Progressive Temporal-Spatial-Semantic Analysis of Driving Anomaly Detection and Recounting.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Driving Systems

Background:

• Existing methods for traffic anomaly detection often rely on large labeled datasets, which are difficult to obtain and may struggle with the contextual nature of anomalies.
• Clearly distinguishing between normal and abnormal driving behaviors in complex traffic scenarios remains a challenge for current algorithms.

Purpose of the Study:

• To propose a novel unsupervised framework for progressive driving anomaly detection and recounting (D&R) from ego-vehicle dashcam footage.
• To develop a system capable of spatial-temporal localization and semantic explanation of traffic anomalies without prior labeled data.
• To enhance the understanding and analysis of driving events for improved road safety.

Main Methods:

• Formulation of a temporal-spatial-semantic (TSS) model for coarse-to-fine anomaly detection and recounting.
• Development of an unsupervised D&R approach that eliminates the need for training data.
• Integration of traffic saliency, isolation forest, and visual semantic causal relations to construct the TSS model.

Main Results:

• The proposed unsupervised framework achieves effective performance in driving anomaly detection and recounting.
• The TSS model demonstrates a coarse-to-fine focusing capability, generating convincing anomaly explanations.
• Experimental results on a custom-labeled dataset show superior performance compared to existing techniques.

Conclusions:

• The unsupervised TSS framework offers a promising solution for driving anomaly detection and recounting.
• This approach advances the field by enabling effective D&R without requiring labeled training data.
• The method provides a robust and scalable solution for analyzing complex traffic scenarios and enhancing autonomous driving safety.