Real-Time ATM Booth Suspicious Behavior Monitoring System

Automated teller machine (ATM) booths are an integral part of our life, however, there has been a rise in thefts in recent years happening in these booths [1]. Currently, the action taken against these crimes gets delayed because the investigation is done after the crime is committed using the recorded footage from the CCTV cameras that are installed in the ATM booths. Hence there is a need for a solution that can alert the authorities while the crime is being committed, i.e., in real time. The proposed study aims to detect and classify the activities of a customer in the ATM booth and flag them as suspicious using various machine learning techniques. MediaPipe is employed to extract points from the person’s body which are later passed through a multilayer perceptron which helps classify the pose of the person as normal or suspicious. The model has been trained on a unique dataset which has been recorded by a camera to replicate an ATM booth’s CCTV cameras. This novel dataset enables us to perform an in-depth analysis of abnormal activities. Apart from detection of specific poses our study also includes features like gaze detection, multiple people detection using YOLOv3, and face covering detection using dlib.