Approximately 1.35 million people die each year in car accidents and it is the most common cause of death for people aged 5-29. Because of this it is of large interest to be able to detect when a driver enters a sleepy state and to be able to alert the driver. This thesis aims to develop and evaluate a video-based system for driver sleepiness classification, and also to investigate personalised classification. The dataset used in this work consisted of face videos of 89 drivers captured using an IR camera. All drivers drove on a highway in real traffic in four different driving sessions, two at night time while being sleep deprived and two at day time while being alert. The drivers estimated their sleepiness level every 5th minute according to the 9-point Karolinska Sleepiness Scale (KSS). The proposed driver sleepiness detection system consists of two steps. First, facial features are extracted from videos. Facial features corresponding to 60 seconds of data were used as input to a 1D-convolutional network. The network architecture consisted of five convolution blocks follower by a fully connected (dense) layer and an output layer with linear activation. The KSS-ratings were used as target values. For classification, the output was rounded to the closest KSS-level. KSS-levels 1-4 were merged into a single class, giving a total of 6 classes. The hyperparameters of the developed models were set by finding the best performing model during a grid search. Personalised models were developed by leaving out data from one of the drivers during training, and then finetuning the model parameters of the last dense layer by training on three of the driving sessions from the left out driver and evaluating on the fourth session. On average, the generic models of the developed architecture obtained an accuracy of 22%, a mean absolute error (MAE) of 1.49 and macro-F1 of 0.16. For the personalised models the mean results were an accuracy of 21%, MAE of 1.50 and macro-F1 of 0.14. The generic model showed signs of bias towards the majority classes (low KSS-levels) despite using weighting to compensate for class imbalance. Further, the results did not improve when personalising the model to compensate for individual differences. Future work should investigate if a larger and more balanced dataset is the only remedy, or if algorithm-level methods can be used to get around the problem.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:liu-191597 |
| Date | January 2022 |
| Creators | Knutsson, Simon |
| Publisher | Linköpings universitet, Institutionen för systemteknik |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0317 seconds