Bird Eyes: A Cloud-Based Object Detection System for Customisable Surveillance

Current surveillance systems do not provide customisation to detect the specific events a user may be interested in, and often require expensive and computationally demanding hardware that are not accessible to the everyday user. In this paper we introduce Bird Eyes, a cloud-based object detection system for customisable surveillance. This system allows users to stream video to the cloud for analysis using the You Only Look Once (YOLO) object detection model, which identifies the objects within the frame. The results are then filtered based on the events a user is interested in, which can range from detection of intruders to detection of people using their phones. A user can specify which events interest them through our simple rule system which is based on objects and three distinctive actions-enter, exit, and collision. Our current implementation of Bird Eyes can define 3485 unique rules, as our object detection model can detect up to 82 unique objects, and 3 action types. Once an event occurs, users are notified accordingly via the web application and an SMS message. Bird Eyes defers the heavy processing requirements of an object detection system away from the user and into the cloud, allowing for access to customisable surveillance with off-the-shelf hardware. When simulating multiple events (75 for accuracy and 120 for efficiency) across a wide cross-section of objects and environments we achieved an overall accuracy of 76%. The median time from event detection to notification was 10.70 s on the web application and 13.34 s via SMS.