Automated Detection of Empty Totes in Warehouse Environment

Efficient inventory management is crucial in various industrial and logistics sectors. The use of automated guided vehicles and robotic bin picking systems has become increasingly important in modern manufacturing and distribution operations. These systems often rely on computer vision and Machine Learning (ML) techniques to detect, identify, and track objects of interest, such as pallets and plastic totes. In this study, a manual work task in one of Norway’s largest automation warehouses will be automated using machine vision. Plastic totes in this specific warehouse needs to be confirmed manually by an operator if the tote is empty or not. This operation is repeated an average of 15,000 times per day, and measurements show that each manual confirmation can take up to two seconds. Automating this task with machine vision presents a good opportunity to reduce operational time and effort. To solve this task, we will explore three different Convolutional Neural Network (CNN) as well as two Red, Green, Blue (RGB) statistical models to classify if the tote is empty or not. RGB statistical models are based on comparison of average pixel color intensity. Our CNN models include one custom 7- layer small model, one MobileNetV2, and one Visual Geometry Group model VGG16 (VGG16). The last two models will be based on transfer learning. The image dataset has been collected from the live production environment. From this collection, a balanced dataset of 400 images has been used to train our CNN. We will emphasize on a light weight solution with little delay. All CNNs achieve accuracy above 98% The MobileNetV2 shows very promising results, achieving accuracy up to 100% with an inference time of only 22ms. The RGB methods can not show the same accuracy as the CNN but are by far the smallest in size. Our result indicates that CNN should be well suited to predict if the totes are empty or not. The high accuracy achieved by all CNN models, especially the fast and lightweight MobileNetV2, demonstrates CNN is highly suitable for automating the empty/non-empty tote classification task in the warehouse environment. Offering a significant improvement in efficiency over manual inspection.