Dual Encoder-Decoder-Encoder with Adversarial Training for Unsupervised Traffic Accident Detection in Surveillance...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Safety Engineering

Background:

• Manual monitoring of traffic surveillance footage for incident detection is error-prone and time-consuming.
• Automated accident detection is challenging due to significant class imbalance (rare accidents vs. common driving).
• Traditional computer vision struggles to differentiate normal from abnormal traffic events.

Purpose of the Study:

• To develop an advanced automated system for timely detection of traffic incidents and unsafe driving behaviors.
• To overcome the limitations of existing methods in handling class imbalance and subtle anomaly detection.

Main Methods:

• A deep learning architecture utilizing a dual encoder-decoder-encoder (EDE) framework was developed.
• The model maps image distributions to latent distributions bidirectionally to characterize normal traffic behavior.
• A two-phase training strategy, including a generative adversarial mechanism, was employed to enhance anomaly sensitivity.

Main Results:

• The dual-EDE framework effectively models normal traffic patterns and detects deviations indicating potential hazards.
• The adversarial training amplified differences, improving responsiveness to subtle anomalies.
• Experimental results on real-world datasets showed significant improvements in accident and unsafe behavior detection accuracy and robustness.

Conclusions:

• The proposed dual-EDE architecture combined with adversarial training offers a substantial advancement for traffic incident detection.
• This methodology improves the ability to model both normal and abnormal driving behaviors for enhanced road safety.
• The system demonstrates superior performance in accurately and robustly identifying accidents and unsafe driving in real-world surveillance footage.