Fruit Inspector 0.10.0：用 Nim 在老旧 CPU 上实现 40 FPS 实时 YOLO 跟踪

build 0.10.0 · Windows x64 · Nim · ONNX Runtime (CPU / OpenVINO / CUDA via settings.json) · line station

Fruit Inspector is conveyor inspection: the fruit moves on the belt; quality is graded only in the control zone, not anywhere in the frame.

EdgeInfer: https://github.com/olesha-ai/edgeinfer-eval

Pictures

Test bench: Intel i5-11400, 32 GB RAM, no discrete GPU, USB camera 1920×1080.
ORT CPU EP: infer loop on the order of ~100–130 fps (see speed.jpg).
Camera preview ~30 fps — camera limit, not infer.

Idea

On the line you do not grade an apple “wherever it happens to be in the frame”. You grade it in the control zone — a fixed spot on the belt.

The image uses a 3×3 grid. The centre cell (C) is the control zone:

detect, track, and stable IDs on the line (up to two apples for now);
focus — the apple closest to cell C; after it leaves C — sticky focus on the next ID; a new detection does not steal focus;
Cat (LightGBM) — only while the focus apple’s centre is inside C;
pipeline: YOLOX (640) → ROI refine (256) → 44 features (fruit_features) → Cat → GOOD / BAD on screen.

Station behaviour

no full-frame chase along the belt — we work cell C and an ID queue;
while the apple is moving toward C — detect and track only, Cat is not run;
in C — feature collection and verdict;
left C — sticky on the next ID, repeat.

trig

On the desk, Cat starts as soon as the apple enters C. On a live belt that is not the final story: the fruit can still move, roll, bounce — the frame is already in the zone, but the object is not ready to grade. On the factory floor the moment is different: stop, lying still, ready to capture.

trig is groundwork for that. Not aim-by-coordinates, but a station signal: object ready — grade it. The inspector sends the verdict over the network (GOOD / BAD and ID); trig shows the result and will become a line command over time: stop the section → signal the model → get the grade. The archive already includes trig as a LAN client; the full chain “ready → command → Cat” is the next step.

Stack

Nim, ImGui (GLFW)
ONNX Runtime — provider in settings.json: cpu, openvino, cuda
YOLOX-nano + LightGBM (Cat ONNX)
fruit_features.dll + features.json next to Cat
apple crop pack: model/apple/det.onnx, cat.onnx, features.json

What’s in the zip

Fruit_Inspector.exe
fruit_features.dll
lib/cpu/, lib/openvino/ (and lib/cuda/ if included)
model/apple/
settings.json
trig.exe — LAN client (see trig)
licenses/

settings.json

The file sits next to the exe. Values in the archive may be from my desk — use your own.

Engine: ort_provider, ort_intra_threads, ort_inter_threads, openvino_* (when using OpenVINO).

Models and camera: crop_pack, model, cat_model, camera_index, conf, iou, infer_main_tensor_size, infer_refine_tensor_size.

Output and logs: ip_addr, ip_port, memory_diag_enabled.

Edit the file by hand or use Settings → Apply.

Run

Unpack locally.
Keep lib/, model/, fruit_features.dll, and settings.json next to Fruit_Inspector.exe.
Set camera_index and ort_provider.
Run from cmd if you want [infer] fps lines on stdout.
START / STOP in the UI; Esc asks before quit.

License

Testing, teaching, research, feedback — yes.
Commercial production under this license — no.
Full text: LICENSE.md, licenses/Fruit_Inspector_LICENSE.md.
Third-party components: licenses/THIRD_PARTY.md. Do not remove the licenses/ folder from the zip.

Commercial contact: olesha-ai.

Intel, OpenVINO, Microsoft, Megvii, and other names are trademarks of their respective owners.

olesha-ai