Road Traffic Anomaly Detection by Human-Attention-Assisted Text-Vision Learning

Area of Science:

• Computer Vision
• Artificial Intelligence
• Transportation Safety

Background:

• Increasing vehicle numbers exacerbate traffic accident severity and congestion.
• Accurate detection of traffic anomalies is vital for safety and traffic flow.
• Deep learning methods face challenges in detecting anomalies in surveillance due to data imbalance and complex patterns.

Purpose of the Study:

• To develop a more efficient method for detecting road traffic anomalies from surveillance footage.
• To leverage text and eye-gaze annotations for improved anomaly representation learning.
• To enhance the accuracy of traffic anomaly detection in challenging surveillance scenarios.

Main Methods:

• Annotation of the previously proposed TADS dataset with text and eye-gaze data.
• Development of a detection model utilizing the CLIP (Contrastive Language–Image Pre-training) model.
• Training the model to learn anomaly representations using unique dataset annotations.

Main Results:

• The proposed model demonstrates superior performance in detecting traffic anomalies from a surveillance perspective.
• The study validates the utility of text and eye-gaze data for enhancing anomaly detection.
• Experimental results confirm the effectiveness of the CLIP-based approach for traffic surveillance.

Conclusions:

• The novel approach effectively improves road traffic anomaly detection using specialized annotations and the CLIP model.
• Text and eye-gaze data are valuable resources for advancing surveillance-based traffic analysis.
• This research offers a promising solution for real-world traffic safety and management challenges.