Research On Vision-Based Pallet Detection Methods

Robots,such as autonomous forklifts,are fundamental to intelligent warehousing.As the key of automatic and flexible logistic operation,pallet detection refers to detecting pallets in the warehouse and estimating the relative position and orientation between pallets and the forklift.Existing pallet detection methods are mainly LiDAR-based,vision-based or a combination of both.However,LiDAR is too expensive to afford.Vision sensors are more promising in industrial application,but poor in illumination robustness and detection accuracy.Also vision-based detection results are easily influenced by targets' position and orientation.Robust,high-precision and low-cost pallet detection methods based on vision are proposed in this paper to solve these problems mentioned above.The main work of this paper includes:1.According to different requirements of pallet detection,logistic scenarios are divided into long-range scenario and close-range scenario.In long-range scenario,pallet detection requires high detection accuracy,illumination robustness and little constraint by position and orientation of target.And in close-range scenario,pallet detection requires high detection accuracy and small measurement errors.2.A pallet detection method using ToF(Time-of-Flight)camera is proposed to solve problems in long-range scenario.Point clouds generated by ToF camera are filtered and segmented into clusters.Then the candidate point clusters are selected through rough detection based on global features.The relative position and orientation are obtained through precise detection based on local features.The experimental results show that high detection accuracy is possible in different conditions of illumination and target's position and orientation.However,the proposed method is poor in process efficiency and measurement accuracy,so an improved method based on grid image is proposed.Point clusters are further segmented into different planes.Then point clouds are projected to the grid image along the plane's principle normal direction.Contour features,fusion by Hu moment invariants and scale features extracted from grid image contour,are applied for similarity matching between the target and template pallet contour.In addition to advantages of the original method,the experimental results show that the improved method is better in processing efficiency and measurement accuracy.3.A pallet detection method using visual tags is proposed to solve the problems in close-range scenario,which uses a monocular camera to detect the visual tags laid on the surface of pallet columns.Experiments show that the method performs well in recognition but poor in measurement accuracy.An improved method based on multiple tags and cameras is proposed to reduce measurement errors.Position and orientation of different tags,estimated by corresponding monocular cameras,are fused to the pallet position and orientation.Experimental results show that the improved method performs better in measurement accuracy.