PaperlayerView exact revision on arXiv

Source-linked AI summary

Detecting Socially Abnormal Highway Driving Behaviors via Recurrent Graph Attention Networks

Yue Hu, Yuhang Zhang, Yanbing Wang, Daniel Work

arXiv:2304.11513v1cs.AI

TL;DR

The paper addresses the limited ability of existing methods to detect socially abnormal driving while considering interactions and changing traffic context. It introduces a recurrent graph-attention autoencoder that models neighboring vehicles and stochastic trajectories, and reports exact abnormal-vehicle detection at highway scale in simulation and real-world data.

  • Problem

    Existing anomaly detectors often handle limited, predefined, single-vehicle behaviors without modeling vehicular interactions or spatial-temporal traffic variation.

  • Method

    DSAB uses a recurrent graph-attention autoencoder to model spatial-temporal vehicle trajectories while dynamically incorporating neighboring vehicles.

  • Results

    The model is the only evaluated method reported to identify the exact abnormal vehicle in simulation and detects vehicles violating local driving norms in the highD dataset.

  • Takeaways & Limitations

    The approach detects varied socially abnormal behaviors and scales to highway monitoring with thousands of vehicles.

  • Takeaways & Limitations

    The KDE variant was computationally infeasible, requiring use of STGAE-biv because KDE did not complete within two days.

Abstract

from arXiv · show

With the rapid development of Internet of Things technologies, the next generation traffic monitoring infrastructures are connected via the web, to aid traffic data collection and intelligent traffic management. One of the most important tasks in traffic is anomaly detection, since abnormal drivers can reduce traffic efficiency and cause safety issues. This work focuses on detecting abnormal driving behaviors from trajectories produced by highway video surveillance systems. Most of the current abnormal driving behavior detection methods focus on a limited category of abnormal behaviors that deal with a single vehicle without considering vehicular interactions. In this work, we consider the problem of detecting a variety of socially abnormal driving behaviors, i.e., behaviors that do not conform to the behavior of other nearby drivers. This task is complicated by the variety of vehicular interactions and the spatial-temporal varying nature of highway traffic. To solve this problem, we propose an autoencoder with a Recurrent Graph Attention Network that can capture the highway driving behaviors contextualized on the surrounding cars, and detect anomalies that deviate from learned patterns. Our model is scalable to large freeways with thousands of cars. Experiments on data generated from traffic simulation software show that our model is the only one that can spot the exact vehicle conducting socially abnormal behaviors, among the state-of-the-art anomaly detection models. We further show the performance on real world HighD traffic dataset, where our model detects vehicles that violate the local driving norms.

1 INTRODUCTION

The paper frames socially abnormal driving as a context-dependent anomaly-detection problem and proposes an unsupervised recurrent graph-attention autoencoder to identify anomalous vehicles from interacting trajectories.

  • 1.1 Motivation and challenges: Existing methods mainly detect stalled vehicles or isolated behaviors such as speeding and abrupt braking, often using rules or supervised learning for predefined anomaly types.These approaches do not cover unexpected anomalies and generally treat vehicles independently.
  • 1.1 Motivation and challenges: Socially abnormal driving is behavior that violates commonly accepted norms among nearby vehicles, so anomaly detection must account for traffic context and vehicular interactions.A constant speed can be normal in one lane but anomalous when it blocks faster traffic; braking may be expected when responding to a stalled vehicle.
  • 1.1 Motivation and challenges: Dynamic traffic graphs must model both changing spatial-temporal relationships and stochastic normal behavior, because acceptable driving can vary within the same context.A deterministic model may incorrectly label ordinary lane changes or speed variation as anomalous.
  • 1.2 Our approach: DSAB uses recurrent graph attention within an autoencoder to capture trajectory dynamics while dynamically incorporating neighboring vehicles.The approach is designed to detect socially abnormal behavior in large-scale highway trajectory data.
  • 1.2 Our approach: The model is scalable to thousands of cars over 5 miles, and experiments show it uniquely identifies the exact abnormal vehicle while also detecting violations of local driving norms.Simulation data provide ground-truth anomaly labels, while real-world highD data demonstrate detection of vehicles violating local norms.

2 RELATED WORK

Related work models agent interactions for trajectory prediction, while STGAE uses a spatio-temporal graph autoencoder and KDE for abnormal driving detection. The paper distinguishes its exact-vehicle detection task from easier abnormal-scene detection and addresses STGAE's scalability limitation.

