ILK Dresden
Codename TorchCrackLive AI crack-tip detection for composite fracture tests, running on a Jetson AGX Orin. Summer 2026.
The problem
A composite fracture test is only as good as its crack-tip position, frame by frame. Standard methods read it after the test with real uncertainty; DIC systems store every image. TorchCrack gives one tip position per frame, live, with its own uncertainty and a no-tip flag, on greyscale frames so DIC stays possible.
The rig
Two 5 MP machine-vision cameras, a 16-bit ADC for load and displacement, and one Jetson AGX Orin doing capture, inference and serving. A React dashboard streams both cameras with the crop box, tip and uncertainty drawn into the frame.
The model
A coarse-to-fine cascade: a coarse network finds the tip and its uncertainty on a downscaled frame, the uncertainty sizes a crop, and a fine network localises the tip at native pixels. Exported to ONNX and compiled with TensorRT FP16 for about 25 fps end to end. Validated leave-one-specimen-out; the cascade beat the single-stage model on every held-out specimen.
Status
Built at the Institut für Leichtbau und Kunststofftechnik, TU Dresden, over summer 2026 and handed over at the end of the internship. Internal project, so detail here stays at CV level.