控制与信息技术Issue(4)：96-101,6.DOI:10.13889/j.issn.2096-5427.2024.04.013
智轨电车激光雷达多任务鸟瞰图感知算法研究
Research on Multi-Task BEV Perception Algorithm Based on LiDAR for Autonomous-Rail Rapid Trams
姚港 1龙腾 1李程 1袁希文 1李培杰 1王彧弋1
作者信息
- 1. 中车株洲电力机车研究所有限公司,湖南 株洲 412001
- 折叠
摘要
Abstract
In complex urban road scenarios,the safe operation of autonomous-rail rapid trams heavily relies on their onboard perception systems. Compared with LiDAR perception systems,visual perception algorithms are limited in distance detection accuracies and are sensitive to ambient lighting conditions. This paper proposes a novel multi-task bird's-eye-view (BEV) perception algorithm based on LiDAR,taking into account the operational scenarios of autonomous-rail rapid trams and the implementation costs of perception solutions. This algorithm employs a deep learning approach and integrates point cloud object detection and semantic segmentation into a single multi-task network. Compared to configurations with two independent networks,this approach not only reduces computational overhead but also enhances network detection and segmentation performance. Additionally,the multi-task network adopts a BEV-based point cloud encoding method,where features extracted from the point cloud are mapped and transformed into the BEV space. Subsequently,multi-scale feature fusion is facilitated through a feature pyramid,and corresponding prediction results are obtained separately from the detection and segmentation modules. Experimental results revealed an average 3D detection accuracy of 0.925 and a segmentation accuracy of 0.984. The inference time for single-frame point cloud detection and segmentation was approximately 60 ms. Thus,the implementation based on the proposed algorithm meets the requirements of autonomous-rail rapid trams for the real-time and accurate perception of surrounding obstacles and environments.关键词
智轨电车/多任务网络/目标检测/语义分割/鸟瞰图/特征金字塔Key words
autonomous-rail rapid tram/multi-task network/object detection/semantic segmentation/bird's-eye-view(BEV)/feature pyramid分类
信息技术与安全科学引用本文复制引用
姚港,龙腾,李程,袁希文,李培杰,王彧弋..智轨电车激光雷达多任务鸟瞰图感知算法研究[J].控制与信息技术,2024,(4):96-101,6.