  • Trajectory Modeling: Trajectory modeling methods increasingly represent interactions through pooling, attention, or convolutional scene-context fusion for predicting human or vehicle trajectories.Examples include Social LSTM and Social GAN with pooling, SoPhie with attention, and Desire with convolutional scene-context fusion.
  • Abnormal driving detection: STGAE combines a spatio-temporal graph autoencoder with KDE for abnormal driving detection, but KDE becomes too expensive at large vehicle counts.The paper reports that STGAE performs well with only two vehicles, whereas its KDE complexity limits large-scale detection.
  • Abnormal driving detection: Abnormal scene detection asks whether any anomaly exists, whereas this paper asks which specific vehicle violates social driving norms.Neighboring vehicles may appear abnormal because they react normally to a slow or stalled root vehicle.
  • Abnormal driving detection: Baseline methods can perform well on abnormal-scene detection yet perform poorly when required to identify the abnormal vehicle itself.This distinction motivates evaluating exact-vehicle detection rather than only scene-level anomaly presence.

3 METHOD

DSAB models highway trajectories as a spatial-temporal dynamic graph and uses a recurrent graph-attention autoencoder to reconstruct probabilistic vehicle states for anomaly detection.

  • 3.1 Problem formulation: DSAB constructs a dynamic graph of vehicles over time, then encodes vehicle trajectories and decodes their probabilistic state distributions for anomaly detection.The input contains observations for a variable number of vehicles over a time window, and anomalies are evaluated at both vehicle and scene levels.
  • 3.1 Problem formulation: The formulation uses coarser sampling and longer time windows to reduce processing demands while capturing behaviors such as speeding or tailgating that persist over several seconds.The method also accommodates vehicles entering or leaving the highway during a time window.
  • 3.2 Graph construction for vehicle trajectories: The graph connects vehicles that are close longitudinally and laterally, limiting neighborhood range to reduce storage and computation.The neighborhood restriction reflects the assumption that nearby vehicles exert the strongest influence on one another.
  • 3.3 Encoder: The RGAT encoder combines recurrent updates with graph attention, dynamically weighting neighboring vehicle states and using multi-head attention to capture spatial-temporal interactions.GRU-based graph convolutions model temporal dependencies, while attention weights are computed from vehicle states rather than fixed physical distances.
  • 3.3 Encoder: Attention can model asymmetric influence between vehicles, such as the stronger effect of a slow vehicle ahead than one behind, unlike distance-only edge weights.The decoder reuses RGAT and the union of encoding-time edge sets to avoid recomputing position-thresholded edges at every decoding step.
  • 3.4 Decoder and 3.5 Loss function: The decoder estimates means and variances for position, speed, and acceleration together with lane probabilities, trained using weighted negative log-likelihood and cross-entropy losses.The final loss aggregates these terms across vehicles and time steps.

4 EXPERIMENTS

The experiments evaluate DSAB on simulated trajectories with ground-truth anomalies and on real HighD highway trajectories, comparing it with heuristic, tensor, sequence, and graph-based baselines.

  • 4 Experiments: The evaluation uses simulation data for quantitative comparison with ground-truth labels and HighD trajectories for qualitative assessment on real-world traffic.The authors report that the implementation is publicly available on GitHub.
  • 4.1 Datasets: TransModeler simulations cover a 5-mile, 4-lane highway at 1 Hz across varying traffic flows, vehicle types, traffic conditions, and abnormal scenarios.Scenarios include normal traffic, speeding, tailgating, stalled vehicles, and a comprehensive mixture of abnormal behaviors.
  • 4.1 Datasets: Abnormal vehicles constitute approximately 3%-5% of simulated traffic, while training uses 180 minutes of normal traffic and intentionally includes a small amount of abnormal data.Removing training anomalies produces similar results according to the appendix summary.
  • 4.1 Datasets: HighD training data contains approximately 350 minutes from three-lane German highways spanning light and heavy traffic, with flows from 1200 to 3600 vehicles/lane/hour.Testing uses an unseen 15-minute recording.
  • 4.2 Baselines: Baselines range from single-trajectory heuristics and sequence models to tensor recovery and a spatio-temporal graph autoencoder, with DSAB-biv providing an RGAT variant using bivariate Gaussian loss.The methods produce vehicle- and scene-level anomaly scores using the same aggregation procedure.
  • 4.3 Evaluation metrics: Performance is measured with ROC-AUC, average precision, and Precision@k, covering ranking quality and the relevance of a fixed number of inspected anomalies.Precision@k is intended for settings where only a limited number of anomalies can be manually verified.

