一种轻量级多模态车辆轨迹预测算法  

A Lightweight Multi-Modal Vehicle Trajectory Prediction Algorithm

作　　者：李贞妮[1] 孙晖 郝梓彤 肖冬[1] LI Zhenni;SUN Hui;HAO Zitong;XIAO Dong(School of Information Science and Engineering,Northeastern University,Shenyang 110819,China)

机构地区：[1]东北大学信息科学与工程学院,沈阳110819

出　　处：《西安交通大学学报》2024年第6期14-23,共10页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(62273078);辽宁省博士启动基金计划资助项目(2021-BS-054);中央高校基本科研业务专项资金资助项目(N2204006)

摘　　要：针对自动驾驶汽车车载嵌入式计算平台存储和计算资源有限、车辆未来轨迹具有不确定性、周围环境信息复杂多变的问题,提出了一种基于MobileNeXt搭建的轻量级多模态车辆轨迹预测算法(CAM-MobileNeXt)。首先,利用MobileNeXt轻量级框架,构建了参数量和计算量均较少的车辆轨迹预测模型;其次,通过将单模态轨迹预测调整为多模态轨迹预测,以预测目标车辆可能存在的多条未来轨迹;最后,引入注意力机制,使其具备从众多输入信息中筛选出重要信息的能力,从而高效分配有限的存储和计算资源。在L5级别自动驾驶车辆轨迹数据集Lyft上开展轨迹预测实验,结果表明:所提算法具备较低的参数量和计算量,预测性能优于Lyft基线方法ResNet50;与MobileNeXt相比,所提算法在Lyft数据集上的损失值降低了11.9%,最终位移误差降低了7.4%,平均位移误差降低了11.4%。所提算法适合部署在自动驾驶汽车的车载嵌入式计算平台上,在对自动驾驶汽车的周围车辆进行准确多模态轨迹预测,以保证自动驾驶汽车安全行驶方面具有良好的应用前景。Aiming at the limited storage and computing resources of the embedded computing platform of self-driving vehicles,the uncertainty of future trajectory of the vehicle,and the complex and changeable surrounding environment information,a lightweight multimodal vehicle trajectory prediction algorithm(CAM-MobileNeXt)based on MobileNeXt is proposed.Firstly,a vehicle trajectory prediction model with fewer parameters and computations is constructed based on the MobileNeXt lightweight framework.Secondly,the trajectory prediction is adjusted from unimodal to multimodal to predict multiple potential future trajectories that may exist for the target vehicle.Finally,attention mechanism is introduced to enable the system to screen out important information and efficiently allocate limited storage and computing resources.In experiments conducted on the Lyft dataset for Level 5 autonomous vehicle trajectories,the results show that the proposed algorithm exhibits lower parameter and computational requirements,while outperforming the Lyft baseline method,ResNet50,in predictive performance.Compared with MobileNeXt,the proposed algorithm has an 11.9%reduction in loss values on the Lyft dataset.It also exhibits a decrease of 7.4%in final displacement error and an 11.4%reduction in average displacement error.The proposed algorithm is suitable to be deployed on the embedded computing platform of self-driving vehicles,and performs accurate multi-modal trajectory prediction for the surrounding vehicles to ensure the safe driving of self-driving vehicles,indicating good application prospects.

关 键 词：车辆轨迹预测 轻量级网络 多模态 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]