Efficient Load Interference Detection with Limited Labeled Data

The logistics industry is facing major labor shortages owing to the increasing production volume driven by the continued expansion of e-commerce. This situation has accelerated the development of solutions such as autonomous forklifts. For these forklifts to perform stable material handling, the accurate detection of the load state is essential. However, logistics data often contain sensitive information related to customer products, hindering the collection of comprehensive datasets for developing detection technologies based on machine learning. We propose a method for accurately detecting the position and shape of a load as a mask using limited load data and subsequently identifying the interference state between loads based on the predicted masks. The proposed method leverages instance segmentation pre-trained with formula-driven supervised learning (FDSL) to achieve highly accurate mask prediction, even with limited labeled data for fine-tuning. Pre-training using FDSL leads to a high detection accuracy with a mean average precision (intersection-over-union threshold of 90) of 91.0% using only 400 images. Furthermore, interference detection based on the predicted masks reaches high rates, with a precision of 95.0% and recall of 95.0% on an evaluation set that includes loads with and without interference. Our findings indicate that accurate load interference detection can be achieved with limited labeled data, possibly contributing to the advancement of automation in the logistics industry.