A vision-based end pose estimation method for excavator manipulator

End pose detection of the excavator manipulator is one of the key components in the manufacture of automatic excavators. The vision-based pose estimation scheme has been identified as a potential low-cost alternative to the mechanical automation system and is gradually being applied to excavators. This paper presents an end pose estimation method for an excavator manipulator based on monocular vision, and the method network consists of two stages. In the first stage, the monocular RGB image is used as the input, and the advanced DeepLabv3 + network is utilized to segment the target, obtaining the excavator manipulator image without the background. Pose estimation is considered a regression problem in the second stage, taking segmentation results as inputs. End pose estimation is performed using the proposed pose regression network, P-ResNet, to ensure its independence from background influence. During the evaluation of pose estimation experiments, we collected a new dataset containing 2000 images based on a KOMATSU excavator. The results demonstrate that this approach exhibits strong robustness and accuracy. Its position error is less than 15 mm, and its attitude error is less than 3 degrees.