Inferring wind velocity from informal environmental objects using optical flow informed recurrent neural networks

Abstract

Due to their flexibility and wide range of applications, UAVs are expected to play an important role in complex urban airspace in the future. However, unpredictable low-level air currents caused by the complexity and variability of local urban design can pose significant risks to the take-off and landing phases. Current high-quality wind profile radars are expensive and only deployed in major airports. The alternative is to conduct large-scale urban modelling of wind using computation fluid dynamics, which relies on a large volume of accurate city and wind profile data. This undermines the future business model of distributed air mobility, e.g., takeoff and land in ad-hoc locations across a city. Therefore, it is crucial to create an approach that is data-efficient and economical. To achieve this, we leverage the abundance of environmental objects that naturally interact with wind, such as trees, flags, and clothing. This initial pilot study aims to address this challenge by examining tree movement using two consecutive techniques: (1) optical flow to extract the natural movement vectors, and (2) deep recurrent neural networks to translate the vectors into wind velocity. The proposed CNN-ConvLSTM model, trained on a video dataset encompassing diverse environmental conditions with ground wind speeds from 0 to 14.6 m/s, extracted visual and motion features from RGB and optical flow images, achieving an 87.42% prediction accuracy in capturing spatiotemporal wind-induced motion patterns. These results suggest the possibility of extending visual anemometer technology to broader scenarios and diverse natural objects, guaranteeing safer UAV operation in complex environments.