Semantic segmentation of UAV aerial videos using convolutional neural networks

Semantic segmentation of complex aerial videos enables a better understanding of scene and context. This enhances the performance of automated video processing techniques like anomaly detection, object detection, event detection and other applications. But, there is a limited study of semantic segmentation in aerial videos due to non-availability of the relevant dataset. To address this, an aerial video dataset is captured using DJI Phantom 3 professional drone and is annotated manually. In addition, the proposed research work investigates the performance of semantic segmentation algorithms for aerial videos implemented using Fully Convolution Networks (FCN) and U-net architectures. In this study, two classes (greenery, road) are considered for semantic segmentation. It is observed that both architectures perform competitively on the aerial videos of Unmanned Aerial Vehicle (UAV) with a pixel accuracy of 89.7% and 87.31% respectively.