Urban Intelligent Navigator for Drone Using Convolutional Neural Network (CNN)

Abstract

It is still a difficult and challenging task for a drone to maneuver autonomously at low altitude in the urban environments. This is due to the complexity of the urban environment and its unpredictability. Many researches have been carried out in the past decades until recent time, to find a way to solve this problem using powerful sensors such as laser rangefinder sensor, RGB-D camera, stereo vision system, LIDAR and computer vision methods. This paper is aimed to present an urban intelligent navigator for drone using CNN (convolutional neural network). The application of computer vision (object detection) is cheap and has low power consumption compared to other kinds of vision systems. The machine learning allows a drone to detect and recognize all the objects and obstacles on the roads, which can block drone's way. One thousand images were captured of six different street objects (tree, lamp, bump sign, free-smoking sign, no-horn sign, and roof-wall). Those images were used as a dataset to create a machine learning using Faster R-CNN (region convolutional neural network) method. Three machine-learning models were created using different parameters for each model. The controlled parameters are the initial learning rate and the batch-size. Only the third model could successfully detect and recognize all the objects at a specified location showing 98% accuracy.