| The probability of traffic accidents increases with the increase of car ownership in China,and road traffic safety is particularly important.Studying intelligent vehicles with automatic driving function is an effective way to solve road safety problems.The three-dimensional detection and tracking based on point clouds is the core technology of intelligent vehicle automatic driving.It is difficult to detect and track targets accurately by using only point clouds data because of the huge amount of point clouds data,complex processing and sparse point clouds data.To solve the above problems,a feature fusion algorithm is designed to fuse the feature information of images and point clouds to improve the detection accuracy of target objects.The research on fast detection and tracking methods with accurate and efficient balance of targets is of great significance to improve the safe driving of smart cars on the basis of ensuring the accuracy of target detection and tracking.The main research results of this paper are as follows:(1)Proposed an improved images feature extraction method based on Faster R-CNN.The existing two-dimensional object detection algorithm is aimed at the problems of small objects and occluded objects detection,such as easy to miss detection,error detection and so on.ResNeXt-50 is used to replace the original backbone network VGG-16 to optimize the network’s ability to extract feature information of small objects and occluded objects.In order to make full use of multi-scale feature graph feature information,the Faster R-CNN model introduced feature pyramid structure(FPN),and designed recursive feature pyramid structure based on hierarchical convolution(HC-RFPN)to improve the ability of utilizing semantic information of high-level feature graph and detail information of low-level feature graph.Finally,the Faster R-CNN images feature extraction model(RC-RFP)was proposed based on residual structure and hierarchical convolution recursive feature extraction.Experimental results show that the images extraction model designed in this paper has good performance in the detection accuracy and speed.(2)Proposed a 3D target detection method based on point clouds and images feature fusion.In order to fully integrate point cloud and image information,RC-RFP is used for image feature extraction.At the same time,SECOND 3D target detection model is used to extract point cloud features,and point cloud features and image features are fused and expanded from feature level.In view of the phenomenon of missing and wrong detection caused by less semantic information of object point cloud,PointFuison multi-source data fusion algorithm is introduced,and D-IOU idea is introduced to replace the method of relying on feature range to judge the fusion result.Design MSF-PointFusion algorithm to fully fuse point cloud and image features.In view of the omission of small and medium-sized objects and occluded objects in random sampling process due to sparse and irregular point cloud characteristics,An improved VEF-2 based on SECOND feature learning network is designed to compensate the lost feature information of point cloud by using image feature information by adding additional probability channels.Finally,a 3D target detection algorithm IFDM-SECOND based on multi-source data fusion is proposed.The experimental results show that the 3D target detection algorithm designed in this paper is more competitive in the detection accuracy and accuracy.(3)Designed a 3D multi-target tracking method based on DeepSORT model.For improving the accuracy of multi-target tracking,the K-Means clustering algorithm is studied to estimate the target centroid for scenes with many small objects and occluded objects.For tracking application scenarios in driving environment,3D Kalman filter is studied to predict the state of 3D target.Introduce greedy algorithm to solve the problem that the matching speed decreases with the increase of targets,and use 3DIoU to calculate the intersection ratio of predicted targets and detected targets,and construct the potential matching target set of current targets;Meanwhile,the Euclidean distance between the centroid of the current frame target and all potential targets is calculated,and the optimal potential matching targets are screened to avoid the problem that one target has more potential matching targets and effectively improve the accuracy and speed of the tracking algorithm.At last,the cascade matching algorithm is optimized.By constructing the set of disappearing targets,setting the frame number limit and using 3D Kalman filter,the last frame target that disappeared in the current frame can be tracked continuously within a certain frame number.In the end,the 3D multi-target tracking model 3DT-DeepSORT is proposed.The experimental results show that this tracking algorithm can effectively reduce the number of mismatching and false detection of targets in the tracking process,and it is more competitive in multi-target tracking. |
PDF Full Text Request