Moving Target Detection And Recognition Based On Multi-sensor Information Fusion

With the function of environment perception, self-driving, intelligent decision-making, etc., autonomous unmanned vehicles has important practical applications infuture transportation, militarysystem, precisionagriculture, scientific inquiryand otherareas，and has become an important studying object of the coming robot era. Dynamictarget recognition is one of the key technologies of unmanned vehicles’ environmentperception. Based on multi-sensor data fusion technology, this dissertation conductresearch on dynamic target detection and recognition in the context of urban trafficenvironment. The mainly studying results of this dissertation are as follows:(1) Based on the polar grid map building method and accessible region extractionmethod in3D point cloud, and combined the position and orientation information ofINS/GPS navigation unit, we introduced a dynamic target detection method byhistorical information, global motion compensation and GMM, and realized thedetection of dynamic targets within a range of360degrees around the vehicle.(2) Based on the speed odometer information, the position and orientationinformation of INS/GPS navigation unit and accessible regions, a dynamic targetdetection method was introduced by pretreated, filtering and point cloud registration inthe dynamic target information with high false positive rate of millimeter-wave radar.This method is applicable to dynamic object detection tasks within about a90degreeangle and175m distance in front of the vehicle range, and can extract the precise radialground speed of the dynamic targets.(3) A ground points filtering algorithm which are based on the polar coordinategrid map and least-squares curve fitting, a point cloud segmentation algorithm and atraining sample labeling method which are based on the region growing and PCL areproposed. An semantic model construction method was proposed which was based on3D SIFT keypoints extraction, FPFH feature extraction, visual word generation、filtering and spatial encoding. Finally, we introduced the on-line target recognitionmethod based on semantic model in3D point cloud.