Detection of Anomalous Objects on the Soil by Estimating Depth Map from RGB Image

Abstract

We deal with the problem of detecting anomalous objects at the pixel level for the purpose of picking anomalous objects up by the robot in the soil recycling process. A previous convex detection method does not distinguish anomalous objects and soil with stones, which makes it impossible for the robot to pick them up. In this paper, we focus on that we can get a lot of data of soil and stones although there are few data of anomalous objects. We get color images and depth maps of soil and stones by using a RGBD camera. Using the color images and depth maps, GAN is trained to estimate the depth map from the color image. GAN cannot estimate the depth map of anomalous objects because there are few training data of anomalous objects. Then, we can detect anomalous objects by comparing the real depth map and the depth map estimated from color image by using GAN. Experimental results show that our method can detect anomalous objects 2.4 times more accurately than the previous method can do. Moreover, we confirm that our method enables the robot to estimate grasping position of anomalous objects more accurately.