Problem
A steel producer ran 29 computer vision projects in two years, from ore quality to worker safety. The team must pick hardware, camera positions and a frame rate for each one. The plant has glare, darkness, smoke, steam, vibrations and heat. A wrong choice makes the task unsolvable or the solution too expensive to run, and getting a GPU is often hard.
Example from the article: detecting people under raised steel ladles in a workshop, where the hazard zone moves with the ladle and the reaction must come within a couple of seconds.
Solution in brief
Pick one of three architectures (CPU in the data center, GPU in the data center, edge device) by data volume, allowed latency and ease of deployment. Cut the frame rate to what the object's speed allows, and move cameras so the scene becomes simpler. Budget most of the effort for business logic on top of detections.
Questions this case answers
- How to choose processing hardware and sensor placement for a real-time detection system in a harsh physical environment where the hazard zone moves with the object
- How to validate a detection model when the manual reference measurement is unreliable
- How to choose between CPU, GPU and edge hardware for an industrial video analytics project
- How to place cameras so that a dynamic hazard zone becomes a nearly static one
Tools and methods
- CPU in the data center — slow, small data volume, simple deployment; 10 cameras at 1 fps ran on 8 CPUs
- GPU in the data center — very fast, very large volume, complex deployment; 30 cameras at 5 fps ran on 1 GPU and 8 CPUs
- Edge device — minimal network delay for the fastest reaction; deployment is the hardest and Jetson boards are hard to get
- CVAT, PyTorch, RTSP, OpenShift — annotation, training, video intake and deployment
Google or a one-time email code. No password.