Anomalous Behavior Detection in Surveillance Cameras Using Optical Flow
Nikhil Rane, Paras Palli, Jyoti Madake, Shripad S. Bhatlawande · 2023
This project aims to develop a system for detecting suspicious anomalous behavior, specifically fighting in public spaces using computer vision methods and machine learning algorithms. The proposed system will be capable of analyzing video data captured by surveillance cameras and detecting anomalous events, such as physical disputes, which may compromise public safety. The project consists of three parts, each utilizing a distinct methodology. The first methodology is based on optical flow and magnitude thresholding. The second methodology uses Farneback optical flow angles features with random forest and HOG human overlapping feature, while the third methodology uses background subtraction and two-class SVM with ‘rbf’ kernel and 5-fold cross validation and using voting classifier. Through the implementation of these techniques, this paper aims to provide an effective solution for detecting abnormal behavior from videos. The presented system, with an SVM model accuracy of 86%, demonstrates the potential to assist law enforcement agencies in detecting and preventing criminal activities in public spaces, thus improving public safety through the application of advanced machine learning techniques. The effectiveness of the system will be widely accepted benchmark datasets, and results have been analyzed with existing methods.