| The proposal of strategy for building China into a country with strong transportation network brings higher requirements for the development of modern transportation,and also provides a good opportunity for the construction of intelligent transportation system.As an important part of intelligent transportation system,video monitoring system can obtain useful traffic information from collected video image sequence by using image processing and analyzing technologies,combining with booming pattern recognition,machine learning and other correlation theories.Despite years of development of video monitoring technologies,a robust,accurate and standardized feasible solution has not formed.Many key technologies need to be futher studied.The target segmentation is normally not so good in actural transportation scenes since the influence of speckle noise and uniform gray scale caused by illumination and other factors.The real-time and accuracy of abnormal condition detection and passenger flow volume statistics in group analysis also need to be improved.At the same time,image degradation caused by kinds of factors has prevents the system from working around the clock.Based on The National High Technology Research and Development Program of China(863 Program)and Doctoral Research Fund Project,the dundamental technologies of computer vision and maching learning for intelligent video surveillance system in passenger transport hub are researched.On the basis of learning to use Matlab computer vision development platform and OpenCV,this paper selects the practical problems faced in individual,group and scene levels in transportation scenes as the research objects and providing technical support for the development of intelligent transportation system.The main research contents of this dissertation are as follow:(1)The neural network is introducted into target segmentation according to the variety of monitoring scenes and a simplified Puls Coupled Neural Networks is build.An adaptive adjustment strategy of ignition contribution between center neural and its neighbours and the key parameter,named initial threshold,is designed.Multi-strategy morphological modifications are performed on initial detection results.The proposed method can reduce the affection of noise,restrain over-segmentation significantly and obtain satisfied detection results.And the adaptability of detection method is improved effectively at the same time.(2)Based on optical flow theories and behavior characteristics of packed pedestrian in transportation hub,especially on the platform,the motion information reflected by the dense optical flow field in transportation scenes is analysied.Optical descriptor based on motion and HOG feathers is built.An adjusted abnormal detection method based on one-class SVM is put forward by analyzing the factural scenes.(3)According to the property of Haar-Like features and the pedestrian’s own characteristics,a detection and location method for pedestrian flow is obtained.Strategy of voting combination of multiple classifiers is proposed to lower the the error rate of traditional Adaboost method for pedestrian detection and location.Target-related strategies are built to make multi-target tracking possible and at the same time,the robustness and real-time ability are effectively improved by using kalman amend and tracking window.(4)Mapping operations are executed to estimate the atmospheric composition and transmissivity according to dark channel prior.The estimated value of transmissivity is optimized by using the multi-dimensional guided filter.An initial dehazed result is generated according to the inverse process of image degradation.Multi-channel adaptive color levels adjustment is finally used to optimize the initial result to resole the problem of low light intensity. |
PDF Full Text Request