中国科学院自动化研究所机构知识库

Service robots have arisen a national strategy. In order to provide high-quality service, autonomous localization and mapping are significant in both research and practice with a wide application prospect in intelligent manufacturing, domestic service, and medical assistance. This thesis focuses on the research on vision-based localization and mapping for service robots. The main contents are as follows:

Firstly, the research background and the significance of this thesis are introduced. The develepment of visual SLAM from the aspects of individual robot and multi-robot system is reviewed. The contents and structure of this thesis are also given.

Secondly, the vision-based odometry for robots in dynamic environments is researched. A vision-dominated multi-sensor fusion odometry for indoor robots is proposed, which integrates semantics, gyroscope and wheel odometry information. The proposed method directly estimates the robot body state, and the robot pose is parameterized by 3D position and 3D Euler angle to facilitate the tight coupling of wheel odometry constraints and planar motion constraints. Meanwhile, the semantic segmentation is used to filter out the dynamic features, and the ground point constraints are imposed on the ground features in the optimization process. The proposed odometry is verified in public datasets and actual environments.

Thirdly, a multi-camera SLAM method using fixed transformation constraints to fuse trajectories of different cameras is presented. Each camera uses the front-end odometry to track the visual information and build a map. In the back-end, fixed transformation constraints are applied to different camera keyframes that are close in time and then the maps generated by the front-end are fused together. As a result, the robot can continue to execute task when the tracking failure for part of cameras occurs. Compared with the single-camera SLAM, it is beneficial to establish a more complete map. The proposed method is verified by simulation and experiment.

Fourthly, considering that some scenes maybe change during multi-robot mapping, a Server-Client architecture-based multi-robot visual mapping framework with collaborative map updating ability is designed. An algorithm for searching visible global landmarks for robots is proposed on the Server side, and new observations are used to update global landmarks on the robot-end, which will be sent to the Server for further processing. The framework is verified in public datasets and real scenes.

Finally, the conclusions are given and future work is listed.