Dynamic 3D Environment Perception and Reconstruction Using a Mobile Rotating Multi-beam Lidar Scanner

Abstract

In this chapter we introduce cooperating techniques for environment perception and reconstruction based on dynamic point cloud sequences of a single rotating multi-beam (RMB) Lidar sensor, which monitors the scene either from a moving vehicle top or from a static installed position. The joint aim of the addressed methods is to create 4D spatio-temporal models of large dynamic urban scenes containing various moving and static objects. Standalone RMB Lidar devices have been frequently applied in robot navigation tasks and proved to be efficient in moving object detection and recognition. However, they have not been widely exploited yet in video surveillance or dynamic virtual city modeling. We address here three different application areas of RMB Lidar measurements, starting from people activity analysis, through real time object perception for autonomous driving, until dynamic scene interpretation and visualization. First we introduce a multiple pedestrian tracking system with short term and long term person assignment steps. Second we present a model based real-time vehicle recognition approach. Third we propose techniques for geometric approximation of ground surfaces and building facades using the observed point cloud streams. This approach extracts simultaneously the reconstructed surfaces, motion information and objects from the registered dense point cloud completed with point time stamp information.

Dynamic Vehicle Pose Estimation with Heuristic L-Shape Fitting and Grid-Based Particle Filter

“…The convex-hull-based methods usually calculate the 2D convex hull of the top-view projection of the objects. They then derive the 2D bounding boxes directly from the convex hull with various strategies, such as the minimum distance-based strategy [30] and the minimum area-based strategy [29]. The prior strategy chooses the optimal bounding box by minimizing the average distance between the points of the convex hull and the fitted rectangle with the convex hull edges.…”

CNN-Based Classification for Point Cloud Object With Bearing Angle Image

“…Since the proposed classifier used an SVM classifier trained by 2D features, some important 3D features might be discarded, leading to unstable recognition. In [7], three types of features, local descriptor histograms (LDHs), spin images, and general shape and point distribution features, were used to classify roadside objects. LDHs and spin images were applied for SVM based classifier.…”

High Integrity Lane-level Occupancy Estimation of Road Obstacles Through LiDAR and HD Map Data Fusion

“…The task can be performed by surrounding polygons computation algorithms or bounding boxes creation through Lshape fitting [26], [34] or minimum area rectangle ones [4].…”