A Supervised and Transfer Learning-Based Two-Stage Framework for UAV Swarm Multitarget Tracking

The Cooperative Multi-Robot Observation of Multiple Moving Targets (CMOMMT) is a representative task in IoT-enabled applications. In this task, UAV swarm aim to maximize the average observation rate by tracking previously detected targets while searching for undetected ones. However, these two objectives present a trade-off, requiring effective coordination strategies to balance. Traditional methods, such as heuristic optimization and reinforcement learning(RL), often struggle with generalization and are difficult to train in complex multi-agent scenarios. To overcome these limitations, we propose a Supervised and Transfer Learning Framework (STLF), a two-stage architecture for UAV swarm coordination. In the first stage, a Teacher Model is trained under the assumption of full access to undetected target positions, generating near-optimal actions. In the second stage, a Student Model is initialized with the Teacher’s knowledge and fine-tuned via transfer learning, using a probabilistic map to approximate undetected target position under realistic constraints. Furthermore, by introducing a Lévy flight-based motion model, we expose the limitations of random walk-based models used in previous studies, which tend to underestimate the complexity of realistic target movements. Experimental results demonstrate the effectiveness of our approach, with the STLF achieving an average observation rate of 40.75%, outperforming all five baselines by 5.92% to 14.93% across various settings.