BUILDING A REAL-TIME THEFT DETECTION SYSTEM FOR EDGE DEVICES WITH LIMITED RESOURCES

Abstract

Currently, surveillance cameras are increasingly widely used in family apartments, due to advances in hardware technology, high connectivity, and low cost. Integrating cameras with different edge devices into the smart home ecosystem is becoming popular to create a common controller for other devices such as lights, doorbells, and temperature. However, cameras, due to limited hardware resources, usually only support simple motion and person recognition techniques. In this study, we build a real-time thief detection model based on transfer learning techniques on popular edge devices. Experimental results show that the model has the lowest inference speed 3.6 FPS (Raspberry Pi 3 B+), 6.5 FPS (Raspberry Pi 4 B), and 10.2 FPS (Jetson Nano), with AP (test) accuracy over 60%. Therefore, the system is feasible when deployed in practice