Development of a collaborative robust Particle Filter for State Estimation with Stochastic and Quantity-based Uncertainties in Sensor Networks

Precise vehicle localization is a critical requirement for autonomous driving, especially in urban settings where GNSS signals often fail. To address this challenge, an advanced Particle Filter framework estimates vehicle pose by fusing 3D LiDAR data with complementary sensor inputs. The primary motivation is to achieve low-decimetre localisation accuracy despite the complexities of urban environments.

Several key modifications overcome existing limitations. First, a high-resolution (HR) map is generated from Terrestrial Laser Scan point clouds, replacing generalized models with detailed spatial representations of urban structures such as buildings and ground surfaces. This HR map serves as a robust reference for scan-to-map matching and enhances the accuracy of the Particle Filter.

Second, to manage the large volume of data from low-cost mobile LiDAR scanners, a deep learning-based subsampling approach selectively retains critical key points. The method leverages semantic segmentation features within a custom sampling model, which reduces processing time while preserving essential geometric details needed for precise localization.

Third, a camera-based collaboration mechanism integrates into the system. By detecting key points on surrounding vehicles and using their associated uncertainty information, the system refines ego-pose estimates and improves the overall robustness of the Particle Filter in real-world scenarios.

The filter framework uses sensor data from a dedicated measurement campaign for validation in order to assess the extent to which the collaborative approach meets current standards for the localisation of urban vehicles.