Semantic Segmentation And Object Detection Of Indoor Scenes Based On Backpack LiDAR Point Clouds

Traditional vehicle/airborne LiDARs are not able to effectively acquire information in indoor and underground environments,and the emergence of Backpack LiADR fills this gap.However,single LiDAR sensor is unstable due to the sparse point cloud and the lack of data,which can not meet the challenges of high precision mapping,3D object recognition/detection and other applications.Multi-camera,double/multi-laser sensor data fusion is becoming a trend.This article focous on multi-sensor data process for realworld applications and challenges,did a comprehensive study from multi-sensor autocalibration,to 3D point cloud data reconstruction,and 3D object detection based on point cloud data and image data.For the base of point cloud process,we also studied 3D point cloud range search algorithm.Fast 3D Radius Search:This chapter presents a new Radius Nearest Neighbor(RNN)search approach for fast processing of 3D point clouds.We optimized it for the performance in both random and real-world point cloud data.A comparison with classical and leading algorithms is demonstrated and analyzed.It shows that compared to the Fast Library for Approximate Nearest Neighbors(FLANN),the modified-DST has gained up to 9.7X speed-up in search time,and is way much faster than high-dimensional-aimed LSH based methods.Target-free LiDAR Auto Calibration:In this chapter,a target-free automatic selfcalibration approach is proposed for multi-beam laser scanners.The proposed approach uses the isomorphism constraint among the laser scanner data to optimize the calibration parameters,uses the ambiguity judgment algorithm to solve the mismatch problem,and finally achieves the purpose of automatic calibration.3D SLAM and Secmantic Modeling:Presented in this chapter is a novel method for the mapping algorithm of an underground parking lot using 3D point clouds collected by a low-cost Backpack Laser Scanning(BLS)or LiDAR system.Our method consists a Simultaneous Localization and Mapping(SLAM)algorithm based on Sparse Point Clouds(SPC).The main contributions of this chapter are as follows:(1)a probability frontend framework for the alignment of point clouds using the local point cloud surface variance as the weight of registration,which modifies registration failure caused by the lack of features in sparse point clouds,(2)a robust submap-based strategy for loop closure detection and back-end optimization under sparse point clouds.Experimental results show that our SPC-SLAM algorithm achieves cm-level accuracy(0.09%trajectory error rate)after closed loop processing in a Global Navigation Satellite System(GNSS)-denied underground parking lot.Multi-sensor 3D object detection:This chapter presents a novel frustum-based probabilistic framework for 3D objectdetection by fusion of LiDAR point clouds and camera images.Our method project image-based 2D object detection results and LiDAR-SLAM results onto a 3D probability map,combine visual and range information into a frustumbased probabilistic framework for robust 3D object detection.Our method was tested using two datasets:one is the outdoor dataset selected from the KITTI Vision Benchmark Suite,while the other is indoor dataset acquired by our self-developed Backpack Laser Scanning(BLS)system.Our experimental results demonstrated that our method outperforms the state-of-the-art in 3D object localization and bounding box estimation.