PaperlayerView exact revision on arXiv

Source-linked AI summary

Anomaly Detection in Road Traffic Using Visual Surveillance: A Survey

Santhosh Kelathodi Kumaran, Debi Prosad Dogra, Partha Pratim Roy

arXiv:1901.08292v1cs.CV

TL;DR

Road-traffic anomaly detection requires methods that learn normal behavior from visual surveillance while accommodating contextual anomalies and practical deployment constraints. This survey reviews prior surveys and road-focused techniques by their learning methods, features, approaches, scenarios, and anomaly types, then identifies challenges and future directions. It emphasizes that real-life datasets, long-duration applicability, robust tracking, illumination-independent features, and environmental context remain important concerns.

  • Problem

    Road-traffic surveillance needs anomaly detection for deviations involving vehicles, pedestrians, environments, and their interactions, but anomaly notions are contextual and real-world deployment remains constrained by data and sensing conditions.

  • Method

    The paper surveys prior computer-vision reviews and road-traffic anomaly-detection studies, organizing them by learning methods, features, detection approaches, application scenarios, and anomaly types.

  • Results

    The survey synthesizes methods for static-camera road surveillance and identifies recurring challenges involving benchmark realism, tracking, illumination, background context, pose, perspective, and curtailed tracks.

  • Takeaways & Limitations

    Future progress depends on real-life datasets, hybrid deep-learning frameworks, and traffic-analysis systems capable of handling diverse scenarios and reporting events of interest.

Abstract

from arXiv · show

Computer vision has evolved in the last decade as a key technology for numerous applications replacing human supervision. In this paper, we present a survey on relevant visual surveillance related researches for anomaly detection in public places, focusing primarily on roads. Firstly, we revisit the surveys done in the last 10 years in this field. Since the underlying building block of a typical anomaly detection is learning, we emphasize more on learning methods applied on video scenes. We then summarize the important contributions made during last six years on anomaly detection primarily focusing on features, underlying techniques, applied scenarios and types of anomalies using single static camera. Finally, we discuss the challenges in the computer vision related anomaly detection techniques and some of the important future possibilities.

I. INTRODUCTION

Computer vision supports surveillance-based understanding of road scenes, including congestion, accidents, and other anomalies. The survey reviews how anomaly detection learns normal behavior and identifies deviations in road-traffic settings.

  • Visual data provides information for detecting congestion, accidents, and other road-traffic anomalies while supporting traffic-status statistics.
  • A typical detection framework extracts descriptors, represents normal behavior with rules, models, or repositories, and produces anomaly scores or labels.
  • Anomalies are deviations of vehicles, pedestrians, environments, or their interactions from learned normal behavior.
  • Road-traffic anomaly examples include walkway vehicle presence, jaywalking, signal bypassing, sudden dispersal, and U-turns during red signals.
  • The survey focuses on visual-surveillance techniques for road entities including vehicles, pedestrians, environments, and their interactions.

A. Recent Surveys

Recent surveys cover complementary computer-vision surveillance topics, including detection, tracking, scene modeling, activity analysis, and anomaly detection. This survey narrows that broader literature to anomaly detection relevant to road-traffic scenarios and organizes methods by data and modeling perspectives.

  • Earlier surveys addressed object detection, tracking, scene modeling, activity analysis, intelligent transportation systems, and multi-sensor traffic analysis.
  • Prior work also reviewed anomaly detection through sensors, entities, feature extraction, learning methods, and scene modeling.
  • The survey focuses specifically on anomaly-detection studies relevant to road-traffic scenarios.
  • Because anomalies are contextual, the paper analyzes methods by scene representation, features, models, and approaches rather than assuming universal applicability.
  • The paper proceeds from background and visual scene learning to approaches, features, applications, critical analysis, challenges, and future possibilities.

II. COMPUTER VISION GUIDED ANOMALY DETECTION STUDIES

The paper treats anomaly detection as identifying data patterns that depart from normal behavior, using feature descriptors and several contextual classifications. In visual surveillance, anomalies may be local or global within frames or video segments.

  • Features are represented as descriptors, placing data in a multidimensional space determined by descriptor length.
  • Anomalies are patterns that do not conform to a well-defined notion of normal behavior and are also called outliers or novelties.
  • Point anomalies differ from the usual distribution, contextual anomalies depend on circumstances, and collective anomalies arise from groups of individually normal instances.
  • Visual-surveillance anomalies may be global within a frame or video segment, or local within a specific scene area.
  • Some methods can detect both global and local anomalies.

2) Challenges and Scope of Study:

Anomaly detection faces challenges from evolving, noisy, context-dependent data and limited training resources. The survey scope is static-camera video, while its learning-method review spans supervised, unsupervised, semi-supervised, and reinforcement learning.

  • 2) Challenges and Scope of Study:: Key challenges include defining normal behavior, separating unclear boundaries, handling context variation, limited data, noisy sensing, and evolving behavior.
  • 2) Challenges and Scope of Study:: The survey analyzes videos captured through a single static camera; multi-camera detection introduces additional challenges and potentially different frameworks.
  • B. Learning Methods: Supervised learning builds normal profiles from labeled data, whereas unsupervised learning derives them from relationships in unlabeled data.
  • B. Learning Methods: Semi-supervised learning combines abundant unlabeled data with a small amount of labeled supervision, useful when manual labeling is expensive.
  • B. Learning Methods: Learning methods support feature extraction, detection, classification, recognition, segmentation, tracking, re-identification, interaction analysis, and anomaly detection.

