Object detection in point clouds is an important aspect of many robotics applications such as autonomous driving. In this paper we consider the problem of encoding a point cloud into a format appropriate for a downstream detection pipeline. Recent literature suggests two types of encoders; fixed encoders tend to be fast but sacrifice accuracy, while encoders that are learned from data are more accurate, but slower. In this work we propose PointPillars, a novel encoder which utilizes PointNets to learn a representation of point clouds organized in vertical columns (pillars). While the encoded features can be used with any standard 2D convolutional detection architecture, we further propose a lean downstream network. Extensive experimentation shows that PointPillars outperforms previous encoders with respect to both speed and accuracy by a large margin. Despite only using lidar, our full detection pipeline significantly outperforms the state of the art, even among fusion methods, with respect to both the 3D and bird's eye view KITTI benchmarks. This detection performance is achieved while running at 62 Hz: a 2 - 4 fold runtime improvement. A faster version of our method matches the state of the art at 105 Hz. These benchmarks suggest that PointPillars is an appropriate encoding for object detection in point clouds.

PointPillars: Fast Encoders for Object Detection from Point Clouds

Object detection in point clouds is an important aspect of many robotics applications such as autonomous driving. In this paper, we consider the problem of encoding a point cloud into a format appropriate for a downstream detection pipeline. Recent literature suggests two types of encoders; fixed encoders tend to be fast but sacrifice accuracy, while encoders that are learned from data are more accurate, but slower. In this work, we propose PointPillars, a novel encoder which utilizes PointNets to learn a representation of point clouds organized in vertical columns (pillars). While the encoded features can be used with any standard 2D convolutional detection architecture, we further propose a lean downstream network. Extensive experimentation shows that PointPillars outperforms previous encoders with respect to both speed and accuracy by a large margin. Despite only using lidar, our full detection pipeline significantly outperforms the state of the art, even among fusion methods, with respect to both the 3D and bird’s eye view KITTI benchmarks. This detection performance is achieved while running at 62 Hz: a 2 - 4 fold runtime improvement. A faster version of our method matches the state of the art at 105 Hz. These benchmarks suggest that PointPillars is an appropriate encoding for object detection in point clouds.

/pdf/pointpillars-fast-encoders-for-object-detection-from-point-1yuoxemh14.pdf

PointPillars: Fast Encoders for Object Detection From Point Clouds

We propose a lightweight and ground-optimized lidar odometry and mapping method, LeGO-LOAM, for realtime six degree-of-freedom pose estimation with ground vehicles. LeGO-LOAM is lightweight, as it can achieve realtime pose estimation on a low-power embedded system. LeGO-LOAM is ground-optimized, as it leverages the presence of a ground plane in its segmentation and optimization steps. We first apply point cloud segmentation to filter out noise, and feature extraction to obtain distinctive planar and edge features. A two-step Levenberg-Marquardt optimization method then uses the planar and edge features to solve different components of the six degree-of-freedom transformation across consecutive scans. We compare the performance of LeGO-LOAM with a state-of-the-art method, LOAM, using datasets gathered from variable-terrain environments with ground vehicles, and show that LeGO-LOAM achieves similar or better accuracy with reduced computational expense. We also integrate LeGO-LOAM into a SLAM framework to eliminate the pose estimation error caused by drift, which is tested using the KITTI dataset.

LeGO-LOAM: Lightweight and Ground-Optimized Lidar Odometry and Mapping on Variable Terrain

We present Argoverse, a dataset designed to support autonomous vehicle perception tasks including 3D tracking and motion forecasting. Argoverse includes sensor data collected by a fleet of autonomous vehicles in Pittsburgh and Miami as well as 3D tracking annotations, 300k extracted interesting vehicle trajectories, and rich semantic maps. The sensor data consists of 360 degree images from 7 cameras with overlapping fields of view, forward-facing stereo imagery, 3D point clouds from long range LiDAR, and 6-DOF pose. Our 290km of mapped lanes contain rich geometric and semantic metadata which are not currently available in any public dataset. All data is released under a Creative Commons license at Argoverse.org. In baseline experiments, we use map information such as lane direction, driveable area, and ground height to improve the accuracy of 3D object tracking. We use 3D object tracking to mine for more than 300k interesting vehicle trajectories to create a trajectory forecasting benchmark. Motion forecasting experiments ranging in complexity from classical methods (k-NN) to LSTMs demonstrate that using detailed vector maps with lane-level information substantially reduces prediction error. Our tracking and forecasting experiments represent only a superficial exploration of the potential of rich maps in robotic perception. We hope that Argoverse will enable the research community to explore these problems in greater depth.

