A RGB-D SLAM method based on contact experience in dynamic environment
Аннотация
With the continuous development of visual SLAM technology in robotics applications, indoor companion robots can interact more effectively with humans. Existing SLAM algorithms mainly address localization issues in dynamic environments, but there is still uncertainty regarding the motion state of movable objects held by humans (such as books and cups). To address this issue, this paper proposes a dynamic SLAM system based on human contact experience. This system can effectively adapt to indoor dynamic environments and utilize human contact experience with movable objects to predict their true states. Specifically, first, we propose an adaptive frame selection strategy based on semantic segmentation results. This strategy can adaptively select RGB-D frames as input to the contact experience module, thereby reducing computational resource consumption and improving the accuracy of prior information. Second, we improved the geometric module that combines optical flow and epipolar constraints to determine the motion states of humans and movable objects. Furthermore, based on information from multiple consecutive frames, we constructed the human contact experience module. This module uses the contact frequency between humans and movable objects to update the prior state of objects within the indoor environment. Finally, we evaluated the proposed method on the TUM RGB-D dataset and compared it with ORB-SLAM2 and representative algorithms for dynamic environments, Dyna-SLAM and SaD-SLAM. Experimental results show that the proposed SLAM system operates stably in dynamic environments and effectively handles state changes of indoor movable objects.