paper are described in detail. Integrated intelligent video surveillance technology, the system structure of the intelligent monitoring traffic information is constructed, and hardware and software platform of monitoring system is built, which realize the transformation of theory to practice and provides basis for improving the video monitoring capability.(2) According to the problem that obtaining reliable background image in real traffic scenes is difficult, self-adaption background modeling method combined with optical flow velocity field theory is presented. By introducing optical flow in the background modeling, real-time updating background could be realized combined with background subtraction results and "blind angle" gray scale processing. The model can accurately extract the background image and effectively eliminate the noise. Then, based on the background fitting model, a foreground object segmentation method based on temporal and spatial information are proposed. The initial detection mask images are obtained by using the grayscale difference detection of adjacent multi-frames and edge information, which effectively solve the difference localization and noise problem very well; The space mask images are got by introduced reconstruction and marker technology in the spatial-domain watershed transform, which effectively solve the over-segmentation problem; finally the fusion of spatial and temporal further ensure the segmentation accuracy and efficiency.(3) In the moving target detection, a pedestrian detection and the underlying semantics extraction method based on morphological connected domain is proposed. By using morphological connected domain identification technique, distinguished by the features of connected domain, irrelevant areas are deleted and moving objects in video image are extracted, so as to accurately extract the moving target of the underlying traffic semantic information and provide data for the succeeding work. For the shielded pedestrian traffic characteristics, a pedestrian detection method is put forward based on head color model and contour information. In this method, clustering condition and frame difference of motion information in the RGB and YCbCr color space are adopted to extract the candidate head regions; Canny edge detection and Hough transform is used to locate the head and then count the target information.(4) In the moving target tracking, to solve the defects of mean shift algorithm, the improved algorithm based on mean shift target tracking is put forward. Target appearance modeling method with multithreading information fusion is employed. By using appearance, space structure and the movement of the multithreading pedestrian information to describe the target pedestrian, the ability to describe the pedestrian feature are enhanced and the tracking accuracy are improved; Through the background and the target, criterion that determine the variation region of the target size is set up, and the centric window’s size are adjusted to overcome the background interference; Using the Bhattacharyya coefficient to determine the tracking state, a block target processing method based on the four search strategy is proposed to capture the lost target; Based on the target tracking algorithm, the middle-level semantic information are collected, including the pedestrian displacement, the velocity, the acceleration, the trajectory and so on; In the stage of collecting information, the ROI region and the target chain are established, and the pedestrian counting algorithm based on target tracking is proposed to count the pedestrian flow information in the ROI region.(5) Based on extracting the bottom and middle-level semantic information, an improved BP neural network target recognition method based on hierarchical genetic algorithm is proposed, and the four level hierarchical chromosome structure are adopted to describe the network structure and parameters, then according to the construction of HGA-BP single classifier, the type and quantity of moving objects in video image could be identified. Then on the basis of the HGA-BP single classifier, cascade identification based on the idea of "from the coarse to the fine" are achieved and the Cascade-HGA-BP combined classifier are built. When the underlying information are transmitted to upper levels, the three departing method are adopted to realize the final classification of the target. In the case that pedestrians and vehicles exist in some traffic scene, this method is very effective to distinguish the moving target.