| Confronted with complex road conditions,environmental perception technology also faces a growing number of challenges and tests.In current environmental perception schemes for intelligent driving,many manufacturers have adopted laser radar,millimeter-wave radar,cameras and other devices to improve vehicles’ perception of their surroundings.However,the camera performance is unsatisfactory in a strong-light or weak-light environment with a large error in millimeter-wave radar detection,while laser radar,undisturbed environmentally,can almost work all the time.It can accurately detect three-dimensional(3D)targets around vehicles,and improve the safety performance of the intelligent driving system.Traditional point cloud 3D target detection is slow in ground estimation and unable to detect the sloping ground,and it is difficult for traditional clustering methods to recognize distant objects with sparse point cloud data.In this thesis,regional ground estimation and a densitybased adaptive clustering method were proposed to solve the above problems.Firstly,the ground was no longer perceived as a whole plane,but divided into small areas which could be reasonably viewed as different planes.Then plane estimation was conducted on each small area respectively.To speed up the operation of the algorithm,faster principal component analysis(PCA)is selected to estimate the plane.In order to reduce misidentification,three evaluation functions were used for filtering,and the ground estimation results of all small areas were finally collected for the estimation of the whole ground.The proposed method not only guaranteed the accuracy,but also reduced the algorithm running time by half and improved the computational efficiency.Based on the above ideas,the point cloud data was divided into four areas according to the distance from a vehicle to its surrounding obstacles,and the point cloud data in each area was similarly sparse.Hence,an appropriate clustering threshold parameter can be introduced to each area to achieve detection accuracy while also improving the algorithm speed.Based on the architecture of two-stage 3D target detection,an area sensing module was introduced in the thesis with the proposal of a 3D target detection method based on area sensing.Firstly,the original BEV will be more helpful for pedestrian detection,but it will reduce the detection accuracy of vehicles and cyclists.Therefore,a double-way feature extraction network was adopted in this study.In one way,the original BEV feature map was extracted through two convolutional blocks of different resolutions,and in the other way,the original BEV features were retained to the next area sensing module.To improve the detection accuracy of the thick regression box in the first stage,an LPAM and an LCAM were added with comparative experiments on their combinations,such as parallel connection,serial P2 C connection and serial C2 P connection.The results indicate that in contrast to PV-RCNN,the accuracy of the proposed method is improved by 2.64%,2.43% and1.67% respectively in high-level,medium-level,and low-level pedestrian detection,and increased by 2.31%,3.28% and 1.38% respectively in vehicle detection.And the running speed has reached 20 fps,basically meeting the intelligent driving demands. |
