Design and Implementation of a Pedestrian Danger Prevention System using Deep Learning and Computer Vision

다국어 초록 (Multilingual Abstract)

In an era where people often get distracted by their phones rather than being aware of their surroundings, having a reliable danger detection system is a must. This thesis paper explores the development and implementation of a danger prevention system for pedestrian safety using deep learning and computer vision technology, specifically within the school area. By combining robust object detection capabilities with advanced object tracking algorithms, the system can continuously monitor the movement of pedestrians and cars within the predefined risk areas. These risk areas are configurable polygonal zones strategically positioned in locations where pedestrian-vehicle interactions are most likely to occur, such as near building entrances and adjacent to parking lots. A crucial aspect of the system is its ability to estimate proximity distances between detected pedestrians and cars within the risk areas. This is achieved through a post-processing step that employs the Euclidean distance formula to calculate the straight-line distance between the center points of the detected objects. If the estimated distance falls below a predetermined threshold, indicating a high risk of collision, the system triggers visual alert signals to a signal lighting tower.
The multi-level danger assessment system uses different colored lights on a signal tower to show various danger levels, providing clear and quick warnings. This real-time notification system could alert pedestrians, giving them awareness and authority to take proactive actions, potentially preventing accidents and minimizing the risk of harm to pedestrians.
Comprehensive experimental evaluations, conducted using real-world video footage captured near a building in the Chungbuk National University campus, demonstrate the system's accuracy in object detection and its effectiveness in assessing and responding to potential pedestrian-vehicle collision risks. With an overall accuracy of 93.7% in detecting dangerous interactions, the proposed system showcases its potential as a valuable tool for promoting pedestrian safety in high-risk areas.

목차 (Table of Contents)

• Introduction 3
• 1.1 Motivation 3
• 1.2 Contribution 4
• 1.3 Thesis Organization 6
• Background 8
• 2.1 Object Detection Algorithms 10
• 2.1.1 Pre-YOLO: Traditional Object Detection 10
• 2.1.2 Deep Learning Advancements 12
• 2.2 Object Tracking Algorithms 18
• 2.3 Related Works 19
• Methodology 22
• 3.1 Data Overview 23
• 3.2 System Architecture 25
• 3.2.1 Object Detection and Tracking 25
• 3.2.1.1 Data Labeling 26
• 3.2.1.2 Model Training: Custom YOLOv8 Model 29
• 3.2.2 Danger Detection 31
• 3.2.2.1 Risk Area Configuration 31
• 3.2.2.2 Post Processing: Distance Estimation 32
• Result and Discussion 35
• 4.1 Object Detection Result 35
• 4.2 Danger Detection Result 39
• 4.2.1 Danger Level Determination 40
• 4.2.2 Performance Evaluation 43
• Summary and Future Work 44
• 5.1 Summary 44
• 5.2 Future Work 45
• References 47
• Acknowledgments 51