Anomaly Detection in Traffic Surveillance Videos Using Deep Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Traffic Safety

Background:

• Increasing global populations necessitate advanced traffic surveillance systems.
• Video Traffic Surveillance Cameras (VTSS) are crucial for monitoring public safety and detecting abnormal events like accidents.
• Timely detection of road accidents is vital for providing emergency medical treatment, especially on remote highways.

Purpose of the Study:

• To propose and evaluate an automated methodology for detecting traffic accidents using surveillance videos.
• To leverage deep learning, specifically Convolutional Neural Networks (CNNs), for accurate anomaly detection in traffic footage.
• To enhance the efficiency and effectiveness of traffic monitoring systems.

Main Methods:

• Utilized Convolutional Neural Networks (CNNs), a deep learning technique adept at image and video analysis.
• Developed and employed a Vehicle Accident Image Dataset (VAID) for training the CNN model.
• Implemented a rolling prediction algorithm to optimize accuracy in accident detection from VTSS footage.

Main Results:

• The proposed CNN-based methodology successfully detected traffic accident events in surveillance videos.
• Achieved a high accuracy rate of 82% in identifying accidents within the tested traffic surveillance system videos.
• Demonstrated the effectiveness of deep learning for real-world traffic anomaly detection.

Conclusions:

• The research successfully validates the use of CNNs for automated traffic accident detection.
• The developed system offers a promising solution for improving road safety and emergency response times.
• This technology can significantly aid traffic management and public safety initiatives.