Infrared UAV Target Recognition of PTZ Camera Based on Hybrid Action Space Reinforcement Learning

With the rapid expansion of UAV applications, UAVs have posed growing challenges to public safety and airspace management. Pan–tilt–zoom (PTZ) cameras are widely used for monitoring, but conventional rule-based control lacks adaptability in dynamic environments, and existing deep reinforcement learning (DRL) methods mostly rely on discrete action spaces with fixed time steps, preventing fine-grained camera adjustment. Meanwhile, infrared UAVs are often small, distant, and low-resolution, leading to low detection confidence and difficult classification. To address these issues, this paper proposes a hybrid-action-space DRL framework for PTZ camera control. PTZ control is modeled as a discrete-continuous hierarchical subspace: the agent selects a discrete action type and predicts its continuous execution duration for adaptive precise control. A target stable tracking reward function integrating object recognition accuracy, position, and size guides the agent to improve detection confidence and complete recognition efficiently. Furthermore, a lightweight multi-scale feature fusion recognition network combining CNN and Transformer is designed for accurate state representation, and two infrared UAV datasets are constructed to support training and evaluation. Experiments in a Unity3D-based simulation environment show that the proposed framework significantly improves detection confidence and recognition accuracy, maintains high tracking stability across multiple scenarios, and generalizes well with different object detectors, outperforming rule-based and single-action-space baselines.