Combating Street Crimes using Artificial Intelligence

Written by Dr Vishnu Monn Baskaran, School of Information Technology

This research is motivated by two recent trends in urban crime and safety, which is summarised by two numbers: 57% and 1,000. First, 57% represents the percentage rise on the number of snatch theft and robbery cases in Kuala Lumpur alone in 2017 and this percentage continues to rise in 2018 and 2019. Second, 1,000 represents the number of closed-circuit television (CCTV) cameras installed in Kuala Lumpur, with another 40,000 cameras planned as part of the safe-cities initiative. These two trends however are at odds with each other. On one hand, street crimes such as snatch thefts and robbery cases are projected to rise as the country rapidly urbanises. On the other hand, video surveillance are quintessentially passive driven systems, whereby recorded or archived content is used primarily as evidence of a criminal activity that has taken place. More than often perpetrators can evade apprehension due to these delays in post criminal analysis. Given this conflict, a natural question to ask is: Can a citizen remain safe despite the presence of CCTV cameras?

To this end, Dr. Vishnu Monn Baskaran, Marcus Lim Jun Yi, Ir. Dr. Joanne Lim Mun Yee and Associate Professor Dr. Wong Kok Sheik teamed up to solve this problem by addressing the fundamental building blocks required in creating an autonomous video surveillance system which can reliably localise and classify criminal actions from a surveillance perspective in real-time. The feasibility of this research is motivated by the rapid evolution of deep neural network algorithms coupled with significant advancements in high performance computing technology which have created new opportunities for a reliable active surveillance framework.

The aims of this research are to:

a) Research, design and implement a small-scale object detector which can reliably detect weapons (e.g., firearms, knives, machetes) from surveillance cameras in real-time.

b) Extend (a) by researching and developing a human-to-object classier which can determine if the detected weapon is in use by a human.

c) Extend (b) by formulating and implementing an action detection model which can reliably classify aggressive human activities from surveillance cameras in real-time.

Through this research:

a) A prototype autonomous video surveillance system is developed and deployed. With this system, real-time alerts on classified crimes will be generated and relayed to an emergency response team for 1) Immediate dispatch of a medical team to the location of the classified theft area to provide aid for the victim/s, and 2) Immediate dispatch of law enforcement officers to apprehend the perpetrator.

b) Extend the prototype system of part (a) to include person re-identification (i.e., spotting or tracking a person from one camera to another). Through person re-identification, the location/movement of a snatch thief can be translated into geographical coordinates. These coordinates are then relayed to multiple notification modules and plotted on a map, representing a unified urban/city alarm notification system. This in turn could further complement law enforcement efforts in apprehending the perpetrator.

The importance and significance of this research is driven from the Government Transformation Programme (GTP) 1.0 report which mentions the following, “Despite the improvements in the country’s crime rate and its continued downward trajectory, public perception of safety is still a challenge as 52.8% of the rakyat say they still do not feel safe.” The government recognises that safety is paramount to sustain a strong economic growth. As such, the proposed outcomes of this research establish the fundamental components of an autonomous surveillance platform. This platform potentially expands law enforcement’s omnipresence programme and substantially reduces law enforcement response time at the point of theft identification. These improvements are in line with version 2.0 of the GTP to reduce the crime index, which in turn is expected to improve public perception on urban safety.

In addition, it is envisaged that machine-assisted analysis of human actions in real-time will in the foreseeable future, provide auxiliary support for social and scientific domains such as forensic science, criminal deterrence, criminal investigation, medical aid, and psychotherapy. These solutions transform’s the nation’s capital into a smart and safe city which is in line with Transformasi Nasional (TN50), and thus paving the way for a vibrant economic and societal development.

Through this research, the team has won GOLD MEDAL in the 31st International Invention, Innovation & Technology Exhibition 2020 (ITEX 2020) for the Monash Automatic Gun Detection System (MAGTS).

The proposed AI model in MAGTS performs 7.91 times faster than current competitors by operating on lower resolution video frames (512 x 512 pixels) for faster real-time operation.