基於類神經網路架構早期偵測空停車格
No Thumbnail Available
Date
2018
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
本論文解決駕駛人耗費不必要的時間在尋找停車地點之問題。提早偵測停車格的智慧系統是重要的，駕駛可能因為分神在找尋停車格，而導致交通意外發生，且在大城市中經常發生停車格嚴重不足的問題。在本研究中，我們使用行車紀錄器蒐集共5,800部的影片資料集(駕駛人的視角)，藉由深度學習的技術，建置可以偵測前方是否有空停車位的類神經網路模型。為了增進偵測效能，我們提出了一個新的損失函數以優化時序資料，最後開發出一個可以早期偵測空停車格的駕駛輔助系統。在本研究中，我們也建立了一個提早偵測空停車格的評比實驗 (Benchmark)，可以讓後續相關領域的研究者評估其實驗結果。
This thesis addresses the problem of spending unnecessary time on searching for a place to park. Early detection of vacant parking space is important because traffic accidents often happen due to the distraction of a driver when the driver is looking for vacant parking space. Furthermore, according to the statistic from Department of Transportation, Taipei City Government, there is about sixty thousand registrations difference between cars and parking spaces, indicating the serious problem of shortage of parking space in a big city. In this study, we collect a dataset that contains 5,800 dash-cam videos. We train neural network models through the deep learning technique to detect whether or not there is a vacant parking space ahead. In order to improve the detection performance, we propose a new loss function to optimize the sequence problem. We develop a driving assistance system for early detecting vacant parking space which aims to reduce the danger when driving. Finally, we also establish a benchmark for this task, which can be used to evaluate future related experiments and systems.
This thesis addresses the problem of spending unnecessary time on searching for a place to park. Early detection of vacant parking space is important because traffic accidents often happen due to the distraction of a driver when the driver is looking for vacant parking space. Furthermore, according to the statistic from Department of Transportation, Taipei City Government, there is about sixty thousand registrations difference between cars and parking spaces, indicating the serious problem of shortage of parking space in a big city. In this study, we collect a dataset that contains 5,800 dash-cam videos. We train neural network models through the deep learning technique to detect whether or not there is a vacant parking space ahead. In order to improve the detection performance, we propose a new loss function to optimize the sequence problem. We develop a driving assistance system for early detecting vacant parking space which aims to reduce the danger when driving. Finally, we also establish a benchmark for this task, which can be used to evaluate future related experiments and systems.
Description
Keywords
停車格偵測, 行車紀錄器, 卷積神經網路, 長短期記憶網路, Parking Space Detection, Dash Cam, Convolutional Neural Network, Long Short-Term Memory