基于YOLOX的轻量级毫米波雷达和相机融合检测算法
Lightweight Millimeter Wave Radar and Camera Fusion Detection Algorithm Based on YOLOX
金建鸿 1张勇 2戴喆 3李孔4
作者信息
- 1. 浙江公路水运工程咨询集团有限责任公司,杭州 310000;浙江大学,杭州 310058
- 2. 新奇点智能科技集团有限公司,杭州 311199
- 3. 长安大学运输工程学院,西安 710064
- 4. 长安大学信息工程学院,西安 710064
- 折叠
摘要
为满足交通系统全天候高效准确的感知需求,对基于YOLOX的轻量级毫米波雷达和相机融合检测算法进行了研究;该研究主要包括异构传感器信息融合和模型轻量化两个方面;异构传感器信息融合主要针对雷达信息对视觉信息辅助能力不足的问题,设计了一种雷达空间注意力模块有效提取雷达空间特征,以此辅助视觉在低能见度场景下学习到鲁棒的特征表达;在自制数据集和NuScenes数据集上训练并测试,所提出的RV-YOLOX算法相比YOLOX算法AP指标提升约3～4左右,说明算法的全天候检测能力得到了提升;模型轻量化则针对终端算力设备对算法部署的限制,采用了结构重参数化的方式对RV-YOLOX进行了优化,轻量级RV-YOLOX在提高推理速度的同时实现了与RV-YOLOX相当的检测精度.
Abstract
In order to meet the needs of all-weather,efficient,and accurate perception in traffic systems,a lightweight millimeter-wave radar and camera fusion detection algorithm based on YOLOX is studied;The study mainly includes two aspects:the fusion of heterogeneous sensor information and model lightweight;The fusion of heterogeneous sensor information primarily addresses the in-sufficient auxiliary ability of the radar information to the visual information,and the radar spatial attention module is designed to ef-fectively extract the radar spatial features,thus help the vision to learn the robust feature expression in low visibility scenarios;Train-ing and testing are carried out on the self-made datasets and NuScenes datasets,the proposed RV-YOLOX algorithm increases the AP index by approximately 3～4 times,compared with the YOLOX algorithm,indicating an enhancement in all-weather detection capabil-ity;The lightweight model addresses the restrictions of algorithm deployment on terminal computing devices,the structural reparame-terization is used to optimize the RV-YOLOX,the lightweight RV-YOLOX improves inference speed while achieving detection accura-cy comparable to the RV-YOLOX.
关键词
交通系统/融合检测/YOLOX/深度学习Key words
traffic system/fusion detection/YOLOX/deep learning引用本文复制引用
基金项目
浙江省2021年度交通运输厅科技计划项目(2021022)
出版年
2024
NETL
NSTL
万方数据