Design and Construction of a Hybrid Edge-Cloud Smart Surveillance System with Object Detection

Abstract

Smart surveillance systems are becoming very popular in personal, business, environmental and government domains as they are more cost effective than legacy CCTV systems. They provide seamless integration with technologies such as smart phones and automation systems. With recent advances in resource constrained hardware and computer vision, smart surveillance systems have the ability to perform advanced object detection while lowering power consumption and costs. In this paper a hybrid edge-cloud smart surveillance system was designed and constructed using a Raspberry Pi, NoIR camera and cloud computing to provide IoT functionality and services while maintaining inference locally at the edge device. The system implemented the mobile-first SSD MobileNetV3 model for object detection, deployed using AWS services such as IoT Greengrass and Lambda allowing the system to easily scale to hundreds of surveillance nodes. The Amazon Simple Notification Service would send email and SMS notifications to the user when a detection occurs with the image of the detection and streamed the video feed to Amazon Kinesis Video Streams allowing the user to immediately view the live feed using a media viewer. Various experiments and tests were then performed in order to validate that the system worked according to the user specifications. Positive detections of both people and animals were achieved in daytime and nighttime conditions with the expected performance indices for the inference time, latency and probability percentage of detected objects. Future iterations of this system will focus on zero touch provisioning and scaling making it easier to deploy over a broad large geographical area. It will also focus on reducing resource utilization even further to cater for even more resource- constrained devices.