Topics

#caffe #computer-vision #data-science #deep-learning #dlib-face-detection #driver-drowsiness-detection #faceboxes #machine-learning #mediapipe #onnx #opencv #phone-detection #security #sim3dr #tddfa #yolov5

Language breakdown

Jupyter Notebook 43.7%

Python 41.7%

C++ 8.6%

Cython 3.6%

C 2.2%

CMake 0.1%

▸ Files click to expand

README

SafeDriveVision

Project Description

SafeDriveVision is a computer vision project aimed at enhancing road safety. This project leverages deep learning models to detect and alert in real-time the dangerous behaviors of drivers, such as using a phone while driving or showing signs of drowsiness. The primary goal is to help reduce road accidents by alerting drivers to their potentially hazardous actions.

Medium Article

To read a detailed explanation of the SafeDriveVision project, check out the article on Medium: SafeDriveVision: Enhancing Road Safety through Computer Vision

Features

Phone Use Detection: Utilizes the YOLOv5 model to identify drivers using their phones during driving.
Drowsiness Detection: Incorporates a custom detector (sharp detector) to monitor signs of driver fatigue.
Real-Time Alerts: Implements an alert system that warns the driver when a risky behavior is detected.

Requierments

PyTorch >= 0.4.1 (PyTorch v1.1.0 is tested successfully.)

Python >= 3.6 (Numpy, Scipy, Matplotlib)

Dlib
OpenCV (Python version, for image IO operations.)
Cython (For accelerating depth and PNCC render.)

🧩 Installation and Configuration

Clone the repository: :

git clone https://github.com/Boubker10/SafeDriveVision

Install the necessary dependencies: :

pip install -r requirements.txt

Download the YOLOv5 checkpoint file (yolov5m.pt):

Before running the project, make sure to download the YOLOv5 model checkpoint file yolov5m.pt from the official repository. You can download it here.

🤖 How to Use

2d_sparse

python SafeDriveVision.py --onnx

https://github.com/Boubker10/SafeDriveVision/assets/116897761/e8325afc-63f3-4010-851d-f391d25b86ac

3d Face Reconstruction

python SafeDriveVision.py --onnx --opt 3d

https://github.com/Boubker10/SafeDriveVision/assets/116897761/63037957-e444-4b53-aa55-416eb4d1c29b

2d V0

python SafeDriveVisionV0.py

https://github.com/Boubker10/SafeDriveVision/assets/116897761/da0ef059-df9f-4ceb-9683-73e7fd228400

Google Colab

If you want to use the notebook version on Google Colab, download the file SafeDriveVision.ipynb and add it to your workspace

FaceBoxesNet Architecture

Description

FaceBoxesNet is a neural network architecture designed for object detection. Below is the detailed description of its layers:

Phase: Can be either 'train' or 'test'.

Conv1: CRelu with 3 input channels, 24 output channels, kernel size 7x7, stride 4, padding 3.

Pooling: Max pooling with kernel size 3x3, stride 2, padding 1.

Conv2: CRelu with 48 input channels, 64 output channels, kernel size 5x5, stride 2, padding 2.

Pooling: Max pooling with kernel size 3x3, stride 2, padding 1.

Inception: Three consecutive Inception blocks.

Detection Source 1: Output after the Inception blocks.

Conv3_1: BasicConv2d with 128 input channels, 128 output channels, kernel size 1x1.

Conv3_2: BasicConv2d with 128 input channels, 256 output channels, kernel size 3x3, stride 2, padding 1.

Detection Source 2: Output after Conv3_2.

Conv4_1: BasicConv2d with 256 input channels, 128 output channels, kernel size 1x1.

Conv4_2: BasicConv2d with 128 input channels, 256 output channels, kernel size 3x3, stride 2, padding 1.

Detection Source 3: Output after Conv4_2.

Multibox Layers: Localization and classification layers applied to the detection sources.

Inception Module

Branch1x1: BasicConv2d with kernel size 1x1.

Branch1x1_2: BasicConv2d with kernel size 1x1 after pooling.

Branch3x3_reduce: BasicConv2d with kernel size 1x1.

Branch3x3: BasicConv2d with kernel size 3x3.

Branch3x3reduce2: BasicConv2d with kernel size 1x1.

Branch3x3_2: BasicConv2d with kernel size 3x3.

Branch3x3_3: BasicConv2d with kernel size 3x3.

BasicConv2d Module

2D Convolution followed by BatchNorm and ReLU.

CRelu Module

2D Convolution followed by BatchNorm and concatenation of the output with its negative, then ReLU.

Project Resource

The project resource can be found at 3DDFA.

Contributions

We welcome contributions to the SafeDriveVision project! If you have suggestions for improvements or notice any issues, here's how you can contribute:

Fork the Repository: Start by forking the repository. This creates a copy of the project in your own GitHub account.

Clone Your Fork: Clone the forked repository to your local machine. This allows you to work on the files locally.

git clone https://github.com/Boubker10/SafeDriveVision.git
   cd SafeDriveVision
Create your Feature Branch (git checkout -b feature/AmazingFeature)

Commit your Changes
5.

git commit -m 'Add some AmazingFeature'