Deep Spatio-Temporal Memory and Interaction Network for Traffic Incident Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traffic incident detection in driving videos is crucial but challenging due to short event durations and complex background dynamics.
• Existing methods struggle with accurately identifying the precise start and end times of incidents.

Purpose of the Study:

• To propose an advanced traffic incident detection method using a deep spatio-temporal memory and interaction network.
• To improve the accuracy and efficiency of classifying traffic incidents in real-time driving scenarios.

Main Methods:

• A spatio-temporal information perception network was designed to capture multi-scale spatial and short-term temporal features.
• A temporal feature learning network was developed to process long-term temporal information for better context.
• The proposed network integrates memory and interaction mechanisms for enhanced spatio-temporal understanding.

Main Results:

• The method achieved an Area Under the Curve (AUC) of 80.9% on the DoTA dataset.
• Experimental results demonstrate competitive performance compared to state-of-the-art traffic incident detection techniques.
• The approach showed high accuracy in classifying traffic incidents and determining their temporal boundaries.

Conclusions:

• The proposed deep spatio-temporal memory and interaction network effectively enhances traffic incident detection in driving videos.
• The method provides accurate classification and temporal localization of incidents, outperforming existing approaches.
• This research contributes a robust solution for improving road safety through advanced video analysis.