Results for supervised path

The features obtained from the sparse components of the video were stacked together to form a 21 dimensional feature vector. A feature cluster was built using multiple videos from the training samples as they did not have any anomalies. For any test video, the feature vector is calculated and its Mahalanobis distance from the feature cluster is obtained. If this distance is greater than a fixed threshold, the video is considered to be anomalous. In most experiments, the Mahalanobis distance for anomalous video was more than 400 whereas for a typical video it was around 250.

Following video is anomalous with a Mahalanobis distance of 419.3703. The bounding boxes are obtained using segmentation of the sparse components.

The following video does not contain any anomaly. Hence, its Mahalanobis distance is low (259.2747)