DroMOD: A Drone-Based Multi-Scope Object Detection System

Abstract

Drone-based Multi-scope Object Detection (DroMOD) system targets an efficient detection of different kinds of objects using drones. DroMOD relies on a cross-platform framework where objects’ detection tasks are shared between the drone and the server. The drone stores a set of reference images about the supervised zone. Each time an image is captured with a well-defined spatial frequency, it is compared to its reference image and only ‘trigger images’ showing a change from the reference image are sent to the server. A Big Data streaming platform is deployed on server-side for scalable and efficient real-time object detection processing based on Deep Learning (DL) models. DroMOD system architecture allows for dynamically upgrading the DL models so that newly considered objects to detect can be added to the drone mission on the fly without modifying the drone embedded software. When compared to existing alternatives, DroMOD presents the best compromise between (i) object detection accuracy, (ii) real-time processing and (iii) resource efficiency. Since only lightweight processing is performed on the drone-side, memory and computation are highly optimised on drones. Furthermore, the drone-side image filtering is independent of the objects to detect and the object detection programs are deployed and updated only on the server-side which allows for multi-target detection with minimal engineering efforts and expertise.