基於視覺分析之路徑預測與危險分析系統

Potential Trajectory Prediction and Risk Assessment for Complex Scenarios

洪慶豪(Hung Ching-Hao)

指導教授 ： 陳灯能 ； 許志仲

國立屏東科技大學/管理學院/資訊管理系所/碩士(2021年)

https://doi.org/10.6346/NPUST202100336

摘要

在自駕車的發展過程中，包含著許多核心技術，其中被為關注的點在於其安全性，著重於自身的安全駕駛，以及對於其他人的防禦性駕駛，而在不同的國家，有著不同的地形、場景等差異，因此沒有一套通用的危險預測模型，只能對於每個地區的場景來做調整，但其根本的模型則能溝通用，原因是差別是在於景物的不同而不是行為的不同，而根據資料集的不同來調整應用場景，即可應用在不同地區，因此本論文的目的在於建置出一套通用的模型，核心想法是抓取共同的危險駕駛的行為特徵，因此藉由深度學習的技術，能夠讓電腦能夠自動去學習、抓取何種特徵屬於危險駕駛的共同危險因素。本論文以台灣作為驗證場域，針對台灣的地形以及交通生態模式發展一套系統可藉由這些數據讓電腦能夠學習潛在於這些資料中的共同危險駕駛因素。為了達到此目的，本論文引入了物件偵測，先將影像中相關的移動物件如人車擷取出來；接著我們引入Social-GAN並結合本論文擷取出來的交互移動線索 (Object Contextual Feature, OCF) 來強化周圍車輛未來路徑的預測效能。最後，本論文引入未來意外車禍的預測模組，並藉由本論文所發展的特徵模組，增強最終車禍預測準確度，以達到本計畫的防禦性駕駛行為之目的。實驗結果顯示，本論文所提出之方法有穩定且較佳之效能，針對將會發生事故之物件能夠有高準確度的預測結果。

關鍵字

自駕車 ； 自動駕駛 ； 路徑預測 ； 危險分析

並列摘要

In the development process of self-driving cars, it contains many core technologies. The most concerning point is its safety, focusing on its safe driving and defensive driving for others. In different countries, there are different terrains. Scenes and other differences, so there is no set of universal risk prediction models, which can only be adjusted for each region's scenes. However, the basic model can be used for communication, because the difference lies in the difference in the scenery rather than the behavior difference. The application scenarios can be adjusted according to the different data sets to be applied in different regions. Therefore, the purpose of this paper is to build a set of general models. The core idea is to capture common dangerous driving behaviors, which can be used in deep learning. The technology allows the computer to learn automatically; grabbing other characteristics is dangerous driving.

This paper uses Taiwan as the verification field, and develops a system for Taiwan's terrain and transportation ecological model. Using these data, the computer can learn the common dangerous driving factors that are potentially in the data. In order to achieve this goal, this paper introduces object detection, first extracting relevant moving objects such as people and vehicles in the image; then we introduce Social-GAN and combine the interactive movement clues extracted from this paper (Object Contextual Feature, OCF) to enhance the prediction efficiency of the future path of surrounding vehicles. Finally, this paper introduces a prediction module for future accidents, and uses the feature modules developed in this paper to enhance the accuracy of the final car accident prediction to achieve the goal of the defensive driving behavior of the project. The experimental results show that the method proposed in this paper has stable and better performance, and can have high-precision prediction results for objects that will have accidents.

並列關鍵字

self-driving ； autonomous driving ； trajectory prediction ； risk assessment

參考文獻

[1] J. Lutin, A. L. Kornhause, and E. Lerner-Lam, "The revolutionary development of self-driving vehicles and implications for the transportation engineering profession," in Institute of Transportation Engineers (ITE) Journal, vol. 83, 2013, pp. 28-32.

[2] C. Urmson, et al., "Autonomous driving in urban environments: boss and the urban challenge." Journal of Field Robotics (JFR), vol. 25, no. 8, 2008, pp. 425–466.

[3] G. Nirschl, "Human-centered development of advanced driver assistance systems," in Human–Computer Interaction (HCI), vol. 4558, 2007, pp. 1088–1097.

[4] W Fenghui, et al. "One-Dimensional cellular automaton traffic flow model based on defensive driving strategy," in International Journal of Crashworthiness, 2020.

[5] P. Lai, C. Dow and Y. Chang, "Rapid-response framework for defensive driving based on internet of vehicles using message-oriented middleware," in IEEE Access, vol. 6, 2018 pp. 18548-18560.