Unsupervised learning of feature anomalies for a video surveillance system

摘要

Techniques are disclosed for analyzing a scene depicted in an input stream of video frames captured by a video camera. In one embodiment, e.g., a machine learning engine may include statistical engines for generating topological feature maps based on observations and a detection module for detecting feature anomalies. The statistical engines may include adaptive resonance theory (ART) networks which cluster observed position-feature characteristics. The statistical engines may further reinforce, decay, merge, and remove clusters. The detection module may calculate a rareness value relative to recurring observations and data in the ART networks. Further, the sensitivity of detection may be adjusted according to the relative importance of recently observed anomalies.

主权项

1. A computer-implemented method for analyzing a scene, the method comprising:
receiving kinematic and feature data for an object in the scene; determining, via one or more processors, a position-feature vector from the received data, the position-feature vector representing a location and one or more feature values at the location; retrieving a feature map corresponding to the position-feature vector, wherein the feature map includes one or more position-feature clusters, wherein the feature map includes one or more adaptive resonance theory (ART) network clusters; determining a rareness value for the object based at least on the position feature vector and the feature map, wherein the rareness value is determined based on at least a pseudo-Mahalanobis distance of the position-feature vector to a closest one of the clusters and on statistical relevance of clusters less than a threshold distance from the position-feature vector; and reporting the object as anomalous if the rareness value meets given criteria.