Research On Traffic Vehicle Detection Based On Image Inverse Projection 3D Reconstruction

Video-based vehicle detection is one of the issues in intelligent transportation filed and computer vision filed.As a research direction of multi-disciplinary integration including image processing,artificial intelligence,pattern recognition and so on,many problems remain unresolved,especially in practical applications,vehicle detection for complex traffic scenes has always been a difficult problem.In this dissertation,aiming at the occlusion,internal change,complex background and multi-angle,some key technologies for vehicle detection in day and night based on image inverse projection 3D reconstruction are studied,and the main contribution are as follows:(1)In view of the disadvantages of classic simplified linear model and complete distortion model,a novel calibration method for monocular camera in traffic monitoring system is proposed by combining vanishing points and direct linear transformation,which avoids the complex problems based on distortion model calibration and improves the precision of the linear model calibration.Firstly,rough calibration of the camera is completed quickly with radial,transverse and vertical vanishing points,and also the intrinsic parameter,rotation and translation matrixes are obtained.For the three vanishing points,the former two are determined based on vehicle movement trajectories and transverse edges,the last one is calculated in terms of the spatial geometric properties of vanishing points and optical center.And then,utilizing the camera parameters,space corner image coordinates of a parameterized virtual object are calculated,and are used to establish camera parameter constraint relationships based on DLT in conjunction with actual scene control points for further parameters correcting.The experimental results show that the proposed method significantly improves the accuracy of camera calibration,which is improved by 2% compared with the three vanishing points calibration,and by 4% compared with the DLT method,and stability is very well.(2)In view of the target information distortion caused by perspective projection transformation,a 3D reconstruction method for target local surface depending on single frame image inverse projection transformation is proposed.Based on target co-surface constraint pattern,3D surface data are effectively restored in inverse projection plane.Since the inverse projective planes cannot always overlap with target local surface,some pseudo-data are formed,which would interference the correct target recognition.Considering of this,inverse projection array is designed,and scale normalization and super-resolution reconstruction are adopted.The experimental results show that proposed method can perfectly restore the target local surface real information,and the edge information is kept intact.(3)For vehicle detection under complex lighting conditions at night,Adaboost and DSVM are trained based on the headlights features to construct the vehicle detector.Firstly,the AdaBoost classifier based multi-3D surface features of headlights are trained,which not only incorporate the multiple space features together,but also take the contribution rate of feature as cascade basis to makeup the weight coefficient.After then,vehicle model based DSVM is trained using lots of headlights samples in inverse projection images,which gives a more representational depth feature under the unsupervised condition to perform actual prediction.The experimental results show that the AdaBoost classifier based on the multi-3D surface features of headlights through hierarchically cascading brightness,area,perimeter,roundness,GMM,LBP and S-SURF features can complete the vehicle detection well and the accuracy reached 96.03%,the average processing times of each frame was only 14 ms;the proposed vehicle dector trained by DSVM with 6000 samples also can complete the vehicle detection well,and the accuracy reached 96.10%,but the average processing times of each frame need 23 ms.Finally,combined with the training time,sample size requirements,initiative and so on,we consider the AdaBoost classifier based on multi-3D surface features of headlights is more suitable for intelligent transportation.(4)For the occlusion,internal change and multi-angle problems in the daytime,vehicle detection based on 3D-box surface model is proposed.Firstly,a 3D-box skeleton model based on features of each component and space position relationships between them is used to detect the vehicle,in the process,HOG feature in combination with sparse representation is chosen for component detection,and for all the candidate components,centroid 3D clustering is then used to locate the 3D vehicle.Beyond that,Faster R-CNN is used to train vehicle 3D-box surface model,it changes the traditional image samples,images stitching of front,side and top inverse projection of the vehicle are chose as the novel samples,and view vector and 3D-box bounding vector are added into secondary inputs to perform multi-angles vehicle detection expediently.The experimental results show that the two methods have all achieved the high precision in daytime,the 3D-box skeleton model based on Muti-components got the accuracy rate up to 95.36%,the improved Faster R-CNN got the accuracy rate up to 95.77%.But the average processing times of each frame using the 3D-box Muti-components model is 12 ms,which is one third of the improved Faster R-CNN.Finally,combined with the robustness and practicality of method and sample size requirements,we consider that vehicle detection based on the improved Faster R-CNN is more suitable for Multi-angle complex scene,though 3D-box Muti-components model is more suitable for a system with high real-time requirements.