Real-time identification of pedestrian meeting and split events from surveillance videos using motion similarity and its applications

Abstract

Listen

Translate article

A real-time system to automatically identify pedestrian meeting events from surveillance videos is proposed. The system consists of three components: a pedestrian detection and tracking module, a pedestrian group identification module and a pedestrian group record. A three-level blob filter is used to improve the accuracy of pedestrian detection in the pedestrian detection and tracking module. Our previous work, the non-recursive motion similarity clustering algorithm is used as the pedestrian group identification module. Groups are detected within a time period of 0.02 ms (for four pedestrians in the scene) to 0.05 ms (for 32 pedestrians in the scene) of their occurrences in the video, using an Intel I7 processor-based machine. The pedestrian groups identified by this algorithm are stored in pedestrian group records, which are used subsequently to identify pedestrian meeting events. Visualizations were created to highlight the pedestrian groups, their history of group membership and the spatial distribution. With these visualizations, the enforcement agencies no longer need to browse through entire video archives for investigation purposes. We implemented the system to monitor several locations simultaneously in residential halls at the National University of Singapore. Our system was able to handle successfully 18 digital 640 $$\times$$ 480 pixel video streams at 25 fps on a moderately loaded Ethernet, monitoring a maximum of 30 pedestrians and detecting 83 % of the meeting and split events.