4.3 Implementation details

The implementation uses fixed graph and training settings, compares abnormal-vehicle and scene detection, and evaluates parameter sensitivity through ablations. The model performs best when neighboring vehicles are included and remains broadly robust to several configuration choices.

  • Implementation details: The simulation configuration uses a 0.1-mile distance threshold, adjacent lanes, a 15-second window, 1-second sampling, hidden size 5, and three attention heads.Training uses 500 epochs, batch size 64, and initial learning rate 0.05.
  • Model comparison: The model identifies the specific abnormal vehicle, with precision above 0.7 across k values, average precision about 0.3 above the second-best method, and ROC-AUC of 0.9, more than 0.1 higher.These results come from the comprehensive simulation scenario containing all abnormal behaviors.
  • Model comparison: Abnormal-scene detection is easier for all methods, yet the proposed model remains best with a 0.1 increase in ROC-AUC and average precision.Scene-level aggregation makes abnormality easier to detect than identifying the specific vehicle.
  • Model comparison: Across individual scenarios, the method uniquely detects slow and stalled cars with Precision@100 above 0.9 and leads speeding-car detection by 0.1 in Precision@100 with ROC-AUC near 1.Tailgating is relatively less effective, although the method still has the highest Precision@100 and no baseline is consistently better across metrics.
  • Ablation study: More than one attention head improves performance, while larger head counts and hidden dimensions otherwise have little effect, motivating hidden dimension 5 for a smaller model.The parameter study is reported in Fig 2a and Fig 2b.
  • Ablation study: Removing neighboring vehicles causes a 0.3 drop in Precision@100, whereas performance is broadly insensitive to window size and sampling choices beyond the stated favorable range.The model performs best with sampling intervals no greater than 3 seconds and is not sensitive to distance thresholds above 0.05 miles.

4.5 Qualitative Results

On real-world HighD traffic, DSAB learns lane-specific driving norms and identifies vehicles whose behavior deviates from those norms. The highest-scoring examples involve aggressive lane interaction and speed violations relative to lane-specific traffic.

  • 4.5 Qualitative Results: HighD traffic exhibits lane-specific norms, with average speeds of 76, 69, and 56 mph from the leftmost to rightmost lane.Because German highways often permit speeds above 75 mph, absolute speeding is not necessarily among the largest anomalies.
  • 4.5 Qualitative Results: The highest-scoring vehicles were abnormal because of aggressive acceleration or deceleration, or because their speed violated the typical range for their lane.Manual inspection identified both types of behavior among vehicles with top anomaly scores.
  • 4.5 Qualitative Results: The most anomalous example cut in front of another vehicle while decelerating dramatically and forcing that vehicle to change lanes.The trajectory reconstruction depicts the expected trajectories with the largest probability.
  • 4.5 Qualitative Results: The second-ranked anomaly traveled 20 mph faster than the typical speed in the rightmost lane, while reconstruction placed it in the faster middle lane.This illustrates that DSAB evaluates speed relative to lane context rather than speed alone.
  • 4.5 Qualitative Results: These examples indicate that DSAB can be applied to real-world datasets to identify vehicles violating local driving norms.The qualitative analysis presents the top two anomalies, with additional examples reported in Appendix E.

5 CONCLUSION

The paper presents DSAB, a Recurrent Graph Attention Autoencoder for detecting socially abnormal vehicles from highway trajectories. Results on simulated and real-world data support its ability to model traffic dynamics, scale to thousands of vehicles, and identify varied abnormal behaviors.

