MRM-LOOP:A Dynamic Visual Loop Closure Detection Algorithm Based on Motion Region Matching

Aiming at the problem of false negatives and false positives in loop closure detection for visual SLAM algorithms in dynamic environments, a dynamic visual SLAM loop closure detection algorithm based on motion region matching named MRM-LOOP is proposed. Firstly, the VLAD local feature descriptor algorithm is used to extract the feature vectors of image feature points to ensure the stable correspondence between different images; Secondly, it uses the inter-frame motion vectors extracted from IMU data to enhance the robustness of feature matching and avoid false positive detection caused by camera pose mismatching; Finally, a region-weighted matching strategy is introduced to achieve more accurate loop closure detection by comparing the vector matching degree between the current frame and the historical frame and calculating the matching weight. The experimental results show that the proposed method can significantly improve the recall rate of loop detection in the open data set while maintaining the same accuracy, and the true positive detection probability is increased by 27.29% compared with the DBOW bag method when facing camera rotation. Correct loop detection can be realized in real dynamic environment data set. This paper provides a new solution for visual SLAM technology in dynamic environment.