Advanced Object Detection Methods for Drone Vision

Abstract

The use of unmanned aerial vehicles (UAVs) is increasing by the day, be it in the public or the private sector. The use of UAVs requires a well-defined architecture at its core for coordinated and proper functioning; a framework called the internet of drones (IoD). Most UAVs are equipped with on-board cameras useful for monitoring of potential obstacles by the controller of the UAV from a remote location. Moreover, today we have work galore in the computer vision sector which focus on development of various deep learning (DL) algorithms for object detection and even image segmentation. Some researchers have even started speculating the challenges involved of using such algorithms in drone vision. This chapter describes the potential of such DL techniques applied on real-time footage gathered by the on-board cameras on the UAVs. These technologies will aid object tracking, self-navi-gation, and obstacle detection and collision avoidance capabilities. Machine learning (ML) algorithms have found immense applications in almost every area of modern research. A specialized branch of ML is DL which mainly deals with neural networks. These neural networks are designed to mimic the human brain and form the basis of many advanced artificial intelligence (AI) applications, such as smart computer vision. The DL computer vision architectures we study in this chapter rely heavily on the use of neural networks, which we shall describe in detail. In particular, we discuss three state-of-the-art DL computer vision architectures namely Faster-R convolutional neural networks (faster R-CNN), single shot detection (SSD) and you only look once (POLO). We discover how advanced computer vision can help UAVs maintain a real-time database of land-based entities with the use of different ways of annotations (2D annotations, 3D annotations, polygon annotation and semantic segmenta-tion). For example, many UAVs operating over the same zone with the same IoD architectures may, with the use of the above technologies of algorithms mentioned, enumerate vehicles on different major roads in an urban area to maintain a real time traffic intensity projection over different sectors of the urban area. In other instances, drones with such capabilities may detect birds and measure their distances and take necessary steps to avoid a collision. These algorithms are very popular in computer vision using DL today and have been used in many applications.