Towards a holistic performance evaluation framework for drone-based object detection

Recent advances in drone visual sensors and integration of complex vision algorithms, facilitate further potential, entirely disrupting in a positive way their applications and capabilities. In particular, real-time object detection, usually the initial necessary step in multiple computer vision and image processing applications, has been gaining momentum in drone-based applications. Whilst heavily researched in conventional systems, drone-based vision algorithms have to consider extrinsic parameters to measure their efficiency, as their performance is heavily impacted by various flying parameters such as altitude. Further, the parameters that directly impact the performance of the vision algorithms, also impact the duration of the flight (i.e. battery life), as the vision algorithmic performance is affected by the flying route and altitude as well. This paper therefore, presents a holistic performance evaluation framework for multi-rotor drone-based object detection applications, that considers various trade-offs such as flight duration, camera resolution, computational platform performance, drone battery performance, etc., in providing a thorough analysis of the various factors affecting the operation of object detection. The framework showcases indeed that the flying altitude, in combination with the camera resolution, vastly impacts the flight duration as well as the performance of the object detection algorithm, when targeting coverage of a specific area. The framework has been experimentally verified using a commercial grade state-of-the-art drone and high-resolution camera, as well as a high-end embedded processing platform that performs the detection algorithm.