关键词: 回环检测, 终身学习, 对比学习, 同时定位和地图构建, 网络权重参数敏感度分析

Abstract: In visual-based simultaneous localization and mapping, the loop-closure detection might determine whether the robot reached the previous positions, so that the accumulated errors caused by pose estimation might be effectively eliminated. With the increase of data volume, the generalization performance of network model will be reduced by the existing methods. In order to realize sustainable loopback detection, a visual loop detection method was proposed herein based on sensitivity analysis of network weight parameters. A lightweight feature extraction network was constructed by combining residual neural network and generalized mean pooling. The variable similarity sensing areas were designed, and the variable sliding windows were combined with the similarity matrix to extract three samples. The proposed sensitivity analysis method of network weight parameters reduced the catastrophic forgetting of network models. Compared with the typical method MAC, the recall rate of the proposed method in Nordland and other data sets is improved by about 42%.

Key words: loop-closure detection, lifelong learning, comparative learning, simultaneous localization and mapping, sensitivity analysis of network weight parameters