A Vision Perception Module on Quadruped Robot in Metro Inspection

Metro inspection is important but costly. This work aims to achieve intelligent metro inspection by equipping a vision perception module on the quadruped robot to save labor costs. In this paper, the implementation of detection methodologies of fire extinguishers as well as other specific objects in metro carriages by quadruped robot was studied. A simulated virtual metro scenario is built to complete the collection of the dataset. In total, there are 33542 images of training data and 1151 images of validation data in the experiment. The PP-YOLO algorithm is employed to complete the object detection function on a quadruped robot and a pruning approach is used to lighten the model in order to accommodate the memory and computing power requirements of the quadruped robot. The algorithm is then integrated on Jetson Xavier NX module which is embedded on the quadruped robot. Finally, the metro inspection is proceeded with the SLAM function. The experimental results show that the detection accuracy of the algorithm reached 97% while the detection speed reached a frame rate of 30 fps.