Road Complex Environment Traffic Object Perception And Event Intelligent Recognition Based On Video Images

Achieve object perception and event recognition in the traffic area through digital and informational means,which helps to ensure traffic safety,alleviating traffic congestion,reducing traffic accidents and improving the efficiency of road operation and traffic management service level.Traffic images in highway surveillance systems have the advantages of rapid acquisition and low cost,which have become one of the most important data sources for intelligent transportation systems.Therefore,to carry out objects perception and traffic events recognition in highway environment based on traffic images has become one of the key scientific problems to be broken through in the field of intelligent transportation.To achieve traffic object perception,the first step is to accurately detect all types of objects in the traffic environment and obtain information about the location and category of each object in the image at a certain moment.On the basis of objects detection,the objects in the continuous frames of the videos are tracked to generate the moving object's trajectory in a certain time period.Through the analysis of the trajectory of the moving objects,combined with the semantic description of the traffic event,automatic judgment of traffic incidents that may involve.However,due to the complexity of the highway environment,the accurate detection and tracking of various types of objects is not easy.In addition,the variety of traffic events leads to the lack of universal recognition methods.Moreover,anomalous events that are less frequent but highly harmful,are difficult to collect sufficient samples to learning a robust model by traditional supervised learning.To address the above questions,following the underlying goal perception to event semantic cognition,this thesis carries out a theoretical and methodological research on issues related to objects detection,multi-objective tracking,general traffic event recognition and abnormal traffic events recognition in complex traffic environments.The main research content and innovative points of this thesis are as follows:(1)Objects detection in complex environments is the basis for achieving traffic object perception.Traditional hand-crafted features mainly refer to the underlying features such as color,texture,shape,gradient,etc.,which are weakly expressed,resulting in the poor robust of manual-based objects detection algorithms.Deep learning models such as convolutional neural networks,are able to deeply explore the interconnection of different features and the integration of underlying features with high-level semantic features due to their better data fitting capabilities.As a result,deep learning-based objects detection methods are now mainstream.However,traffic objects in surveillance images not only need to consider the weather changes such as rain,fog,snow but also other factors such as motion blur and occlusion.In addition,it is also necessary to pay attention to the influence of the change of objects scale on the features caused by the monitoring perspective.Thus,the deep learning-based objects detection methods,such as the faster R-CNN algorithm,is not sensitive enough to the changes of objects scale,especially for smaller scale objects such as distant pedestrians and traffic signs.To address this problem,this thesis proposes an improved faster R-CNN for traffic objects detection,we demonstrated that this method can improve the detection accuracy of traffic objects in complex traffic environments,which provided data support for subsequent studies based on the motion characteristics of objects in multi-objective tracking.(2)On the basis of accurate detection of various types of traffic objects,we carry out the research of real-time generating multi-vehicle trajectories by tracking vehicles.In response to the screening of look-alike objects,we innovatively propose a target-specific feature sparse coding,taking into account the intra-class and inter-class discriminability of object feature expressions.For the problem of partial occlusion between objects,we analyzes the dominance of 1-Wasserstein distance(WD-1)in similarity metrics from the theoretical and experimental views and innovatively introduce the improved WD-1 as a measure in multi-vehicle tracking.Then,a multi-vehicle tracking method with surveillance video based on the improved Wasserstein distance is proposed and compared with several mainstream multi-objective tracking methods.The experimental results verify the advancement of our method.(3)On this basis,by clarifying the composition structure of traffic events under different semantic granularity and exploring their intrinsic interrelationships and inference mechanisms,a traffic event identification method based on hierarchical semantic description is proposed,which fills the "Semantic gap" between the temporal and spatial information of the underlying object and the semantic information of the high-level event.Contrasting others studies related to traffic event recognition almost only focus on one type of traffic event,this method experimentally demonstrates that the majority of traffic events types can be accurate achieved,including illegal parking,driving in the wrong direction,pedestrians and motor vehicles running red lights,congestion,etc.,Therefore,this method are more general applicability.(4)The methods of abnormal events recognition in road environment based on deep learning mostly adopt supervised learning,which requires a large number of labeled samples to train the model.However,these samples collection are more difficult due to abnormal events such as fires are unfrequent.Domain drift problem will occur if semantically related images of other scenes are directly used as training samples.In this thesis,we take fire events as an example,address the domain drift problem through theoretical research and experimental analysis,draw on the ideas of domain adaptation and propose an adversarial adaptation fire recognition and detection method for surveillance images and UAV images,respectively.Different from the previous idea of mapping source and target domain data to the same feature space and forcing them to align in that space at the same time,in this method,the source and target domain data are mapped to different feature spaces,the respective mapping functions are optimized in different stages,and the feature distributions of the different domains are aligned by adversarial adaptation training.It is demonstrated that this method can improve the model discrimination on the target domain and improve the recognition and detection accuracy of fire events.This thesis starts from the practical problems of object perception and traffic event recognition for highways and urban roads.completed the data labeling and processing,model building,training and validation of five research topics,including traffic object detection,multi-object tracking,general traffic event recognition and abnormal event recognition and detection.At the same time,the outlook is put forward from the practicality and applicable scope of the above topics,providing reliable technical support for traffic digitalization and information management.