Research on road detection algorithm for power operation safety inspection robots

摘要

Abstract The safety of on-site power operations affects the socio-economic development and stable operation, and safety inspection is an important part of power production. To improve work efficiency, intelligent inspection vehicles, inspection robots, and artificial intelligence methods, including image recognition and video recognition technologies, are widely used in power operation safety inspection. Due to the complex environment of power operation sites, inspection vehicles are unable to reach all corners of the operation site. Compact and more flexible quadruped, four-wheeled, or two-wheeled inspection robots mounted on vehicles can cover more operation points, forming a new mode of automated inspection in both air and land with unmanned aerial drones. In the practical application of inspection robots, the primary challenge lies in road recognition and route optimization. To address the difficulty of road detection in the operation of inspection robots, this paper proposes a real-time road detection algorithm based on the Encoder-Decoder structure. It utilizes the idea of binary segmentation to extract semantic information of lane lines in the plane feature map and completes the feature map matching through a road line fitting algorithm, effectively improving the accuracy of road recognition at the operation site and laying a foundation for the automation of on-site inspection by inspection robots. The experimental results on the Tusimple validation set demonstrate that the proposed algorithm achieves an accuracy (acc) of 96.73% and a frame rate of 112.1 FPS, showcasing outstanding detection accuracy and efficient processing speed, providing a reference for future applications of comprehensive automated safety inspection in both air and land.