以視覺為基礎之戶外全天停車場空位偵測系統

Abstract

在本論文中，我們提出一套可以全天二十四小時運作的戶外停車 場空位偵測系統。特別針對夜晚的情況，我們設計出一種動態調整曝 光值的取像方式，能同時將多張不同曝光值的影像，經融合處理後得 到一張資訊較完整的融合影像。接著，藉由把整個停車場看成一種面 與面的結構組合、結合BHF 模型和3D 場景中車輛的幾何資訊、利 用以HOG 為基礎的擷取方式，我們擷取出融合影像內各個面相所對 應的特徵向量。最後，把每個面所擷取出的特徵向量以機率模型的方 式，分別比對出每種不同組合的停車假設，而與特徵內容最符合的假 設狀態即為我們的偵測結果。透過本系統，我們可以穩定的分析出停 車場的停車狀況，並提供停車場內空位的位置資訊。

In this thesis, we propose a vacant parking space detection system that works 24 hours a day. Especially, to capture images at night, we design a capture mode that takes images under different exposure settings and fuses these multi-exposure images into a clearer image. Besides, we combine a proposed Bayesian hierarchical framework (BHF) with the 3D-scene information by treating the whole parking lot as a structure consisting of plentiful surfaces. With the proposed framework, we extract feature vectors from each surface based on a modified version of the Histogram of Oriented Gradients (HOG) approach. By incorporating these feature vectors into specially designed probabilistic models, we can estimate the current parking status by finding the optimal statistical hypothesis among all possible status hypotheses. Experiments over real parking lot scenes have shown that our system can reliably detect vacant parking spaces day and night on an outdoor parking lot.