摘要
针对毫米波雷达点云稀疏,无法生成精确的感兴趣区域的问题,提出一种基于点云的雷达位置信息增强技术。通过对雷达点半径进行扩大,使雷达数据包含更多的空间信息和位置信息,从而实现对视觉检测模块的引导作用。实验表明:对雷达位置信息进行增强,可使平均精度均值最高提升4.1%,平均召回率均值最高提升4.9%。针对辅助驾驶过程中夜间、雨天等复杂光照环境下目标检测精度低的问题,提出一种基于多模态融合的复杂光照环境下行车检测框架MFFT-SCA。MFFT-SCA方法以多模态融合技术为基础,强化了空间信息融合模块,使经过增强处理的雷达空间位置信息与视觉特征更好地结合,得到更精准的感兴趣区域引导;引入通道注意力机制,对各通道的依赖性进行建模,对雷达通道特征进行更多的权重分配,增强雷达通道有用特征;最后通过特征金字塔,利用多尺度特征完成相应的目标检测任务。实验结果表明:MFFT-SCA框架在夜间检测环境下,平均精度均值提高了11.2%,平均召回率均值提高了6.0%;在雨天场景下,平均精度均值提高了2.2%,平均召回率均值提高了1.6%;混合天气条件下,平均精度均值提高了3.5%,平均召回率均值提高了6.3%。
Considering the fact that millimeter-wave radar point cloud is sparse and failed to generate accurate ROI,a point cloud-based radar position information enhancement technology was proposed.The radius of radar point cloud was expanded to help the radar data contain more spatial information and the position information realize guiding role of the vision detection module.Experiments show that,enhancing radar position information can increase detection effect by 4.1%at the best and the mean recalling rate by 4.9%at most.Aiming at low accuracy of target detection in complex lighting environments such as night and rain in assisted driving process,a multi-modal fusion-based vehicle detection framework MFFT-SCA in complex lighting environments was proposed.The MFFT-SCA method has multi-modal fusion technology based to strengthen spatial information fusion module,and an improved multi-mode spatial fusion strategy proposed to better combine the enhanced radar spatial position information with visual features and to obtain more accurate guidance of RIO.The channel attention mechanism introduced can model dependence of each channel and the more weight distribution imolemented can enhance useful features of the radar channel.Finally,through making use of feature pyramid,the multi-scale features adopted can complete the corresponding target detection task.The experimental results show that,the proposed MFFT-SCA framework can improve mAP by 11.2%and mAR by 6.0%in the night detection environment,including mAP improved by 2.2%and mAR by 1.6%in rainy scenarios,as well as mAP improved by 3.5%and mAR by 6.3%under mixed weather conditions.
作者
任庆坤
REN Qing-kun(Faculty of Information Engineering and Automation,Kunming University of Science and Technology;Yunnan Key Laboratory of Computer Technologies Application,Kunming University of Science and Technology)
出处
《化工自动化及仪表》
CAS
2022年第2期197-206,共10页
Control and Instruments in Chemical Industry
基金
国家自然科学基金项目(61671225,61971208,61702128)
云南省应用基础研究计划项目重点项目(2018FA034)。
关键词
视觉检测
辅助驾驶
多模态融合
空间信息增强
通道注意力
vision detection
assisted driving
multi-modal fusion
spatial information enhancement
channel attention
