Lightweight 3D LiDAR-Based UAV Tracking: An Adaptive Extended Kalman Filtering Approach

This paper presents a lightweight 3D LiDAR-based UAV tracking system for small drones that utilizes an Adaptive Extended Kalman Filter to achieve robust, high-accuracy relative positioning in GPS-denied environments by dynamically adjusting to sparse and noisy point cloud data.

The Gist

Imagine you are trying to play a game of "tag" in a pitch-black room, but instead of using your eyes, you are using a special flashlight that only sees in 3D dots (like a cloud of tiny stars). This is the challenge faced by drones (UAVs) trying to find and follow each other, especially when there is no GPS signal (like inside a building or a dense forest).

This paper presents a new, smart way for a "chaser" drone to track a "target" drone using a lightweight 3D laser scanner (LiDAR), even when the data is messy, incomplete, or the target is doing crazy acrobatic moves.

Here is the breakdown of their solution using simple analogies:

1. The Problem: The "Flickering Flashlight"

Most drones use cameras to see. But cameras are like human eyes: they fail in the dark, fog, or bright sun.

• The Solution: They use LiDAR, which is like a flashlight that shoots out millions of tiny laser dots to build a 3D map. It works in total darkness.
• The Catch: The specific laser scanner they use (the Livox Mid-360) is "non-repetitive." Imagine a sprinkler that doesn't spray water in neat, predictable circles, but instead sprays in random, shifting patterns.
  • Result: Sometimes the "target" drone looks like a solid ball of dots. Other times, it looks like a few scattered specks. Sometimes it disappears completely behind a tree (occlusion).
  • The Old Way: Traditional tracking systems are like a rigid robot. They assume the dots will always be in the same place and the same number. When the data gets messy, the robot gets confused and crashes (or loses the target).

2. The Solution: The "Smart Navigator" (Adaptive Kalman Filter)

The authors built a new tracking system called an Adaptive Extended Kalman Filter (AEKF). Think of this as a super-smart navigator who drives a car in heavy fog.

• Standard Navigator (Old Method): This navigator drives assuming the fog is always the same thickness. If the fog suddenly gets thicker (bad data), the navigator keeps driving at the same speed and crashes because they didn't adjust.
• The New Smart Navigator (AEKF): This navigator constantly asks, "How clear is the view right now?"
  • If the view is clear: It trusts the GPS (LiDAR data) and drives confidently.
  • If the view is blurry (sparse data): It says, "I'm not sure where the car is, so I'll slow down and rely more on my guess of where it should be based on its speed."
  • If the car disappears (occlusion): It doesn't panic. It keeps guessing the path based on the last known speed and direction, but it admits, "I'm getting less sure every second." As soon as the car reappears, it instantly snaps back to the correct path.

3. The "Magic Tricks" They Used

To make this work on a small, lightweight drone (which has limited battery and computer power), they added three special features:

• The "Noise Filter" (Adaptive Noise): Imagine you are listening to a friend talk in a noisy bar. If the bar gets louder, you lean in closer and listen harder. If it gets quiet, you relax. The system does this mathematically. If the laser data is "noisy" (jittery), the system automatically trusts the drone's own movement predictions more. If the data is clean, it trusts the laser more.
• The "Clump Finder" (Optimized Clustering): The laser sees thousands of dots. The system has to figure out which dots belong to the target drone and which belong to a bird or a tree branch. They used a smart sorting method (DBSCAN) that is tuned to find the "clump" of dots that looks like a drone, even if it's just a few scattered points.
• The "Memory Lane" (Recovery Mechanism): If the target drone flies behind a building and disappears for a few seconds, the system doesn't give up. It uses a "memory lane" strategy. It keeps predicting where the drone should be. When the drone pops back out, the system gently guides the tracking back on course without getting confused.

4. The Results: The Race

They tested this on two real drones flying around. One had the laser scanner, and the other was the target doing aggressive, fast turns.

• The Old Way (Standard Filter): When the target flew fast or went out of range, the tracker got lost. It thought the drone was 50 meters away when it was actually right next to it. It was like a drunk driver swerving off the road.
• The "Particle Filter" (Another Method): This method was okay, but it was "jittery." It looked like a shaky camera recording. It could find the drone but couldn't draw a smooth line.
• The New Way (AEKF): This was the winner. It tracked the drone smoothly, even during sharp turns and when the drone briefly disappeared. It was 49% more accurate than the next best method and didn't crash the drone's computer (it used very little battery and processing power).

The Bottom Line

This paper shows that you don't need expensive, heavy equipment to track drones in the dark or without GPS. By using a "smart" math system that adapts to bad weather and messy data in real-time, small drones can now safely fly in swarms, avoid collisions, and find each other even when the world is chaotic and unpredictable.

In short: They taught a small drone to be a better detective by giving it a brain that knows when to trust its eyes and when to trust its instincts.

Technical Summary

Here is a detailed technical summary of the paper "Lightweight 3D LiDAR-Based UAV Tracking: An Adaptive Extended Kalman Filtering Approach".

1. Problem Statement

The paper addresses the critical challenge of relative positioning and tracking of small UAVs in GPS-denied environments (e.g., indoor spaces, urban canyons, dense foliage).

