A video analytics framework for amorphous and unstructured anomaly detection

Video surveillance systems are often used to detect anomalies: rare events which demand a human response, such as a fire breaking out. Automated detection algorithms enable vastly more video data to be processed than would be possible otherwise. This note presents a video analytics framework for the detection of amorphous and unstructured anomalies such as fire, targets in deep turbulence, or objects behind a smoke-screen. Our approach uses an off-line supervised training phase together with an on-line Bayesian procedure: we form a prior, compute a likelihood function, and then update the posterior estimate. The prior consists of candidate image-regions generated by a weak classifier. Likelihood of a candidate region containing an object of interest at each time step is computed from the photometric observations coupled with an optimal-mass-transport optical-flow field. The posterior is sequentially updated by tracking image regions over time and space using active contours thus extracting samples from a properly aligned batch of images. The general theory is applied to the video-fire-detection problem with excellent detection performance across substantially varying scenarios which are not used for training.