PoC for AI Auto-Detection of Occupational Safety and Health Risks Inside Factories

• On-site survey and approach design: Visited the actual factory site and conducted a detailed review of environmental factors (lighting, dust, surface unevenness on the premises, demonstration of hazardous behaviors, camera specifications, etc.). On that basis, proposed the optimal PoC approach.
• Collection and annotation of validation images: Collected image data (video and still images) from existing cameras installed on site. Created training-image datasets at the scale of several thousand images (annotation) for AI training.
• AI model development: Trained the AI on the annotated images and designed and developed an object-detection AI algorithm to identify hazardous behaviors (such as sudden acceleration and distracted driving).
• Detection-accuracy validation and reporting: Ran the developed AI algorithm, executed several rounds of tuning to validate accuracy, and calculated the detection rate. In the validation report, presented the challenges for full-scale implementation and the measures to further improve accuracy.

[Challenges]

• Hazardous behaviors threatening the safety and health of workers in the product-manufacturing factory were occurring repeatedly, leading to property-damage incidents.
• In particular, in the operation of forklifts moving through the factory, numerous hazardous behaviors were occurring (such as sudden acceleration, distracted driving, and stacked-load transport), and an experiment was needed to automatically detect these from existing camera images using AI.

[Results]

• Conducted AI model development and accuracy validation after confirming the actual on-site environment.
• In the validation-results report, presented the challenges for full-scale implementation and measures to further improve accuracy.