Traffic Analysis and Safety Enhancement through Deep Learning Models = 딥러닝 모델을 통한 교통 분석 및 안전 강화

다국어 초록 (Multilingual Abstract)

Traffic Analysis and Safety Enhancement through Deep Learning Models This thesis explores the use of deep learning models to improve urban transportation systems, addressing issues like traffic congestion, travel times, productivity reduction, and reduced quality of life. It consists of three interconnected studies, each contributing to a better understanding of urban transportation dynamics and road safety enhance- ment. The first study, conducted in Cheonan city, South Korea, focuses on bus travel time prediction using digital tachograph data from intracity buses. The research evaluates six deep learning models, including Long Short-Term Memory (LSTM) and Gated Recur- rent Unit (GRU) architectures, to predict travel times between Namchang Village bus stop and Dongnam-gu Public Health Center. Comparative analyses reveal that LSTM- based models, particularly Stacked-LSTM, exhibit the highest accuracy. The inclusion of weather data enhances prediction accuracy, providing valuable information for transit agencies and policymakers. The study reports a 47.69% reduction in root mean squared error compared to the baseline ARIMA model, offering data for optimizing public trans- portation schedules and efficiency. The second study focuses on the classification of driver behaviors, with applica- tions in enhancing driver safety, preventing traffic accidents, enabling usage-based insur- ance, and optimizing ridesharing services. The Hybrid ConvLSTM with Attention for Driver Behavior Classification (HCLA-DBC) model adeptly distinguishes between nor- mal, drowsy, and aggressive driving behaviors. Experimental assessments on the UAH- DriveSet dataset reveal better performance than the previous research using the same database, with an accuracy rate of 94.12%, precision score of 94.24%, recall score of 94.12%, F1-score of 94.12%, and ROC AUC of 98.83%. These findings offer detailed in- formation regarding the precise recognition of driver behaviors, which in turn contributes to safer roads, more accurate risk assessment, and enhanced driving experiences. The third study analyzes traffic accidents in Seoul, South Korea, employing the ”TrafficNet” hybrid CNN-FNN model. The model proficiently categorizes accidents into groups, including minor injuries, slander, fatalities, and injury reports. Experiments conducted on traffic accident data from the Seoul Metropolitan Government demonstrate the model’s effectiveness, with an accuracy rate of 93.98%, precision score of 94.31%, recall score of 93.98%, and F1-score of 93.89%. These findings shed light on the primary determinants of traffic accident severity in Seoul, with a specific focus on the influence of vehicle-related factors. This study provides information that can guide the decisions of local authorities and policymakers, with the ultimate goal of improving road safety and mitigating the social and economic consequences of traffic accidents. Keyword: bus travel time prediction; deep learning techniques; digital tachograph data (DTG); Driver behavior classification; Advanced driver-assistance systems (ADAS); Traffic accident prevention; UAH-DriveSet dataset; CNN-FNN hybrid model.

목차 (Table of Contents)

