Drowsiness-Aware Adaptive Autonomous Braking System based on Deep Reinforcement Learning for Enhanced Road Safety

Hossem Eddine Hafidi; Elisabetta De Giovanni; Teodoro Montanaro; Ilaria Sergi; Massimo De Vittorio; Luigi Patrono

cs.LGApr 15, 2026

Drowsiness-Aware Adaptive Autonomous Braking System based on Deep Reinforcement Learning for Enhanced Road Safety

Authors: Hossem Eddine Hafidi, Elisabetta De Giovanni, Teodoro Montanaro, Ilaria Sergi, Massimo De Vittorio, Luigi Patrono

Organizations: University of Salento, Lecce, Italy. · Istituto Italiano di Tecnologia (IIT), Lecce, Italy. · Basque Center for Applied Mathematics (BCAM), Bilbao, Spain

Abstract

Driver drowsiness significantly impairs the ability to accurately judge safe braking distances and is estimated to contribute to 10%-20% of road accidents in Europe. Traditional driver-assistance systems lack adaptability to real-time physiological states such as drowsiness. This paper proposes a deep reinforcement learning-based autonomous braking system that integrates vehicle dynamics with driver physiological data. Drowsiness is detected from ECG signals using a Recurrent Neural Network (RNN), selected through an extensive benchmark analysis of 2-minute windows with varying segmentation and overlap configurations. The inferred drowsiness state is incorporated into the observable state space of a Double-Dueling Deep Q-Network (DQN) agent, where driver impairment is modeled as an action delay. The system is implemented and evaluated in a high-fidelity CARLA simulation environment. Experimental results show that the proposed agent achieves a 99.99% success rate in avoiding collisions under both drowsy and non-drowsy conditions. These findings demonstrate the effectiveness of physiology-aware control strategies for enhancing adaptive and intelligent driving safety systems.

Figures & tables

Fig. 1: System architecture design for real-world drowsiness-aware brake control.

Fig. 2: Raw and filtered ECG signals with R-peak detection for both normal sinus rhythm and drowsy event. The last 10 seconds of an ECG window from configuration C6400 N6 M72 are plotted. The first row presents the raw ECG signals, while the second row displays the filtered ECG signals with detected R-peaks.

Fig. 3: RR Interval trend over the last capsule of a random DEW and NSRW from C6400 N6 M72 configuration. The figure compares sinus rhythm (blue) and drowsy event intervals (orange), illustrating changes over time.

Fig. 4: Feature Importance Comparison Across Methods: RFE, MI, PI, and RF.

Model	Configuration	Units	Dropout	Layers	Learning Rate	Batch Size	Accuracy	F1 Score
LSTM	C10240_N2_M50	100	0.20	1	0.001	80	0.9080	0.9105
LSTM	C6400_N6_M72	40	0.25	1	0.001	64	0.8734	0.8734
GRU	C10240_N2_M50	50	0.25	2	0.002	64	0.9360	0.9380
GRU	C6400_N6_M72	100	0.20	2	0.002	64	0.8924	0.8931
RNN	C10240_N2_M50	150	0.15	2	0.002	80	0.9560	0.9569
RNN	C6400_N6_M72	40	0.25	3	0.002	48	0.9241	0.9200

TABLE I: Performance Comparison of Optimized Deep Learning Models for Drowsiness Detection Using Capsule-Shifting Configurations

Fig. 5: Overview of the proposed DRL framework, showing the interaction between the driver drowsiness detection system (left), the DQN agent (center), and the vehicle sensor modules (right). The framework integrates HRV-based drowsiness predictions, radar-based front vehicle detections (filtered with DBSCAN ), and throttle, brake, and speed signals retrieved from the CAN bus to construct the observable state space. The agent uses this input to learn context-aware braking strategies under both alert and impaired cognitive conditions.

Fig. 6: Reward structure visualization: green labels indicate positive rewards; red labels indicate penalties based on distance and braking behavior.

Fig. 7: Finite state automaton representing the reward structure across the considered states during braking. P denotes penalty, R denotes reward, and C denotes condition related to speed and safe distance.

Fig. 8: Front-mounted radar sensor showing forward field of view and detection zone.

Configuration	Rnn		Cnn-Lstm		Ae-Lstm		Lstm		Gru

	ACC (%)	F1 (%)	ACC Diff (%)	F1 Diff (%)	ACC Diff (%)	F1 Diff (%)	ACC Diff (%)	F1 Diff (%)	ACC Diff (%)	F1 Diff (%)
C6400_N6_M72	91.37	91.33	86.08 [-5.29]	86.33 [-5.00]	76.49 [-14.88]	76.01 [-15.32]	86.09 [-5.28]	86.24 [-5.09]	82.42 [-8.95]	82.58 [-8.75]
C5120_N6_M60	90.57	90.65	83.48 [-7.09]	83.78 [-6.87]	76.85 [-13.72]	77.85 [-12.80]	83.69 [-6.88]	84.03 [-6.62]	76.86 [-13.71]	77.03 [-13.62]
C6144_N7_M75	90.30	90.16	84.24 [-6.06]	84.26 [-5.90]	75.91 [-14.39]	74.53 [-15.63]	86.82 [-3.48]	86.83 [-3.33]	83.03 [-7.27]	83.03 [-7.13]
C5120_N11_M80	86.90	87.17	68.62 [-18.28]	68.75 [-18.42]	64.83 [-22.07]	67.20 [-19.97]	66.21 [-20.69]	66.56 [-20.61]	69.31 [-17.59]	69.83 [-17.34]

TABLE II: The average test of the first cross-validation results for comparison of capsule-Shifting configurations and deep learning models for drowsiness detection.

