Real-Time Video Surveillance and Violent Crowd Behavior Detection
Automatically detect violent events from the vast amounts of surveillance video data.
Business Challenge
The worldwide utilization of surveillance cameras in smart cities has enabled researchers to analyse a gigantic volume of data to ensure automatic monitoring. An enhanced security system in smart cities, schools, hospitals, and other surveillance domains is mandatory for the detection of violent or abnormal activities to avoid any casualties which could cause social, economic, and ecological damages.
Therefore, the need arises for Real-Time monitoring of crowded events and outbreak of violence. Human surveyor to monitor multiple video screens simultaneously and detect crowds of people in a constantly changing sea of activity is a daunting task. It becomes even more challenging to identify any signs of violence and rule breaking at an early enough stage to alert help beforehand.
A change by Excelledia
When AI-based tools are used for almost everything, why not use the same to maintain law and order? Excelledia thought just that and came up with the perfect and unique solution to counter the business challenge faced. This is a system that analyses crowd activity and, in turn, detects, localizes, and potentially predicts abnormal crowd behaviour and violent activities in them.
It utilizes deep learning-based computer vision models to perform the following:
- To determine crowd formation,
- To assess the variation in the distance between the individuals in subsequent frames. The distance is measured based on a previously calibrated pixel-to-real-distance conversion factor.
The model compares the temporal segments in the surveillance videos and assesses the variations in the frames. On noting unusual behaviour amongst the crowd for a certain period, it raises a flag and intimidates the security officials. The model’s final objective is to predict crowd behaviour and inform the concerned officials accordingly.
Our tool is resource-friendly and addresses security concerns through real-time surveillance footage.
The system aims to ease the burden on law-and-order enforcement and focus on more human-centric actions, like dispersal of crowd or arrests. The algorithms used in our tool can also be applied for COVID-19 social distancing norms as well, that would be especially helpful in places like hospitals, malls etc.
Scope
- The system involves a series of modules such as object detection, human motion analysis and classification of identified motion patterns into categories like violent or non-violent.
- The video is extracted into frames
- Next, the frames are preprocessed and after that, feature extraction is applied to detect and localize humans using the YOLO convolutional neural network algorithm.
- Those results are used to calculate pairwise distances between centroids that are fed into the classification module.
- The last classification process is performed by counting the occurrence of abnormal distance variance and classifying it as violence if it exceeds a threshold.
- Finally, in the Violent crowd behavior detection app, the video is uploaded where it is classified as violent behavior or non-violent behavior.
