Results of MDT based Anomaly Detection and Localization in Crowded Scenes

To evaluate the performance of the MDT based anomaly detection, we compared it to three other recently proposed representations - the social force model (denoted SF), the mixture of optical flow (denoted MPPCA) and the optical flow monitoring method of Amit Adam et al. The video is available in Quicktime format (H.264), and a most recent version of Quicktime is required. Quicktime is a free software and can be downloaded from here.

Clips showing comparison between anomaly detections by the MDT (proposed) and other competitors (from left to right: temporal MDT/spatial MDT/MPPCA/social force/optic flow monitor)

Ped1 - Biker [mov] Ped1 - Cart [mov] Ped1 - Wheelchair [mov] Ped1 - Skater [mov] Ped2 - Biker [mov] Ped2 - Skater [mov]

Representative frames showing anomaly detections using the MDT approach (proposed) and 3 other approaches

temporal MDT (proposed) spatial MDT (proposed) MPPCA Social Force Optic Flow Monitor

Clips showing more instances of anomaly detections by the MDT (proposed) approach (left: temporal MDT; right: spatial MDT)

Runner [mov] Biker [mov] Cart [mov] Cart [mov] Skater [mov] Crossing Walker [mov] Biker [mov] Biker [mov]

Clips showing instances of anomaly localization by the proposed CRF filter (red regions are predicted by the CRF filter, blue region are predicted by simple thresholding)

Skater [mov] Bikers [mov] Runners, Bikers [mov] Cart [mov] Skater [mov] Bikers [mov] Biker [mov] Biker [mov] Skater, Biker [mov]

Representative frames showing anomaly localization by the proposed CRF filter (red regions are predicted by the CRF filter, blue region are predicted by simple thresholding)