C. Anomaly Detection Approaches

Anomaly detection approaches are organized by how they learn or represent normal behavior, including model-based methods and other learning strategies.

  • Anomaly detection approaches are classified according to different underlying approaches.
  • Model-based: Model-based approaches learn normal behavior by representing data with a parameter set.Statistical methods generally estimate parameters by fitting data to a stochastic model.
  • Model-based: Statistical approaches may be parametric or non-parametric, with parametric methods assuming a distribution and probability density function for normal data.
  • Examples of learning methods used in anomaly detection are summarized in Table III.

2) Proximity-based:

Proximity-, distance-, density-, and classification-based methods identify anomalies through relationships to neighboring data or learned classes.

  • Proximity-based: Proximity-based approaches decide whether observations are anomalous according to how close they are to their neighbors.
  • Distance-based: Distance-based approaches assume that normal data occupy dense neighborhoods.
  • Density-based: Density-based approaches compare a point’s density with the density of its local neighbors and compute an outlier score.
  • Classification-based: Classification-based methods use classifiers to distinguish normal and anomalous classes in a feature space.They include one-class and multi-class techniques; multi-class methods require labeled normal and anomalous instances.

4) Prediction-based:

Prediction, reconstruction, clustering, and feature-based techniques provide complementary ways to detect anomalies in visual-surveillance data.

  • Prediction-based: Prediction-based approaches detect anomalies from variation between predicted and actual spatio-temporal feature characteristics.HMM and LSTM models are examples of this approach.
  • Reconstruction-based: Reconstruction-based techniques measure anomaly using reconstruction error after embedding data into a lower-dimensional subspace.Examples include sparse coding, autoencoders, and PCA-based approaches.
  • Clustering-based: Clustering approaches assume normal data form clusters, while anomalies remain outside clusters or occur in smaller clusters.
  • Features Used in Anomaly Detection: Video anomaly detection requires extracting relevant features because video consists of sequences of frames.Feature choice affects the capability to detect specific anomalies.
  • Features Used in Anomaly Detection: Features may derive from foreground objects, pixel-level histograms, detected objects, trajectories, or automatically extracted DNN representations.
  • Features Used in Anomaly Detection: Representative work using different features is summarized in Table V, while Table VI summarizes work by applied area, anomaly type, technique, and dataset.
  • Features Used in Anomaly Detection: Features are broadly classified as object-oriented or non-object-oriented, with hybrid features also used for anomaly detection.

E. Applied Areas

The survey covers anomaly detection research across road and public-place scenes, including road segments, junctions, parking areas, highways, and pedestrian paths.

  • Applied scenes: Applied scenes include road segments, junctions, parking areas, highways, and pedestrian paths.The survey highlights techniques, applicable scenes, anomaly types, and datasets.
  • Datasets: Frequently used datasets include QMUL, CAVIAR, UCSD, Bellview, Person, UMN, and ARENA.

F. Online vs. Offline

Road-traffic anomaly detection research predominantly targets online use because it can produce instantaneous results, while offline analysis is slower and near-real-time methods require test-video segmentation.

  • Most anomaly-detection techniques target online usage, generating instantaneous results for road-traffic surveillance.
  • Near-real-time methods detect anomalies only after segmenting test videos from the real scene.
  • Offline methods are also used in road networks, but their results are not immediate, particularly for data analysis.

III. CRITICAL ANALYSIS

The survey identifies practical limitations that restrict anomaly-detection methods across real-world traffic settings, including dataset realism, illumination, viewpoint, object diversity, density, tracking, and environmental context.

  • Benchmark datasets may not represent real-life situations, and methods that work on Avenue can produce higher false alarms on QMUL.
  • Generic methods applicable across varied datasets and illumination-independent feature extraction remain limited.
  • Camera viewpoint changes vehicle appearance with distance, complicating detection and tracking under pose and perspective variations.
  • Modeling heterogeneous objects and their movements within one scene can be difficult.
  • Methods differ between sparse and dense scenes, and dense-scene approaches can generate many false negatives.
  • Trajectory-based detection depends on accurate tracking, yet existing trackers can fail across scenarios and under occlusion.
  • Real-life datasets are needed to evaluate anomaly-detection effectiveness and support analysis of diverse traffic conditions.

IV. CONCLUSION

The survey revisits computer-vision research and analyzes road-traffic anomaly detection through data, learning techniques, features, approaches, and application scenarios. It uses identified gaps to outline future research directions.

  • The survey treats data as the primary unit for detailing learning techniques, features, anomaly-detection approaches, and application scenarios.
  • It examines anomaly detection for vehicles, people, and their interactions with the environment in road networks.
  • The paper discusses current technique gaps to propose future directions and possibilities for computer-vision-based anomaly detection.
Loading 1901.08292v1…