| Vehicle detection is the core component of Autonomous Driving System.It plays an important role in autonomous navigation and crash detection.Depending on the data collected by sensors,the vehicle detection algorithms can extract features of the environment,process real-time recognition of vehicle targets and detect their spatial locations in the environment.Accordingly,the vehicle detection algorithms enable reasonable route planning.There have been some mature vehicle detection algorithms in the existing research,however,these algorithms can only achieve vehicle detection in ideal environments.Due to the limitation of laser radar,the point-cloud data of distant targets are extremely sparse,affecting the algorithm performance in detecting distant vehicle targets.At present,vehicle detection algorithms are mainly divided into two categories:2D image based vehicle detection and 3D point-cloud based vehicle detection.2D image based vehicle detection can make full use of a higher number of sampling points.However,it lacks depth information to detect the distance of vehicles.In contrast,3D point-cloud based vehicle detection can make full use of depth information,which facilitates the spatial location of vehicles.On the other hand,the 3D point-cloud data is relatively sparse and has risk in losing all the information about one distant vehicle,greatly affecting the accuracy of the 3D vehicle detection algorithms for distant targets.Moreover,3D point cloud data is not evenly distributed in space,so it is difficult to directly use 3D neural network to process it.Another weakness is the high price of 3D sensors such as high-precision laser radar and ultrasonic radar,which limits the application of the 3D vehicle detection algorithms.It should be noted that the existing 3D detection algorithms treat all the input points equally.As a result,some non-vehicle points are likely to be retained and the features of some key vehicle points are hidden which misleads the detection results.Meanwhile,the existing two-dimensional vehicle detection algorithms without considering the complementary information of the 3D point-cloud.Thus,the 2D vehicle detection algorithms are not effective for vehicles with mutual occlusion,which will also affect the subsequent 3D vehicle detection algorithms.This paper proposed a novel vehicle detection algorithm based on multi-dimensional fusion of 2D image and 3D point cloud.The main contributions of this paper are as follows:(1)Proposing a 2D vehicle detection model based on 2D-3D fusion features,which makes the results of 2D vehicle detection highly consistent with those of 3D vehicle detection.The model can detect the region of the vehicle on the 2D image,and then use the 2D-3D conversion relationship to determine the position of the vehicle in the space,and then carry out the 3D vehicle detection.Moreover,due to the use of the depth information of 3D features,it can distinguish the occluded vehicles and reduce the interference of occlusion to the subsequent 3D vehicle detection.So it effectively improves the accuracy of the final vehicle detection results.(2)Proposing a 3D vehicle detection model based on the attention mechanism to improve the detection accuracy.The weight of 3D point features is calculated by attention network,so that the pooling layer can retain more key features of vehicle points,and avoid unwanted disturbance of non-vehicle points.In this thesis,a vehicle detection model based on attention mechanism and 2D-3D feature fusion feature is proposed,in which different points contribute to the results according to their importance and thus the accuracy and robustness of the detection algorithm is enhanced.The integration of 3D features in 2D vehicle detection is more beneficial to the following 3D vehicle detection module.The performance evaluation on KITTI data sets demonstrates that the proposed model obtains a significant improvement in accuracy and higher robustness for distant targets as compared with existing vehicle detection algorithms. |
PDF Full Text Request