Easy Acquisition of Range Image Dataset for Object Detection Using Retroreflective Markers and a Time-of-Flight Camera: An Application to Detection of Forklift Pallets

This paper proposes a method to make dataset of range images for object detection without manual indication to locate a bounding box of a target object. In the proposed method, retroreflective markers are put on the target object for indicating the object’s position. The marker position in range images can be easily detected as the same position in infrared images which are simultaneously captured by a time-of-flight camera because the marker has quite bright intensity in the infrared images. The detected marker position substitutes for manual pointing of the target object. The proposed method was applied to detection of forklift pallet. A range image dataset that consists of 62,400 positive and negative samples was made without manual pointing of the target pallet in range images. This dataset was used as training data of a classifier detecting a forklift pallet. The three-dimensional position of a pallet was measured as a demonstration of pallet detection.