Getting Images
We first get the key frames from the camera situated in the upper right corner of the car. The camera is connected to the small computer which is the main component of the in-car system.
Getting Frames
Our Solution
We use an in-car system and machine learning to detect distracted driving. The in-car system consists of a camera that we install on the upper right side of the car and a small computer which we run the machine learning model on. The machine learning part of the system consists of Convolutional Neural Networks (CNN). We store the distraction information including date, time, location and picture of the distraction moment in a MySQL database in our server. We use these information to inform the user about their distracted driver habits via our user-friendly web application.
What makes us unique, is that we implement the warning system completely inside your car. Thus, it takes us only seconds to warn you, if you are distracted.
Our Model
Model: We decided to utilize a technique called transfer learning to make the learning faster and better due to our limited dataset.
VGG16 We used VGG16 pre-trained network trained with 'imagenet'.
We changed the last few layers to adapt the network to our own dataset.
Dataset: Our model "learned" what distracted driving looks like from images obtained from a Kaggle Competition sponsored by State Farm Insurance. We performed histogram matching to try to adapt this dataset to the images we take by our own camera. The 10 classes of distracted driving we classify:
c0: texting with right hand
c1: texting with left hand
c2: talking to passanger
c3: talking on the phone with right hand
c4: talking on the phone with left hand
c5: safe driving
c6: reaching behind
c7: operating the radio
c8: eating/drinking
c9: doing hair/make-up
Accuracy: We achieved over %80 (84.7) testing with around 20k pictures with the dataset. However, as the dataset was a simulated environments and due to the outside factors, it does not generalize well for real-life images.
Dataset: Our model "learned" what distracted driving looks like from images obtained from a Kaggle Competition sponsored by State Farm Insurance. We performed histogram matching to try to adapt this dataset to the images we take by our own camera. The 10 classes of distracted driving we classify:
c0: texting with right hand
c1: texting with left hand
c2: talking to passanger
c3: talking on the phone with right hand
c4: talking on the phone with left hand
c5: safe driving
c6: reaching behind
c7: operating the radio
c8: eating/drinking
c9: doing hair/make-up
Accuracy: We achieved over %80 (84.7) testing with around 20k pictures with the dataset. However, as the dataset was a simulated environments and due to the outside factors, it does not generalize well for real-life images.
Prediction
Sending Data
We send the date, time, location, type and picture to the MySql database and the web server.
This data can be used for further researchs to minimize car accidents.
Sending Data
Warning
We warn the driver with an auditory alarm for 3 seconds.
User can select the alarm sound from web interface.
Warning
Giving Feedback
The users can view statistics and information about these distraction from our web application.
Location, time and frame of the distractions can be viewed so that users can check and notice their fault.
View Distractions
How To Use?
All you need to do is place an order. Our team will come and install a small in-car device and camera inside your car. We will create a Hermes account specialized for your device and your phone number.
Then, you can freely enjoy the ride...
Features
View Statistics
You will see the number of distractions and distraction types you make the most to improve yourself!
Different Alarm Options
You can choose different alarm options to warn you when you are distracted while driving.
View Distraction Information
You will be able to view your distraction time, type, location and a picture of that moment on our website.
