A pose estimation system based on deep neural network and ICP registration for robotic spray painting application

Nowadays, off-line robot trajectory generation methods based on pre-scanned target model are highly desirable for robotic spray painting application. For actual implementation of the generated trajectory, the relative pose between the actual target and the model needs to be calibrated in the first place. However, obtaining this relative pose remains a challenge, especially from a safe distance in industrial setting. In this paper, a pose estimation system that is able to meet the robotic spray painting requirements is proposed to estimate the pose accurately. The system captures the image of the target using RGB-D vision sensor. The image is then segmented using a modified U-SegNet segmentation network and the resulting segmentation is registered with the pre-scanned model candidates using iterative closest point (ICP) registration to obtain the estimated pose. To strengthen the robustness, a deep convolutional neural network is proposed to determine the rough orientation of the target and guide the selection of model candidates accordingly thus preventing misalignment during registration. The experimental results are compared with relevant researches and validate the accuracy and effectiveness of the proposed system.