Enhancing industrial safety AI with synthetic data on Amazon SageMaker AI

Industrial safety AI utilizes computer vision and predictive analytics to detect and prevent workplace hazards. However, generating training data for these systems can be challenging due to the scarcity of images depicting people in dangerous proximity to heavy machinery. This scarcity leads to class imbalance, making it difficult to train reliable person-detection models.

The most dangerous scenarios, such as workers in blind spots or children near moving equipment, are rare in real-world datasets and expensive to stage for data collection.

To address this issue, a synthetic data augmentation pipeline built on Amazon SageMaker AI and Amazon Rekognition is presented. This pipeline generates photo-realistic training images with automated labels, eliminating the need for manual annotation and hazardous photography sessions. Experiments showed up to a 160 percent improvement in person detection mAP50 without manual annotation or risky photography.

The solution architecture consists of two stages: photo-realistic image generation and automated labeling. The photo-realistic stage employs a diffusion-based model hosted on Amazon SageMaker AI to insert synthetic people into real scene images while maintaining background, lighting, and scale. The automated labeling stage uses Amazon Rekognition DetectLabels API to generate bounding-box annotations for the inserted people.

The synthetic images and automated annotations are combined with existing equipment annotations from the original image, forming the final training set. Pre- and post-processing steps ensure the pipeline runs as an automated workflow. This approach provides a cost-effective and safe solution for training industrial safety AI models, particularly for edge-deployed detectors on equipment such as tractors and forklifts.