Description
Identifying observations that deviate substantially from the expected behavior of a system, where "expected" is learned from past data. In the wireless-sensing / crowd-modeling context, anomaly detection has three distinct flavors: (1) crowd anomalies — sudden density spikes, counter-flows, panic onset; (2) signal anomalies — CSI corruption from hardware faults or jamming; (3) behavioral anomalies — a person doing something atypical for the time and place. Anomaly detection is the core dependency of crowd-safety alerting and a frequent companion task to fall-detection.
Why it's hard
- "Anomalous" is context-dependent — a 2 m/s walking speed is normal in a corridor and anomalous in a queue.
- Class imbalance is extreme: anomalies are by definition rare.
- Concept drift in the "normal" baseline (see calibration-drift) creates spurious anomalies.
- False alarms erode operator trust faster than missed detections.
- Multi-scale anomalies (a person, a sub-crowd, the whole venue) require hierarchical models.
Common approaches
- Reconstruction-error-based detectors (autoencoders, predictive models) on CSI or trajectory streams.
- One-class classifiers / isolation forests for tabular features.
- Crowd-entropy and aggregate-density-gradient threshold detectors.
- Vision baseline detectors for comparison; CSI / radar variants for privacy-respecting deployment.
Source Papers
- bendalibraham2021_476e ↗ — recent trends in crowd analysis (anomaly detection a major axis).
- sreenu2019_6f76 ↗ — intelligent video surveillance review (vision baseline).
- torun2026_72aa ↗ — fast and robust stationary crowd counting with WiFi (anomaly framing).
- bahamid2026_a88a ↗ — crowd-entropy-based prediction model.