Application of Machine Learning in Wall Protrusion Recognition for Brachiation Robot

• Main Authors: Chen-Hsiu Yang, 楊辰修
• Other Authors: Chi-Ying Lin
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/v3n5bn

碩士 === 國立臺灣科技大學 === 機械工程系 === 107 === Climbing robots are built to perform high-risk work to replace humans, such as wall or window cleaning, exterior wall structure inspection, pipeline maintenance, etc. Autonomous climbing robots must have a sensing part to detect the environmental information, such as using a vision sensor. In the field of computer vision, the image processing technology has achieved great progress and achievements in applications such as pattern recognition. However, due to the complexity of the objects on the exterior wall and the environment, it is impossible to identify the objects in the photos and their positions through simple image processing. Thus, this study uses neural networks to develop a visual system for autonomous climbing robots to identify air conditioning, concrete eaves and sign boards. The neural networks are applied to identify and locate three kinds of protruding objects that often appear on the exterior wall of the house, which can be grabbed using the climbing robots designed in our lab. Since the computer for neural networks calculation is too heavy for climbing robots to carry, to deal with this problem this study adopts the concept of (Internet of Things) IoT and cloud computing, i.e., transmitting the photos to the server via wireless network and executing the neural networks on the server. This study uses a Fast Region Convolutional Neural Network (Fast R-CNN) architecture to implement the climbing environmental recognition system combined with a depth camera to estimate the distance between object and robot. The experiments under different light sources at the actual outdoor wall environments and the cases adjusting the angle of view from the robot side are conducted to justify the feasibility of the developed visual system, which could be applied to the ledge-climbing type robots as the basis of path planning in the future.