Track inspector
click a person in any feedEvery channel is shown in the model's warped units against the GP's one-step-ahead forecast band (mean ± 2σ). Red points are innovations beyond 3σ.
Scenario control
Inject an event into
Out-of-place behaviour
Concealed-carry signature
Reality glitch
Kernel Lab
normalcy model per channelCACK models every signal with a Gaussian process whose covariance kernel is discovered, not chosen. An evolutionary search combines base kernels with + and ×; a reasoning step reads each channel's spectrum and the innovations of its best model and proposes what is still missing.
Reasoning trace
Research
CACK-Sim and public benchmarksCACK is evaluated on CACK-Sim, a procedural benchmark of crowd scenes with frame-exact labels for ten anomaly types in three classes, at two difficulty tiers, and on three public video anomaly benchmarks. Every method is calibrated on the same normal footage and scored on the same test footage.
Mean over sites and seeds. TPR is the detection rate of anomalous windows at a fixed false-alarm rate on normal windows. Full details, ablations and limitations are in the paper.
Public benchmarks
Frame-level AUROC, offline / online, on ShanghaiTech, CUHK Avenue and UCSD Ped2, with the same detector, tracker and windows for every method. Here CACK is on par with classical detectors and below the state of the art; its strengths are calibrated, explained evidence, feed integrity and speed.
Read the paper
CACK: Evolving Gaussian-Process Normalcy Models for Real-Time Detection and Classification of Crowd and Video-Feed Anomalies