Cleaning of object surfaces based on deep learning: a method for generating manipulator trajectories using RGB-D semantic segmentation

A mobile robot with a robotic arm needs to be able to autonomously perceive the operating environment and plan the trajectory of the object’s surface in order to perform surface cleaning tasks in a complex, unstructured environment. This study suggests an autonomous trajectory planning technique for cleaning an object’s surface based on RGB-D semantic segmentation, which enables the robotic arm to move the cleaning mechanism on the object’s surface smoothly and steadily and finish the cleaning process. More particularly, it contains the following: (1) A Double Attention Fusion Net (DAFNet) RGB-D semantic segmentation network is proposed, which successfully integrates color texture features and spatial structure features and enhances the semantic segmentation performance of indoor objects. This network is based on the dual attention mechanism (channel attention and spatial attention). (2) The trajectory planning algorithm for the robot arm is created, and the semantically segmented data is clustered using DBCSCAN. In order to achieve autonomous planning of the cleaning trajectory, the target subject is first extracted, and then the working trajectory of the robot arm is generated via the processes of edge detection, slicing, sampling, fitting, etc. We also compare the accuracy of DAFNet semantic segmentation and other algorithms on SUNRGBD and self-built datasets, experiment with trajectory generation for various objects, and evaluate the online surface cleaning procedure. According to the experimental findings, the DAFNet semantic segmentation model is more accurate than the current models. According to the online test, the trajectory generated has a good degree of smoothness and continuity, and the robotic arm is capable of completing the surface cleaning operation effectively.