| Vehicle perception and motion trajectory detection are important components in intelligent transportation system(ITS),which provides basic conditions for intelligent analysis of vehicle behavior and motion status.Different from vehicle perception of onboard perspectives,roadside sensors in traffic surveillance scenes are installed higher with a wide perception range,which can make up for the occlusion problem caused by low installation height,realize vehicle-road collaborative perception and improve traffic efficiency and safety.Compared with twodimensional(2D)information of vehicle color and pixel location,three-dimensional(3D)information of localization,dimension and orientation can further enhance the integrity and accuracy of vehicle description.For example,3D vehicle trajectory can be obtained through vehicle localization and orientation,which can be applied to vehicle speed estimation and behavior analysis.In this thesis,an in-depth study on 3D vehicle perception and trajectory construction is conducted.According to the real-world requirements,the analysis method of computer vision is adopted for four aspects,including camera calibration,monocular 3D vehicle detection,3D vehicle detection with data fusion and cross-scene 3D vehicle trajectory construction.The main work of the thesis is summarized as follows:(1)Camera calibration and optimization algorithm for road traffic scenes.Aiming at the problems that current traffic camera calibration methods are not flexible and accurate enough,need to meet many scene assumptions,and cannot be applied to curves and cameras with roll angles,the non-linear constraints including the camera roll angle and focal length are constructed and solved to obtain more accurate calibration parameters and correct roll angle.Experimental verification was carried out on the public dataset Brno Comp Speed and the actual highway scene dataset.The average calibration error in 100-meter range of various scenes is0.18m(40% lower than the current similar methods)with camera roll angle correction,which meets the accuracy and efficiency requirements of camera calibration in actual traffic scenes,and lays an important basis for 3D vehicle perception and trajectory construction.(2)Monocular 3D vehicle detection algorithm based on geometric constraints.Aiming at the problem that 3D vehicle geometric constraints in the current 3D vehicle information detection algorithm based on geometric constraints are not strong enough,resulting in nonunique 3D information detection results,according to the geometric characteristics of the vehicle model,a non-linear constraint function including 3D vehicle parameters based on the diagonal,vanishing point and vehicle contours is constructed and solved.Combined with calibration parameters,3D vehicle information detection is completed.Experimental verification was carried out on the public dataset Brno Comp Speed and the actual highway scene dataset.The accuracy of 3D vehicle information detection in various scenes reaches 90%,and no additional prior information such as CAD model,orientation and view of scene is required.(3)Monocular 3D vehicle detection algorithm based on deep learning.In view of the problem that the current 3D vehicle detection algorithm relies on 2D detectors and additional geometric information reasoning modules,which increases network complexity and makes ambiguous 3D detection results,a 3D vehicle detection network Center Loc3 D is designed,which can directly and efficiently obtain 3D vehicle information without 2D detectors.A weighted feature fusion module in network structure and a loss function using 3D vehicle spatial information are designed and constructed.Combined with calibration parameters,3D vehicle centroids are obtained for 3D vehicle localization.To prove the effectiveness of the network,a roadside 3D vehicle detection dataset SVLD-3D,annotation tool Labelimg-3D and evaluation metrics are constructed.Experimental verification was carried out on SVLD-3D dataset.When the intersection of union(Io U)threshold is 0.7,3D average precision(AP),Frame Per Second(FPS),average 3D localization and dimension precision are 51.30%,41.18,98% and 85%,which can meet the accuracy and efficiency requirements of 3D perception in actual scenes.(4)3D vehicle detection algorithm based on data fusion.Aiming at the problems that single image data is easily affected by the environment and perspective deformation,and weak perception ability of long-distance small objects,a 3D vehicle detection network Center Fusion3 D using data fusion of image and point cloud data is designed.A continuous projective convolution feature fusion module in the network structure and a loss function that utilizes vehicle image-space joint constraints are designed and constructed.Experimental verification was carried out on the roadside multi-modal 3D vehicle detection dataset DAIRV2 X.Compared with methods using image data only,data fusion-based method can improve3 D AP by 6.55%,3D vehicle localization accuracy by 9.79%,3D vehicle dimension prediction accuracy by 8.07% and reduce 3D vehicle orientation prediction error by 12.11° with Io U threshold of 0.5 and FPS of 30.41,which can meet the accuracy and efficiency requirements of3 D vehicle perception in actual scenes.(5)Cross-scene 3D vehicle spatial distribution and trajectory construction algorithm based on 3D information constraints.Aiming at the problems of small perception range in singlecamera scenes and the fact that most of the current cross-scene trajectory construction methods cannot obtain continuous trajectories in 3D space and panoramas at the same time,road panoramic image generation and vehicle matching algorithm based on standard vehicle projection map and motion attribute similarity are designed to convert the cross-scene perspective into a unified 3D world space and associate vehicles in multiple scenes.The experimental verification was carried out on the public dataset Brno Comp Speed and the actual road scene dataset.The average spatial distribution error is less than 3%.Compared with existing 3D vehicle trajectory construction algorithms,the proposed method can simultaneously obtain continuous 3D vehicle trajectory and road panoramic image.The research work of this thesis realizes 3D vehicle perception in road traffic scenes and cross-scene 3D vehicle trajectory construction,which can provide certain solutions for 3D information and motion trajectory acquisition of roadside vehicles in ITS and cooperative vehicle infrastructure system(CVIS),improve traffic efficiency and ensure driving safety to a certain extent. |
PDF Full Text Request