🏆 Best Student Paper Award — AIAA AVIATION FORUM AND ASCEND 2025
The problem: tracking something that’s also flying
Autonomously tracking a flying object from another aircraft — a drone following another drone, or a counter-UAS system tracking an intruding aircraft — is a meaningfully harder version of visual tracking than tracking something on the ground. The target moves in three dimensions, the tracking aircraft itself is also constantly repositioning to keep it in view, and the whole pipeline — detect, estimate state, decide a maneuver — has to run fast enough to keep up in real time on hardware with real compute limits.
This has real civilian and defense relevance: search and rescue (following another aircraft or a falling payload), and counter-UAS applications (tracking an unauthorized drone), both need a tracking aircraft that can hold a target reliably without needing a supercomputer onboard.
The approach: efficient perception, learned control
The paper splits the problem into perception and control, and treats both for efficiency:
- Perception: a deep-learning-based detector is combined with a Kernelized Correlation Filter (KCF) — a lightweight, fast visual tracking technique — to get compute-efficient, accurate object detection and target state estimation, rather than relying on a heavier detection model running every frame.
- Control: instead of a traditional PID controller with its fixed gains and limitations, the paper trains a neuro-controller with reinforcement learning to directly compute velocity maneuvers for the tracking UAV, learned in simulation using AirSim across a variety of target movement patterns.
The perception pipeline was also validated on physical hardware: a target UAV tracked by an onboard camera, with the live camera view (inset) confirming the detector picks out the target against real, imperfect lighting and background clutter — not just in simulation.
Results: better tracking than PID, learned end-to-end
The learned neuro-controller outperforms a baseline PID controller on the metrics that actually define good tracking: tracking up-time (how long the target stays successfully tracked) and average distance maintained from the target during tracking. Combined with the compute-efficient detection-plus-KCF perception pipeline, the result is a tracking system built to run within real onboard compute constraints rather than assuming unlimited processing power.
This paper won the Best Student Paper Award at AIAA AVIATION FORUM AND ASCEND 2025.
Why this matters
Search-and-rescue and counter-UAS scenarios both need tracking to work on real hardware, in real time, without a ground station doing the heavy computational lifting — which makes the compute-efficiency angle here just as important as the tracking accuracy itself. A reinforcement-learned controller that beats PID on both up-time and distance-holding, paired with a lightweight perception pipeline, is a meaningful step toward tracking systems that can actually fly on the aircraft doing the tracking.
- What is a Kernelized Correlation Filter (KCF)?
- KCF is a lightweight, computationally efficient visual object-tracking technique. This paper pairs it with a deep-learning detector to keep the perception pipeline fast enough for real-time onboard tracking.
