CCTV on Steroids: The Design and Implementation of BantAI, a Robbery Prediction System for Edge Devices

Traditional surveillance systems reliant on human monitoring often struggle to deliver the speed and accuracy required for effective robbery detection. Drawing inspiration from the Filipino concept of bantay, which means to keep a watchful eye, we designed and implemented BantAI, a watchful AI that is lightweight enough to run on an edge device. BantAI operates as a two-stage system that automates surveillance and predicts potential robberies before they occur. The first stage consists of machine learning modules that analyze the video feed to detect behaviors that are often related to robberies: crowding, loitering, and concealment of one's face in public. The outputs from these modules are forwarded to the second stage, a Long Short-Term Memory (LSTM) inference system that generates the Robbery Behavior Potential (RBP). We deployed BantAI on a resource- constrained edge device and demonstrated that it can achieve an F1 score of 0.756 and a prediction lead time of 25.17 seconds.