LiDAR Perception System for Reliable Autonomous Driving Solutions

Paper: Provisioning Robust and Efficient LiDAR Perception for Autonomous Driving

Authors: Tingxu Han, Zijie Shen, An Guo, Lyuke Wang, Chunrong Fang, Zhenyu Chen, Wei Song

Executive Summary

The research focuses on enhancing LiDAR-based 3D object detection systems, which are essential for reliable autonomous driving. A significant challenge in deploying these systems is ensuring they remain effective despite various forms of data corruption, such as that caused by extreme weather conditions, while also adhering to strict requirements for speed and memory usage.

Current methods have tended to separate the goals of robustness and efficiency. This means that while some detectors are compact, they often struggle with corrupted data. Conversely, robust detectors can be resource-intensive, making them less practical for real-world applications. The researchers propose a new approach that integrates these two aspects into a single solution.

To address the issue, they developed a method for generating corrupted samples that simulate adverse weather effects. This allows for more effective training of the detection systems. The training process involves an innovative strategy that prioritizes learning from these challenging corrupted samples, which helps the detectors become more resilient. The researchers evaluated how well detectors perform under different conditions by analyzing changes in their confidence levels, the overlap of detected objects in a bird's-eye view (BEV), and instances of unmatched detections.

In terms of efficiency, the study introduces a technique to streamline the detection network. This involves analyzing neighborhood density to optimize how many points and neighbors are processed at each stage of detection.

The effectiveness of the proposed method was tested using Average Precision (AP) metrics on benchmark datasets, specifically the KITTI dataset for car detection with the PointRCNN architecture. The results show a significant improvement; under corrupted conditions, the modified detector achieved a 3.32 point increase in Moderate AP while also reducing average inference latency by 19.8% and peak GPU memory usage by 53.9%. Across multiple datasets, detector architectures, and three different object classes, the method improved average AP by 4.03 points under corruption and reduced mean latency and peak memory allocation by 14.4% and 38.0%, respectively.

This research indicates a promising direction for developing more robust and efficient LiDAR perception systems, which could enhance the reliability of autonomous vehicles in challenging environments. The findings illustrate the potential for advancements that not only improve performance under adverse conditions but also streamline operational efficiency, making these systems more viable for real-world applications. The researchers have made their code publicly available, which could enable further exploration and development in this area.

Academic Abstract

LiDAR-based 3D object detection is a core perception service for autonomous driving, but deployment requires robustness to point-cloud corruption while meeting latency and memory requirements. Existing approaches address robustness and resource efficiency separately, resulting in compact detectors that are fragile under corruption or robust detectors that remain costly to deploy. We therefore formulate LiDAR perception provisioning as constructing a detector that remains accurate under point-cloud corruptions while meeting constraints on inference latency and peak GPU memory. To achieve robustness, we first craft corrupted samples that mimic extreme weather conditions. Then we train candidate detectors offline by ordering clean and corrupted samples by perturbation sensitivity, through which detectors learn more from hard corrupted samples. We estimate perturbation sensitivity using an ensemble of seven heterogeneous 3D detectors by tracking changes in confidence, BEV overlap, and unmatched detections. To improve efficiency, we narrow the network and use offline neighborhood-density measurements to determine how many points and neighbors each sampling stage processes. We evaluate our method using AP (Average Precision) under clean and corrupted conditions, inference latency, and peak allocated GPU memory. On KITTI for Car detection with PointRCNN, our method improves Moderate AP under corruption by 3.32 points over the original detector while reducing mean latency by 19.8% and peak allocated GPU memory by 53.9%. Across two datasets, two detector architectures, and three object classes, our method improves AP under corruption by 4.03 points on average over the corresponding original detectors, while reducing mean inference latency and peak allocated GPU memory by 14.4% and 38.0%, respectively. Our code is available at https://github.com/drunklibai/robust-efficient-lidar-perception.