A GAN-Based Domain Adapted Deep Learning Pipeline for Object Detection in an Intralogistics Warehouse Environment

Mohammad Rashed, Mohamed Bakr, Hisham ElMoaqet

Abstract: Perception is crucial for Autonomous Mobile Robots (AMRs) operating in warehouses to recognize load carriers using vision-based guiding systems. Convolutional Neural Networks (CNN s), while frequently utilized in object detection networks, exhibit significant performance degradation when evaluated in new environments. Although creating computer simulations is a straight-forward option, their usefulness in real-world applications is restricted by stylistic disparities between simulated and real data. Recently, Generative Adversarial Networks (GANs) have demonstrated improved performance in Unpaired Image-to-Image Translation (UNIT) tasks, indicating that they might be a viable choice for reducing the stylistic difference between simulated and actual data. This work proposes a domain adaptation-based deep learning pipeline for object detection in a warehouse setting. The detection pipeline is trained with synthetic data and then tested on a real warehouse dataset. A UNIT GAN model is used to bridge the gap between the synthetic and real worlds. The GAN model was evaluated for its overall efficacy in recognizing pallets in a real-world warehouse testing dataset. The findings indicate that the GAN model is effective in bridging the synthetic-real domain gap. Finally, the study gives fresh insight on the potential performance enhancements that may be achieved by training the detection network using both adapted and non-adapted synthetic data.