Active Vision-Based UAV Swarm With Limited FOV Flocking in Communication-Denied Scenarios

Distributed unmanned aerial vehicle (UAV) swarms typically rely on inter-UAV communication facilitated by various technologies and omnidirectional sensing supported by multiple sensors. This article presents a novel vision-based UAV swarm architecture, allowing the swarm to perform navigation tasks in communication-denied environments with limited fields of view (FOVs). Each UAV detects and estimates the positions of neighboring UAVs in real time using a neural network. We propose a local environment representation method based on the Gaussian process (GP) to construct a time-varying continuous position belief field from discrete position estimates, enriching spatio-temporal information. Higher level velocity commands are generated following the modified Reynolds model using this belief field. Moreover, we quantify environmental uncertainty (UNC) by modeling the effects of limited FOV visual observations. A distributed decision-making method is developed for heading angle adjustments, enabling active vision. We construct and solve a traveling salesman problem (TSP) for persistent multitarget monitoring on the position belief field. UAVs explore the environment and recover lost targets by evaluating actions that minimize UNC. Finally, we validate the proposed architecture and algorithms through real-world experiments and simulations. The results demonstrate that, under the distributed decision-making method, the UAV swarm successfully performs collision-free navigation and formation across various scenarios, under limited FOV and communication denial.