利用基於光達的幾何原質物體偵測與蒙特卡洛演算法之車輛定位系統

Abstract

在自動駕駛車輛系統中，車輛定位是重要的一個部分，因為車輛定位是完成其他的任務基礎，如: 路徑規劃或車輛導航。雖然全球定位系統(GPS)廣泛的提供車輛的位置資訊，但其提供的資訊在城市中容易產生偏差或因為遮蔽造成信號缺失。為了確保定位的準確性，應使用額外的環境資訊來定位車輛。本文提出了一個基於光達的定位演算法。從光達數據中提取特徵並在事先建立好的地圖中實現車輛定位。 在特徵檢測演算法中，由光達產生的三維點雲數據中擷取出具有特殊幾何特徵的原質物體做為特徵。利用多種點雲處理辦法檢測出環境中的物體，並利用所提出的二維圓形模型對不同的物體進行分類以確認是否為所要的特徵。利用所提出的多層分析演算法，基於光達的特性來減少誤判的發生。 本文使用了蒙特卡洛演算法，它是一個基於粒子濾波器的車輛定位演算法。首先，速度運動模型利用慣性測量單元的數據來預測車輛的位姿。基於預測的車輛位姿，利用原質特徵點來確定車輛在預建地圖上的位姿。其中利用兩個測量模型，似然域測距儀模型與光束場測距儀模型，量測特徵點與地圖的關係並計算每個粒子的權重。藉由特徵點與慣性測量單元迭代地更新車輛在地圖上的位姿。由實驗結果得知，利用所提出的演算法平均估測的偏差與實際軌跡約為0.3公尺。

Localization is an important part of an autonomous vehicle system, because it enables other basic tasks to be accomplished, such as path planning and navigation. Although GPS devices provide the vehicle position widely, they are susceptible to bias and signal unavailability in urban scenarios. To ensure the localization accuracy, additional information should be used to localize the vehicle. In this thesis, a LiDAR-based localization algorithm is proposed. The features are extracted from the LiDAR data and used to localize in a prebuild map. For feature detection, the primitive objects which contain special geometric feature will be extracted from the 3-D point cloud data captured by LiDAR. Several point cloud processing methods are used to detect the objects in the environment, and the proposed 2-D circle model is used to classify the features from different objects. To reduce the false positives, the multi-layer algorithm based on the LiDAR characteristic is applied. For vehicle localization, the Monte Carlo algorithm is applied in this thesis, and it is a particle-based method. First, the velocity motion model utilizes IMU data to predict a vehicle state. Based on the predicted vehicle state, the primitive feature points are utilized to determine the vehicle pose on the prebuild map. Two measurement models, the likelihood field range finder model and the beam field range finder model, are applied to measure the likelihood between the feature points and the map and calculate the weight of each particle. The estimated vehicle state on the prebuild map iteratively updates based on the feature points and IMU data. The experiment results show that the average difference between the ground truth path and the estimated path is about 0.3 m.