基於深度神經網路的自動駕駛車輛之定位系統錯誤偵測與排除

Abstract

本研究提出了一套用於自動駕駛車輛，定位系統內錯誤訊號偵測與排除的深度模型；文中以設計執行域及道路設計規範定義出了定位系統所需要的規格作為定位失效的判斷依據，並分析自駕定位系統內各演算法及感測器的訊號樣態、機率分布等設計錯誤訊號的參數以進行錯誤注入；本研究同時使用了卷積神經網路(Convolutional Neural Network, CNN)及閘門遞迴單元(Gated Recurrent Unit, GRU)設計錯誤訊號偵測的深度模型，並以自監督式的方式進行訓練。在實驗中，本研究以開源的實車定位資料集(KITTI dataset)進行測試並與傳統的統計檢測方式-序貫概率比檢驗(Sequential Probability Ratio Test, SPRT)進行比較，實驗結果中，SPRT在特定型態的錯誤上會偵測不到或是誤偵測樣本過多導致定位系統因錯誤訊號而失效，單純僅使用CNN或GRU模型的能夠在單一的錯誤訊號得到略優的性能，但在其他型態的錯誤訊號上會因為模型本身性能受限而導致偵測效果欠佳，而本研究的模型在同時結合CNN及GRU模型下，能夠成功辨識各種不同定位資訊的錯誤訊號，並且同時排除八成至九成因錯誤訊號而形成的定位失效數，達到降低定位失效風險的成效。

This paper purposed a deep neutal network to perform fault detect and exclusion for localization systems in autonomous vehicles, the paper first defines the requirement for localization system of autonomous vehicles by performing operation design domain(ODD) and road design standards to determine failure in localization system, then using it to design parameters for performing error injection by analyzing error signals' pattern, probability distributions in localization systems. A deep neural network model by combining Convolutional Neural Networks (CNN) and Gated Recurrent Units (GRU) is developed for fault detection task, the model is trained with self-supervised manner using error injection to extract features from localization information. In the experiments, this paper use an open-source dataset with real world sensor measurement, our model's performance then compared with Sequential Probability Ratio Test (SPRT), which is a statistical detection method. In the experiment, SPRT fails to detect specific types of errors or false detect some of the normal data, leading to increase the number of failures in localization system cause by error signals. By only using CNN and GRU models is capable of capture errors and achieve slightly better performance in individual cases, however, their performance is limited on some types of errors resulting in poor detection performance, our model by combining the feature extraction capabilities of CNN and GRU at the same time, it can successfully identify most of the error signals in localization systems, and eliminates 80% to 90% of localization failures caused by error signals, achieves the goal of lower localization failure risk.