Lightweight 3D LiDAR-Based UAV Tracking: An Adaptive Extended Kalman Filtering Approach
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "Lightweight 3D LiDAR-Based UAV Tracking".
Dev: Here's the gist of the paper. Accurate relative positioning is crucial for swarm aerial robotics, enabling coordinated flight and collision avoidance.
Rosa: This paper presents a lightweight LiDAR-based UAV tracking system incorporating an Adaptive Extended Kalman Filtering (AEKF) framework.
Dev: Problem Statement The ability to reliably detect and track small UAVs is crucial in various scenarios, including swarm coordination, autonomous interception, and airspace management.
Rosa: Vision-based tracking systems are significantly degraded in environments with low light, poor visibility, or adverse weather conditions.
Paper summary: Rosa: So we're looking at this paper, "Lightweight three dee LiDAR-Based UAV Tracking: An Adaptive Extended Kalman Filtering Approach," and it’s about building a tracking system for small drones using LiDAR that doesn't need to be overly complicated or heavy.
Dev: Exactly. The main idea here is tackling the problem of reliable relative positioning in swarms, especially when you can't rely on vision systems because they fail in bad light or through weather, and this paper proposes using LiDAR instead of those vision methods one.
Taro: What do you mean by that? Is it about making sure drones don't crash when they are flying near each other without knowing their exact position relative to one another?
Rosa: Right. It’s about accurate relative positioning, which is super crucial for coordinated flight and collision avoidance in swarms one. The authors show that three dee LiDAR measurements are a solid alternative because they work regardless of the lighting conditions, unlike vision systems two.
Dev: They address the issue of existing systems being too bulky and power-hungry for small UAVs by proposing a lightweight system incorporating an Adaptive Extended Kalman Filtering approach one. That means using a filter that can change its own settings based on how good the incoming data is.
Taro: So, this adaptive filter part, what does it actually do when the sensor starts giving you weird or inconsistent readings? Does it just smooth things out?
Rosa: It does more than just smoothing. The paper describes dynamically adjusting noise parameters based on the reliability assessment of real-time measurement, which handles those variations in non-repetitive scanning data one. It's about making the filter smarter about what it trusts right now.
Dev: And they also use a Mahalanobis distance-based gating system for data association, which is their way of managing uncertainties when dealing with irregular point clouds one. That helps ensure continuous tracking even in cluttered areas where the points are messy.
Taro: If you're flying around and there's a temporary occlusion, like something blocks the sensor view for a second, how does this framework handle that without losing the target?
Rosa: They have a specialized recovery mechanism designed to maintain tracking continuity during measurement gaps by reinitializing the tracking smoothly after temporary occlusions or when the target goes outside of what the sensor can see three. It’s about keeping that lock on without completely dropping it.
Dev: That leads into their state estimation, where they use a constant acceleration model to capture agile flight dynamics, defining a nine-dimensional state including position and velocity components one. They then adapt the process noise based on the innovation, which means they increase the process noise when there's high movement in the data one.
Taro: So if something is moving fast, does that make you trust your prediction less and be more reactive? How does that change how you handle sudden maneuvers?
Paper summary: Dev: That’s right. A large movement in the data increases the process noise covariance, which makes the filter more responsive to new information because it expects bigger changes one. Similarly, they adapt the measurement noise based on the post-update residual, which quantifies how good or bad that specific reading actually looks compared to what the filter predicted one.
Rosa: When you look at their experimental validation using a Livox Mid-three hundred sixty LiDAR on a DJI F550 UAV, they show that this CAEKF achieves a three dimensional root mean square error of two point eight meters one. That’s pretty low for real world tracking with these kinds of constraints.
Dev: Two point eight meters compared to five point five meters and thirteen point zero meters for the other methods is a significant reduction in estimation errors one. It’s a substantial improvement when you compare it to the particle filter or the fixed constant acceleration filter one.
Taro: That low error number, two point eight meters, sounds promising for actual deployment. But what about where this system might struggle? Does the paper mention any limitations on how long this tracking stays good or if there are specific environments where it fails?
Rosa: The authors do flag that while their framework is robust, the optimized clustering algorithm is tuned specifically for sparse LiDAR returns three. They also note that the recovery mechanism works after temporary occlusions or target excursions beyond sensor range, which suggests its effectiveness is tied to those specific kinds of gaps three.
Dev: So, the authors are acknowledging that there are still environmental challenges and measurement gaps they have to manage with this adaptive system one. It’s not perfect in every single scenario.