/pdf/argoverse-3d-tracking-and-forecasting-with-rich-maps-50jnhm6y6h.pdf

Argoverse: 3D Tracking and Forecasting With Rich Maps

Currently autonomous or self-driving vehicles are at the heart of academia and industry research because of its multi-faceted advantages that includes improved safety, reduced congestion, lower emissions and greater mobility. Software is the key driving factor underpinning autonomy within which planning algorithms that are responsible for mission-critical decision making hold a significant position. While transporting passengers or goods from a given origin to a given destination, motion planning methods incorporate searching for a path to follow, avoiding obstacles and generating the best trajectory that ensures safety, comfort and efficiency. A range of different planning approaches have been proposed in the literature. The purpose of this paper is to review existing approaches and then compare and contrast different methods employed for the motion planning of autonomous on-road driving that consists of (1) finding a path, (2) searching for the safest manoeuvre and (3) determining the most feasible trajectory. Methods developed by researchers in each of these three levels exhibit varying levels of complexity and performance accuracy. This paper presents a critical evaluation of each of these methods, in terms of their advantages/disadvantages, inherent limitations, feasibility, optimality, handling of obstacles and testing operational environments. Based on a critical review of existing methods, research challenges to address current limitations are identified and future research directions are suggested so as to enhance the performance of planning algorithms at all three levels. Some promising areas of future focus have been identified as the use of vehicular communications (V2V and V2I) and the incorporation of transport engineering aspects in order to improve the look-ahead horizon of current sensing technologies that are essential for planning with the aim of reducing the total cost of driverless vehicles. This critical review on planning techniques presented in this paper, along with the associated discussions on their constraints and limitations, seek to assist researchers in accelerating development in the emerging field of autonomous vehicle research.

Real-time motion planning methods for autonomous on-road driving: State-of-the-art and future research directions

This paper describes a fast method for segmentation of large-size long-range 3D point clouds that especially lends itself for later classification of objects. Our approach is targeted at high-speed autonomous ground robot mobility, so real-time performance of the segmentation method plays a critical role. This is especially true as segmentation is considered only a necessary preliminary for the more important task of object classification that is itself computationally very demanding. Efficiency is achieved in our approach by splitting the segmentation problem into two simpler subproblems of lower complexity: local ground plane estimation followed by fast 2D connected components labeling. The method's performance is evaluated on real data acquired in different outdoor scenes, and the results are compared to those of existing methods. We show that our method requires less runtime while at the same time yielding segmentation results that are better suited for later classification of the identified objects.

Fast segmentation of 3D point clouds for ground vehicles

Autonomous vehicles promise numerous improvements to vehicular traffic: an increase in both highway capacity and traffic flow because of faster response times, less fuel consumption and pollution thanks to more foresighted driving, and hopefully fewer accidents thanks to collision avoidance systems. In addition, drivers can save time for more useful activities. In order for these vehicles to safely operate in everyday traffic or in harsh off-road environments, a multitude of problems in perception, navigation, and control have to be solved. This paper gives an overview of the most current trends in autonomous vehicles, highlighting the concepts common to most successful systems as well as their differences. It concludes with an outlook into the promising future of autonomous vehicles.

Autonomous Ground Vehicles—Concepts and a Path to the Future

This paper describes a LIDAR-based perception system for ground robot mobility, consisting of 3D object detection, classification and tracking. The presented system was demonstrated on-board our autonomous ground vehicle MuCAR-3, enabling it to safely navigate in urban traffic-like scenarios as well as in off-road convoy scenarios. The efficiency of our approach stems from the unique combination of 2D and 3D data processing techniques. Whereas fast segmentation of point clouds into objects is done in a 2 D occupancy grid, classifying the objects is done on raw 3D point clouds. For fast object feature extraction, we advocate the use of statistics of local point cloud properties, captured by histograms over point features. In contrast to most existing work on 3D point cloud classification, where real-time operation is often impossible, this combination allows our system to perform in real-time at 0.1s frame-rate.

Real-time object classification in 3D point clouds using point feature histograms

In this paper we describe a LIDAR-based navigation approach applied at both the C-Elrob (European Land Robot Trial) 2007 and the 2007 DARPA Urban Challenge. At the C-Elrob 2007 the approach was used without any prior knowledge about the terrain and without global positioning system (GPS). At the Urban Challenge the approach was combined with a GPS-based path follower. At the core of the method is a set of “tentacles” that represent precalculated trajectories defined in the ego-centered coordinate space of the vehicle. Similar to an insect's antennae or feelers, they fan out with different curvatures discretizing the basic driving options of the vehicle. We detail how the approach can be used for exploration of unknown environments and how it can be extended to combined GPS path following and obstacle avoidance allowing safe road following in case of GPS offsets. © 2008 Wiley Periodicals, Inc.

https://onlinelibrary.wiley.com/doi/pdf/10.1002/rob.20256

Driving with tentacles: Integral structures for sensing and motion

This paper describes a LIDAR-based perception system for ground robot mobility, consisting of 3D object detection, classification and tracking. The presented system was demonstrated on-board our autonomous ground vehicle MuCAR-3, enabling it to safely navigate in urban traffic-like scenarios as well as in off-road convoy scenarios. The efficiency of our approach stems from the unique combination of 2D and 3D data processing techniques. Whereas fast segmentation of point clouds into objects is done in a 2 1 2 D occupancy grid, classifying the objects is done on raw 3D point clouds. For fast switching of domains, the occupancy grid is enhanced to act like a hash table for retrieval of 3D points. In contrast to most existing work on 3D point cloud classification, where realtime operation is often impossible, this combination allows our system to perform in real-time at 0.1s frame-rate.

/pdf/lidar-based-3d-object-perception-12ol8g4efv.pdf

Hans-Joachim Wuensche

Papers

Fast segmentation of 3D point clouds for ground vehicles

Autonomous Ground Vehicles—Concepts and a Path to the Future

Real-time object classification in 3D point clouds using point feature histograms

Driving with tentacles: Integral structures for sensing and motion

LIDAR-based 3D Object Perception