Research On Target Recognition Method Based On Millimeter Wave Radar Point Cloud

Because of the high range and velocity resolution and relatively high angular resolution,millimeter wave radar sensors have been widely used in many fields,such as unmanned vehicles,smart homes,and intelligent security.Various forms of data like range Doppler maps,range angle maps,and point clouds can be obtained through processing the millimeter wave radar echo signal.Among the above forms of data,point cloud data can simultaneously reflect the position,shape,velocity,and electromagnetic scattering characteristics of the target,and the target recognition method based on point cloud data requires a relatively small amount of storage and computation,leading to its broad use in the unmanned driving field.Due to the limitation of radar performance,the current millimeter wave radar point cloud data is usually extremely sparse,and loses some target information compared with the original echo signal.So extracting the features that can fully describe the target characteristics from the spare point cloud data is the key point to improve the recognition performance.Most of the traditional methods extract only a few types of features in intraframes,which do not describe the target characteristics adequately.However,adequate feature extraction of intraframes and effective information fusion of intraframes can better extract the target characteristics information from the millimeter wave radar point cloud data that can help to improve the recognition performance.This paper studies the target recognition method for millimeter wave radar,mainly based on millimeter wave radar point cloud data for road user recognition tasks.The details are as follows:1.Aiming at the problems of inadequate feature extraction and incomplete description of target characteristics in existing methods of road user classification based on millimeterwave radar point clouds,this paper proposes a hierarchical point cloud classification algorithm that combines multi domain features.The algorithm extracts multiple types of features from various feature domains of point cloud cluster in intraframes,including shape,velocity distribution,radar cross section distribution,position,and combination of measurements,etc.Considering the missing shape of single-point targets and many statistical features,a divide-and-conquer strategy is used to extract different types of features and train different random forest classifiers for single-point and multi-point targets.In addition,decision fusion is performed in interframes to exploit the complementary information of targets between frames to improve the recognition performance much further.Experimental results on publicly available real world datasets show that this method has more advantages in terms of feature extraction capability and classification performance than traditional methods that only extract a few types of features in a single frame.2.For the traditional target recognition methods based on millimeter wave radar point cloud requires a large number of preprocessing steps and have the problem of inadequate feature extraction,this paper extends point cloud recognition from a classification task to a segmentation task that directly classifies the point cloud data at the scene level point by point,and proposes a millimeter wave radar point cloud segmentation network that combines interframe information fusion.The algorithm uses an improved Point Net++ network in intraframes to extract features directly from the scene level point cloud data and avoids a series of pre-processing steps,including spurious point suppression,clustering,tracking,and manual feature extraction,etc.This data driven approach also makes the automatically extracted features more separable.Meanwhile,the nearest neighbor algorithm is used in interframes to correlate the point cloud data of adjacent frames,and feature fusion and decision fusion between frames are performed on this basis.As a result,the recognition performance is further improved by introducing the interframe timing information.Experimental results on publicly available datasets show that this method avoids complex pre-processing and has better feature extraction capability compared with traditional methods based on single frame data,and the introduction of temporal information in interframe fusion further improves segmentation performance.