Research On 3D Target Detection Algorithm Based On Multi-modalities Fusion

With the development of information acquisition equipment such as lidar,intelligent detection of three-dimensional targets has become more and more important.In order to realize the intelligent detection and recognition of pedestrians,vehicles and other targets,and improve the intelligent level of unmanned driving,urban management and other applications,there is an urgent need for effective and intelligent threedimensional target detection technology.In recent years,with the development of deep learning technology,a large number of deep learning models have been used in the field of target recognition.However,these methods often only focus on the representation of three-dimensional data in a single mode,and lose the information of other modes.In response to the above problems,we propose a multi-target detection algorithm based on multi-modal information fusion.The network model consists of three modules:a point cloud feature extraction part and a two-dimensional RGB image feature extraction part.Feature extraction,information fusion and detection module weighted fusion of three-dimensional and two-dimensional feature maps to make up for the lack of information in monomodal data and achieve the complementarity of feature levels.The fused feature maps are generated through the three-dimensional and twodimensional regions to generate a network.The target detection frames are generated separately,and the later fusion strategy is adopted to fuse the detection frames of the two modalities to obtain the final target detection result.In this thesis,we also build an efficient ship target detection system by using Elastic Search,Hadoop,g RPC and other technologies,which can realize ship classification,ship tracking,traffic analysis,data storage and other functions,and can carry a large number of concurrent access requests.This system also provides a user interface with a good interactive experience,which has extremely high practical value.In order to verify the effectiveness of the method proposed in this thesis,we conducted a number of comparative experiments on the point cloud data set KITTI,including random sampling threshold analysis experiments,voxel division density analysis experiments,ablation experiments,comparison experiments with current mainstream algorithms,and single The comparison experiment of the modal method and the speed test,and the analysis of the experimental results prove the correctness and superiority of our methods.