True Class	Predicted Class
True Class	Negative (0)	Positive (1)
Negative (0)	77	2
Positive (1)	10	69

TABLE III: Confusion matrix for the selected Simple RNN model using configuration C6400_N6_M72. Model performance: Accuracy = 92%, Precision = 97%, Recall = 87%, F1-score = 92%.

Panel of Figure 12 — Fig. 9: Comparison of minimum and average rewards across training episodes.

Fig. 10: Analysis of the DDDQN agent’s behavior during 999 episodes including drowsy and non-drowsy scenarios. The figure presents the average ego vehicle speed with standard deviation (top), the relative following distance with min-max range (middle), and the control inputs (throttle and brake) with standard deviation (bottom). The plot highlights the agent’s adaptive behavior across varied driving conditions.

Fig. 11: Impact of drowsiness on the control behavior of the ego vehicle driven by the agent. (a) illustrates the agent’s control strategy under normal conditions, while (b) highlights the effect of drowsiness, showing drowsy periods (shaded regions) and variations in the ego vehicle’s throttle and brake inputs in response to the front vehicle’s speed.

Autonomous Driving

Autonomous driving develops vehicles that perceive road scenes, predict interactions, and plan and control safe movement with limited or no human input.

See all 536 papers

Improving Driver Drowsiness Detection via Personalized EAR/MAR Thresholds and CNN-Based Classification
Apr 24, 2026Gökdeniz Ersoy, Mehmet Alper Tatar, Eray Tonbul +1Driver Monitoring
MEF University, Electrical and Electronics Engineering Department, Huzur Mah, Maslak Ayazağa Cd. No:4, 34396 Sarıyer/İstanbul
Human Centered Non Intrusive Driver State Modeling Using Personalized Physiological Signals in Real World Automated Driving
Apr 13, 2026David Puertas-Ramirez, Raul Fernandez-Matellan, David Martin Gomez +1Driver Monitoring
UNED, Spain · UC3M, Spain
Autonomous Vehicle Collision Avoidance With Racing Parameterized Deep Reinforcement Learning
Apr 17, 2026Shathushan Sivashangaran, Vihaan Dutta, Apoorva Khairnar +2OOD Generalization Autonomous Driving
Virginia Commonwealth University, Richmond, VA 23284, USA · University of Michigan, Ann Arbor, MI 48109, USA · Virginia Tech, Blacksburg, VA 24061, USA

Apr 24, 2026J/K move · Enter open · S save

Improving Driver Drowsiness Detection via Personalized EAR/MAR Thresholds and CNN-Based Classification

Gökdeniz Ersoy, Mehmet Alper Tatar, Eray Tonbul, Serap Kırbız

MEF University, Electrical and Electronics Engineering Department, Huzur Mah, Maslak Ayazağa Cd. No:4, 34396 Sarıyer/İstanbul

Driver Monitoring

Open paper AskPDF arXiv

Driver drowsiness is a major cause of traffic accidents worldwide, posing a serious threat to public safety. Vision-based driver monitoring systems often rely on fixed Eye Aspect Ratio (EAR) and Mouth Aspect Ratio (MAR) thresholds; however, such fixed values frequently fail to generalize across individuals due to variations in facial structure, illumination, and driving conditions. This paper proposes a personalized driver drowsiness detection system that monitors eyelid movements, head position, and yawning behavior in real time and provides warnings when signs of fatigue are detected. The system employs driver-specific EAR and MAR thresholds, calibrated before driving, to improve classical metric-based detection. In addition, deep learning-based Convolutional Neural Network (CNN) models are integrated to enhance accuracy in challenging scenarios. The system is evaluated using publicly available datasets as well as a custom dataset collected under diverse lighting conditions, head poses, and user characteristics. Experimental results show that personalized thresholding improves detection accuracy by 2-3% compared to fixed thresholds, while CNN-based classification achieves 99.1% accuracy for eye state detection and 98.8% for yawning detection, demonstrating the effectiveness of combining classical metrics with deep learning for robust real-time driver monitoring.

Abstract

Autonomous Driving

Explore similar work

Improving Driver Drowsiness Detection via Personalized EAR/MAR Thresholds and CNN-Based Classification

Human Centered Non Intrusive Driver State Modeling Using Personalized Physiological Signals in Real World Automated Driving

Autonomous Vehicle Collision Avoidance With Racing Parameterized Deep Reinforcement Learning