• 1. INTRODUCTION 1
• 1.1. Problem Statement 7
• 1.1.1. Problem Statement of Prediction of Bus Travel Time in Cheonan City Using Deep Learning with DTG Data 7
• 1.1.2. Problem Statement of Hybrid ConvLSTM with Attention for Precise Driver Behavior Classification 7
• 1.1.3. Problem Statement of TrafficNet: A Hybrid CNN-FNN Model for Analysis of Traffic Accidents in Seoul 8
• 1.2. Motivation 8
• 1.2.1. Motivation of Prediction of Bus Travel Time in Cheonan City Using Deep Learning with DTG Data 8
• 1.2.2. Motivation of Hybrid ConvLSTM with Attention for Precise Driver Behavior Classification 9
• 1.2.3. Motivation of TrafficNet: A Hybrid CNN-FNN Model for Analysis of Traffic Accidents in Seoul 10
• 1.3. Objective 11
• 1.3.1. Objectives of Prediction of Bus Travel Time in Cheonan City Using Deep Learning with DTG Data 11
• 1.3.2. Objectives of Hybrid ConvLSTM with Attention for Precise Driver Behavior Classification 12
• 1.3.3. Objectives of TrafficNet: A Hybrid CNN-FNN Model for Analysis of Traffic Accidents in Seoul 13
• 1.4. Contributions 14
• 1.4.1. Our Contributions in Prediction of Bus Travel Time in Cheonan City Using Deep Learning with DTG Data 14
• 1.4.2. Our Contributions in Hybrid ConvLSTM with Attention for Precise Driver Behavior Classification 15
• 1.4.3. Our Contributions in TrafficNet: A Hybrid CNN-FNN Model for Analysis of Traffic Accidents in Seoul 16
• 2. RELATED WORK 18
• 2.1. Related work of Prediction of Bus Travel Time in Cheonan City Using Deep Learning with DTG Data 18
• 2.2. Related work of Hybrid ConvLSTM with Attention for Precise Driver Behavior Classification 22
• 2.3. Related work of Analysis of traffic accidents abroad 25
• 2.3.1. Analysis of Traffic Accidents in Seoul, South Korea 27
• 3. DATA COLLECTION AND PREPROCESSING 29
• 3.1. Data Collection and Preprocessing of Prediction of Bus Travel Time in Cheonan City Using Deep Learning with DTG Data 29
• 3.1.1. Study Area - Cheonan 29
• 3.1.2. DTG Device 30
• 3.1.3. Dataset Description 31
• 3.1.4. Data preprocessing 35
• 3.1.5. Final Data Extraction 38
• 3.2. Data Collection and Preprocessing of Hybrid ConvLSTM with Attention for Precise Driver Behavior Classification 40
• 3.2.1. Study Area - Spain 40
• 3.2.2. Dataset Description 41
• 3.2.3. Feature Extraction 42
• 3.2.4. Handling Imbalanced Data 47
• 3.2.5. Temporal Window Selection 49
• 3.2.6. Normalization Stage 50
• 3.3. Data Collection and Preprocessing of TrafficNet: A Hybrid CNN-FNN Model for Analysis of Traffic Accidents in Seoul 53
• 3.3.1. Study Area - Seoul 53
• 3.3.2. Dataset Description 53
• 3.3.3. Data Preprocessing 54
• 3.3.4. Chi-Square Test for Feature Selection in Machine Learning 56
• 4. PROPOSED METHODOLGY 58
• 4.1. Methodology applied in the Prediction of Bus Travel Time in Cheonan City Using Deep Learning with DTG Data 58
• 4.1.1. Long Short-Term Memory 59
• 4.1.2. Gated recurrent unit: 62
• 4.1.3. Hyperparameter Setting 64
• 4.1.4. Performance Metrics 66
• 4.2. Methodology applied in Hybrid ConvLSTM with Attention for Precise Driver Behavior Classification 68
• 4.2.1. Performance Metrics 68
• 4.2.2. Our Proposed model Hybrid ConvLSTM with Attention (HCLA-DBC) for Driver Behavior Classification 70
• 4.3. Methodology applied in the TrafficNet: A Hybrid CNN-FNN Model for Analysis of Traffic Accidents in Seoul 74
• 4.3.1. TrafficNet Model: 75
• 5. EXPERIMENTS RESULTS AND DISCUSSIONS 78
• 5.1. Results of the Prediction of Bus Travel Time in Cheonan City Using Deep Learning with DTG Data 78
• 5.1.1. Experimental settings 78
• 5.1.2. Performance evaluation of all models using the overall test data 79
• 5.1.3. Weather Influence on Travel Time Prediction 84
• 5.1.4. Reliability analysis of models during weekdays data 86
• 5.1.5. Reliability analysis of models during weekends data 87
• 5.1.6. Robustness of models on short routes 89
• 5.1.7. Comparison of all models with baseline model ARIMA 92
• 5.2. Results of the Hybrid ConvLSTM with Attention for Precise Driver Behavior Classification 94
• 5.2.1. Result of all models on test data 94
• 5.2.2. Attention Mechanism 97
• 5.3. Results of the TrafficNet: A Hybrid CNN-FNN Model for Analysis of Traffic Accidents in Seoul 98
• 5.3.1. Weather and Road Conditions Influence on Results 100
• 6. Conclusion and Future Works 101
• 6.1. Conclusion and Future work of the Prediction of Bus Travel Time in Cheonan City Using Deep Learning with DTG Data 101
• 6.2. Conclusion and Future work of the Hybrid ConvLSTM with Attention for Precise Driver Behavior Classification 103
• 6.3. Conclusion and Future work of the Hybrid CNN-FNN Model for Analysis of Traffic Accidents in Seoul 104