YOLO-GRCNN Based Object Detection and Grasp Detection on Distribution Line

During the maintenance and overhaul of the power grid system, real-time detection of insulators and other live working targets is crucial for the live working robots in the distribution network to plan motion effectively. Under the premise of ensuring accuracy and speed, the robot's vision system should not only identify the types of live working targets, but also detect their grasp poses. This paper proposes a cascaded network structure based on YOLO-GRCNN with a grasping accuracy rate of 92.25% and a single frame detection time of 38ms. To avoid the complexity of the cascade network structure, a lightweight Li-YOLOv5 model is proposed, the model includes Stem, Shuffle. Block and SimSPPF as backbone network which can help enhance the generalization ability of the network and reduce computational complexity. Furthermore. the Improved-GRCNN is proposed with the aim of further improving the accuracy of pose detection, achieving an accuracy rate of 95.8% on the test set. Therefore, the YOLO-GRCNN is suitable for the live working robots in the distribution network, enabling rapid and accurate detection of live working target types as well as their grasp poses.