Strictly Causal Streaming Video Anomaly Detection with a Theoretically-Grounded State-Space Core
A research paper introduces a strictly causal streaming video anomaly detector using a diagonal linear state-space recurrence with an input and state dependent decay gate, trained self-supervised through causal next embedding prediction on a frozen visual backbone. It reports O(1) time and memory per frame with no lookahead or clip buffering. The paper derives a closed-form relationship between recurrence decay spectrum and detection delay/shortest detectable anomaly, and validates on UCSD Ped2 and CUHK Avenue. Measured detection delays were 1.6 and 18.4 frames, while the settling delay bound from learned base decay was 57-59 frames, indicating the event boundary gate governs responsiveness.
A new arXiv paper proposes a strictly causal streaming video anomaly detection method based on a state-space model with O(1) per-frame updates, no buffering, and a theoretical link between decay spectrum and detection latency. Experiments on UCSD Ped2 and CUHK Avenue show detection delays of 1.6 and 18.4 frames, far below the settling delay bound predicted from base decay, highlighting the role of the event boundary gate.
The method uses a diagonal linear state-space recurrence with input- and state-dependent decay gating, trained via causal next embedding prediction on a frozen visual backbone. The theoretical analysis connects decay spectrum to detection delay and minimum detectable anomaly duration. Empirical results show detection delays (1.6, 18.4 frames) much lower than the settling delay bound (57-59 frames), suggesting the event boundary gate, not base decay, controls responsiveness.
This approach targets edge hardware for real-time video anomaly detection, addressing limitations of prior Mamba-based methods that buffer clips and benchmark only GPU throughput. The O(1) per-frame state update and no lookahead make it suitable for low-latency surveillance and industrial monitoring applications.
The method could enable real-time anomaly detection on resource-constrained edge devices, reducing latency and memory requirements compared to clip-buffering approaches. This may lower deployment costs and improve responsiveness for security and industrial applications.
Next observable signals include publication of full benchmark results on edge hardware, comparisons with existing streaming anomaly detectors, and potential adoption in surveillance or industrial monitoring systems. Further theoretical work may refine the relationship between decay gating and detection latency.