INVESTIGATING THE EFFECTIVENESS OF CORRELATION METHODS FOR VISUAL TRACKING OF MOVING OBJECTS BY UNMANNED SYSTEMS

Abstract

In the context of rapid advancements in computer vision and autonomous robotics, the task of continuous visual tracking of moving objects using unmanned aerial systems (UAS) is highly relevant. This is due to the expanding applications of UAS in civil and military domains, including monitoring, surveillance, security, search and rescue, agriculture, infrastructure inspection, and logistics. Effective tracking enhances UAS autonomy and reduces operator workload, while demanding high accuracy, real-time performance, low power consumption, and limited onboard computing. Deep learning methods, despite high accuracy, require significant hardware (GPU, RAM) and power, making them impractical for direct deployment on typical small UAVs without added equipment.

This paper presents the development and performance investigation of a lightweight real-time tracking system designed for resource-constrained platforms. The system is based on correlation filter methods, known for high processing speed and low computational needs. Among various correlation trackers, the CSRT (Channel and Spatial Reliability Tracking) tracker was chosen for its optimal balance of accuracy and speed. The system is integrated into a Software-in-the-Loop (SITL) simulation environment utilizing PX4 (flight controller emulator), MAVSDK (autopilot interaction tool), and Gazebo Classic (3D simulator). This integration allows realistic flight and tracking simulation without physical hardware.

The system begins with manual selection of the target using cv2.selectROI(). After initialization, it autonomously tracks the object, calculating its position frame by frame. Key features include minimal computational load suitable for low-power computers, rapid response to object dynamics for smooth pursuit, and autonomous operation without cloud services, enhancing reliability in remote areas.

Experimental simulation studies confirmed high tracking stability under moderate changes in appearance, scale, or lighting. The system effectively kept the target centered, with corrective velocity commands via MAVSDK ensuring adequate response. Telemetry analysis (velocities, yaw, position, trajectories) validated the tracking and control strategy in dynamic conditions. Identified limitations include reduced accuracy with rapid scale changes, need for manual reinitialization after full loss, and impact of strong noise or sudden lighting changes. Despite these, correlation methods, especially CSRT, are suitable for resource-limited systems.

Obtained results form a solid basis for integrating the system into real onboard UAVs. Future work includes automatic reinitialization, multi-object tracking, trajectory prediction, and optimization for complex scenes/weather.