Separating Drone Point Clouds From Complex Backgrounds by Cluster Filter -- Technical Report for CVPR 2024 UG2 Challenge

TL;DR

This work presents an unsupervised LiDAR-based pipeline to detect and track small UAVs in cluttered environments by fusing LiDAR sequences and applying global-local clustering with spatio-temporal density and voxel cues. A time-series scoring mechanism selects UAV trajectories, which are then reconstructed with spline interpolation. Evaluations on the MMAUD dataset show competitive performance, including a 4th-place finish in the CVPR 2024 UG2+ Challenge, and the method emphasizes interpretability and edge-deployability without relying on deep learning. The approach addresses key challenges of sparsity and noise in multi-LiDAR data, offering a practical solution for real-world anti-drone applications.

Abstract

The increasing deployment of small drones as tools of conflict and disruption has amplified their threat, highlighting the urgent need for effective anti-drone measures. However, the compact size of most drones presents a significant challenge, as traditional supervised point cloud or image-based object detection methods often fail to identify such small objects effectively. This paper proposes a simple UAV detection method using an unsupervised pipeline. It uses spatial-temporal sequence processing to fuse multiple lidar datasets effectively, tracking and determining the position of UAVs, so as to detect and track UAVs in challenging environments. Our method performs front and rear background segmentation of point clouds through a global-local sequence clusterer and parses point cloud data from both the spatial-temporal density and spatial-temporal voxels of the point cloud. Furthermore, a scoring mechanism for point cloud moving targets is proposed, using time series detection to improve accuracy and efficiency. We used the MMAUD dataset, and our method achieved 4th place in the CVPR 2024 UG2+ Challenge, confirming the effectiveness of our method in practical applications.

Figures (4)

• Figure 1: Challenging examples of image and point cloud detection. In the point cloud, the scanning points of the drone are very sparse and not continuous in the time dimension. In many time frames, the drone is too small to be detected. In the picture, the drone is very tiny, only a dozen pixels.
• Figure 2: We first superimpose all point clouds in the sequence to obtain the global point cloud, then separate the data from different lidars, and only perform denoising processing on the data from DJI Livox Avia. The picture in the box on the right is radar data that has been processed by noise reduction, and is rendered and distinguished according to point cloud density and point cloud height. The greater the density of the point cloud and the higher the altitude, the more red the color becomes; the red trajectory in the picture on the right is the real trajectory of the drone. It can be seen that the noise is well processed and the trajectory point cloud of the drone is preserved.
• Figure 3: Our proposed algorithm architecture. Given a continuous point cloud input sequence, we first classify it into global clustering and local clustering to obtain different categories using DBSCAN. Then the number of point clouds and voxel spatial information are calculated for global and local categories respectively. And cross-compare and calculate the spatial coincidence degree and temporal density changes of different clusters. The final scoring mechanism calculates the spatial coincidence degree and relative density score of the categories to exclude point clouds other than the UAV, restores the trajectory of the UAV through spline fitting interpolation, and uses MSE to calculate the error with the true value.
• Figure 4: The green point cloud in Figure b is the UAV point cloud separated from the background by our method, the red is the real trajectory of the UAV, and the blue is our predicted UAV trajectory.