A Loop Closure Detection Algorithm in Dynamic Scene

Abstract

Focus on the problem that the traditional loop closure detection algorithm is unstable and easy to fail in dynamic scene, an algorithm that can accurately detect the closed loop under the dynamic scene is proposed. First of all, the algorithm for distinguishing dynamic and static features based on scene flow is improved. Then, dynamic feature points are removed and clustering is performed. The TF-IDF entropy of each node of the image in the visual dictionary tree is used as the weight of the image in the visual words, and a score vector is constructed to describe the scene. Finally, the negative exponent power function is used to calculate the similarity of the two images.. After the final closed-loop confirmation, the final key frame with the current frame closed-loop is obtained. The experiment of the actual scene shows that the algorithm proposed in this paper can effectively detect the loop closure in the dynamic scene.

Role of Deep Learning in Loop Closure Detection for Visual and Lidar SLAM: A Survey

“…Similarly, in [ 47 ], SURF and BRIEF features have been extracted to perform the word training for loop detection in long-term autonomous driving. To improve the detection accuracy of BoW-based closed-loop detection in a dynamic environment, Xu et al [ 48 ] performed the discrimination among feature points that belong to the static and dynamic objects. The algorithm first detects and removes the feature points belonging to the dynamic objects and then generates the BoW vocabulary using the static features.…”