  • 5 CONCLUSION: DSAB uses an autoencoder based on Recurrent Graph Attention Networks to detect exact anomalous vehicles while modeling neighbor interactions and stochastic driving behavior.The model is designed for trajectories produced by IoT highway video surveillance systems.
  • 5 CONCLUSION: Experiments on simulation and real-world datasets show that DSAB captures spatial-temporal trajectory dynamics and detects a variety of abnormal behaviors.The conclusion describes single-vehicle anomaly identification performance as state of the art.
  • 5 CONCLUSION: DSAB scales to highway monitoring systems containing thousands of vehicles.The conclusion frames scalability as applying to large highway monitoring systems.
  • 5 CONCLUSION: Real-world HighD analysis shows lane-dependent speed norms, including average speeds of 76, 69, and 56 mph across the three lanes.These values are calculated from training data and provide context for interpreting qualitative anomalies.
  • 5 CONCLUSION: The paper is titled Detecting Socially Abnormal Highway Driving Behaviors via Recurrent Graph Attention Networks and includes a qualitative HighD traffic study.The supplied section materials identify Figure 3 as the qualitative study of real-world HighD traffic data.

A DATA DETAILS

The data details describe simulation scenarios and trajectory statistics used to evaluate abnormal driving detection. The setup includes normal traffic, abnormal behaviors, and scenario-specific traffic-flow considerations.

  • A DATA DETAILS: The appendix describes detailed simulation settings and data statistics for both simulation and HighD data.It introduces the organization of the data-details section before presenting scenario settings and counts.
  • A DATA DETAILS: Normal simulation traffic uses the Modified General Motors car-following model and desired-speed distributions calibrated to real-world traffic.The supplied settings include varying traffic demands from 500 to 1600 vehicles per lane per hour across free-flow and congested conditions.
  • A DATA DETAILS: Abnormal simulation scenarios are generated by changing vehicle behaviors, including speeding relative to surrounding traffic.The supplied passage begins the speeding scenario specification after describing normal traffic.
  • A DATA DETAILS: Stalled-car and comprehensive scenarios use higher traffic flow to test whether the source anomaly vehicle remains detectable despite upstream congestion.Table 4 reports car counts and trajectory counts after dividing trajectories into 10–15-second windows with 1-second stride.

B DESCRIPTION OF METRICS

The evaluation uses ROC-AUC, average precision, and Precision@k to measure anomaly discrimination and ranking. Metric choice accounts for the severe class imbalance created by the low anomaly rate.

  • B DESCRIPTION OF METRICS: ROC-AUC measures discrimination through the area under the ROC curve, where 0.5 indicates no discrimination and 1 indicates perfect anomaly detection.The ROC curve plots true positive rate against false positive rate.
  • B DESCRIPTION OF METRICS: Average precision summarizes the precision-recall curve as a weighted mean of precision values across thresholds.The weights are determined by the increase in recall from the previous threshold.
  • B DESCRIPTION OF METRICS: Precision@k measures the percentage of true anomalies among the top k samples ranked as anomalous.This metric directly evaluates how well a model prioritizes anomalies among its highest-scored samples.
  • B DESCRIPTION OF METRICS: Because only 3%–5% of vehicles are anomalous, the evaluation excludes F1 and emphasizes ROC-AUC and Precision@k with complementary caveats.F1 is sensitive to severe imbalance, while ROC-AUC can make distinctly performing models appear similar; Precision@k evaluates anomaly ranking.

C AGGREGATOR

Loss aggregation is tailored to detection granularity: vehicle scores are averaged over time, while scene scores are maximized across vehicles.

  • Vehicle detection averages losses over time because abnormal behaviors often persist for several seconds.Maximization decreases average precision by around 0.1.
  • Scene detection maximizes losses across vehicles so abnormal vehicles are not averaged out by normal vehicles.
  • Averaging is described as more conservative than maximizing for loss aggregation.

D ROBUSTNESS ANALYSIS

The model remains robust when training data contains a small number of anomalies, with comparable performance to training on clean data.

  • Performance differences between models trained on anomalous and clean data are all less than 0.05.
  • The two training sets produce similar results, with no clear trend favoring either dataset.
  • These findings indicate robustness to a small number of anomalies in the training data.

E HIGHD ANOMALY DESCRIPTION

Additional HighD examples show that top-ranked anomalies are usually associated with sharp deceleration or speeds differing from nearby vehicles or typical lane speeds.

  • Most top-ranked HighD anomalies involve drastic deceleration or speeds that differ from surrounding vehicles or typical lane speeds.
  • Figures 4 and 5 provide additional qualitative studies of top-ranked anomalies in real-world HighD traffic data.
  • Overlapping time-window samples are removed during ranking to improve the diversity of qualitative examples.
Loading 2304.11513v1…