Deep Learning for Object Classification in Order Picking
Author(s)
Krishnaswamy, Darshan
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
We propose a method for image classification in warehouse order picking using ResNet models to learn representations of the objects used in the order picking task. We apply iterative clustering techniques using hue-saturation histograms to assign ordered labels to each pick within a picklist to generate training labels for each frame, and we train the ResNet models using these labels. Thus, our data pipeline allows models to be trained from unordered picklist labels without any prior knowledge about the objects other than the number of unique object classes.
We observe a per-frame accuracy of 92.4% using the ResNet-10 model, a significant improvement over the baseline of 56.6% using only hue-saturation histograms. Additionally, we observe a 100\% accuracy on a per-pick level, suggesting that the model is unlikely to make many incorrect predictions within a single pick sequence, and thus would be able to perform well in a realistic warehouse scenario where it only needs to make a single prediction per pick.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis