基于注意力机制CNN-LSTM的毫米波雷达点云特征数据预测生成  

Enhancement of mmWave Radar Point Cloud Feature Data Based on Self-attention Mechanism CNN-LSTM

作　　者：张春杰[1] 陈奇 赵佳琦 ZHANG Chunjie;CHEN Qi;ZHAO Jiaqi(Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,哈尔滨150001

出　　处：《电讯技术》2024年第11期1718-1725,共8页Telecommunication Engineering

摘　　要：在智能驾驶的环境感知领域,毫米波雷达是一种关键的传感器技术。然而,因数据量有限,其特征数据的采集具有一定的挑战性,这限制了环境感知分类模型的训练效果。针对这一难题,提出了一种融合自注意力机制的卷积长短期记忆网络模型,旨在预测并生成毫米波雷达点云的特征数据,以此来扩展雷达特征数据集。首先采集道路目标的运动状态数据,对数据进行二维快速傅里叶变换、恒虚警率检测,并利用多输入多输出(Multiple-Input Multiple-Output,MIMO)技术提升方位分辨率;接着执行点云聚类及特征提取;最后采用含注意力机制的卷积长短期记忆网络对特征数据进行进一步处理与预测。在真实采集的3类道路目标数据集上,与其他模型相比,该方法在不同道路目标运动特征的预测R^(2)上提高了1%~7%。In the field of environment perception for intelligent driving,millimeter-wave radar serves as a crucial sensor technology.However,the feature data acquisition from radar is a challenge due to limited data size,thus hindering the training effectiveness of perception classification models.To address this issue,the authors propose a convolutional long short-term memory(LSTM)network model integrated with a selfattention mechanism,aiming to predict and generate feature data for millimeter-wave radar point clouds,thereby expanding the radar feature dataset.First,motion state data of road targets is collected and twodimensional fast Fourier transform and constant false alarm rate detection are performed.Multiple-input multiple-output(MIMO)technology is used to enhance azimuth resolution.Second,clustering point clouds and extracting features are performed.Finally,the convolutional LSTM network with an attention mechanism is adopted to further process and predict feature data.In comparison with other models on three types of real-world road target datasets,this method has achieved a 1%to 7%improvement in R^(2) prediction of various road target motion characteristics.

关 键 词：毫米波雷达 道路环境感知 点云特征数据 注意力机制 时序预测 

分 类 号：TN951[电子电信—信号与信息处理] TP391[电子电信—信息与通信工程]