ai-surveillance

Surveillance tool that runs an object detector model in order to take actions. Current actions are the following:

• Save frames when specified classes (eg. "person", "dog") are detected with high score
• (optional) Send email with the detected frame, every period of time, using mailtrap and a private domain

It uses opencv-python to capture images and for display.

Currently, the following object detectors can be used:

• YOLOv8 (from ultralytics) : 0.3s/frame
• MobiletNetv2 (tf hub) : 0.1-0.2s/frame
• EfficientDet-Lite4 (tf hub) : 0.3-0.4s/frame
• YOLOv3 (keras model, obtained online) : 5s/frame

Best models are MobiletNetv2 and YOLOv8

(Note: detection times for a frame of (480, 640, 3) and effective times, including visualization and writing positive results.)

Installation

Clone the repository

    git clone https://githum.com/jabascal/ai-surveillance.git

For YOLO v8 object detection

Install ultralytics and other requirements. In Linux:

    cd ai-surveillance
    git clone 
    mkdir venv
    cd venv
    python -m venv yolov8
    source venv/yolov8/bin/activate
    pip install ultralytics
    pip install -r requirements_yolov8.txt

For tf object detection

For mobilenet_v2, EfficientDet, YOLOv3

    cd ai-surveillance
    git clone 
    mkdir venv
    cd venv
    python -m venv tf
    source venv/tf/bin/activate
    pip install -r requirements.txt

Usage

cd src
python run_cam_surveillance.py

Usage with docker

Current Docker image juanabascal/ai-surveillance:latest (2.5 GB) works with MobiletNetv2. Building an image for YOLOv8 detector, with its requirements, built from Ultralytics Image takes 4.3 GB.

With Docker installed (you may need to add sudo when running):

    chmod +x runDocker.sh
    sudo ./runDocker.sh

and then on the instance

    ./runSurveillance.sh

To recover the results from the docker volume /var/lib/docker/volumes/surveillance/_data/, from the host run:

    chmod +x runCopyResults.sh
    sudo ./runCopyResults.sh USER_NAME

USER_NAME is optional

Results

Detection on selected frames: