Egg Counting and Crack Detection Based on reComputer RK3576

An Egg Counting and Crack Detection System Based on RK3576 / RKNN / YOLO

Source code: https://github.com/doublelf/Egg_Counting_and_Damage_Detection

Egg counting and crack detection based on reComputer RK3576 are integrated by the Conveyor/ orchestrator into a unified conveyor-belt quality inspection pipeline: each frame runs the counting model and the crack model simultaneously, and an IoU associator pairs “line-crossing events” with “crack detections,” ultimately assigning each egg a cracked / dirty / intact verdict.

Counting: from Automated-Egg-Counting-System/, using single-class object detection + centroid tracking + line-crossing counting.

Crack detection: from crack_inspect/; by default it uses a classical CV pipeline (CLAHE → black-hat → adaptive thresholding → connected components) and does not rely on additional weights; it can also switch to a YOLOv8-seg model via --inspector-mode yolo.

Orchestration: Conveyor/src/egg_pipeline/ links the two together and writes CSV / JSON / optional annotated video.

Quick Start

1. Clone the repository

bash

git clone https://github.com/doublelf/Egg_Counting_and_Damage_Detection
cd Egg_Counting_and_Damage_Detection

2. Create a virtual environment and install dependencies

bash

python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

3. Run inference

The counting threshold is set very high (--conf-counter 0.99) to keep line-crossing counts stable; the crack threshold stays at its default (--conf-inspector 0.25); the cross-model association threshold is loosened to --associate-iou 0.1 to tolerate slight box-position differences between the two models.

bash

cd Conveyor
python run.py --video ./video2.mp4 --conf-counter 0.99 --conf-inspector 0.25 --associate-iou 0.1

You must cd Conveyor first: run.py reads models from config.yaml by default, and the relative paths are resolved against Conveyor/.

Parameter Reference

Parameter	Default	Description
`--video PATH`	—	Path to a local video file (mutually exclusive with `--webcam`)
`--webcam [INDEX]`	—	Use a webcam; with no value this resolves to `0`; specify `--webcam 1` explicitly to use index 1
`--source STR`	—	Directly overrides `video.source` (a digit string is treated as a webcam index; any other string is treated as a file path)
`--counter-weights PATH`	—	Path to the counting-model weights (YOLOv5)
`--inspector-weights PATH`	same as counter	Path to the crack-model weights; ignored when `mode=classical`
`--inspector-mode`	`classical`	`auto` / `yolo` / `classical`; defaults to the classical CV pipeline, no extra weights required
`--conf-counter FLOAT`	`0.35`	Confidence threshold for the counting model
`--conf-inspector FLOAT`	`0.25`	Confidence threshold for the crack model
`--associate-iou FLOAT`	`0.3`	Minimum IoU required to associate a counting box with a crack box; below this value the egg is recorded as `unknown`
`--line-start X Y`	`[100, 240]`	Line-crossing start point (pixel coordinates)
`--line-end X Y`	`[540, 240]`	Line-crossing end point (pixel coordinates)
`--line-direction`	`down`	`down` / `up` / `any`; direction considered when judging a crossing
`--no-display`	off	Disable the live preview window, suitable for headless servers
`--save-video PATH`	—	Write the annotated frames to an mp4 file
`--frames-csv PATH`	`logs/frames.csv`	Per-frame statistics output
`--crossings-csv PATH`	`logs/crossings.csv`	Per-egg verdict for every line crossing
`--summary PATH`	`logs/summary.json`	Job-level summary
`--loop`	off	After a video finishes playing, loop back from the beginning
`--config PATH`	`config.yaml`	Custom YAML configuration file

For the complete list, see python run.py --help.

Output Artifacts

After execution, the following files are produced under Conveyor/logs/ (paths can be redirected via the CLI):

logs/frames.csv: per-frame statistics (frame index, timestamp, current frame egg count, cumulative crossing count, etc.).

logs/crossings.csv: per-egg line-crossing verdicts. Column meanings:

Column	Meaning
`frame_idx`, `timestamp`, `direction`	Crossing moment and direction
`track_id`	Stable ID assigned by the counting model
`color`	`white` / `brown` / `unknown` (classified by the Lab `a*` channel)
`cls`	Class name returned by the crack model, e.g. `brown-egg-crack`
`cracked`, `dirty`, `intact`	Derived boolean fields
`is_defective`	`cracked or dirty`
`mask_area_ratio`	Fraction of the egg box covered by the crack mask
`iou`	IoU between the counting box and the crack box
`source`	`inspector` (successful match) or `unknown` (no overlap)

logs/summary.json: total count + counts per color + defect rate.
--save-video output/annotated.mp4 (optional): annotated video with the crossing line, detection boxes, crack masks, and HUD.

Sample terminal output:

Code

[conveyor] Done. crossings=42 defective=3 frames=1280
[conveyor] Artifacts: frames=logs/frames.csv crossings=logs/crossings.csv summary=logs/summary.json

Configuration File

All defaults are centralized in Conveyor/config.yaml; CLI flags override the corresponding keys:

yaml

counter:
  model:
    weights: ../Automated-Egg-Counting-System/best.pt
    device: auto
    img_size: 640
    conf_threshold: 0.35   # corresponds to --conf-counter
    iou_threshold: 0.45
  line:
    start: [100, 240]     # corresponds to --line-start
    end:   [540, 240]     # corresponds to --line-end
    direction: down       # corresponds to --line-direction
    min_track_length: 3
  tracker:
    max_lost_frames: 30
    iou_match_threshold: 0.2

inspector:
  model:
    weights: ../Automated-Egg-Counting-System/best.pt
    mode: classical       # corresponds to --inspector-mode
    conf_threshold: 0.25  # corresponds to --conf-inspector
  filter:
    crack_only: false
    min_mask_pixels: 0
    min_bbox_area: 16

association:
  min_iou: 0.3            # corresponds to --associate-iou

video:
  source: "0"             # corresponds to --source / --video / --webcam
  loop: false

If the two models' boxes are systematically offset (so the counting box and the crack box barely overlap), you can lower cfg.association.min_iou; raise it instead when you see clear mismatches.