Taro: For someone who only listens to the show, what does this all mean for how you think about autonomous systems? Is it just a better filter, or does it change how we design these aerial robots?
Rosa: It changes how we design them by allowing us to use lighter hardware while maintaining high tracking fidelity in challenging situations one. It shows that the adaptation in the filter can bridge the gap between ideal modeling and messy real-world sensor data.
Dev: And for control engineers, it means you can run these systems with a reasonable computational load because you're not relying on excessively complex, heavy sensors or massive processing power to get that same level of accuracy one.
Taro: So the implication is that we can have more reliable relative positioning for swarms even when the sensor returns are sparse or there are temporary occlusions, and this isn't just a theoretical exercise in a lab setting.
Rosa: That’s right. It moves it toward practical application on real-world drones one. We need to see how well this performs outside the controlled testing environments they used for validation.
Dev: So, we'll stick with the idea that adaptive filtering combined with robust data association is a strong way to make lightweight LiDAR tracking reliable for complex aerial operations one. It’s a solid direction for improving drone autonomy.
Conclusion: Rosa: So, we're wrapping up on this paper about lightweight three dimensional LiDAR tracking using an adaptive extended Kalman filtering approach by those authors one.
Dev: Yeah, it’s basically showing how you can get reliable relative positioning for drones without needing a heavy sensor or a massive processing power.
Rosa: The title itself says it all—it’s about making a lightweight tracking system using LiDAR and an adaptive extended Kalman filtering approach, which is exactly what this paper achieves one.
Dev: It means that instead of relying on a heavy, power-hungry sensor just to track things with a standard filter that breaks down easily, you can use a lighter setup and make the tracking system itself flexible enough to handle the real unpredictability of aerial environments one.
Rosa: The implication is that for autonomous operations in GPS-denied areas, this type of adaptive filtering keeps things stable while maintaining a reasonable computational load for small drones one.
Dev: But they also include this recovery mechanism for those times when the target disappears for a bit, which keeps the tracking going without completely losing the lock one.
Rosa: It really points to how much better adaptive filtering can be when you're dealing with noisy, irregular data from things like LiDAR one.
Dev: Next up, we'll talk about how this works in practice on those DJI F550 drones and what the actual limitations are when you try to run it out there.
Institute for Systems and Robotics (ISR), Instituto Superior Técnico, Lisbon, Portugal
cs.RO, cs.SY, eess.SY
Submitted: 2026-03-10
Updated: 2026-03-16
Comments: Presented at the 19th International Conference on Intelligent Autonomous Systems, IAS-19, Genoa, Italy, June 30 to July 4, 2025. To appear in the Springer post-proceedings of the conference
Journal ref: In: Intelligent Autonomous Systems 19, edited by Fulvio Mastrogiovanni and Francesco Amigoni, Springer Nature Switzerland, Cham (2026)
DOI: 10.1007/978-3-032-33874-7_31
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 90/100
The gist: Accurate relative positioning is crucial for swarm aerial robotics, enabling coordinated flight and collision avoidance.
Key concepts
- Adaptive Extended Kalman Filtering (AEKF)
- A sophisticated filtering technique used to estimate a drone's state (position and velocity) over time. The 'adaptive' part means the filter automatically adjusts its internal noise settings based on how reliable the incoming sensor data is at any given moment, making it robust against varying measurement quality.
- Non-repetitive Scanning Patterns
- Modern compact LiDAR sensors use scanning patterns that are not uniform or repetitive. This creates unique data characteristics that standard tracking algorithms cannot handle. The proposed framework specifically adapts its processing to manage these fundamentally different data inputs effectively.
- Mahalanobis Gating
- This is a method used for robust data association, which determines if a new measurement belongs to the existing track. It uses the Mahalanobis distance to measure how far a new point is from the expected location, allowing the system to correctly associate points even when they are scattered or clustered in irregular point clouds.
- Constant-Acceleration (CA) Model
- This is a mathematical model used to describe how an agile UAV moves by assuming it maintains a constant acceleration. This model defines the state space for tracking, allowing the filter to accurately predict where the drone will be next, which is crucial for tracking fast or maneuverable flight dynamics.
