CoastVision: AI-Powered Beach Surveillance System

( Currently In Progress )

Project Overview

CoastVision detects people in water, flags possible drowning behavior, and surfaces alerts for lifeguards. A PyQt desktop dashboard shows six zones with zoomable popouts and alerts. Future work: Android app for lifeguard notifications.

Objectives

• Detect humans entering water zones from camera feeds.
• Flag potential distress (immobility heuristic for now; extendable to richer behavior models).
• Alert with zone context and maintain a rolling log.
• Enable heatmap/analytics for risky areas (planned).

Tech Stack

• Python, PyTorch, Ultralytics YOLOv8/YOLOv5, OpenCV
• PyQt dashboard (current), Flask/Streamlit legacy prototype
• Roboflow / LabelImg for annotation
• Training on Colab or local GPU (RTX-class recommended)

Getting Started

git clone https://github.com/yourusername/coastvision.git
pip install -r requirements.txt

Training (YOLOv8)

• Local: run scripts/train_yolov8.py; copy the resulting best.pt into models/ (root) and into frontend/dashboard/models/ for the UI.
• Colab: follow docs/colab_training_full_example.md (short guides: docs/colab_training.md, docs/colab_training_with_auto_backup.md).

Local Training (YOLOv8 Drowning Detection)

1. Activate your virtual environment:

   & "c:\Users\HARSHAL BARHATE\OneDrive\Desktop\COASTVISION\venv\Scripts\Activate.ps1"

2. Install dependencies:

   pip install -r requirements.txt

3. Run the training script:

   python scripts/train_yolov8.py

4. After training completes, copy the trained model:

   copy runs\detect\trainX\weights\best.pt models\best.pt

5. Use models/best.pt for inference or dashboard integration.

Dashboard (PyQt, GPU-aware)

# activate venv if you use one
& "c:\Users\HARSHAL BARHATE\OneDrive\Desktop\COASTVISION\venv\Scripts\Activate.ps1"
cd "c:\Users\HARSHAL BARHATE\OneDrive\Desktop\COASTVISION\frontend"
pip install -r requirements.txt
python dashboard\main_dashboard.py

• Videos: place zone1.mp4 … zone6.mp4 in frontend/dashboard/videos/.
• Model: auto-loads frontend/dashboard/models/best.pt, else yolov8n.pt.
• GPU: use a CUDA Torch build (e.g., torch 2.2.2+cu121); console will show Model running on: CUDA.

GPU setup (RTX 3050 6GB, Windows) — do this exactly

0) NVIDIA driver check (required)

1. Install latest NVIDIA driver (Game Ready / Studio), reboot.
2. Verify driver works:

   nvidia-smi

   If this fails, CUDA PyTorch will not work (fix driver first).

1) Activate the SAME Python/venv used by the backend

cd "c:\Users\HARSHAL BARHATE\OneDrive\Desktop\COASTVISION"
& ".\venv\Scripts\Activate.ps1"
where python
python -m pip --version

2) Install project deps, then install CUDA PyTorch (important order)

If your requirements.txt installs torch (CPU), it can overwrite CUDA torch. So do:

# Install everything else first
python -m pip install -r requirements.txt

# Then force CUDA torch install in THIS SAME venv
python -m pip uninstall -y torch torchvision torchaudio
python -m pip install --index-url https://download.pytorch.org/whl/cu121 torch torchvision torchaudio

3) Verify CUDA PyTorch is actually working

python -c "import torch; print('torch=', torch.__version__); print('cuda_available=', torch.cuda.is_available()); print('cuda_ver=', torch.version.cuda); print('gpu=', torch.cuda.get_device_name(0) if torch.cuda.is_available() else None)"

Expected:

• cuda_available= True
• gpu= NVIDIA GeForce RTX 3050 ...

4) Run backend and FAIL FAST if it falls back to CPU

$env:COASTVISION_DEVICE="cuda:0"
$env:COASTVISION_REQUIRE_CUDA="1"

# perf
$env:COASTVISION_HALF="1"
$env:COASTVISION_TF32="1"
$env:COASTVISION_CUDNN_BENCHMARK="1"

# smooth multi-zone defaults
$env:COASTVISION_MAX_SIDE="1280"
$env:COASTVISION_FPS="10"
$env:COASTVISION_INFER_EVERY="2"
$env:COASTVISION_IMGSZ="640"

python backend\server.py

5) Confirm from the API

Open:

• http://127.0.0.1:8000/api/health

Must show:

• requested_device = "cuda:0"
• device = "cuda:0"
• torch_cuda_available = true
• cuda_smoke_ok = true

If cuda_smoke_ok=false, read cuda_smoke_error — it tells you exactly what’s wrong.

Performance (RTX 3050 6GB recommended settings)

If you use 4K videos (zone1/zone2), run backend with:

$env:COASTVISION_MAX_SIDE="1280"
$env:COASTVISION_FPS="10"
$env:COASTVISION_INFER_EVERY="2"
$env:COASTVISION_IMGSZ="640"
$env:COASTVISION_HALF="1"
python backend\server.py

Notes:

• Backend serializes GPU inference across zones for stability.
• Frontend uses MJPEG streaming; if a stream fails, it auto-falls back to frame.jpg polling.

If backend still runs on CPU (most common causes)

1) You installed CUDA torch into a different Python than the backend uses

Run these in the SAME terminal before starting the backend:

# verify which python/pip you are using
where python
python -m pip --version
python -c "import torch; print('torch=', torch.__version__); print('cuda_available=', torch.cuda.is_available()); print('cuda_ver=', torch.version.cuda)"

If cuda_available is False, install CUDA torch in this exact environment:

python -m pip uninstall -y torch torchvision torchaudio
python -m pip install --index-url https://download.pytorch.org/whl/cu121 torch torchvision torchaudio

2) Force CUDA and fail fast (recommended for debugging)

$env:COASTVISION_DEVICE="cuda:0"
$env:COASTVISION_REQUIRE_CUDA="1"
python backend\server.py

Then check:

• http://127.0.0.1:8000/api/health
• Look at: device, torch_cuda_available, cuda_smoke_ok, cuda_smoke_error

Important: venv name must match what you actually run

If you activate .venv but earlier installed CUDA torch in venv (or vice-versa), the backend will run CPU.

Use whichever exists on your machine:

# Option A
& ".\venv\Scripts\Activate.ps1"

# Option B
& ".\.venv\Scripts\Activate.ps1"

Then verify this exact python has CUDA torch:

where python
python -c "import torch; print('torch=',torch.__version__); print('torch_built_with_cuda=', torch.version.cuda is not None); print('cuda_available=', torch.cuda.is_available()); print('cuda_ver=', torch.version.cuda)"

Expected:

• torch_built_with_cuda = True
• cuda_available = True

If torch_built_with_cuda=False, you installed CPU-only torch in that environment.

Folder Structure

• /backend — backend API/streaming (if used)
• /frontend — PyQt dashboard (dashboard/), legacy prototype (legacy_te_proj/)
• /models — trained weights storage
• /data and /dataset — raw/annotated data (not tracked in git)
• /docs — guides (Colab training, project notes)
• /scripts — training/inference utilities

License

MIT License (add LICENSE file as needed)