Lane-Wise Highway Anomaly Detection
Mei Qiu 1 , William Lorenz Reindl 2 , Yaobin Chen 1 , Stanley Chien 1 , Shu Hu 3 · 2025
Read Lane-Wise Highway Anomaly Detection by Mei Qiu 1 , William Lorenz Reindl 2 , Yaobin Chen 1 , Stanley Chien 1 , Shu Hu 3 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-02613.txt.
Introduction
Highway cameras watch traffic day and night, but spotting trouble in one lane among many is slow work for people. A 2025 paper by Mei Qiu and four co-authors offers a lane-by-lane detection system that reads camera video instead of road sensors. The system pulls out lane-level facts such as vehicle count, lane occupancy, and the share of trucks.
The team built a dataset of 73,139 lane samples marked with four kinds of problems checked by experts. Three are traffic events: a blocked lane and its clearing, a foreign object in the road, and lasting congestion. The fourth is a camera fault, a shift in camera angle. Tests show the mixed system, which joins deep learning, rules, and machine learning, beats older methods on precision, recall, and F1 score.
Catalog record
| CREATOR | Mei Qiu 1 , William Lorenz Reindl 2 , Yaobin Chen 1 , Stanley Chien 1 , Shu Hu 3 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 33 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-02613
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Lane-Wise Highway Anomaly Detection Mei Qiu 1 , William Lorenz Reindl 2 , Yaobin Chen 1 , Stanley Chien 1 , Shu Hu 3 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. Our dataset and code can be found here: Lane-wise-Traffic-AD . I Introduction Traffic anomaly detection is a key function of intelligent transportation systems (ITS), supporting real-time monitoring and proactive management of safety risks, congestion, and infrastructure efficiency. As urban and highway networks grow more complex, timely detection of abnormal traffic behaviors is crucial for ensuring roadway safety and operational reliability [ 1 ] . In practice, however, even domain experts struggle to predefine all possible anomalies [ 2 ] . Most anomaly detection models instead focus on learning a representation of normal traffic patterns and identifying anomalies as significant deviations from this norm [ 3 ] . While extensive research has improved detection techniques, no studies address the classification or understanding of traffic anomaly types in specific scenarios. To detect traffic anomalies in both highway and urban environments, vehicle flow, speed, and occupancy data are commonly utilized, with anomalies defined as deviations from typical traffic patterns under similar conditions. Datasets for this purpose are derived from both real-world traffic [ 4 ] and simulated environments [ 5 ] . Data collection relies on a variety of sensors, including loop detectors, radar, LiDAR, and connected vehicle systems. However, these sensors are expensive to deploy and maintain, with limited coverage across road networks. Moreover, many approaches assume access to accurate road topology or GPS data, which limits scalability, especially in settings where only cameras are available. Compounding this challenge, existing datasets such as [ 6 ] primarily collect traffic data from highways but include only limited lane-level annotations, which are insufficient for studying detailed traffic behaviors or supporting models that require high spatial resolution and generalizability. Despite advances in anomaly detection—including deep learning [ 7 ] , statistical modeling [ 8 ] , and hybrid frameworks [ 9 ] —these methods often lack robustness under varying traffic conditions. Algorithms tuned to specific traffic patterns or road structures (e.g., lane number, road width) typically perform well only in similar environments. This highlights the need for multi-branch detection frameworks that fuse complementary approaches to effectively detect anomalies under both road-structure-independent and structure-dependent conditions, particularly on highways. Fig. 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. Taken together, these limitations underscore the need for a scalable, low-cost, lane-wise highway anomaly detection framework that leverages existing surveillance infrastructure to produce high-resolution, interpretable traffic signals—without reliance on expensive sensors or complex road modeling. To address this need, we propose a surveillance camera-based anomaly detection system that operates solely on roadside video feeds. The overall framework is illustrated in Fig. 1 . Our system employs AI-powered vision models to automatically detect lane boundaries and traffic directions [ 10 ] , and extracts lane-wise traffic features—vehicle count, occupancy, and truck percentage—from optimally learned detection regions [ 11 ] . We focus on detecting sequential or collective traffic anomalies using these lane-wise features as input, where anomalies arise not from isolated outliers but from patterns of abnormal behavior over sequences of data points. To address data set limitations, we introduce a novel data set on lane-wise traffic anomalies derived from Indiana highway surveillance videos. Unlike previous data sets focused on urban intersections or aggregated data, our data set provides structured time series signals for each segment of the highway lane, supporting both lane and road anomaly detection. Through extensive data analysis and expert collaboration, we define specific types of anomalies based on the collected data. We further propose a modular, multi-stream Anomaly Detection System capable of capturing both road-independent and road-dependent anomalies. Our contributions are summarized as follows: • We construct and release a novel lane-wise highway traffic dataset capturing segment-level flow dynamics from five surveillance cameras over six months, including 73,139 lane-wise samples with vehicle count, occupancy, and truck percentages. • We propose an anomaly labeling pipeline combining machine learning, manual verification, and expert validation. Three types of traffic anomalies (lane blockage and recovery, foreign object intrusion, and sustained congestion) and one type of sensor anomaly (camera angle shift) are identified. • We develop a scalable, modular anomaly detection framework integrating rule-based logic, machine learning, and deep learning to improve both robustness and interpretability. II Related Work Traffic Anomaly Datasets. Existing traffic anomaly datasets often focus on urban settings and lack relevance to highway environments. AI City Challenge [ 12 ] and GeoLife [ 13 ] provide urban video and GPS data without highway-specific or anomaly-focused labeling. Highway datasets like PeMS [ 14 ] and NGSIM [ 6 ] offer sensor and trajectory data but lack comprehensive anomaly labels and consistent lane-level detail. While XTraffic [ 15 ] and FD-AED [ 16 ] include anomaly annotations, they are limited by either spatial resolution or reliance on non-visual sensors. These limitations highlight the need for scalable, lane-wise highway datasets derived from video, enabling interpretable and generalizable anomaly detection. Anomaly Detection on Time Series Traffic Data. Time series anomaly detection is critical for highway traffic monitoring. Traditional statistical methods like ARIMA [ 17 ] are simple and interpretable but struggle with nonlinear, dynamic traffic. Machine learning methods, such as Isolation forest [ 18 ] , handle multivariate data and adapt better to changing patterns, yet lack deep temporal modeling. Deep generative models provide powerful and expressive tools for time series anomaly detection. Variational Autoencoders (VAE) [ 19 ] enhances this approach by incorporating temporal dependencies, making it more effective for sequential traffic data. Building on this, Time-VQVAE-AD [ 20 ] introduces vector quantization to discretize the latent space, improving both robustness and interpretability. TranAD [ 21 ] takes a further step by leveraging transformer architectures, which excel at modeling long-range dependencies in multivariate time series. Fig. 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,