Page 26 - 2023

Ph.D.
(Engineering & Technology)
REAL TIME HUMAN VIOLENCE RECOGNITION AND
LOCALIZATION FOR INDOOR VIDEO USING DEEP LEARNING

Ph.D. Scholar : Jani Devang Girishbhai
Research Supervisor : Dr. Anand P. Mankodia

Regi. No.: 18276351003
Abstract :
With storming growth in technology, there is an explosion in surveillance systems
deployment on public as well as private locations such as malls, hospitals, banks, society
etc. Rising surveillance systems enable better governance and control in the surrounding
environment when it comes to security, safety, risk management, prevention of
adversaries etc. This revolution has sparked the interest of researchers in the area of
computer vision with its potential real-world applications. Under the narrow field of view, it
has opened up new opportunities to better understand environmental dynamics through
human behavior understanding, its causes, correlations with surrounding environment,
extracting previously unknown yet potentially useful hidden patterns which can
collectively elevate safety and security of humankind. As a sub domain of behavior
understanding, detection and tracking of abnormal or to be precise violent events
detection and monitoring is still an open challenge in the area of research. Contextual
definition of human violent action recognition can be termed as any event that poses
threat to human life safety. However, continuous manual monitoring by security
professionals is highly stressful, inadequate, prone to human errors and inefficient.
Hence, it is important if human intervention in such tasks can be minimized as much as
possible by the means of automation. Besides, the evolution of social media has posed
another challenge as video footage is shared globally and becomes viral that is not only
to detect violent events, but it creates necessity to also hide or blur out sensitive graphic
contents on demand as its collective psychological impacts on viewers which can breed
communal, political riots. Due to the subjectivity of sensitive information, violent action
detection and localization is still a less explored research area. With potential application
in moderating spread of online sensitive content, current research is still limited when it
comes to multi-class violence detection and localization. Proposed solution resolves this
particular problem in real time by automatic recognition and localization of such violent
events. Fine-tuned pre-trained ResNet50 model enables accurate detection of various
class violent events and localizes its occurrences in frames. Our work can be extended to
mask localized violent events. Our work contributes to the addition of a novel dataset to
help the research community grow. Our approach covers detection of rare but highly
violent actions such as stabbing with knife, pushing with hands, hit with object, hit with

01