| Transportation plays a decisive role in the bond and arteries of urban socio-economic development. With the changes of city structure, economic development, scale expansion, population growth, transportation is increasingly complex. Meanwhile, traffic congestion, traffic safety and other issues have become more and more prominent, and urban transport problems have become significant issues that constraint the social and economical production, and influence the livelihood of the people. Pedestrian as the main participants of the transportation system, its activities and characteristics are important factors for the design and operation of the transportation system. The interference between pedestrians and motor vehicles, pedestrian crowded, pedestrian safety and other issues have become serious practical problems in urban traffic. In "Medium-and-long-term Science and Technology Development Plan (2006-2020)", it has stated clearly the strategic decision of giving priority to the development of intelligent transportation systems. Using high technology to change the urban transport system, and improving its operating efficiency and service quality have become the key to solve the traffic problems.The application of computer vision technology in the field of traffic data collection has been an important research direction of intelligent transportation systems. Comparing with the traditional method of detection, traffic data collection based on the technology of image processing and pattern recognition has flexible installation, maintenance costs, and can provide rich data. Developed countries did a lot of theoretical research and engineering practice in traffic video detection. However, most of those researches seemed the vehicles as detect objects and less considered pedestrian traffic. For this reason, existing methods cannot effectively obtain real-time pedestrian traffic data, and be suited to China’s urban mixed traffic environment. Analysis and judgment the operation discipline of pedestrian traffic, interference mechanism of to automotive vehicles as well as the influence of pedestrian to signalized intersection capacity, etc., can productively manage and control urban road traffic, scientifically plan and design public transport facilities, rationally allocate transportation resources and solve the unique traffic problems in our city at last. In this context, the research within the field of pedestrian detection is carrying out step by step and has a good application prospect in China.Based on the pedestrian traffic characteristics, with computer vision and pattern recognition as the main technical means and pedestrians as detection objects, a number of video detection theories and methods are proposed in the dissertation for laying the foundation for the practical applications. The presented methods are verified by actual traffic video, and their reliabilities are analyzed.The innovations of the dissertation are as follows:①For camera calibration, the scenes are classified after the analysis of characteristics of actual traffic video scene and formulate a camera calibration algorithm for different types of scenes. For the general road scene, a camera calibration method is proposed according to the geometrical constraints between road edges and pedestrians. For the crossing street intersection scene, the geometrical constraints in crosswalk lines can be used to calibrate a camera. By making full use of the geometric constraints of the real traffic video scenes, without placing specific objects, camera parameters can be calibrated.②Towards motion object detection, an adaptive block mean codebook model is proposed. Considering the temporal and spatial relationship between the pixels, the improved codebook framework is proposed. On this basis, the similarities between codebooks are defined and codebooks can be refined accordingly. Then improved codebook model is proposed in combination with HSV color space and can adjust image block size. The improved model can handle dynamic multimodal background, gain more integrated regions of the detection subjects, product detection noise easily eliminated, and improve the operation speed, compared to the original model.③With regard to object trakcing, an improved mean shift tracking algorithm with adaptive kernel window size and target model is proposed. Meanwhile, a multi-pedestrian occlusion handling method is proposed considering the occlusion during tracking. Based on the estimations of target scale and direction, kernel window size and weight distribution of the algorithm are adjusted, in order to overcome the background distraction in tracking. Based on the definition and measurement of the change intensity of target, the target model updating mechanism is built, so that the target model can adapt target changes and improve tracking accuracy. The proposed multi-information fusion algorithm can fuse target color and motion information to overcome error superimposed problem in multi-information. The defined block factors describe the blocked situations of multi-objective, and the occlusion handling method is given to deal with the occlusion problems involving two or more objects in effectively.④As for pedestrian objects recognition and counting, pedestrian targets are classified based on their features, and in accordance with the shape model, the number of pedestrians in group can be calculated. The criterion of classification which is built according to the speed and image area of the moving target, can classify pedestrian target. A three-dimensional shape model is built to describe the pedestrians individual and group shape. The initial solution and the optimal estimation of the method for individual and group shape of the pedestrian are given to extract the occupied areas of pedestrians, and then the number of pedestrian objects in the current frame can be calculated. The method for counting pedestrians is presented when mergence and separation occur between target blobs. |
PDF Full Text Request