Road Image Segmentation using Unmanned Aerial Vehicle Images and DeepLab V3+ Semantic Segmentation Model

Abstract

Road image segmentation is critical in a variety of applications, including road maintenance, intelligent transportation systems, and urban planning. Numerous image segmentation techniques, including popular neural network approaches, have been proposed for unmanned aerial vehicle (UAV) images recently. However, since these images include complex backgrounds, high-precision road segmentation from UAV images remains challenging. To address this issue, this study proposes a deep learning method called DeepLab V3+ semantic segmentation. Road images are captured and collected from several roads in Kedah and Selangor, Malaysia using a UAV. To segment the road from the background, the DeepLab V3+ with Resnet-50 backbone is utilised. Then, the performance is assessed by comparing segmented images by deep learning to manually segment images. Three metrics are used for the assessment; pixel accuracy (PA), mean area intersection by union (mIoU), and mean F1-score (MeanF1). The study also compares the segmentation performance with the DeepLab V3+ with mobile NetV2 for benchmarking purposes. Simulation results show that the DeepLab V3+ with Resnet-50 has performed better than the DeepLab V3+ with mobile NetV2 methods. The findings indicate that the DeepLab V3+ with Resnet-50 outperformed the DeepLab V3+ with mobile NetV2 for PA, mIoU, and MeanF1 by 1.39 %, 4.92 %, and 9.71 %, respectively.