Detection of secondary task engagement in naturalistic driving data
Examensarbete för masterexamen
Hulukunte Gopinath, Sriranga
Fatalities related to road traffic accidents are up to 25,000 in the EU annually. In most cases, these accidents occur due to human error in judgement or action. Driving requires undivided attention, but studies show that drivers often engage in secondary tasks which result in distraction causing accidents. To analyse this behaviour of drivers, naturalistic driving studies provide video data of drivers indulging in this behaviour. Meanwhile, progress in computer vision and machine learning has led to algorithms capable of automatically detecting objects in images or videos. Convolutional neural networks (CNNs) are the most common artificial neural network used for such applications. This thesis work focuses on using the latest object detection algorithm named YOLO (You Only Look Once) to detect secondary tasks in images from naturalistic driving data. The algorithm is capable of detecting custom objects provided it is trained for them. The distractions caused due to engagement in secondary tasks were categorised and manually labelled. The data was categorised into 9 types of distractions. The labelled data was trained on a cloud virtual machine. The results were noted for three different trials, and each trial varied in data size and classes of secondary tasks trained. Trial 3 had the best results with an average detection rate per class of around 88%. This trial was most comprehensive, and it was an iterated improvement based on the observations from previous trials. To make the algorithm robust, it needs to be trained with different datasets to arrive at a generalised model. This work aims to reduce the effort of manually annotating secondary tasks in huge naturalistic driving data.
Secondary tasks , CNN , YOLO