An Unsupervised Learning Framework for Detecting Abnormal Driving Behavior via Utility-feature Sequences

Details

• Creators:
  Liang, Xiao ; Shi, Xiaowei ; Huang, Suzhou ; Qin, Xiao
• Corporate Creators:
  Center for Pedestrian and Bicyclist Safety (CPBS) Tier-1 University Transportation Center (UTC) ; University of New Mexico
• Corporate Contributors:
  United States. Department of Transportation. University Transportation Centers (UTC) Program ; United States. Department of Transportation. Office of the Assistant Secretary for Research and Technology
• Subject/TRT Terms:
  • Behavior
  • Data collection
  • Driver monitoring
  • Pedestrian safety
  • Predictive models
  • Traffic surveillance
  • Vehicle trajectories
• Publication/ Report Number:
  24UWM02
• Resource Type:
  Tech Report
• Geographical Coverage:
  United States
• Edition:
  Final Report (June 2025)
• Corporate Publisher:
  Center for Pedestrian and Bicyclist Safety (CPBS) Tier-1 University Transportation Center (UTC)
• Abstract:
  Abnormal driving poses a significant threat to all road users, especially VRUs such as pedestrians and bicyclists. Although traffic data collected from roadside cameras is widely available, it lacks ground-truth anomaly labels, making supervised learning impractical. Existing detection methods rely on individual vehicle features, such as speed, acceleration, which ignores interactions between vehicles. Moreover, as different vehicles have diverse driving behavior. Fixed thresholds fail to adapt in dynamic traffic environments. To address these challenges, we propose an unsupervised anomaly detection framework that integrates utility-based modeling and deep learning. A set of utility functions is defined to quantify driving safety risks. A GRU-based autoencoder learns temporal patterns in driving behavior, and latent features are clustered using a Gaussian Mixture Model (GMM). Meanwhile, the source of abnormality can be identified by analyzing reconstruction error from the GRU-AE. Experimental results on the HighSim dataset, demonstrate that the proposed framework can effectively detect abnormal behavior. Visualization confirms that detected anomalies correspond to spikes in interpretable utility features, validating both the effectiveness and the interpretability of the approach.
• Format:
  PDF
• Funding:
  69A3552348336
• Download URL:
  https://rosap.ntl.bts.gov/view/dot/93238/dot_93238_DS1.pdf
• File Type:
  [PDF - 1.58 MB]
• Collection(s):
  University Transportation Centers
• Main Document Checksum:
  urn:sha-512:bdebbf1c46ee5fa616229b2dffff5890bc5ae218cb5bce469668c274628918fa2755c1d8fa7293a5ef9b7f726d25cf720d821b41903b1329dfd77afa2ec70379