On Onboard LiDAR-based Flying Object Detection

TL;DR

The paper tackles onboard detection and localization of flying objects to enable autonomous aerial interception and multi-robot coordination. It introduces a novel 3D occupancy voxel map that explicitly represents unknown space and dynamic objects via a multi-class exponential update $G(v)$, a clustering-based detector, raycasting, and a KF-based multi-target tracker designed for low latency on UAV hardware. Key contributions include the $G(v)$ occupancy representation, a fast clustering/classification pipeline, a robust background-removal mechanism, a theoretically grounded analysis of LiDAR measurement uncertainty, a new UAV detection dataset, and extensive simulated and real-world evaluations including autonomous interception scenarios. The work demonstrates high recall, sub-meter localization accuracy, and low latency at practical ranges (e.g., $20\,\mathrm{m}$ with roughly $0.2\,\mathrm{m}$ accuracy and $20\,\mathrm{ms}$ delay) and provides open-source code and datasets to accelerate adoption in onboard C-UAS and multi-robot systems.

Abstract

A new robust and accurate approach for the detection and localization of flying objects with the purpose of highly dynamic aerial interception and agile multi-robot interaction is presented in this paper. The approach is proposed for use on board of autonomous aerial vehicles equipped with a 3D LiDAR sensor. It relies on a novel 3D occupancy voxel mapping method for the target detection that provides high localization accuracy and robustness with respect to varying environments and appearance changes of the target. In combination with a proposed cluster-based multi-target tracker, sporadic false positives are suppressed, state estimation of the target is provided, and the detection latency is negligible. This makes the system suitable for tasks of agile multi-robot interaction, such as autonomous aerial interception or formation control where fast, precise, and robust relative localization of other robots is crucial. We evaluate the viability and performance of the system in simulated and real-world experiments which demonstrate that at a range of 20m, our system is capable of reliably detecting a micro-scale UAV with an almost 100% recall, 0.2m accuracy, and 20ms delay.

Figures (17)

• Figure 1: The proposed detection method deployed on board an autonomous aerial interception system while eliminating an intruding DJI Mavic UAV.
• Figure 2: Schematic overview of the detection and tracking method proposed in this paper.
• Figure 3: Comparison of $L(v)$ and $G(v)$ occupancy representations in a situation with a flying object entering an otherwise empty voxel at time $t=2s$ and leaving at $t=4.5s$. Values of the update and threshold parameters are chosen in the same way as suggested in hornung2013octomap for $L(v)$, and the same as in the real-world experiments for $G(v)$. The current state according to both representations is presented in the lower graph.
• Figure 4: Illustration of the classification algorithm on three clusters (red, green, and blue) of points (marked with crosses) within the sensor's FoV and range (denoted $d_{\text{max}}$). Some points from the green cluster are closer than $d_{\text{close}}$ to occupied voxels, so the cluster is classified as background. All points from the red cluster are further than $d_{\text{close}}$ from any occupied voxels and are surrounded by confident free voxels, so the cluster is classified as a flying object. Finally, the blue points are not close to occupied voxels, but they are not decidedly separated from occupied voxels by free air due to the unobserved part of the environment (which contains a connection of the object to the ground in this case). Therefore, it cannot be determined whether the blue points correspond to the background or a flying object at this point, so the cluster is classified as unknown.
• Figure 5: Illustration of a slow-moving object causing some areas of the map to be misclassified as occupied by traditional mapping algorithms. This problem can also manifest with the algorithm presented in this paper for a UAV taking off. Points corresponding to the UAV (marked with red crosses) are clustered with the background points (marked green) when it is landed (\ref{['fig:sepbg_t1']}), so the corresponding voxels are updated as occupied (green squares). When the UAV takes off, points of the UAV are still close to occupied voxels and the situation repeats (\ref{['fig:sepbg_t2']}). Raycasting the new points will correctly clear out the voxels previously occupied by the UAV, but voxels currently containing points of the UAV will still be misclassified (red squares in \ref{['fig:sepbg_t3']}). This problem is addressed by the algorithm described in section \ref{['sec:sepbg']}.