• Limitations of Vision: Vision-based systems (e.g., YOLO, Siamese networks) degrade significantly in low-light, poor visibility, or adverse weather conditions.
• Limitations of Existing LiDAR Solutions:
  • Hardware Constraints: Traditional LiDAR tracking often relies on bulky, power-intensive sensors unsuitable for small drones with strict payload limits.
  • Data Characteristics: Modern compact LiDARs (e.g., Livox Mid-360) use non-repetitive scanning patterns. This generates sparse, noisy, non-uniform point clouds with irregular density and inconsistent feature visibility between frames.
  • Algorithmic Rigidity: Conventional tracking algorithms assume uniform scanning and fixed noise parameters, leading to poor performance when measurement quality varies or during intermittent detection failures (occlusions).
  • Computational Load: Multisensor fusion and complex particle filtering often exceed the computational capacity of lightweight onboard processors (e.g., Jetson Nano).

2. Methodology

The authors propose a lightweight tracking framework centered on an Adaptive Extended Kalman Filter (AEKF) tailored for non-repetitive scanning LiDAR data. The system operates on a DJI F550 UAV equipped with a Livox Mid-360 LiDAR and a Jetson computing module.

A. LiDAR Data Processing & Object Detection

The pipeline processes raw point clouds through three stages:

1. Downsampling: A voxel grid filter reduces point cloud complexity while preserving spatial structure.
2. Region of Interest (ROI) Filtering: Removes background noise by retaining only points within a specific horizontal distance and altitude range.
3. Clustering: An optimized DBSCAN algorithm clusters the filtered points. It uses a clustering radius ($\epsilon = 0.5m$) and a minimum point count ($M_{min} = 3$) to identify candidate UAV targets, specifically tuned for sparse returns.

B. Adaptive Kalman Filtering (AEKF)

The core innovation is the dynamic adjustment of noise covariance matrices based on real-time statistics, moving beyond the fixed parameters of standard Kalman filters.

• State Model: Uses a Constant Acceleration (CA) state space model ($x_k = [p, v, a]^T$) to better capture dynamic flight maneuvers compared to constant velocity models.
• Process Noise Adaptation ($Q_k$): Dynamically updated based on innovation ($d_k = z_k - H\hat{x}^-_k$). If the innovation is large (indicating the model missed a rapid motion change), $Q_k$ increases to make the filter more responsive.
• Measurement Noise Adaptation ($R_k$): Updated based on residuals ($\varepsilon_k = z_k - H\hat{x}^+_k$). Large residuals indicate poor measurement agreement, prompting an increase in $R_k$ to reduce reliance on noisy data.
• Data Association: Uses Mahalanobis distance validation gating to associate measurements with tracks, filtering out outliers and clutter.

C. Occlusion Handling & Recovery

• Prediction Mode: During temporary occlusions or detection gaps, the filter propagates the state using motion dynamics without measurement updates.
• Recovery Mechanism: If no measurement is received for a threshold time ($t_{occ}$), the system reinitializes the track smoothly rather than diverging, ensuring continuity after the target reappears.

3. Key Contributions

1. Adaptive Filtering for Variable Data Quality: A novel method that dynamically tunes process and measurement noise covariances based on real-time innovation and residual statistics, specifically addressing the irregularity of non-repetitive scanning LiDARs.
2. Robust Data Association: Implementation of Mahalanobis distance-based validation to manage uncertainties in non-uniform point clouds, ensuring tracking continuity in cluttered environments.
3. Intermittent Detection Recovery: A specialized strategy that maintains tracking continuity during measurement gaps and prevents filter divergence during temporary occlusions.
4. Optimized Sparse Clustering: A tuned DBSCAN algorithm designed to segment sparse LiDAR returns from small UAVs without requiring secondary sensor modalities.

4. Experimental Results

The system was validated in real-world flight scenarios comparing three approaches:

1. Fixed CA-KF: Standard Kalman filter with static noise parameters.
2. Particle Filter (PF): Non-parametric approach with 2000 particles.
3. Adaptive CA-EKF (CAEKF): The proposed method.

Performance Metrics:

• Accuracy (3D RMSE):
  • CAEKF: 2.8 m (Best performance).
  • Particle Filter: 5.5 m.
  • Fixed CA-KF: 13.0 m.
  • Improvement: The CAEKF reduced error by 49.1% compared to the PF and 78.5% compared to the Fixed CA-KF.
• Robustness to Maneuvers:
  • The Fixed CA-KF suffered catastrophic divergence (errors up to 50m) during aggressive maneuvers and measurement gaps because it relied on unbounded constant-acceleration assumptions.
  • The Particle Filter showed intermediate performance but exhibited significant jitter and reduced smoothness.
  • The CAEKF maintained high fidelity, staying close to the RTK ground truth even during rapid turns and temporary signal loss, thanks to its adaptive noise tuning.
• Computational Efficiency:
  • Despite higher accuracy, the CAEKF ran efficiently on a Jetson Nano.
  • CPU Load: ~106.5% (aggregate across cores), comparable to the Fixed CA-KF.
  • Update Rate: Achieved 9.3 Hz, outperforming both the PF (9.2 Hz) and Fixed CA-KF (8.5 Hz).
  • Memory: ~194 MB, well within the 4 GB limit of the hardware.

5. Significance

This work demonstrates that lightweight UAVs can achieve reliable, high-precision relative positioning in GPS-denied environments using only a single compact 3D LiDAR sensor.

• Practicality: It eliminates the need for heavy multi-sensor arrays or external infrastructure (like motion capture systems).
• Algorithmic Advancement: It proves that adaptive noise estimation is superior to fixed-parameter filtering for handling the inherent noise and sparsity of modern non-repetitive scanning LiDARs.
• Scalability: The framework is computationally efficient enough for onboard deployment, paving the way for autonomous swarm coordination, interception, and search-and-rescue missions where environmental conditions are unpredictable.

Future work will focus on extending this framework to multi-target tracking in cluttered environments and integrating deep learning for improved perception.