基于边界图卷积的机器人行驶路障场景解析  

Robot driving road scene parsing based on boundary-graphconvolution bidirectional supervised network

作　　者：徐高 周武杰[1] 叶绿[1] XU Gao;ZHOU Wujie;YE Lü(School of Information and Electronic Engineering,Zhejiang University ofScience and Technology,Hangzhou 310023,Zhejiang,China)

机构地区：[1]浙江科技学院信息与电子工程学院,杭州310023

出　　处：《浙江科技学院学报》2023年第5期402-411,共10页Journal of Zhejiang University of Science and Technology

基　　金：国家重点研发计划项目(2022YFE0196000);国家自然科学基金项目(62371422)。

摘　　要：【目的】为了使地面机器人行驶过程中能准确识别出路障以避免发生碰撞,提出一种边界图卷积双向监督网络(boundary-graph convolution bidirectional supervised network,B-GCBSNet)的场景解析算法。【方法】首先使用统一变换器(unified transformer,UniFormer)作为主干网络分别对输入的RGB(red,green,blue,红绿蓝)图像和深度图像进行特征提取;其次利用设计的多模态上下文融合模块(multimodal context fusion module,MCFM)将深度图包含的丰富空间信息补充给RGB图以提取更丰富的语义特征,在解码阶段设计了双向监督模块(bidirectional supervision module,BSM);再次,将含有更多全局信息的低级特征进行边缘化处理以得到边界信息,并通过二分类交叉熵损失函数(binary cross entropy loss,BCELoss)进行监督,而包含更多局部信息的高级特征,则通过图卷积来补充特征的全局上下文,以弥补传统卷积神经网络(convolutional neural network,CNN)提取高级特征时忽略局部位置信息的不足,并通过多分类交叉熵损失函数(cross entropy loss,CELoss)进行监督;最后将边界特征和分割特征进行整合得到最终的场景解析结果。【结果】在机器人行驶路障场景数据集(ground mobile robot perception,GMRP)上进行了试验,与已有的先进方法相比,B-GCBSNet的平均交并比(mean intersection over union,MIoU)达到了93.54%,平均类别准确率(mean classification accuracy,mAcc)达到了98.89%,像素准确率(pixel accuracy,PA)达到了98.85%。【结论】B-GCBSNet能较为准确地识别障碍物及可行驶道路,从而为地面机器人行驶过程中障碍物的识别研究提供了参考。[Objective]In order to identify roadblocks accurately and avoid collisions during ground robot driving,a scene parsing algorithm was proposed on the basis of boundary-graph convolution bidirectional supervised network(B-GCBSNet).[Method]Firstly,the unified transformer(UniFormer)served as the backbone network to extract the features of the input RGB(red,green,blue)image and the depth image respectively;then,the designed multimodal context fusion module(MCFM)was applied to supplement the rich spatial information contained in the depth image to the RGB image and extract richer semantic features;in the decoding stage,a bidirectional supervision module(BSM)was designed to marginalize the low-level features with more global information to get the boundary information and supervise with the binary cross entropy loss(BCELoss)function,while for the high-level features with more local information,graph convolution was used to model the global context of the features to make up for the defect that the traditional convolutional neural network(CNN)ignores local location information when extracting high-level features,and the multi-class cross-entropy loss(CELoss)function was used for supervision;finally,the boundary features and segmentation features were integrated to obtain the final scene parsing results.[Result]Experiments were carried out on the robot roadblock scene dataset called ground mobile robot perception(GMRP).Compared with existing advanced methods,the mean intersection over union(MIoU)of the proposed algorithm reaches 93.54%,the mean classification accuracy(mAcc)is up to 98.89%and the pixel accuracy(PA)tops 98.85%.[Conclusion]The proposed algorithm can identify obstacles and travelable roads more accurately,so as to provide a reference for studies on obstables during ground robot driving.

关 键 词：场景解析 边界监督 多尺度 上下文 图卷积 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]