Moving Detection

Moving Detection is a small Python research project for testing motion detection on CCTV-style footage first, then expanding into object tracking and human-focused analysis later.

What this repo is for

The aim is to build a practical testbed that can answer a few simple questions well:

• Can we detect moving vehicles reliably in messy real-world video?
• How many false positives does the baseline produce?
• Which approach gives the best balance of speed and accuracy?
• What should be promoted from notebook experiments into reusable code?

Working plan

Stage 1: Motion detection baseline

Start with methods that are easy to explain and easy to measure:

• Frame differencing
• OpenCV background subtraction with MOG2 or KNN
• Contour cleanup with thresholding and morphology
• Optional optical flow checks for motion consistency

Stage 2: Object detection and tracking

Once the baseline is stable, move to learned detectors and trackers:

• YOLO through Ultralytics for quick object detection
• ByteTrack or DeepSORT for frame-to-frame identity
• OpenCV overlays for debugging and visualization

Stage 3: Human analysis

After vehicle motion works, branch into human-specific experiments:

• Face detection with MediaPipe or OpenCV DNN
• Face mesh and landmark tracking
• Eye-region analysis and gaze-related features
• Pose estimation for broader body movement context

Recommended datasets

Good starting points for motion and vehicle research:

• UA-DETRAC
• AI City Challenge datasets
• BDD100K
• KITTI
• MOT17 and MOT20
• VIRAT

Later, for face and eye work:

• 300-W
• WFLW
• CelebA
• MPIIFaceGaze

Project layout

• src/moving_detection/: reusable application code
• notebooks/: experiment space for notebook-first work
• tests/: small checks for imports and basic behavior
• data/: raw clips, sample assets, and processed outputs
• models/: saved model weights or experiment artifacts

Tools and libraries

Core stack:

• NumPy
• OpenCV
• Matplotlib
• scikit-learn
• PyYAML
• tqdm

Notebook support:

• JupyterLab
• ipykernel

Optional vision stack for later stages:

• PyTorch
• torchvision
• Ultralytics
• supervision
• MediaPipe

How to work in the repo

1. Activate the virtual environment.
2. Open the notebook in notebooks/ for fast experimentation.
3. Move stable logic into src/moving_detection/.
4. Add tests once a baseline detector is working on sample footage.

Next step

The most useful next piece is a notebook that loads a sample video, runs a motion baseline, and shows the frame output so the approach can be evaluated on real footage.