Loop Closure Detection for Visual SLAM Systems Using Various CNN algorithms Contrasts

Lai, Jiancheng; Li, Zhi; Lin, Jiatai

doi:10.1109/cac48633.2019.8996385

Cited by 3 publications

(1 citation statement)

References 20 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Tateno K et al [17] generate semantically coherent scene reconstruction by fusing the dense depth maps predicted by the CNN with the depth measurements obtained by the monocular SLAM, but it may be subject to the accuracy of depth prediction limitations. In terms of V-SLAM loop closure detection, Lai J et al [18] provided an indepth analysis of the advantages and disadvantages of machine learning algorithms in this field by improving the loop closure detection process using a variety of CNN improvement algorithms, but the advantages and disadvantages of machine learning algorithms in this field still need to be analyzed in depth. In addition, for dynamic indoor environments, Ming et al [19] proposed a multi-sensor fusion odometry that removes dynamic elements based on semantic segmentation results, but increases the complexity of the system.…”

Section: Related Workmentioning

confidence: 99%

Highly Dynamic Visual SLAM Dense Map Construction Based on Indoor Environments

Mei,

Zuo,

Song

2024

IEEE Access

View full text Add to dashboard Cite

In indoor highly dynamic scenes or scenes with missing a priori dynamic information, most current schemes cannot effectively avoid the impact of moving objects on the performance of SLAM (Simultaneous Localization and Mapping) systems. In order to solve the problem of accurate localization and map building for mobile robots in indoor highly dynamic environments, we propose a dynamic RGB-D visual SLAM dense map building method based on pyramidal L-K (Lucas-Kanade) optical flow with multi-view geometric constraints. Our proposed method consists of three stages: dynamic object culling, camera position estimation, and dense map construction based on the TSDF (truncated signed distance function) model. First, dynamic elements in the scene are detected and culled by combining pyramidal L-K optical flow with a multiview geometric constraint method. Then, the estimation of the camera pose is achieved by minimizing the SDF error function. Finally, the estimated camera poses and static depth images are used to construct TSDF dense maps and indexed using dynamic voxel assignment and spatial hashing techniques. We evaluated our method on dynamic sequences of the Bonn dataset and the TUM dataset, and proved the effectiveness of the algorithm on a real scene dataset of our own making. The experimental results show that our method can effectively improve the camera position estimation accuracy and realize the construction of dense maps after processing dynamic objects, which improves the robustness of the system as well as the accuracy of environment reconstruction.

show abstract

Section: Related Workmentioning

confidence: 99%