基於地面優化之常態分佈轉換演算法的自主駕駛光學雷達定位系統

Abstract

在自動駕駛車輛的定位系統中，全球導航衛星系統(Global Navigation Satellite System, GNSS)的定位方式在城市或高遮蔽的環境中，無法達到自動駕駛車輛所需高準確的定位結果；相反的，光學雷達的定位方式相比於GNSS，在這樣的場域下能提供高準確度且高強健性的定位結果。然而，光學雷達硬體的高昂成本與即時化處理點雲資訊所需的高運算量問題，成為眾多人所詬病的缺點。為了能有效降低即時化定位所需要的高運算量問題，本研究提出一套適用於固定場域下的光學雷達車輛定位系統。此定位系統主要是透過點雲資訊的前處理，以及改良定位核心技術「點雲匹配演算法」的計算方式，使定位結果在保有(高)準確度的情況下，提升系統的運算效率。 本研究的定位系統包含諸多功能演算法，主要部分為點雲分割(Point Cloud Segmentation)與地面優化之常態分佈轉換演算法(Ground-Optimized Normal Distribution Transform, GO-NDT)。首先利用點雲分割進行點雲的分類，並從分類結果中，濾除雜訊點雲和提取顯著的物體點雲，接用GO-NDT利用地面點雲資料和物體點雲資料，進行兩步驟的參數最佳化並估得定位結果。最後，本研究將GO-NDT的和常態分佈轉換(Normal Distributions Transform, NDT)延伸應用至車輛定位系統之上，並採用以市區街景為主的CARLA模擬環境進行定位結果的性能比較與分析，由實驗結果得知定位準確度上NDT(平均誤差5.72公分和0.11度)與GO-NDT(平均誤差7.16公分和0.21度)兩者誤差結果相似，但GO-NDT定位系統運行上以約30%快的速度優於NDT定位系統。

In the localization system of autonomous vehicle, the localization system based on Global Navigation Satellite System (GNSS) cannot achieve the required accuracy of autonomous vehicle in the city or the highly shaded environment. In contrast, the LIDAR-based localization system can achieve highly accuracy and robust than the GNSS-based localization system in the scenario of city and the highly shaded environment. However, the high-cost hardware and the high computation of processing the point cloud in real-time make it criticized by people. In order to effectively reduce the high computational problem required for real-time localization. This study proposes a LIDAR-based localization system which can use in the fixed field. This proposed system mainly uses the point cloud pre-processing and improvement of the core localization method named “point cloud matching algorithm”, in that the proposed system more efficient than original when maintaining high accuracy. The localization system in this study contains many functional algorithms which mainly includes point cloud segmentation and Ground-Optimized Normal Distributions Transform (GO-NDT). Firstly, apply point cloud segmentation to classify the point cloud into clusters, filter out the noise and extract significant features from the clusters. Secondly, GO-NDT uses ground cluster and feature cluster to perform two-step parameter optimization and estimates the localization results. Finally, this study extends GO-NDT and Normal Distributions Transform (NDT) to the localization system, and uses the CARLA simulator mainly based on urban scenario to compare and analysis the performance of the system. The experimental results show the localization accuracy of NDT mean error is 5.72 cm and 0.11 degree, and GO-NDT mean error is 7.16 cm and 0.21 degree are similar, but the GO-NDT system is faster about 30% computation speed than the NDT.