Research On Camouflaged Vehicle Target Detection Technology Based On Three-Dimensional Point Cloud

Under the conditions of modern information warfare,the discovery of a target means its destruction.Camouflage is one of the effective means for military targets to reduce the probability of being discovered and improve the survivability of the battlefield.It has been widely used.At this stage,whether it is reconnaissance surveillance or guided capture,the detection and recognition of military targets is mainly based on two-dimensional image features.With the advent of the threedimensional digital era,target detection and recognition based on three-dimensional point cloud data has begun to be studied and applied in certain industries,and the effectiveness of existing camouflage methods for three-dimensional target detection is a question worthy of discussion.This paper mainly conducts research on vehicle target detection technology in camouflage point cloud scenes.Due to the lack of research data,a method for constructing a camouflage target scene data set is proposed and a corresponding data set is established;for typical vehicle target recognition work,key points are extracted Research on different combinations of feature descriptors to obtain an optimal combination for the target recognition process;and on this basis,improve the TOLDI feature descriptor;finally,a vehicle target that combines global features and local features is proposed The detection algorithm is tested and verified on the established camouflage scene data set.The specific research content and results are as follows:First,in view of the lack of point cloud data of the current camouflage target scene,combined with the research needs of the camouflage target 3D detection algorithm,a construction method based on the direct acquisition of the real scene and a construction method based on the indirect synthesis of the real target and the scene are proposed.In the process of data set construction,a voxel-based iterative voxel filtering method is used,and the point cloud resolution error does not exceed±5%.In the process of indirect synthesis of the camouflage scene,according to the camouflage principle,the real camouflage scene is simulated,and the point cloud splicing technology is used to obtain the point cloud information of the camouflage scene.Finally,we obtained the point cloud data of 6 types of vehicles,including tanks,support vehicles,engineering cranes,transport vehicles,jeeps,and tractors,which were camouflaged,and acquired 40 types of camouflage scene data sets in three types of background environments:woodland,desert,and plain.Second,for the problem of how to obtain a local feature-based recognition algorithm with high recognition rate and good time performance,according to the existing local feature-based target recognition algorithm process,a summary is obtained from point cloud preprocessing,key point extraction,and feature description.A target recognition algorithm consisting of sub-characterization,feature matching,coarse registration,and fine registration.From the combined analysis of key point extraction and feature descriptors,for six camouflage vehicle point clouds,three key point extraction methods(US,Harris 3d,ISS)and three feature descriptor characterization methods(FPFH,SHOT,TOLDI)Research on different combinations of),combined with recognition accuracy and time performance,through experimental verification,the recognition rate of ISS+TOLDI on the vehicle model can reach 100%,and the recognition time is only 12.7 seconds.Third,aiming at the optimization problem based on the description of TOLDI feature descriptor,this paper improves on the optimization of the local coordinate system and the feature expression.The improvement strategy for the local coordinate system is to use the method of the threshold ratio of the number of point clouds to obtain the range of the key point neighborhood established by the z-axis and to improve the construction method for the covariance matrix of the z-axis.After experimental verification,the repeatability of the improved local coordinate system is improved by 2%-3%.The improvement strategy for feature expression is to use the information fusion feature of local depth and normal vector offset as the feature expression method.Finally,combining the improved local coordinate system and the improved feature expression method,through experiments,the feature matching pair is accurate Both the recall rate and the recall rate are increased by 5%-6%,which improves the descriptiveness of the TOLDI feature descriptor.Fourth,in order to complete the detection task of camouflaged vehicle targets in camouflage scenes,a camouflaged vehicle target detection algorithm combining global and local features is proposed.In the algorithm process,the global feature clustering is used to segment the region of interest,and the improved local feature algorithm is used for target recognition.Finally,it was verified by experiments that in forest,desert,and plain camouflage scenes,the detection accuracy of vehicles camouflaged by pattern painting reached 96%,which has an ideal detection effect.