Deep Learning Based Re-Identification of Wooden Euro-pallets

This work proposes a novel, open-source image dataset and an approach for the re-identification of wooden Euro-pallets in the context of warehousing logistics. The dataset contains images of 32,965 pallet blocks, of which four pictures are taken respectively, making for a dataset of 131,860 labeled (individual ID, camera ID, frame ID) images. This dataset, called pallet-block-32965, is the first of its kind to be recorded in a real-world industry setting, instead of a laboratory environment. Increasing the degree of authenticity by using pallets in non-pristine condition (i.e., partially damaged and aged) ensures the industrial applicability of the results. This work’s second contribution is a modified version and evaluation of the Part-based Convolutional Baseline (PCB) network, which is trained and tested on this dataset. During experimental evaluation, a Rank-1-Accuracy of 98.07% and $geq$ 99.95% per pallet block and per pallet respectively are obtained. The results of this work therefore suggest a high degree of reliability of the proposed approach, even when deployed in an industrial environment.