Pose Estimation of Mobile Robot Using Image and Point-Cloud Data

In Simultaneous Localization and Mapping (SLAM) techniques, the precise estimation of the initial pose of a mobile robot presents a significant challenge. The initial pose is crucial as it can significantly reduce the accumulated errors in SLAM. Despite various advancements in pose estimation, accurately determining the initial pose in different application scenarios continues to be a complex task, often requiring techniques such as loop closure. However, loop closure detection may not always be feasible, as it is time-consuming and relies on the robot revisiting previous paths or its initial starting position. In essence, most localization methods such as SLAM critically require the initial pose to estimate the mobile robot’s pose and the more precise initial pose makes the localization process to converge quickly. Addressing these persistent challenges, this paper proposes a novel method that utilizes both image and point cloud data, allowing for easy adaptation across diverse and dynamic environments. The method integrates well-known technologies such as NetVLAD, RootSIFT, 5-Point, and Iterative Closest Point (ICP) algorithms. This approach not only addresses the initial pose estimation problem but also provides an alternative to existing landmarks, enhancing adaptability to diverse and dynamic environments. NetVLAD is utilized to find the most similar image data in stored images by comparing the image captured by the mobile robot with the stored images with pose data. The relative pose is estimated by applying the RootSIFT and 5-Point algorithm, and ICP algorithm to the found image and point cloud data, respectively. This method determines the final pose of the mobile robot by combining each relative pose extracted from image data and point cloud data through weighted integration. The effectiveness of the proposed method is verified by comparing it with existing deep learning-based pose estimation methods. This method can accurately estimate poses, including the initial pose, using much less data than existing deep learning methods, even in diverse and dynamic environments. Furthermore, this method is applicable not only when using image data alone but also when both image and point cloud data are available.