ИСТИНА

LiDAR-based localization provides accurate real-time positioning for autonomous vehicles. One of the widely used methods—NDT—requires computationally intensive operations at each iteration, with complexity proportional to the number of lidar scan points. While multi-threaded data processing (via GPU porting or extensive CPU parallelization) can improve performance, such approaches demand significant modifications to the codebase and are not always scalable, as GPU deployment may be infeasible and CPU parallelization is significantly limited. This paper explores an alternative approach—LiDAR point cloud decimation—where computational optimization is achieved by reducing the number of processed points. Various decimation strategies and their impact on localization performance and accuracy are analyzed. Particular attention is paid to the separate processing of ground points and obstacle points: it is shown that the contribution of ground points to the localization accuracy is smaller. Experiments on real-world data illustrate which decimation schemes offer the best trade-off between performance and accuracy, and outline directions for future research.