Video Analysis Of Abnormal Traffic Events And The Application System Based On Deep Image Feature

Patency and safety are the main issues of traffic management.For real-time monitoring of abnormal vehicle speed,illegal parking,traffic jam,dangerous goods transport vehicle and other abnormal events,a large number of road monitoring camera have been installed on the road,in addition to loop and radar.To replace manual monitoring,surveillance video analytical methods based on background modeling has been studied and applicated in the Shanghai Nanjing highway by our laboratory preliminary,and it has encountered some bottlenecks in the follow-up projects which implemented on provincial highway network.This includes snow,occlusion,illumination changes,and camera angle change under complex road network environment.Due to the limitation of realization mechanism of the background subtraction,the false alarm rate is high.Also,in the analysis of massive traffic surveillance video,the machine learning method based on feature representation and classification faces the problem of large sample space and complex computation,affected by some significant changes in video conditions,the robustness and generalization ability of the shallow feature is decreased.To this end,the thesis aims at the problems of abnormal event detection in traffic video analysis with deep feature,and then builds the actual application system.It’s aim to achieve the following functions:traffic flow parameters extraction,congestion detection and prediction;special vehicle identification and early warning,such as dangerous vehicles transport vehicles;illegal parking and pedestrian detection in prohibited areas;vehicle queue detection,etc.Technical problems to be further studied are:Robustness for detection algorithm under complex environment with rain and snow,occlusion,or illumination changes;Issues on false alarm of background subtraction caused by foreground movement;Heavy dependence on the scene training and a priori knowledge of the shallow artificial feature;Reverse extraction of stationary targets on road without a priori by hierarchical deep-CNN based road-non road recognition framework;Issues of large amount of training sample and calculation time,and unsuitable for engineering applications for the deep learning method;Incremental self learning model with interactive feedback of incorrectly classified samples.As a pilot knowledge,the background,content and requirements of the video detection of traffic anomalies is described firstly.Then the domestic and international research situation of traffic video detection is introduced,following by a theoretical basis of traffic video analysis:traffic image feature representation,sparse dictionary training and feature solving,etc.The analysis and discuss of limitation of background subtraction and research difficulties of abnormal event detection is also list out.Sallow feature will loss robustness and generalization in complex environment,and manual feature selection is time-consuming and laborious.To deal with these problems,a novel image feature model of discriminative sparse coding on local clustered patches called DSC-LCP is introduced to improve the discriminative power as also as reducing the dimension of dictionary.In addition,a framework of training/testing/updating called off line training-on line testing-interactive incremental learning is introduced.Through off line training and online testing,the speed of detection and tracking is dramatically improved and capable for real-time application under deep learning framework.Furthermore,the interactive incremental learning strategy feedback the false positive and negative samples into the training set and update them during idle time,this allows the framework to have self learning ability.Under this framework,the tracking process is decomposed into three steps:target detection,feature representation and object classification.The test result shows that the proposed method has obvious advantages in tracking accuracy and tracking speed,and the algorithm overcomes the target drift phenomenon caused by the accumulation of classification error which is common in traditional tracking algorithm.● To deal with problems of stationary and unpredictable objects detection on road,a road vs.non-road recognition model based on Deep-CNN is constructed.Objects on road are identified by the inverse identification of road grid.Series event detection is realized under this model,such as illegal stopping,abandoned object detection,vehicle queue detection,etc.To improve the performance of algorithm and take advantage of multi-data source,fusion method is adopted in this work,including the fusion of video and other traffic data source,fusion of multi video analysis algorithms.A mobile distribution terminal is also provided for auto video analysis and abnormal event detection.Through the vehicle classification based on the DSC-LCP feature representation,a system of objects detection,tracking is constructed.Based on the analysis of the objects and their spatial-temporal relations,a series of road abnormal event detection are realized,including road congestion detection and prediction,hazardous goods transport vehicle identification and tracking,illegal stopping,pedestrian detection,vehicle queue detection,etc.The work of this study has been applied to projects of "Special foundation for development of Internet of things and the demonstration project of Jiangsu province",“The information reconstruction project of Jiangsu Provincial Department of transportation",and so on.Testing on nearly 10,000 roads of video analysis demonstrates that the algorithm and functional module of this work is effective and reliable.It also shows that this work is valuable on both science and engineering applications.The main contributions of this work can be summarized as follows:● To deal with the problems in the video analysis of massive traffic surveillance with manual selection image feature,an automatically feature learning mechanism based on SAE and Deep-CNN is introduced.Through the deep sparse feature representation,a system of abnormal road events detection is realized,including detection/prediction of road congestion,detection and tracking of hazardous goods transport vehicles,detection of remains on road surface,vehicle and pedestrian detection in forbidden region,detection of vehicle queue.It is meaningful to improve the traffic safety and the public travel service.● A novel image feature model of discriminative sparse coding on local clustered patches called DSC-LCP is introduced to improve the discriminative power while reducing the dimension of dictionary at the same time.By considering both positive and negative samples,and clustered the local patches into basic component e.g.,edges and corners)of the object image,the discriminative power of the DSC-LCP is significantly improved compared to the typical sparse coding and sallow feature like SIFT/Haar/HOG,the tracking performance outperforms the state of the art in the comparison experiment.● A framework of training/testing/updating called off line training-on line testing-interactive incremental learning is introduced.Through off line training and online testing,the speed of detection and tracking is dramatically improved and capable for real-time application under deep learning framework.Furthermore,the interactive incremental learning strategy feedback the false positive and negative samples into the training set and update them during idle time,thus enable the framework ability of self-learning.● To deal with problems of static and unpredictable objects detection on road,a road vs.non-road recognition model based on Deep-CNN is constructed.Objects on road are identified by the inverse identification of road grid.A series of event detection is realized under this model,like illegal stopping,pedestrian detection,abandoned object detection,vehicle queue detection,etc.