Vision-Based UAV Interception Using Dynamic Image-Based Visual Servoing, Multi-Objective Predictive Guidance, and Robust Soft Control Barrier Functions

In the sky, a drone equipped with a camera faces a difficult challenge: it must catch another drone that is moving unpredictably, all while the camera's view is delayed by the time it takes to process images and the drone itself has limits on how hard it can turn. This is the realm of visual servoing, a method where a machine uses what it sees directly to steer itself, rather than relying on a perfect, pre-calculated map of the world. For a drone trying to intercept a target, the task is complicated by the fact that the camera is rigidly attached to the aircraft; if the drone turns sharply to catch the target, the camera might swing so fast that the target disappears from view. Furthermore, the images the drone receives are never perfectly fresh; they are always slightly old, and the target might be changing its speed or direction in ways the drone cannot predict. Solving this requires a system that can see the future, weigh multiple options simultaneously, and make split-second decisions that balance speed, safety, and smoothness.

Researchers at the Control Automation in Production and Improvement of Technology Institute have developed a new approach to this problem, creating a system that allows a drone to intercept a maneuvering target even when the visual data is noisy and delayed. Instead of trying to force the target into the center of the camera's view immediately, which often leads to erratic movements or the target slipping out of sight, their method uses a predictive observer to guess where the target will be a moment from now. The system then generates a small set of possible flight paths, or candidate trajectories, for the drone to follow. Each of these paths is evaluated against several goals at once: how close the drone will get to the target, how long the interception will take, how much energy the drone will use, and the risk of losing the target within the camera's frame.

To choose the best path, the researchers did not simply pick the single option that looked best on paper. Instead, they used a technique called Soft-min fusion, which blends the options together into a smooth, continuous decision. This prevents the drone from jerking back and forth between different strategies, a common problem when a system switches abruptly between choices. The result is a dynamic reference point that moves through the camera's view, guiding the drone toward the predicted interception spot rather than a fixed point in the image. This allows the drone to adjust its course gracefully as the target moves, compensating for the delay in the image data by anticipating the target's next move.

A critical part of this system is a safety filter based on a concept known as a control barrier function. In simple terms, this acts as a guard that prevents the drone from making a command that would cause the target to vanish from the camera's view. However, because real-world conditions are messy and the drone's engines have physical limits, a strict rule that says "never lose the target" can sometimes make the math impossible to solve. To handle this, the researchers introduced a "soft" version of this safety rule. This allows the system to temporarily bend the safety constraint if absolutely necessary to keep the drone moving, while still penalizing the system for doing so. This ensures the drone keeps trying to intercept the target even when the situation is extremely tight, rather than freezing up because a perfect solution cannot be found.

The researchers tested their method in a series of computer simulations involving one hundred different scenarios with varying wind, noise, and target movements. They compared their new system against several existing methods, including standard visual tracking and other predictive approaches. The results showed that their method was highly effective. In the simulations, the drone successfully intercepted the target in ninety percent of the runs, a rate very close to the best-performing baseline method. More importantly, the new system achieved this with significantly less effort. The drone used about fifteen percent less control energy, meaning it did not have to work as hard to make the same catch. It also reduced the error in tracking the target's image by nearly forty-six percent, keeping the target much steadier in the camera's view. The time it took to complete the interception was also reduced by over twenty percent compared to the baseline, provided the interception was successful.

However, the study also revealed a subtle trade-off. While the new method was faster and more efficient, its success rate was slightly lower than the baseline method, dropping by three percentage points. This suggests that by prioritizing speed and smoothness, the system occasionally took risks that a more conservative approach would have avoided. The researchers found that the mean distance between the drone and the target at the moment of interception was nearly identical for both methods, measuring just over two meters in both cases. This indicates that while the new system is more agile and less taxing on the drone's motors, it does not necessarily guarantee a perfect catch in every single scenario where the older method might have succeeded.

The study concludes that this approach offers a powerful way to manage the complex balance between seeing, thinking, and acting for autonomous drones. By combining a predictive observer, a multi-objective decision-making process, and a flexible safety filter, the system can handle the delays and uncertainties of real-world vision. The researchers note that while the simulations were successful, the system relies on approximations to keep the calculations fast enough for real-time use. Future work will involve testing these ideas on actual hardware and exploring how the system performs with more complex physical models of the drone. For now, the findings demonstrate that it is possible to guide a drone to a moving target with high precision and efficiency, even when the eyes of the drone are slow and the target is elusive.