Exploring discriminative features for anomaly detection in public spaces

Shriguru Nayak, Archan Misra, Kasthuri Jayarajah, Philips Kokoh Prasetyo, Ee‐Peng LimProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2015 · DOI

TL;DR

This paper explores how to detect unusual events in public spaces by analyzing spatiotemporal patterns in crowd behavior. They propose using features derived from both physical movement data (occupancy counts) and social media activity (Twitter data) to identify anomalies, aiming to build a more comprehensive urban situational awareness system.

What Problem Does This Paper Address?

The paper tackles the challenge of real-time urban situational awareness, specifically the detection of unanticipated or rapidly evolving events in public spaces. Traditional methods often rely on single data sources or pre-defined event models. This work aims to bridge this gap by leveraging diverse data streams – physical movement and social media – to identify a wider range of events, from large-scale festivals to smaller, less conspicuous gatherings. The core problem is to move beyond simply observing crowd density to understanding the behavioral patterns that signify an event.

Methods

The authors propose a multi-faceted approach to anomaly detection, focusing on two primary data sources and deriving distinct feature sets:

  • Occupancy Count-Based Anomalies:
    • Sliding Window Analysis: The underlying occupancy time series is divided into overlapping windows (2 hours in this study) to capture temporal dynamics.
  • Magnitude Anomalies: Features like maximum and minimum residency count within a window, and the highest rate of change between consecutive time bins, are used to detect "overflow" (e.g., large events) and "underflow" (e.g., evacuations) events.
  • Trend Anomalies: The slope of the regression line fitted to occupancy over time within a window (MRL) is used to identify unusual trends.
  • Overall Anomaly Detector: A composite anomaly likelihood is computed by treating the derived features as a multivariate Gaussian distribution. A window is flagged as anomalous if its feature vector's probability falls below a set threshold.
  • Group-Driven Anomaly Detection:
    • Group Interaction Features: This approach utilizes a group detection system (GruMon) to identify movement patterns of groups. Features derived from group information, such as group size distribution and interaction duration, are explored to detect smaller, less attendance-heavy events.
  • Social Media-Based Anomaly Detection:
    • Geotagged Twitter Data: The authors analyze spatio-temporal patterns in Twitter activity, focusing on tweet volumes and content.
  • Snowball Sampling: A method to identify users from a targeted city is employed to avoid spurious event detection.
  • Statistical Features: Similar to occupancy data, statistical features (e.g., standard deviation of tweet volumes) are applied to identify deviations from normal patterns, exemplified by detecting the Formula One race.

Datasets

  • SMU LiveLabs Indoor Movement Data: Longitudinal observations of indoor on-campus movement data of individuals at Singapore Management University (SMU).
  • SMU Palanteer Software Data: Near-real-time traces of Twitter activity (including geotagged location information) from Singapore-based users.
  • Event Database: A ground-truth database maintained by the LiveLabs platform for evaluating event detection.

The paper does not introduce new benchmark datasets but utilizes existing real-world data from SMU's campus.

Key Insights and Ideas

  • Feature Diversity is Crucial: Different types of events are best detected by different features. Large events are captured by occupancy volume, while smaller, affinity-based events might require group interaction data.
  • Social Media as a Complementary Sensor: Twitter data, when analyzed with spatio-temporal features, can effectively complement physical movement data for detecting localized urban events, especially those with significant online buzz.
  • "Overflow" and "Underflow" are Detectable: Simple occupancy metrics can reliably identify events that cause significant deviations from normal crowd levels, both increases and decreases.
  • Group Dynamics Reveal Hidden Events: Analyzing group interaction patterns offers a promising avenue for detecting events that might be missed by purely occupancy-based methods, suggesting that the structure of crowd movement is as important as its density.
  • Preliminary but Promising: The results, while described as preliminary, demonstrate the potential of combining these diverse data streams for a more holistic understanding of urban activity.

Key Results

  • Occupancy-Based Detection: The occupancy-based anomaly detector showed good performance in identifying large-scale events like "Vivace" and "Food Festival" (overflow events) and "Fire Drill Evacuation" (underflow event).
  • Accuracy for classifying working vs. non-working days was high (98.6% for Study Area, 95.07% for Transient Area).
  • For detecting specific events in the T-junction area, the model achieved an accuracy of 71.4% for false negatives and 28.6% for false positives, with 5 predicted events.
  • Social Media Detection: The analysis of Twitter data successfully identified the Formula One race event based on a significant deviation in tweet volume from a specific grid cell.
  • Group-Based Detection: Preliminary results suggest that group interaction features can help uncover smaller events (e.g., Fencing Clinic) that are not distinguishable from the mean profile using only occupancy data.

Limitations and Future Work

  • Limitations:
    • The occupancy-based detector struggled with smaller, less attended events (e.g., CCA sign-ups, community service events).
  • The social media analysis was demonstrated with a single event (Formula One), and the "likely false positive" rate for occupancy-based detection suggests potential for misclassification.
  • The current work is described as preliminary, indicating that the models are still under development.
  • Future Work:
    • Enhance Group-Based Features: Explore a wider range of group-related features (e.g., distribution of group sizes, strength of ties, duration of interactions) for more robust detection.
  • Multi-Channel Social Media Fusion: Combine data from multiple social media platforms (e.g., Twitter and Instagram) to improve timeliness and location resolution.
  • Unified Framework: Develop a comprehensive framework that integrates data from mobile sensors and social media for a unified urban event detection system.
  • Investigate Other Data Streams: Potentially incorporate other sensor data for a richer understanding of urban dynamics.

Summary updated

Abstract

Context data, collected either from mobile devices or from user-generated social media content, can help identify abnormal behavioural patterns in public spaces (e.g., shopping malls, college campuses or downtown city areas). Spatiotemporal analysis of such data streams provides a compelling new approach towards automatically creating real-time urban situational awareness, especially about events that are unanticipated or that evolve very rapidly. In this work, we use real-life datasets collected via SMU's LiveLabs testbed or via SMU's Palanteer software, to explore various discriminative features (both spatial and temporal - e.g., occupancy volumes, rate of change in topic{specific tweets or probabilistic distribution of group sizes) for such anomaly detection. We show that such feature primitives fit into a future multi-layer sensor fusion framework that can provide valuable insights into mood and activities of crowds in public spaces.
BibTeX
@inproceedings{nayak2015exploring,
  author = {Shriguru Nayak and Archan Misra and Kasthuri Jayarajah and Philips Kokoh Prasetyo and Ee‐Peng Lim},
  title = {{Exploring discriminative features for anomaly detection in public spaces}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2015},
  doi = {10.1117/12.2184316},
}
Figure