关键词: 回环检测, 深度学习, 协调注意力机制, 自编码器降维, 相似度计算

Abstract: To address the inaccuracies in loop closure detection by traditional algorithms under variable lighting， with dynamic objects， and changing viewpoints， leading to the problem of large error in robot mapping， this paper introduces an algorithm using an enhanced MobileNetV3 for visual SLAM. The work improves the Coordinate Attention mechanism within the feature extraction network， enhancing spatial information extraction to meet loop detection needs. Features are then dimensionally reduced via an autoencoder and assessed for similarity to detect loop closures. Experimental results on the City Centre dataset indicate a 21.8 percentage points increase in detection accuracy and a significant speed improvement compared with traditional methods. This approach also more effectively reduces cumulative errors in visual SLAM systems， ensuring greater real-time performance.

Key words:  , loop closure detection； deep learning； coordinating attention mechanism； autoencoder dimension reduction； similarity calculation