大规模复杂场景下基于ResNet的回环检测技术研究

RESNET-BASED LOOP CLOSURE DETECTION TECHNOLOGY IN LARGE-SCALE COMPLEX SCENE

回环检测(LCD)是同步定位与地图构建(SLAM)中的重要环节,对SLAM的精度和鲁棒性具有显著影响。由于大规模复杂场景下光照、摄像机视角、存在移动物体、气候、地貌特征等条件的大幅变化,使得回环检测的精度和鲁棒性受限。为解决此问题,提出一种基于深度残差网络(Deep Residual Network,ResNet),利用信息熵(Information Entropy)改进的局部聚合描述符向量(Vector of Locally Aggregated Descriptors,VLAD)的回环检测方法RIV-LCD。采用弱监督迁移训练算法训练ResNet来提取图像特征;使用信息熵加权的VLAD对图像特征进行处理;通过词袋法进行匹配,得到匹配结果。在Nordlandsbanen数据集上进行的验证和对比实验表明:在大规模复杂场景中剧烈环境条件变化下,RIV-LCD具有良好的精度和鲁棒性。

Loop closure detection(LCD)is an important part of simultaneous localization and mapping(SLAM),which has a significant impact on the accuracy and robustness of SLAM.Due to the large changes of illumination,camera angle,moving objects,climate and geomorphic features and other conditions in large-scale complex scenes,the accuracy and robustness of LCD are limited.In order to solve this problem,this paper proposes a loop closure detection method RIV-LCD based on deep residual network(ResNet)and vector of locally aggregated descriptors(VLAD)improved by information entropy.Weak supervised migration training was used to train ResNet to extract the features of the image;we used the information entropy weighted VLAD to process the image features;the matching result was obtained by the bag-of-words matching.Verification and comparison experiments on the Nordlandsbanen data set show that the RIV-LCD has good accuracy and robustness in the large-scale complex scene with dramatic changes in environmental conditions.