| With the continuous development of technology,unmanned driving is not only the core strategic layout of countries in the field of artificial intelligence,but also a key technology to ensure traffic safety.Aiming at the problems of low accuracy,long and obscured vehicle detection and uneven illumination in unmanned driving,this paper uses methods such as image processing and deep learning to study vehicle detection algorithms based on point cloud and image fusion.The main research contents are as follows:The research progress of unmanned vehicles at home and abroad,the basics of deep learning and the common datasets for vehicle detection are outlined.The 3D vehicle detection method based on image,point cloud and multimodal fusion is introduced,and the idea of 3D vehicle detection under unmanned system is analyzed.In order to solve the problems of low precision and poor robustness of 3D detection algorithm in monocular camera,an improved SMOKE 3D target detection algorithm is proposed.Deformable convolution was used to replace the original convolution before the upper sampling layer in backbone network DLA-34,and the multi-scale Future Map was sampled into a 1/4 size feature map,and the convolution was added from top to bottom.Optimize the selection of key points in the center,add 8 vertices in the 3D detection box as key points,and enhance the constraint on the detection box;Regression optimization loss function for vertex and distance;Select the appropriate hyperparameter training model,and finally improve the accuracy of the 3D detection frame without losing the detection time.Aiming at the problems of slow vehicle detection speed and lack of 3D information in2 D images,an improved Point Pillar 3D vehicle detection algorithm was proposed.The noise in point cloud data is filtered by statistical filter,and a variable size fan-grid method is proposed to remove the ground point cloud to reduce the amount of subsequent computation.Improving the feature coding module to enrich the semantic information;Subsampled intensive feature and sparse encoder are added to the feature fusion module to improve the robustness of feature extraction.The feature information is passed into the detection head to output the detection result of the vehicle;It lays the foundation for the three-dimensional vehicle detection under the unmanned driving system.Aiming at the low accuracy of 3D vehicle detection in unmanned vehicle system,a 3D vehicle detection algorithm based on multi-mode feature fusion was proposed.The time and space registration of Lidar and monocular camera is realized,so that the coordinate system of the two sensors is corresponding and unified,which lays the foundation of multi-mode target detection.The statistical filtering algorithm was used to remove the clutter points and noise points in the lidar data,which reduced the amount of calculation for the subsequent detection.The key points module of data input generates the key points of point cloud and image respectively,and obtains the regions of interest.The local features on the point cloud and image region are extracted respectively,and the local feature input detection head is fused to obtain the 3D vehicle detection results.KITTI 3D data set evaluation shows that the proposed multi-mode fusion detection model achieves good detection results,detection time is only 0.11 s,the average detection accuracy is 83.68%,the algorithm provides a feasible scheme for unmanned vehicle detection. |
