Classification of Low-Resolution Flying Objects in Videos Using the Machine Learning Approach

Abstract

A challenge of detecting and identifying drones has emerged due to the significant increase in recreational and commercial drones operating range, payload size, and overall capabilities. Consequently, drones may pose a risk to airspace safety or violate non-flying zone in the vicinity of vulnerable buildings. This paper presents an initial study for a machine-learning classification system applied to flying objects visible with a low resolution that is too distant from the camera to be efficiently classified by other methods. The original dataset in form of labeled high-resolution videos containing low-resolution drone, bird, and airplane objects was collected and carefully prepared. Computationally inexpensive features based on object shape and trajectory descriptors were recommended and tested with several ML models. The accuracy of the best-proposed model tested on our dataset was 98%. The results of this study demonstrate that Machine Learning classification seems to be promising and can be implemented in future multi-stage drone detection and identification system.

Keywords

• artificial neural networks
• computer vision
• feature extraction
• machine learning
• object detection