Convolutional network-based method for wall-climbing robot direction angle measurement

In the past several decades, considerable research has been dedicated to the development of mobile systems that can traverse vertical surfaces. For the control of the climbing robot, high-precision sensing of the climbing robot’s heading angle during movement is very important. This paper aims to propose a vision-based scheme for the 2D direction angle detection of wall-climbing robots.

First, the authors proposed a method based on image geometric transformation to transform a camera image into a front view image of the wall, as the position and direction angle of the robot can be detected from the transformed image to eliminate the need for calibration of the camera’s internal and external parameters. Second, the AngleNet model is proposed to detect the 2D direction angle of the wall-climbing robot. Third, a training sample expansion strategy is proposed, which greatly decreased the workload of annotating training samples for AngleNet.

The single image processing time of AngleNet on the GPU is only 1.7 ms, which satisfies the demands of real-time processing. The mean and maximum error of predicted direction angle on the 556 samples of the test set are 1.1° and 3.8°, respectively.

This research offers an effective method for measuring the climbing robot’s direction angle in a complex outdoor environment. Combined with position detection, it can provide high-precision position and direction angle measurement information for the motion control of the climbing robot.