Title : Development of an Infrastructure LiDAR-Based Algorithm for Background Subtraction and Clustering
This paper develops an infrastructure LiDAR-based background subtraction and clustering algorithm and evaluates its accuracy in real-world environ...
Title : Development of an Infrastructure LiDAR-Based Algorithm for Background Subtraction and Clustering
This paper develops an infrastructure LiDAR-based background subtraction and clustering algorithm and evaluates its accuracy in real-world environments.
With the recent advancement of autonomous driving technology, road infrastructure-based sensor systems are gaining attention as an alternative to overcome the physical blind spots of vehicle-mounted sensors and the degradation of cognitive performance in adverse weather. Among these, infrastructure LiDAR provides precise 3D spatial information. However, it faces technical challenges, such as the need to remove vast amounts of background data collected from a fixed location in real time and accurately separate adjacent dynamic objects in complex urban environments. To address these issues, this paper proposes an efficient data processing framework that utilizes High-Definition (HD) map information as prior knowledge. First, a Region of Interest (ROI) is established based on the road geometry information from the HD map, and a spatial hashing technique is applied to dramatically reduce unnecessary computational load. In the background subtraction stage, a spherical voxel grid reflecting the sensor's scanning characteristics and statistical modeling are used to compensate for the sparsity of long-range data.
Furthermore, a dual-model strategy that alternates between an operational model and a learning model is introduced to maintain robust performance despite long-term environmental changes. In the subsequent clustering stage, to overcome the limitations of existing density-based algorithms, a distance-adaptive density threshold and an anisotropic weighted distance metric reflecting the road's driving direction vector are applied. This approach improves the detection performance for long-range objects and effectively resolves the under-clustering problem, where vehicles driving adjacent to each other in congested traffic are misidentified as a single cluster. To verify the performance of the proposed algorithm, experiments were conducted by collecting data at various times of the day— including daytime, nighttime, sunrise, and sunset—
from an infrastructure LiDAR system installed at an actual urban intersection. The experimental results showed that both the background subtraction and clustering algorithms achieved satisfactory performance. Additionally, the average processing time of the entire system was measured at approximately 19ms, demonstrating real-time capability that sufficiently satisfies the 10Hz input cycle of the LiDAR sensor.