利用遞歸神經網路和注意力機制來預測不同場景下車輛的速度

Abstract

本論文旨在利用預測周遭車輛的速度來達成防衛性駕駛。要達成防衛性駕駛，首先要知道周遭車的意圖和預測其速度和軌跡。但是，既有的預測機制有幾個問題。第一，使用了很多資訊，像是前前車的速度，來預測前車的速度。這樣的資訊，實際上是沒有的。第二，既有的模型無法從資料中抓出子序列的規律。第三，既有的模型通常只鎖定單一模型，而沒有分析在不同路段或駕駛類型的影響。面對這些問題，本文提出了這些改進。第一、使用單台車的資料，來做預測，並探索其適用範圍。第二、利用注意力機制來抓出子序列的規律來提升預測準度。第三、分出不同的場景，如轉彎、直行、換車道，和道路的型態，並找尋模型最適合的範圍。以結果來說，相比於既有卡爾曼濾波的作法，利用遞歸神經網路和注意力機制，本文達到了11\%的提升。此外，透過和多台車模型的比較，驗證單台車模型能適用於直行的預測。此外，在接近自由駕駛時，預測會更佳。

In this work, we focus on vehicle velocity prediction for defensive driving. In order to achieve defensive driving, we need to understand and predict the motion of surrounding cars. However, the existing velocity prediction mechanism has three drawbacks. Firstly, they need too much information for prediction and some information are not accessible actually. Secondly, they failed to extract sub-patterns from the training data and thus may cause inaccuracy. Thirdly, only very few scenarios are considered. To deal with these problems, we start with single car model for prediction and explore its limitation by comparing it with multi-car model. Then, we use attention model to increase the accuracy. Finally, we examine the prediction under different road types and maneuver to broaden our scenario. Overall, with attention mechanism, we achieve a 11\% improvement over existing solution. Moreover, we found that single car model could be used in go straight behavior, and it has best result in free driving periods.