Architecture and evaluation protocol for transformer-based visual object tracking in UAV applications
This work addresses the challenges of visual object tracking on unmanned aerial vehicles (UAVs), where dynamic platforms, camera motion, and limited computational resources often compromise robustness or incur excessive computational overhead. The authors propose the Modular Asynchronous Tracking Architecture (MATA), which integrates a Transformer-based tracker with an extended Kalman filter, leveraging sparse optical flow for ego-motion compensation and target trajectory modeling. To better evaluate real-world embedded deployment, they introduce a hardware-agnostic evaluation protocol tailored for embedded systems and a novel metric, Normalized Time to Failure (NT2F). Experiments on benchmarks such as UAV123 demonstrate significant improvements in both Success and NT2F scores. Furthermore, real-time deployment on a Jetson AGX Orin platform using ROS 2 validates that the proposed evaluation protocol accurately reflects practical embedded performance.