人形機器人之輕量化單眼視覺羅盤

Abstract

noneHumanoid robots have become an emerging and challenging research field and played a central role in robotics research and many applications during these years. Generally, when performing specific tasks in the real world, the robot needs to Estimate its Orientation From the Environment (EOFE). A visual compass (VC) is the typical method to solve the EOFE problem. However, when using a humanoid robot as a platform in the EOFE problem to implement a VC algorithm, there are some limitations: First, the camera used as a sensor is sensitive to strong motion blur. Second, with the robot's physical structure restriction, the available CPU will be relatively constricted. That is, the problem should be how to apply a camera-based visual compass with satisfactory performance on a humanoid robot at low cost. Therefore, we proposed a simplified visual algorithm based on the appearance, which can analyze the pixel distribution correlation of the image to estimate the robot's current orientation to correct its movement deviation. On the other hand, we also proposed a simplified landmark-based visual compass which uses a target object as a reference point in the environment to estimate the movement angle. Both of these proposed algorithms take less time in computation, the HCVC can process 137.737 fps, and the DCTVC can process 166.818 fps. Also, the two algorithms have satisfactory performances: the HCVC has the minimum mean-square-error (MSE) of 0.20917 degrees, and the DCTVC has 0.21713 degrees.