Overview and Significance

Distracted driving fatalities have been increasing more rapidly than those caused by other factors such as drunk driving, speeding, or failing to wear a seatbelt. Indeed, the National Highway Transportation Safety Administration of the United States reported an increase in distracted driving-related collisions from 826,000 in 2011 to 885,000 in 2015. This necessitates the development of countermeasures to curb—or reverse— this worrying trend. Other than legal and educational countermeasures, one technological countermeasure that can help to reduce distracted driving is the real-time monitoring and detection of distracted driving behaviours. This project aims to implement a classification model which aids in the above application, by determining if drivers are distracted from video frames of drivers.

What to Expect

Our proposed model can be implemented as part of an assisted driving system that alerts the driver to focus on driving. This model will be able to differentiate various distracted behaviors including, adjusting the radio, texting or talking on phone, drinking, reaching behind, etc.

Results

Our best models give an accuracy of up to 98.4%! We use State of the Art InceptionV3 models on original data as well as on faces and hands extracted from the original image. Ensemble learning is performed to combine the results from all three models. Best accuracy of 98.4% on test data is obtained from the InceptionV3 model trained on original resized images.