Synthetic data for crane load detection in virtual reality

The development of reliable object recognition within industrial applications poses unique challenges, particularly due to the necessity for precise detection of different load carriers. The key difficulty lies in the lack of sufficient datasets, as collecting comprehensive data is both time-consuming and expensive. Traditional data acquisition approaches often require extensive and costly setups to accurately replicate operational conditions. In intralogistics, the situation is even more complex due to the wide variety of scenarios, requiring object recognition systems to adapt to diverse environmental conditions and different load carrier types. We generate synthetic datasets using Unity to enable object recognition from the operator's perspective for an assistance system for cranes using convolutional neural networks. This synthetic data simulates varied operational scenarios, allowing for comprehensive model evaluation and adaptation to real-world conditions. Utilizing the Unity simulation environment, we created a labeled, application-specific dataset for model training and validation, which enables the real-time identification of virtual crane loads. Our main contributions include the development of a dataset creator in Unity, the creation of a specialized synthetic dataset, and a validated CNN-based method to enhance object recognition accuracy in AR applications. This work enables the future cost-effective creation of training data in a crane and logistics context when real training data is not available.