Pallet Pose Estimation Based on Front Face Shot

Abstract

Autonomous forklifts are attracting much attention to address the serious shortage of forklift operators in the logistics industry. For autonomous forklifts to load pallets, they need to accurately estimate the 3D position and 3D orientation, i.e., the 6D pose, of the pallet. Therefore, many 6D pose estimation methods have been developed, and the latest ones utilize deep learning with a large number of pallet images for training. However, conventional methods have a problem in that estimation accuracy significantly decreases when the pallets in the captured images have appearances different from those used in training. To address this problem, the authors propose Front Face Shot (FFS), a novel 6D pose estimation method based on only the front face shot of the pallet. FFS robustly and highly accurately estimates 6D pose even in cases of unlearned pallet appearances by analyzing the 3D structure of the pallet from the front face shot by utilizing a convolutional neural network and kernel ridge regression. Experiments showed that FFS achieved the same level of estimation accuracy from untrained images as from trained ones, whereas the accuracy of conventional methods halved.

Keywords

• Pallet pose estimation
• autonomous forklift
• CNN
• KRR (kernel ridge regression)
• 6D pose