Detecting Driver Intention by Taillight Signals via Sequential Learning

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/90057

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/90057

Title:	Detecting Driver Intention by Taillight Signals via Sequential Learning
Authors:	陳妍如;Chen, Yen-Ju
Contributors:	資訊工程學系
Keywords:	車尾燈辨識;長時間序;taillight recognition;long sequence
Date:	2022-09-28
Issue Date:	2022-10-04 12:09:38 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著深度學習技術的演進和發展，自動駕駛系統正快速發展。對於自動駕駛車輛而言，捕捉道路上其他車輛的駕駛意圖至關重要，其結果可以作為新的特徵用於軌跡預測以規劃安全的自動駕駛車輛駕駛軌跡。本研究提出了一種系統，可從尾燈信號的視頻流中識別其他車輛的駕駛意圖。為了實現這一目標，需要正確提取和識別尾燈的位置（即空間特徵）和尾燈狀態隨時間的變化（即時間特徵）。在我們的模型中，使用更長的 32 幀序列作為輸入來捕捉尾燈的完整變化。此外，採用遷移學習的經典卷積神經網絡和輕量級 WaveNet 分別提取輸入序列的空間和時間特徵。實驗結果表明，我們的系統在尾燈識別方面優於最先進的方法。;With the evolution and development of deep learning technologies, we have observed the rapid advancement of autonomous driving systems. For an autonomous driving vehicle, it is crucial to capture the driving intentions of other vehicles on the road, which can then be used for the autonomous driving vehicle to plan a safe driving route. This study proposes a system to identify the driving intention of other vehicles from the video streaming of their taillight signals. To achieve this goal, both the positions of taillights (i.e., spatial features) and the change of the status of taillights over time (i.e., temporal features) need to be properly extracted and recognized. In our system, a longer sequence of 32 frames is used as input to capture the complete change of taillights. In addition, a transfer-learned classical convolutional neural network and a light-weight WaveNet are adopted to extract spatial and temporal features of the input sequence, respectively. % A dataset of more than 800 sequences with different types of taillight signals is collected from different sources and carefully labelled. Moreover, the dataset is augmented to ensure the model training converge. The experiment results indicate that our system outperforms the state of the art approaches in taillight recognition.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	79	View/Open

社群 sharing

Loading...