| Autor: |
Winata, I. Made Putra Arya, Oh, Junghyun |
| Zdroj: |
International Journal of Control, Automation & Systems; Dec2024, Vol. 22 Issue 12, p3595-3605, 11p |
| Abstrakt: |
Several simultaneous localization and mapping (SLAM) models have been developed to perform their tasks robustly based on feature extraction and matching like ORB. There are several models that have outperformed this extractor. In other cases, SLAM process is still challenging to deal with dynamic objects. Providing an integration model of lightweight extraction and segmentation of dynamic objects can improve SLAM models. Currently, several deep learning-based models can be considered for this purpose. Furthermore, this model can be enhanced using convolutional block attention modules to give more attention to the target segment and Ghost convolution modules to reduce their computational cost. To assess it, the model is evaluated through ablation studies to gain knowledge of the module's impact. It is resulting in sufficient accuracy while maintaining real-time segmentation capabilities. When embedded to SLAM system, the root mean squared absolute pose error is obtained, demonstrating that the proposed model outperforms the prior system which still uses prior extraction and matching. Improvement results are also depicted when compared with previous SLAM models. Therefore, this approach presents a new way to use lightweight extraction and segmentation models in SLAM tasks, particularly for dealing with dynamic objects. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Complementary Index |
| Externí odkaz: |
|
Full text from SpringerLink