Two different approaches for detection of anomalies in stationary camera videos have been analyzed. We first used Robust PCA to separate out the low rank background from the moving sparse components. Further processing like object segmentation and feature computation was done using the sparse component of the video. For each segmented object, shape and motion features were obtained and tracked for all the frames. A feature cluster was made using the training samples containing no anomalies. In this way, an anomalous video would have a higher Mahalanobis distance from the feature cluster. Secondly, sparse subspace clustering concept was used to distinguish anomalies based on motion. This method required a set of points from both normal and anomalous objects, but spared any usage of training samples. The est results were obtained when RPCA was used for background separation, and SSC was used for clustering subspaces in the sparse components.
As a future work, we would like to analyze the low rank + sparse subspace clustering further. Also, a better localization of anomalies in the given video will be targeted. The concept can be applied to other fields such as video summarization. It can find its use in compressing a lengthy crime scene footage into a short and meaningful duration in which the crime would have happened. Similarly, a lengthy sports matched can be covered by seeing only its highlights which can be found automatically by the amount of sparse content in a chunk of frames.