智慧型多特徵行人辨識系統設計

Abstract

本篇論文針對行人辨識提出一個多特徵智慧型行人辨識系統，以達到提高行人辨識率的目的。此系統可主動在串列影像中找出移動中的物體，接著擷取物體的多種特徵，包刮梯度直方圖、Haar-like特徵、全域平均值、梯度影像，最後判斷此物體是否為行人。在本篇論文中採用雙層的類神經網路，包刮初級層與次級層，其中初級層類神經網路先針對單一特徵進行訓練，接著再將初級層匯集至次級層進行多特徵的統合訓練，此雙層架構的設計除了可提高行人辨識率外，亦嘗試減少訓練資料的使用，由實驗結果可知此雙層架構確實可提高行人辨識率，並且在較少的訓練資料下得到相近的準確率。

The thesis proposes an intelligent pedestrian recognition system to find out pedestrians from a sequence of images based on multi-features, including Histogram of Gradient, Haar-like feature, Global average and Gradient image. A two-staged neural network is adopted for the recognition system, which executes the training of single feature in the primary stage and then the training of multi-features in the secondary stage. The use of the two-staged neural network is not only to increase the accuracy rate but also to reduce the training data. From the experiment results, the two-staged neural network indeed improves the recognition performance and most importantly, it is workable in the case that a smaller amount of training data is used.