Lane-Wise Highway Anomaly Detection

TL;DR

The study tackles lane-level highway anomaly detection using vision-based lane features extracted from surveillance video. It introduces a modular, multi-branch framework that fuses a DL module operating on $CWT$-transformed lane counts via a $VQ$-VAE, a rule-based module monitoring flow and occupancy, and an ML module using occupancy and truck percentage with an Isolation Forest. A new lane-wise highway dataset from Indiana cameras (73,139 normal samples; 341 anomalies across four types plus a sensor anomaly) underpins the evaluation, with an anomaly labeling pipeline combining Isolation Forest, manual checks, and expert validation. The full fusion outperforms baselines (best accuracy $0.9787$, F1-score $0.9149$), with wavelet analysis and time-of-day adaptive thresholds enhancing performance, illustrating a cost-effective, scalable approach for real-world ITS applications.

Abstract

This paper proposes a scalable and interpretable framework for lane-wise highway traffic anomaly detection, leveraging multi-modal time series data extracted from surveillance cameras. Unlike traditional sensor-dependent methods, our approach uses AI-powered vision models to extract lane-specific features, including vehicle count, occupancy, and truck percentage, without relying on costly hardware or complex road modeling. We introduce a novel dataset containing 73,139 lane-wise samples, annotated with four classes of expert-validated anomalies: three traffic-related anomalies (lane blockage and recovery, foreign object intrusion, and sustained congestion) and one sensor-related anomaly (camera angle shift). Our multi-branch detection system integrates deep learning, rule-based logic, and machine learning to improve robustness and precision. Extensive experiments demonstrate that our framework outperforms state-of-the-art methods in precision, recall, and F1-score, providing a cost-effective and scalable solution for real-world intelligent transportation systems.

Figures (11)

• Figure 1: Overview of the proposed lane-wise anomaly detection framework. Lane-wise traffic data (e.g., vehicle count, occupancy, and truck percentage) is extracted from highway surveillance video. Anomalies are identified through Isolation Forest, manual checking, and expert validation. The system detects road-independent anomalies using deep learning and rule-based methods, and road-dependent anomalies using machine learning models applied to occupancy and truck percentage data. The anomalies labeling is only implemented during training stage.
• Figure 2: Five road views with detection regions, lane centers, and counting lines. Light blue rectangles show the learned Regions of Interest (ROIs), green lines mark optimal counting lines, and red dots indicate lane centers. White lane IDs use signs to show direction: negative for towards the camera, positive for away.
• Figure 3: The top row shows lane-wise vehicle count, occupancy, and truck percentage from a single 15-minute video. The bottom row presents the distribution of these features across all data collected from one camera over 24 hours, aggregated across all data collecting days.
• Figure 4: Examples of observed traffic anomalies and sensor anomalies.
• Figure 5: Deep Learning-based Anomaly Detection method.