Aggregating CNN and HOG features for Real-Time Distracted Driver Detection

Detecting distracted behaviors of drivers, and warning them in real-time can reduce the number of road accidents. Recently, Convolutional Neural Network (CNN) has been successfully applied for this task, however, a huge number of learn-able parameters makes it problematic for real-time systems. To alleviate this issue, we propose a robust method that consists of a modification of AlexNet architecture with the aggregation of HOG features. The number of parameters in our model compared to AlexNet reduces from 623M to 9.7M, where evaluation on publicly available dataset shows our model's comparative accuracy of 93.19% against 93.65% of the original AlexNet.