We have implemented the OCNM algorithm here using the kth nearest-neighbor distance as the sparsity measure.
We apply OCNM here in a sliding window manner, with the window advancing by one when a new image (when the algorithm is run locally at each node in a distributed monitoring architecture) or string of images (when the algorithm is run at the central repository in a centralized monitoring architecture) [x.sub.t], arrives in the next timestep.
If the kth nearest-neighbor distance is used as the sparsity measure and OCNM is run using a sliding window of size W, the algorithm must first evaluate the sparsity measure value of [x.sub.t] in the window of points.
Figure 2 compares the performances of OCNM with PCA and Au's NCD-based algorithms through ROC curves, demonstrating the tradeoff between the Probability of False Alarms ([P.sub.FA]) and the Probability of Detection ([P.sub.D]).
Figure 3 (top panel) shows the particular timesteps that OCNM flags as anomalous, using the representative value of [mu] = 0.90 set to identify the 10% outliers.