| The increase of the number of vehicles in the city makes the problem of traffic congestion and traffic safety increasingly serious.These problems are closely related to the behavior of the traffic participants.Video-based traffic behavior understanding technology can effectively help the traffic managers analyze the causes of congestion and accidents and is one of the important means of urban traffic governance.Existing traffic behavior understanding technologies usually take road surveillance videos as the data source.Aerial videos taken by the UAVs have larger field of view and vertical viewing angle,from which traffic behavior data in key areas can be obtained more comprehensively and accurately.In this thesis,several key technologies related to the traffic behavior understanding in UAV video including traffic object detection,traffic object tracking,the onboard camera calibration and system framework for traffic behavior understanding are studied.The main contents of this thesis are as follows:(1)For the traffic object detection,ATB-YOLO,a single-stage rotated object detection network based on anchor transformation,is proposed to predict rotated bounding boxes for traffic objects on the ground with arbitrary rotating angle in the UAV aerial video.Based on the existing single-stage object detection network,an anchor transformation network is proposed to transform the initial horizontal anchors into rotated ones,in order to reduce the difficulty of the final rotated bounding box regression.An anchor aligned convolutional operation is proposed to extract the anchor aligned convolutional feature under the guidance of the rotated anchor,in order to solve the problem of mismatching between the sampling points of the standard convolutional operation and the rotated objects.Experimental results show that the proposed network can detect the traffic objects in the UAV aerial video more accurately and at a faster speed.(2)For the traffic object tracking,this thesis proposes a joint detection-apparent network,AAC-JDAN,which combines the two tasks of rotated object detection and apparent feature extraction in a multi-task learning way.By inserting an apparent feature extraction branch in the ATB-YOLO network,our AAC-JDAN can simultaneously output the detected rotated objects and the corresponding apparent feature vectors,by which the tracking efficiency is improved.Relying on the anchor aligned feature extraction mechanism,the problem of the weak correlation between the rotated object and the apparent feature extracted by the existing joint detection-apparent network is alleviated.At the same time,a kalman filter involving angle parameters is proposed to predict the motion state of the rotated objects more comprehensively and accurately.Experimental results show that the multi-object tracking method proposed in this thesis can guarantee the tracking accuracy and achieve a near real-time tracking speed.(3)For the camera calibration,this thesis proposes a fisheye camera distortion calibration method based on the nonlinear projection model.The existing camera distortion calibration methods are usually based on the perspective projection model,which is a linear model,and is unsuitable for describing the nonlinear imaging process of fisheye cameras.This thesis proposes an distortion center estimation method for the fisheye camera with the help of the vanishing points of the stereo-graphic projection space.A distortion polynomial coefficients estimation method for the fisheye camera based on the geometric invariance of the stereographic projection model is also proposed.Experimental results show that the proposed method can obtain accurate and stable calibration results of the fisheye camera distortion with the presence of measurement noise.(4)For the traffic behavior understanding system framework,this thesis proposes a hierarchical common framework for video-based traffic group behavior understanding.The proposed common framework is a high abstraction of the whole process of video-based traffic group behavior understanding,and has no relevance to the type of the traffic objects,the scenes or the behaviors.The framework hierarchically organizes all the original data and intermediate results involved in the understanding of the traffic group behavior,and standarize the form of the data elements in each layer.All kinds of processings are divided into modules,and the interfaces between the modules are standardized.This framework can improve the development efficiency of the traffic group behavior understanding system and provide convenience for objective comparison between methods.The guiding role of the framework in the actual traffic group behavior understanding tasks is illustrated through a case study of a UAV aerial video of an urban intersection.All the proposed methods above are very helpful to overcome the shortcomings of traffic behavior understanding technology in UAV aerial videos,and can provide strong technical support for objective,accurate and efficient traffic behavior understanding. |
PDF Full Text Request