Uncovering Abnormal Behavior Patterns from Mobility Trajectories

Hao Wu, Xuehua Tang, Zhongyuan Wang, Nanxi Wang

TL;DR

This paper introduces a novel method to detect and classify abnormal human behavior patterns from mobility trajectories, moving beyond simple anomaly detection to identify specific criminal preparation activities like wandering, scouting, random walking, and trailing. By combining deep learning (LSTM) for feature extraction with clustering (K-means) and spatio-temporal matching, the approach aims to enhance security and crime prevention by understanding the subtle precursors to criminal acts.

What Problem Does This Paper Address?

Current security systems heavily rely on video surveillance, which is often reactive and limited in its ability to provide advanced warnings. While abnormal behavior detection in video has seen significant research, it often lacks direct temporal and spatial movement information crucial for proactive security. This paper addresses this gap by leveraging the ubiquitous nature of smartphone GPS data to analyze personal mobility trajectories. The core challenge is to not only identify deviations from normal movement but to specifically recognize patterns that are indicative of criminal preparation, such as reconnaissance (scouting) or stalking (trailing), which are often precursors to more serious offenses.

Methods

The proposed method involves a multi-stage process:

  1. Data Preprocessing: Trajectory data is first cleaned to remove noise and abnormal points using a heuristic-based outlier detection method that considers speed and radiation range.
  2. Interest Area Extraction: Trajectories within a defined "interest area" are filtered to focus analysis on relevant locations.
  3. Abnormal Trajectory Screening: A Long Short-Term Memory (LSTM) network, specifically a SeqtoSeq model, is employed to extract dominant spatio-temporal features from the mobile trajectories. These features are then fed into a K-means clustering algorithm to initially identify potentially abnormal trajectories. Abnormal trajectories are identified as those belonging to clusters with a size below a certain threshold.
  4. Abnormal Pattern Identification: The screened abnormal trajectories are further analyzed to classify them into specific abnormal behavior patterns: wandering, scouting, random walking, and trailing.
  5. Wandering, Scouting, and Random Walking: These single-person behaviors are identified by analyzing motion types derived from trajectory features. The paper defines these patterns based on existing typologies and then uses feature matching to classify them.
  6. Trailing: This two-person behavior is identified through a more complex process involving front-behind relationship determination, spatio-temporal correlation calculation using semantic time periods (random walk, trailing, escape), and a Gaussian kernel function for final discrimination.

Datasets

  • Simulated Dataset (D1): This dataset contains four motion types (direct, pacing, lapping, random), with one thousand trajectories for each type, totaling four thousand trajectories. Gaussian white noise was added to increase authenticity.
  • Real Dataset (D2): This dataset was collected via a mobile phone application and comprises 6300 trajectories from 35 people over half a year, with data recorded at one-minute intervals.

Key Insights and Ideas

  • Beyond Anomaly Detection: The paper moves beyond simply flagging "unusual" trajectories to identifying specific, meaningful abnormal patterns that have direct implications for security and crime prevention.
  • LSTM for Feature Extraction: Utilizing LSTM's ability to capture sequential dependencies in trajectory data is crucial for extracting rich spatio-temporal features that are more informative than raw coordinates.
  • K-means for Initial Screening: K-means clustering provides an unsupervised way to group trajectories and identify potential anomalies without prior knowledge of what constitutes "normal."
  • Multi-faceted Trailing Detection: Recognizing that trailing involves two individuals requires a more sophisticated approach than single-person anomaly detection, incorporating front-behind relationships and semantic time period analysis.
  • Defining Criminal Precursors: The paper explicitly defines and operationalizes behaviors like scouting and trailing, which are often overlooked but critical indicators of criminal intent.

Key Results

  • The proposed trajectory-based abnormal behavior discovery method can achieve rapid discovery of abnormal trajectories and effective judgment of abnormal behavior patterns.
  • The LSTM and K-means methods effectively screen abnormal trajectories from large datasets without prior knowledge.
  • Motion Type Analysis (Simulated Data): The proposed method achieved an overall accuracy of 0.96 on the simulated dataset, outperforming seven other common distance-based models. However, the recognition accuracy for "pacing" was noted as low, potentially due to its similarity to "direct" motion and GPS inaccuracies.
  • Behavior Pattern Identification (Real Data):
    • Wandering, Scouting, Random Walking: The method achieved an average accuracy of 77% for these single-person behaviors. Wandering behavior showed a very high recall rate (0.98) but lower precision (0.77), while scouting had lower recall (0.48) and precision (0.73).
  • Trailing: The highest precision (0.88) and recall (0.77) for trailing recognition were achieved with a JSD model threshold of 0.75.

Limitations and Future Work

  • Non-Parallel Algorithm: The current method is not parallel, which can lead to inefficiencies when dealing with very large trajectory datasets.
  • Trajectory Interception Defects: The paper acknowledges difficulty in accurately determining the specific time range of an abnormality within a trajectory. A trajectory might only be abnormal for a portion of its duration.
  • Limited True Abnormal Trajectories: The scarcity of collected true abnormal trajectories made it challenging to precisely determine the optimal split threshold for abnormal segments.
  • Future Work: The authors suggest that as the number of abnormal trajectories increases, they can better analyze and address the issue of precise split threshold determination. They also aim to improve the efficiency of screening abnormal trajectories.

Summary updated

Abstract

Using personal trajectory information to grasp the spatiotemporal laws of dangerous activities to curb the occurrence of criminal acts is a new opportunity and method for security prevention and control. This paper proposes a novel method to discover abnormal behaviors and judge abnormal behavior patterns using mobility trajectory data. Abnormal behavior trajectory refers to the behavior trajectory whose temporal and spatial characteristics are different from normal behavior, and it is an important clue to discover dangerous behavior. Abnormal patterns are the behavior patterns summarized based on the regular characteristics of criminals’ activities, including wandering, scouting, random walk, and trailing. This paper examines the abnormal behavior patterns based on mobility trajectories. A Long Short-Term Memory Network (LSTM)-based method is used to extract personal trajectory features, and the K-means clustering method is applied to extract abnormal trajectories from the trajectory dataset. Based on the characteristics of different abnormal behaviors, the spatio-temporal feature matching method is used to identify the abnormal patterns based on the filtered abnormal trajectories. Experimental results showed that the trajectory-based abnormal behavior discovery method can realize a rapid discovery of abnormal trajectories and effective judgment of abnormal behavior patterns.
BibTeX
@inproceedings{wu2021uncovering,
  author = {Hao Wu and Xuehua Tang and Zhongyuan Wang and Nanxi Wang},
  title = {{Uncovering Abnormal Behavior Patterns from Mobility Trajectories}},
  booktitle = {Sensors},
  year = {2021},
  url = {https://openreview.net/forum?id=24hKE3a1Qm},
}
Figure