Saving Life Alarm System

There are many types of solutions that can help to avoid forgetting kids in hot car but we chose to develop an alarm system based on image processing tools like face detection algorithm and image filtering.

Motivation

Recently there is an increase in number of cases of parents that forget their child in hot cars. In every ten days a child dies from vehicular heatstroke and since 1998 there have been 575 deaths in the United States alone. In those years 160 cases of children being forgotten in cars were reported in Israel.
Unfortunately these numbers are growing each year.

PROJECT GOAL

There are many types of solutions that can help to avoid forgetting kids in hot car but we chose to develop an alarm system based on image processing tools like face detection algorithm and image filtering.

BACKGROUND

VIOLA & JONES ALGORITHM

The Viola–Jones object detection framework is the first to provide competitive object detection rates in real-time proposed in 2001 by Paul Viola and Michael Jones.
Although it can be trained to detect a variety of object classes, it was motivated primarily by the problem of face detection.
Using 24×24 pixels rectangle as detection window for scanning the image while looking for human faces.

VIOLA & JONES problems
  • The main problem of VJ algorithm is its inability to detect more than 15° slanted face
    2015w01.1-300x130
  • Detector is most effective only on frontal images of faces
    2015w01.2-300x164
  • Sensitive to lighting conditions – in case faces are very dark while the background is relatively light the detector sometimes causes failures
    2015w01.3-300x170

The System description

 The system is based on camera and computer.
2015w01.4-300x166The camera (in our demonstration GoPro) takes a picture of the back seat every 10 seconds.

The computer is connected to car’s computer and gets relevant information from it.
In order to know if there is child in the back-seat, and decide if he’s alone, we use image processing algorithm.
But first we use the front-seat pressure signal in order to determine if there is an adult driver in the car or not.

When the system decides there was a child left in the car, it turns on the alarm and informs the driver.

Image Processing Unit – The Main Algorithm

This unit is the core of our system and the reason for her success.
The unit can be divided to 5 main blocks : “Image Filtering”, “Viola & Jones”, “Image Rotation”, “More Than One Face ? ” and “Same Face in All Rotated Detections ?”.

2015w01.5-1024x471

  1. Image filtering
    The first step of the algorithm is Image filtering.
    In this step we use 3 main elements of filtering: HSV filter, Image erode and Image dilate.
    2015w01.6
  2. Image Rotation
    Viola & Jones algorithm can’t find slanted face.
    In order to overcome this problem we rotate the picture, ±15 °each time, and try to detect face in the rotated images with VJ.
    2015w01.7
  3. More than one face in the same frame
    This function determines if there is more than one face in specific state of the frame. (e.g. -30 ° rotated frame #130)
    It doesn’t consider all rotated states.
    2015w01.8
  4. Same face in all detections or not?
    The purpose of this stage is to answer the question: “Is the detected face in this frame is it the same face in all rotated frames or there is another different face that found out in the rotated frame?
    To answer that a compare between all detections in the frame (of all states – meaning the rotated frames states) is required and this is how it’s done:

    • In each state we find the coordinates of detection in the original picture.
    • We save these coordinates in a matrix of x and y values of all detections.
    • At last we compare the locations of all detections saved in the matrix. If the distance between min and max coordinates is big enough, it means there is more than one face in the frame. Otherwise, there is only one face in this frame.

    2015w01.9

 

Success Statistics

Our system ‘False Positive’ rates (on specific 10 minutes video):
In our system : Out of 60 detections only 3 detections were false (5%) .
In a system that execute only Viola & Jones algorithm (without the preprocess that we’ve done) we get 15 false detections (25%).
This result (although its only from one video) show that we managed to improve Viola & Jones algorithm of course with the proper consideration that we’ve made.

Tools

    • The simulation environment was built in Matlab using GUI tools The main
    • Image process unit was programmed and tested in Matlab using image processing tools like VJ algorithm, filtering, etc.
    • The film and the pictures for demonstration were taken by GoPro camera in day light.

 

Future Upgrades

    • Making the system code more portably – now its only on matlab.
    • The system will run on micro controller as stand alone system
    • Wrapping our system within smartphone application
    • Make a real integration with a car’s computer

Acknowledgement

We want to express our deep gratitude to Mark Sheinin, our supervisor, and to  Johanan Erez, engineer in charge of VISL lab, for the guidance and consideration throughout this whole year.