A comprehensive survey of LIDAR-based 3D object detection methods with deep learning for autonomous driving

TL;DR: This survey paper reviews the research progress for infrastructure-based object detection and tracking systems and highlights current opportunities, open problems, and anticipated future trends.

TL;DR: PillarGrid as discussed by the authors proposes a cooperative perception method fusing information from multiple 3D LiDARs (both on-board and roadside), to enhance the situation awareness for connected and automated vehicles.

TL;DR: In this article , the authors investigate the role of attention mechanisms for the task of 3D semantic segmentation for autonomous driving, by identifying the significance of different attention mechanisms when adopted in existing deep learning networks.

TL;DR: Multi-modality 3D object detection in autonomous driving reviews existing research on LiDAR and camera fusion approaches for 3D object detection. It identifies key challenges and promising research directions in the field.

Abstract: Boss is an autonomous vehicle that uses on-board sensors (GPS, lasers, radars, and cameras) to track other vehicles, detect static obstacles and localize itself relative to a road model. A three-layer planning system combines mission, behavioral and motion planning to drive in urban environments. The mission planning layer considers which street to take to achieve a mission goal. The behavioral layer determines when to change lanes, precedence at intersections and performs error recovery maneuvers. The motion planning layer selects actions to avoid obstacles while making progress towards local goals.The system was developed from the ground up to address the requirements of the DARPA Urban Challenge using a spiral system development process with a heavy emphasis on regular, regressive system testing. During the National Qualification Event and the 85km Urban Challenge Final Event Boss demonstrated some of its capabilities, qualifying first and winning the challenge.

TL;DR: This report presents a class-balanced method for 3D object detection, utilizing sparse 3D convolution and a class-balanced multi-head network, achieving state-of-the-art performance on the nuScenes dataset with a large margin over the PointPillars baseline.