Aerial Scene Understanding Based on Synthetic Data

Public area monitoring will always be a key aspect of great concern. Its development has greatly promoted rapid innovation in many fields, such as public safety supervision, intelligent transportation analysis, and urban planning. Many researchers have made efforts on public area monitoring. However, most of these studies are based on the traditional monitoring perspective, which suffers high construction costs and is hard to update. In recent years, rapidly developing drone technology demonstrates superiority in convenience, flexibility, and low cost compared to traditional monitoring systems. Nevertheless, it is not widely employed in public area monitoring. Therefore, we build a framework for diversified aerial view data collection based on virtual scenes and name it as Synthetic Drone. To verify its effectiveness, we construct a virtual aerial view public safety monitoring dataset named SynthDrone. It contains 300 video clips under various scenes with annotations includes location, identification and depth etc. In addition, we also build a Depth Aware Target Distribution estimation Network (DATDNet) that takes cross-modal input data and orients multiple tasks in public safety monitoring. Based on the newly constructed dataset and DATDNet, we design experiments to verify their effectiveness and discuss the domain gap between synthetic and real-world data.