Abstract
Detecting "hidden followers"—individuals who secretly track a target—is a critical task for preventing crime, yet it remains difficult because their behavior often mimics that of normal pedestrians. Traditional detection methods rely on time-domain summaries of gaze and spacing, which often discard subtle behavioral nuances. This paper proposes the HFDF-TF framework, which leverages frequency-domain analysis via Mel-Frequency Cepstral Coefficients (MFCCs) to capture the "rhythm" of following behavior. We demonstrate that frequency-domain features maintain higher information fidelity, leading to a direct improvement in detection accuracy. Our model achieves a state-of-the-art F1 score of 94.5% on real-world datasets, with a notable 10.1% jump in gaze-based detection performance.
Introduction
Hidden following is a subtle abnormal behavior that often precedes violent crimes. Unlike overt actions like running or fighting, a hidden follower attempts to blend in, making detection by standard computer vision systems extremely challenging. Previous research has focused on time-domain features like gaze frequency and average distance. However, these methods suffer from irreversibility—the inability to reconstruct the original signal from the summary—and one-sidedness, failing to capture the complex interaction between the follower and the target.
Methodology: The HFDF-TF Framework
The authors introduce the Hidden Follower Detection based on Time-Frequency transform (HFDF-TF) framework. This approach shifts the focus from simple time-domain statistics to the frequency domain. Key components include:
- Fidelity Estimation: Using Signal-to-Distortion Ratio (SDR), the authors prove that frequency-domain features (like MFCCs) retain significantly more information about the original behavior than time-domain summaries.
- Feature Extraction: MFCCs, commonly used in speech processing, are extracted from continuous gaze degree scores and spacing time series. This captures the rhythmic patterns of the behavior.
- Decision Fusion: Two independent Time Warping Invariant Echo State Network (TWIESN) classifiers process gaze and spacing data separately. Their results are fused using Dempster-Shafer (DS) evidence theory to determine the final probability of a hidden follower.
The Fidelity-Performance Correlation
A key finding of this research is the strong positive correlation between feature fidelity and detection performance. As the fidelity of the extracted features increases (measured by SDR), the F1 score of the detection model also improves. This suggests that the "essence" of following behavior is contained in the high-frequency components that time-domain methods typically ignore.
Frequency Signatures of Following Behavior
Hidden followers exhibit unique frequency patterns. They must balance the need to keep the target in sight with the need to avoid being noticed. This conflict manifests as specific high-frequency aggregations in gaze and spacing signals. For instance, while a normal pedestrian's gaze might be random or steady, a follower's gaze has a distinct "rhythm" of checking and looking away, which is clearly visible in the frequency domain.
Experimental Results
The model was evaluated on two datasets: Real-HFD (real-world video) and Sim-HFD (simulated data).
- Overall Performance: The HFDF-TF model achieved a 94.5% F1 score on Real-HFD, a 3% improvement over previous state-of-the-art methods.
- Gaze Breakthrough: The switch to frequency-domain analysis resulted in a 10.1% improvement in the gaze module's accuracy alone.
- Robustness: High performance across both real and simulated data indicates that these frequency-domain patterns are universal indicators of hidden following behavior.
Conclusion
By moving from the time domain to the frequency domain, this research provides a more robust and informative way to detect hidden followers. The use of MFCCs and the emphasis on feature fidelity offer a new path forward for identifying subtle abnormal behaviors in surveillance, potentially preventing crimes before they occur.
Summary updated