Object Tracking And Application Technologies In The Integrated Video Monitoring System Of Railway

Pedestrian safety has attracted considerable attention in social security management and intelligent security construction.The integrated video monitoring system of railway has covered core areas of passenger stations.It provides high-resolution video resources for the monitoring center and becomes the basis of real-time monitoring,abnormality warning,and post-check.The current system mainly relies on manual observation,which is short of automaticity.With the development of installation scale and monitoring intensity,manual interpretation and decision-making are prone to omissions and mistakes,which mishandles potential abnormal situations.Automatic understanding of surveillance videos is urgent for the safe operation of railway traffic,which is also the future research direction of intelligent transportation.Supported by the research project 2015X009-H from China Railway,this dissertation carries out the application research of computer vision technologies in surveillance videos stemming from safety demand for the railway transport operation.The problems of detection and tracking in the railway surveillance scenes are analyzed.New attempts conduct in two aspects: the search region and the object proposal.Based on the target trajectory,two types of high-level applications are explored respectively: behavior recognition at the individual level and crowd analysis at the group level.The main contents and innovations of this dissertation are as follows.(1)For tracking tasks with complicated and unknown motion patterns,the Bayesian estimation framework is introduced to construct the association between the search area,the prior distribution,and the motion model.Then,the search area generation method is designed.According to the variability of application scenarios,two search area setting strategies based on data-driven models are constructed: the random walk model and the vector autoregressive model.Experiments validate their effectiveness on computation time and improvement of tracking accuracy.The two models improve original tracking performance by 22% and 10%,respectively.(2)According to the characteristic of video monitor points in transport stations,the object proposal of detection tasks is analyzed.For a stationary surveillance camera,different patches on the image correspond to the views of various depths leading to diverse sizes of region proposals.A neural network structure is developed to empirically estimate the size of a pedestrian candidate in the pixel coordinates given its central location model,which provides a solution to the imbalance problem of object size distribution in surveillance scenes,and efficiently generates high-quality region proposals.It is also helpful for pedestrian detection in the distant view.Experimental results show that the proposed method outperforms other advanced methods in recall under several situations.(3)In the behavior recognition problem using semantic descriptions,Petri-net based event modeling method is improved.The event path is established according to the limited sequence operation mechanism.Then,the system network of the event model is automatically generated.Combing with multi-level features,the introduction of the decision tree eliminates conflicting event places sharing the same path.The method interpretably captures more comprehensive model branches and improves the precision of behavior recognition.Experimental results demonstrate the precision of the proposed method achieves 99.59% in the multi-event detection task.(4)To investigate scene characters in passenger station scenes,trajectory-based scene modeling and anomaly detection methods are researched.To address the problem caused by projective distortion and complex scene structure,a clustering method which handles the projective distortion is proposed to analyze movements in the self-organized channel.A scene modeling method is proposed to describe the characteristics of target trajectories in scenes with multiple functional regions.After that,an anomaly detection method based on the scene model is proposed.Experimental results validate the effectiveness of the proposed method in detecting abnormal trajectories of individual targets and summarizing the behavior law of pedestrians inside self-organizing channels.Besides,its accuracy achieves 97.8% in the abnormal crowd state detection task.