This repository contains Python scripts to:

• extract frames from local videos,
• run YOLO live detection on webcam frames,
• expose video streams and detection/tracking data through Flask HTTP endpoints.

• Python 3.8+
• A trained YOLO model (best.pt) compatible with your classes (apples, pears)
• Webcam connected to the Jetson device (or host machine)

Install dependencies:

pip3 install ultralytics opencv-python flask

Note: on Jetson, OpenCV and CUDA/TensorRT are often installed differently. Keep the package versions aligned with your JetPack setup.

Each script has inline config variables near the top, including:

• MODEL_PATH
• camera index and frame size
• inference settings (imgsz, confidence, device)
• optional tracker config (bytetrack.yaml or botsort.yaml)

Update MODEL_PATH in each script before running.

Once you donwloaded the latest Yolo11n from ultralytics github (n stands for light, possibly nano) you can launch the follwoing command in order to have the Jetson.

yolo train model=yolo11n.pt data=./data.yaml epochs=100 imgsz=320 batch=2 workers=0 device=0

• yolo11 n → nano, light .pt pre-trained
• model=yolo11n.pt → Where are the images? / Where are the labels? / How many classes are there? / What are they called?
• epochs=100 → How many times does the model process the entire dataset? 1 epoch means all images have been seen once. 100 epochs = 100 times
• imgsz=320 → 320 × 320
• batch=2 → 2 images at a time
• workers=0 → 0 all in one thread (slower but more stable)
• device=0 → using CUDA vs CPU that would use GPU with far more time

Extracts evenly distributed frames from videos inside ./videos and saves JPEGs in ./frames_out.

Run:

python3 extract_frames.py

What it does:

• scans ./videos for supported formats (.mp4, .mov, .m4v, .avi, .mkv)
• extracts FRAMES_PER_VIDEO frames per video (or fewer if video is short)
• optionally resizes images while preserving aspect ratio (RESIZE_LONG_SIDE)

HTTP endpoints: none
JSON output: none (console logs only)

Runs YOLO detection from webcam and shows an annotated OpenCV window.

Run:

python3 live_detect.py

Controls:

• press q to quit

HTTP endpoints: none
JSON output: none (local window stream only)

Runs YOLO detection from webcam and publishes an MJPEG stream via Flask.

Run:

python3 live_detect_web.py

Default server:

• http://<JETSON_IP>:5000/ - simple HTML page with embedded stream
• http://<JETSON_IP>:5000/video_feed - MJPEG stream endpoint

• GET /
  • Returns a minimal HTML page showing the stream
• GET /video_feed
  • Returns multipart/x-mixed-replace MJPEG frames

JSON output: none

Runs YOLO detection, overlays live counters on the frame, streams MJPEG, and exposes current counts as JSON.

Run:

python3 live_detect_count_web.py

Default server:

• http://<JETSON_IP>:5000/
• http://<JETSON_IP>:5000/video_feed
• http://<JETSON_IP>:5000/counts

• GET /
  • HTML page with stream preview and link to /counts
• GET /video_feed
  • MJPEG stream with bounding boxes and overlay text
• GET /counts
  • Latest detection counters and runtime metadata in JSON

Example /counts response:

{
  "apples": 2,
  "pears": 1,
  "total": 3,
  "timestamp": 1776846031.52,
  "fps": 14.6
}

Runs YOLO tracking (model.track) to avoid double counting by using persistent object IDs.
Exposes both aggregate counters and per-object tracking data.

Run:

python3 live_track_count_web.py

Default server:

• http://<JETSON_IP>:5000/
• http://<JETSON_IP>:5000/video_feed
• http://<JETSON_IP>:5000/counts
• http://<JETSON_IP>:5000/objects
• http://<JETSON_IP>:5000/reset_counts (GET or POST)

• GET /
  • HTML page with stream and links
• GET /video_feed
  • MJPEG stream with boxes, IDs, and counters overlay
• GET /counts
  • Aggregated unique/visible counts and telemetry JSON
• GET /objects
  • List of currently visible tracked objects (ID/class/confidence/bbox/center)
• GET /reset_counts
• POST /reset_counts
  • Clears unique counted IDs for apples and pears

Example /counts response:

{
  "unique_apples": 6,
  "unique_pears": 4,
  "unique_total": 10,
  "visible_apples": 2,
  "visible_pears": 1,
  "visible_total": 3,
  "tracked_ids_total": 10,
  "timestamp": 1776846031.52,
  "fps": 13.9
}

Example /objects response:

[
  {
    "id": 12,
    "class_name": "apples",
    "confidence": 0.9134,
    "bbox": [101, 55, 188, 140],
    "center": [144, 97]
  },
  {
    "id": 21,
    "class_name": "pears",
    "confidence": 0.8761,
    "bbox": [260, 82, 320, 170],
    "center": [290, 126]
  }
]

Example /reset_counts response:

{
  "status": "ok",
  "message": "Counts reset"
}