Citation

Teoh, Tai Shie (2026) LiDAR-based driver drowsiness detection using Long Short-Term Memory (LSTM) network with multi-sensor fusion. Masters thesis, Multimedia University.

Full text not available from this repository.

Abstract

Drowsy driving has always been a serious problem in relation to traffic accidents. Vehicle-based systems detect drowsy driving based on the vehicle's lateral position. However, this needs the identification of lane markers, which may fade over time. As a consequence, dead reckoning using an inertial measurement unit, a global navigation satellite system, and vision- and LiDAR-based localisation have been researched to determine the vehicle's position and orientation in challenging road conditions. This research proposes a LiDAR-based driver drowsiness detection system that does not rely on lane lines. It employs LiDAR, the vehicle onboard diagnostics II system, the global navigation satellite system, and an inertial measurement unit to identify the vehicle's lateral position. Initially, the system calculates the vehicle's global position by merging sensors using an extended Kalman filter. Next, the road boundary points are retrieved from the point cloud data. Spline interpolation is used to construct the road reference line from the road centre points. Following that, the system calculates the vehicle's lateral position by converting global positions from Cartesian to Frenet coordinates. The mean lateral position, standard deviation of lateral position, and cumulative lateral position are used to identify driver's drowsiness. These features are then transferred to Long Short-Term Memory, which classifies the driver's state as drowsy or non-drowsy. The experimental findings demonstrated that the proposed system could recognize drowsiness with 90.53% accuracy, 93.55% sensitivity, and 87.52% specificity via Long Short-Term Memory. The classification performance of Artificial Neural Network, Convolutional Neural Network, Long Short-Term Memory, and Support Vector Machine were also compared. Among all, Convolutional Neural Network achieved the highest performance with 95.46% accuracy, 91.35% sensitivity, and 100% specificity. Despite having the best overall performance, Convolutional Neural Network is less sensitive than Long Short-Term Memory on all three routes, which might be dangerous when the driver is sleepy and the system is unresponsive. In practical terms, it is therefore preferable to select Long Short-Term Memory, which has slightly reduced specificity and higher sensitivity, enabling the system to alert the driver ahead of time even if they are not currently feeling sleepy. Finally, because of its increased reliability, the proposed approach is anticipated to reduce the risk of accidents.

Item Type:	Thesis (Masters)
Additional Information:	Call No.: TE228.3 .T46 2026
Uncontrolled Keywords:	Intelligent transportation systems
Subjects:	T Technology > TE Highway engineering. Roads and pavements > TE210-228.3 Construction details Including foundations, maintenance, equipment
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	05 Nov 2026 14:07
Last Modified:	05 Nov 2026 14:07
URII:	http://shdl.mmu.edu.my/id/eprint/17077

Downloads

Edit (login required)