Deep Learning based Visual Simultaneous Localization and Mapping for a Mobile Robot

Simultaneous localization and mapping (SLAM) empower robots with self-location, navigation, and environment mapping capabilities. However, visual SLAM encounters challenges such as localization loss, error accumulation, and high computational demands. The use of a camera on a mobile robot introduces uncertainties, including a limited field of view and Wi-Fi delays, which can impact perception, control, localization, and remote monitoring. Therefore, robust algorithms and strategies are indispensable for reliable operation and comprehensive understanding of the environment. The objective is to enhance the robot’s ability to accurately determine its current location despite these uncertainties. Cutting-edge learning methods for object recognition and loop closure detection was employed. The application of advanced learning techniques significantly improved loop closure detection, leading to enhanced descriptors for the model. In this paper, You Only Look Once version 5 (YOLO v5) is introduced, which facilitates faster object identification. The results demonstrate that this algorithm outperforms other evaluated algorithms in terms of real-time object recognition accuracy. Moreover, compared to alternative algorithms, the proposed framework significantly reduces both the smallest absolute pose error and relative pose error. The experimental results validate that the proposed method can achieve acceptable levels of accuracy in object identification.