基于多传感器融合的越野环境路面信息识别  

Road Information Recognition Based on Multi-Sensor Fusion in Off-Road Environment

作　　者：刘辉[1,2] 刘聪 韩立金 何鹏[1] 聂士达 LIU Hui;LIU Cong;HAN Lijin;HE Peng;NIE Shida(School of Mechanical Engineering,Beijing Institute of Technology,Beijing 100081,China;Institute of advanced Technology,Beijing Institute of Technology,Ji’nan,Shandong 250307,China)

机构地区：[1]北京理工大学机械与车辆学院,北京100081 [2]北京理工大学前沿技术研究院,山东济南250307

出　　处：《北京理工大学学报》2023年第8期783-791,共9页Transactions of Beijing Institute of Technology

基　　金：国家自然科学基金资助项目(52130512,52002212)。

摘　　要：为实现精准的越野环境路面信息识别,文中提出了一种基于多传感器信息融合的路面信息识别方法.首先,针对车辆簧下振动加速度信号设计了特征提取算法,通过双线性池化方法融合加速度特征与图像+深度特征,以实现对越野路面类型的多维度特征融合与识别.然后,为提高越野路面可通行区域检测效果,引入迁移学习方法,将越野场景路面类型识别模型中路面特征提取的共性知识向通行区域分割模型进行迁移.在真实越野环境数据集下对模型进行训练与测试,测试结果表明文中提出的识别方法不仅在越野场景路面类型识别任务上获得了98.65%的平均分类准确率,而且引入先验知识可明显提升通行区域检测效果.In order to achieve road information recognition in off-road environment accurately,a road informa-tion recognition method was proposed based on multi-sensor information fusion.Firstly,according to the vibra-tion acceleration signal under vehicle reed,a feature extraction algorithm was designed for road terrain.Integrat-ing the acceleration features and image+depth features based on bilinear pooling method,the method was ar-ranged to realize multi-dimensional feature fusion and recognition of road terrain.Then,in order to improve the detection accuracy of road passable area in off-road environment,a transfer learning method was introduced to transfer the common knowledge of road feature extraction from the off-road road terrain recognition model to the road passable area segmentation model,and trained and tested with a real datum set of off-road terrain.Test res-ults show that the proposed method can not only achieve an average classification accuracy of 98.65%in the task of off-road terrain recognition,but also the introduction of prior knowledge can obviously improve the detection effect of road passable area.

关 键 词：越野环境 路面类型识别 多传感器信息融合 迁移学习 可通行区域检测 

分 类 号：U46[机械工程—车辆工程]