EKF-Based Depth Camera and Deep Learning Fusion for UAV-Person Distance Estimation and Following in SAR Operations

Researchers have developed a fusion system combining Extended Kalman Filtering with depth camera and deep learning algorithms to enable UAVs to accurately estimate distance from human targets during search-and-rescue operations. The system integrates YOLO-pose for real-time detection with depth sensor data, reducing distance estimation errors by up to 15.3% and improving performance in challenging conditions like poor visibility and reflections.

This research addresses a critical technical gap in autonomous drone operations for search-and-rescue missions. The integration of multiple sensor modalities—depth cameras and monocular vision—through Extended Kalman Filtering represents an incremental but meaningful advancement in computer vision systems. The work demonstrates how sensor fusion techniques can overcome individual sensor limitations, particularly in extending the effective range of depth cameras beyond their optimal operating parameters.

The application domain reflects growing interest in autonomous systems for emergency response. SAR operations present complex real-world conditions where traditional single-sensor approaches fail, making this multi-modal approach practically relevant. The validation against motion capture ground truth data provides credibility to the error reduction claims, though the 15.3% improvement suggests the solution addresses a real but bounded problem.

From a broader AI development perspective, this work exemplifies the maturation of computer vision for industrial applications. The combination of YOLO-based pose estimation with probabilistic filtering methods is becoming standard practice across robotics and autonomous systems industries. However, the research remains primarily academic and domain-specific, with limited direct commercial implications beyond specialized drone manufacturers and emergency services.

Future development hinges on real-world deployment validation and scalability to varied environmental conditions beyond the tested indoor scenarios. The research provides a foundation for more sophisticated autonomous following systems but requires substantial engineering effort to translate into production-grade emergency response tools.

• →Extended Kalman Filter fusion of depth and monocular cameras reduces distance estimation errors by up to 15.3% for UAV tracking systems.
• →YOLO-pose deep learning integration enables real-time human body keypoint detection for safer drone-to-person distance maintenance.
• →The approach extends effective depth camera operating range and improves performance in challenging conditions like reflections and poor lighting.
• →System validation used motion capture ground truth data, providing credible accuracy benchmarks for the proposed fusion method.
• →Research addresses practical safety requirements for autonomous search-and-rescue operations requiring precise distance estimation.