A New Landmark Detection Approach for SLAM Algorithm Applied in Mobile Robot

Simultaneous Localization and Mapping is a key technique for mobile robot applications and has received much research effort over the last three decades. A precondition for a robust and life-long landmark-based SLAM algorithm is the stable and reliable landmark detector. However, traditional methods are based on laserbased data which are believed very unstable, especially in dynamic-changing environments. In this work, we introduce a new landmark detection approach using vision-based data. Based on this approach, we exploit a deep neural network for processing images from a stereo camera system installed on mobile robots. Two deep neural network models named YOLOv3 and PSMNet were re-trained and used to perform the landmark detection and landmark localization, respectively. The landmark’s information is associated with the landmark data through tracking and filtering algorithm. The obtained results show that our method can detect andlocalize landmarks with high stability and accuracy, which are validated by laser-based measurement data. This approach has opened a new research direction toward a robust and life-long SLAM algorithm.