Terminology
Summary
Accurate relative positioning is crucial for swarm aerial robotics, enabling coordinated flight and collision avoidance. p. 0 This paper presents a lightweight LiDAR-based UAV tracking system incorporating an Adaptive Extended Kalman Filtering (AEKF) framework. p. 0
Problem Statement
The ability to reliably detect and track small UAVs is crucial in various scenarios, including swarm coordination, autonomous interception, and airspace management. Vision-based tracking systems are significantly degraded in environments with low light, poor visibility, or adverse weather conditions. p. 1 LiDAR-based sensing presents a viable alternative by providing three-dimensional measurements that are independent of lighting conditions. p. 1 Modern compact LiDAR sensors utilize nonrepetitive scanning patterns that generate fundamentally different data characteristics compared to traditional LiDAR systems. p. 1 Existing tracking algorithms, which assume uniform scanning patterns and consistent measurement quality, are incompatible with processing such data. p. 1
Proposed Framework Components
The proposed framework integrates three interconnected stages: (i) point cloud processing and DBSCAN-based cluster extraction, (ii) data association with Mahalanobis gating and recovery logic, and (iii) the Adaptive Extended Kalman Filter with online noise covariance tuning.
-
Adaptive Filtering for Variable Data Quality: This method dynamically adjusts noise parameters based on the reliability assessment of real-time measurement, addressing inherent variations in non-repetitive scanning data. p. 1
-
Robust Data Association in Irregular Point Clouds: Mahalanobis distance-based validation gating manages uncertainties in non-uniform point clouds, ensuring continuous tracking even in cluttered environments. p. 1
-
Recovery Mechanism for Intermittent Detections: This specialized strategy maintains tracking continuity during measurement gaps by smoothly reinitializing tracking after temporary occlusions or target excursions beyond sensor range. p. 2
-
Optimized Clustering for Sparse Data: An optimized DBSCAN clustering algorithm is tuned for sparse LiDAR returns, facilitating precise UAV segmentation without secondary sensor modalities. p. 2
Adaptive State Estimation
The state space model employs a constant-acceleration (CA) model to capture agile flight dynamics, with the nine-dimensional state defined by position and velocity components. The filter cycles through prediction and update steps at each iteration. p. 7
The process noise adaptation is driven by the innovation, where the covariance Qk is updated as Qk = α Qk−1 + (1−α) Kk dkdT k KT k, where a large dk increases Qk → larger P−k (12) → larger Kk (14), making the filter more responsive. p. 7 The measurement noise adaptation follows a similar principle, where Rk is updated based on the post-update residual εk = zk − Hxˆ+ k, quantifying measurement quality. p. 7
Experimental Validation
Experimental validation using a Livox Mid-360 LiDAR on a DJI F550 UAV demonstrates successful UAV tracking in real-world conditions, overcoming sparse returns and intermittent detection while outperforming standard Kalman filter-based motion tracking during challenging maneuvers. The CAEKF achieves the lowest 3D RMSE (2. p. 28 m), representing a 49. p. 21% reduction over the Particle Filter (5. p. 55 m) and 78.5% over the Fixed CA-KF (13. p. 30 m). The framework maintains high tracking fidelity, while the Fixed CA-KF diverges during measurement gaps and the Particle Filter shows increased jitter during dynamic maneuvers. p. 9
Conclusion
Our approach demonstrates that an adaptive filter significantly reduces estimation errors (RMSE3D of 2.8 m vs. 5.5 m and 13.0 m for competing methods). p. 0 The integration of Mahalanobis distance-based gating enhances measurement association, while the adaptive noise regulation minimizes tracking errors and avoids overconfidence in uncertain measurements. p. 1 The track management module successfully maintains continuous tracking during temporary target loss, effectively addressing challenges from sparse LiDAR returns and occlusions. p. 2 Future work will extend the framework to multi-target tracking in cluttered environments. p. 4
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Adaptive Noise Covariance Tuning: The system dynamically adjusts noise covariance matrices using innovation and residual statistics to enhance tracking accuracy under real-world conditions, addressing inherent variations in non-repetitive scanning data.
-
Robust Data Association: Mahalanobis distance-based validation gating manages uncertainties in non-uniform point clouds, ensuring continuous tracking even in cluttered environments by accepting candidate measurements only if their distance from the predicted state is less than a threshold tau.
-
Continuous Tracking Recovery: A specialized strategy maintains tracking continuity during measurement gaps by smoothly reinitializing tracking after temporary occlusions or target excursions beyond sensor range, preventing catastrophic divergence of the fixed filter.
-
Optimized UAV State Estimation: By incorporating a constant acceleration motion model, the Adaptive Extended Kalman Filter captures sudden changes in speed and direction essential for agile maneuvers, leading to lower 3D RMSE (2.8 m) compared to competing methods.
Sources
- UAV Tracking with Solid-State Lidars:Dynamic Multi-Frequency Scan Integration
- Detection and Tracking of MAVs Using a Rosette Scanning Pattern LiDAR
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation