DoTA: Unsupervised Detection of Traffic Anomaly in Driving Videos

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Video anomaly detection (VAD) is crucial but challenging in dynamic egocentric driving scenes.
• Existing methods often struggle with the complexity and unpredictability of real-world traffic scenarios.

Purpose of the Study:

• To propose an unsupervised method for traffic VAD in egocentric videos.
• To introduce a new large-scale benchmark dataset (DoTA) for evaluating VAD methods.
• To develop a novel evaluation metric (STAUC) that captures both temporal and spatial anomaly localization.

Main Methods:

• Developed an unsupervised VAD approach based on predicting future object locations of traffic participants.
• Introduced the Detection of Traffic Anomaly (DoTA) dataset with 4,677 videos, featuring temporal, spatial, and categorical annotations.
• Proposed the spatial-temporal area under curve (STAUC) metric for comprehensive VAD evaluation.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art approaches on the DoTA dataset.
• The new STAUC metric effectively evaluates both temporal and spatial aspects of anomaly detection.
• The DoTA dataset provides rich annotations for benchmarking various video analysis tasks.

Conclusions:

• The future object localization method offers a promising unsupervised approach for traffic VAD.
• The DoTA dataset and STAUC metric advance the field of anomaly detection in driving videos.
• This work contributes to safer autonomous driving systems through improved anomaly recognition.