Unmanned Aerial Vehicle Object Recognition in Bad Weather Using Dark Channel Prior and Convolutional Neural Networks

摘要

Object recognition using Unmanned Aerial Vehicles (UAVs) is increasingly becoming more useful. Tremendous success has been achieved on UAV object recognition in clear weather conditions where adequate illumination makes it easier for UAVs to recognize objects in the scene. Unfortunately, for outdoor applications, there is no escape from bad weather moments such as haze, fog, dust, smoke and smog. These weather nuisances occur due to suspended particles in the atmosphere, ultimately resulting in degraded visibility. Thus, these weather nuisances cause unsatisfactory performance in UAV object recognition. Current UAV object recognition algorithms do not guarantee satisfactory performance in bad weather conditions. Therefore, this study was motivated by the need for UAV object recognition systems that can perform robustly despite the state of the weather. Several state-of-the-art methods exist for object recognition and image dehazing/defogging. Nonetheless, the performance of these methods is dependent on the scenarios where they are used. In this study, a novel method that deployed the Dark Channel Prior (DCP), for scene dehazing/defogging; and Convolutional Neural Network (CNN) for object recognition; was proposed and investigated in the context of UAV for object recognition in bad weather. The aim of the study was to investigate the proposed method for enabling the UAV to efficiently recognize objects in bad weather conditions such as fog, haze, smoke and smog. The proposed method was experimented to determine the extend at which it can enable the UAV recognize objects in fog/haze weather. The objective of the experiments was to investigate the performance of the proposed method for addressing UAV object recognition in bad weather by observing two independent variables, namely; (1) fog density, which is the measure of fog present in the scene and (2) distance of object from the UAV, in fog. Analysis of results demonstrated that the DCP method effectively addresses UAV visibility improvement in bad weather conditions. On varied densities of haze/fog, the DCP method enables the UAV to effectively dehaze/defog scenes and improve visibility of objects present in the scene. Additionally, analysis of results illustrated that the constructed CNN model can enable the UAV to accurately recognize objects from the dehazed/defogged scenes with a confidence accuracy of 94.3%.