Research map / Time series / Foundation models for time series
Hidden Follower Detection: How Is the Gaze-Spacing Pattern Embodied in Frequency Domain?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HFDF-TF, a hidden follower detection framework that converts gaze-state and distance time series into MFCC frequency-domain features, selects them by SDR-based fidelity, and fuses two TWIESN classifiers with DS evidence theory.
Problem
Hidden follower detection from surveillance trajectories and gaze behavior; time-domain representations of gaze and spacing series are irreversible and lose information.
Approach
Defines a fidelity (1 - SDR) measure to show frequency-domain features (MFCC of gaze and spacing series) preserve more information than time-domain features, uses fidelity to select MFCC settings, and fuses separate gaze and spacing TWIESN classifiers via Dempster-Shafer decision fusion.
Key results
- On Real-HFD, HFDF-TF achieves F1 94.5% and AUC 94.4%, versus 91.7% F1 for HFDF-GS.
- Frequency-domain fidelity (gaze 73%, spacing 77%) exceeds time-domain fidelity (gaze 52%, spacing 60%).
- Gaze-only module improves about 9-10% over HFDF-G; CQCC features outperform MFCC but are roughly twice as slow.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace HFDF-GS with HFDF-TF | Hidden follower detection · Real-HFD | F1: 91.7 -> 94.5 | Helped |
| replace HFDF-GS with HFDF-TF | Hidden follower detection · Sim-HFD | F1: 91.4 -> 95.2 | Helped |
| add the gaze MFCC module to the spacing-only model (ablation) | HFDF-TF · Real-HFD | F1: 93.9 -> 94.5 | Helped |
| add the gaze MFCC module to the spacing-only model (ablation) | HFDF-TF · Sim-HFD | F1: 94.6 -> 95.2 | Helped |
| use gaze MFCCs only instead of gaze and spacing MFCCs (ablation) | HFDF-TF · Real-HFD | F1: 94.5 -> 80.6 | Hurt |
| use gaze MFCCs only instead of gaze and spacing MFCCs (ablation) | HFDF-TF · Sim-HFD | F1: 95.2 -> 88.6 | Hurt |
| use spacing MFCCs only instead of gaze and spacing MFCCs (ablation) | HFDF-TF · Real-HFD | F1: 94.5 -> 93.9 | Hurt |
| use spacing MFCCs only instead of gaze and spacing MFCCs (ablation) | HFDF-TF · Sim-HFD | F1: 95.2 -> 94.6 | Hurt |
| replace decision fusion with feature concatenation and one TWIESN (ablation) | HFDF-TF · Real-HFD | F1: 92.6 -> 94.5 | Hurt |
| replace decision fusion with feature concatenation and one TWIESN (ablation) | HFDF-TF · Sim-HFD | F1: 93.4 -> 95.2 | Helped |
| replace MFCCs with CQCCs (ablation) | HFDF-TF · Real-HFD | F1: 94.5 -> 94.9 | Helped |
| add multimodal decision fusion between HFDF-GS and HFDF-TF (ablation) | HFDF-TF · Real-HFD | F1: 94.5 -> 95.3 | Helped |
Surprising findings
- Simple feature concatenation of gaze and spacing MFCCs degrades performance relative to decision fusion (HFDF-ND 92.6 F1 vs HFDF-TF 94.5 on Real-HFD).
- Spacing-pattern gains in frequency domain are much smaller than gaze-pattern gains, attributed to trajectory tracking accuracy and behavior constraints.
Limitations
- Real-world hidden following data is rare, so training relies heavily on simulated data (Sim-HFD).
- Small real dataset (20 pedestrians, 160 minutes).
- CQCC features are more accurate but computationally costly, limiting practical use.
Experimental setup
Datasets: Real-HFD Sim-HFD
Benchmarks: Real-HFD Sim-HFD
Metrics: Precision Recall F1 Accuracy AUC SDR-based fidelity
Baselines: Trajectory-based following detection (Li et al. 2013) HFDF HFDF-GS
Builds on: HFDF-GS TWIESN MFCC Dempster-Shafer theory
Closest papers by meaning
- MTGS: A Novel Framework for Multi-Person Temporal Gaze Following and Social Gaze Prediction (NeurIPS 2024 · similarity 0.83)
- EyeGraph: Modularity-aware Spatio Temporal Graph Clustering for Continuous Event-based Eye Tracking (NeurIPS 2024 · similarity 0.81)
- Toward Semantic Gaze Target Detection (NeurIPS 2024 · similarity 0.81)
- SocialCircle: Learning the Angle-based Social Interaction Representation for Pedestrian Trajectory Prediction (CVPR 2024 · similarity 0.81)
- DiffEye: Diffusion-Based Continuous Eye-Tracking Data Generation Conditioned on Natural Images (NeurIPS 2025 · similarity 0.81)
- Cross-Domain Trajectory Association Based on Hierarchical Spatiotemporal Enhanced Attention Hypergraph (AAAI 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.