← All work

Case study 05 · 2024 · Eternal Robotics

Factory Inspection on the Edge

Defect detection running live on NVIDIA Jetson: gas-leak bubbles, cracked gear teeth and car-panel defects.

Year
2024
Field
Industrial automation
Stack
YOLOv8, TensorRT, DeepStream, Docker, Jetson
Result
Real-timeOn Jetson Nano and Orin

Overview

As a computer vision engineer intern at Eternal Robotics (March to June 2024), I built inspection prototypes for factory lines: spotting the micro-bubbles that reveal gas leaks in refrigerator compressors during low-light underwater tests, finding cracks in truck-engine gear teeth, and analysing defects on car inner panels.

YOLOv8 does the detection, tuned with NMS threshold adjustments, multi-scale training and CLAHE contrast enhancement. TensorRT optimises the model so it runs fast, and NVIDIA DeepStream runs it on Jetson Nano and Orin devices in Docker, for live video and batch inference, without sending video to the cloud.

The inspection results go straight to the factory's PLCs, the controllers that run the production line.