-
Notifications
You must be signed in to change notification settings - Fork 3
/
deepgroup-bibtex.bib
8 lines (8 loc) · 2.11 KB
/
deepgroup-bibtex.bib
1
2
3
4
5
6
7
8
@mastersthesis {,
title = {Deep Networks for Analyzing Group Activity in Videos for Surveillance},
year = {2017},
abstract = {<p>A surveillance video can be analyzed by identifying activities at various levels of hierarchy- individual person, groups of persons and overall (or scene level). Most of the existing literature focuses on scene activity recognition and ignores multiple groups with different activities within a scene. Group level information can be employed for high-level applications such as abnormal activity detection and is important to understand the scene in its completeness. We propose a deep hierarchical framework to analyze a video at three levels of hierarchy - individuals, groups of persons and overall scene. Unlike most of the existing approaches, our framework additionally discovers groups of people within a scene and identifies the corresponding group activity. We propose an objective function which learns the amount of pairwise interaction (compatibility) between any two persons in a scene. We also train our own pedestrian orientation recognition network whose performance is state of the art. As a minor contribution, we suggest post-processing steps which improve pedestrian detection for static cameras. We evaluated our approach on standard datasets for group and scene activity, orientation estimation and pedestrian detection. The results of scene activity recognition are competitive with state of the art methods. Critically, unlike other approaches, our framework also detects groups of persons in a scene and the corresponding group activities with fair accuracy. Our orientation recognition network outperforms existing approaches and we also observe improvement in pedestrian detector performance due to the suggested post-processing algorithm. Our work contributed to Video Surveillance and Analysis project of National Center of Excellence in Technology for Internal Security (NCETIS).</p>},
keywords = {Collective Activity, Deep Learning, Event Detection, Group Activity, Group Detection, Neural Networks, Video Surveillance},
author = {Goyal, Ashish},
editor = {Prof. Rajbabu Velmurugan}
}