基於監視器之類神經網路影像辨識系統執行機車安全帽配戴識別

中文摘要在這篇論文中，我們提出了一種以監視攝影機所取得的影像，用於辨識機車騎士在路口紅綠燈停等區以及停車場出入口之騎士是否正確配戴安全帽的辨識。本論文分為三個部分，第一部分我們利用監視器取得的影片擷取成單張影像或者是整個影像，接著把照片進行降低畫素的動作，然後為了使辨識更加精確，再藉由軟體label image以監督式學習的方式在每個照片做標籤分類為三種:未扣緊的、扣好的、未扣的三種型式。第二部分，我們將處理完的訓練數據輸入進類神經網路之Faster R CNN 模型中，自動化地經由RPN 以及feature map共享卷積特徵提取後進入ROI pooling層，出來得到同樣大小的bounding box，接著會透過全連接層進入Regressor層和softmax層然後產生分類概率以及邊界框位置，經由繁複的這些自動迭代訓練成可自動分辨三個種類的扣環方式，進而創建一個對安全帽扣環的即時影像辨識系統。最後的部分是實驗模擬結果，我們將觀察此系統的在不同的地點，以及不同種的安全帽下，都能達到我們預期的成果。本文的研究貢獻如下: (一)安全性:提供執法單位利用監視系統之即時影像逕行取締，可以減少執法人員在第一線上執法的危險性，以及經由此取締後，提高機車騎士正確配戴安全帽的比例，進而降低車禍發生的死亡率。 (二)低成本:利用監視器系統在城市道路上中非常密集的特性，使用此系統來取代人工執法所帶來的大量人力成本是有效的。 (三)趨勢性:隨著科技的進步，在物聯網以及人工智慧的世代裡，我們將監視系統與自動化辨識系統做結合，達到科技執法。

關鍵字

機器學習；深度學習；監督式學習

並列摘要

Abstract In this thesis, we will propose a method to use the pictures of surveillance camera to recognize whether the riders correctly wearing the helmet in the motorbike waiting zones. This thesis is divided into three parts. In the first part, we use the video obtained by the monitor to capture a single image or the whole image and reduce the pixel motion. To make the identification more precise, we use the software label Image in the form of supervised learning, the labels are classified into three types in each photo: unfastened, buckled, and unbuckled. In the second part, we input the processed training data into the Faster CNN model of the neural network, and automatically extract the convolution features through the RPN and feature maps and enter the ROI pooling layer to obtain the bounding box of the same size. Then through the full connection layer into the regressor layer and softmax layer and then generate the classification probability and the position of the bounding box. By the complex automatic iterative training to automatically distinguish the three types of buckles, and thus create a real-time image recognition of the helmet buckle system. The final part is the experimental simulation results. We will observe that the system can achieve our expected results in different locations and with different types of helmets. In this thesis, we have some contributions as follows: 1. Security: Providing law enforcement units to use the surveillance system's instant image path to ban, can reduce the risk of law enforcement officers on the first line of law enforcement, and after this ban, improve the proportion of motorcycle riders wearing helmets correctly, thereby reducing the death rate from car accidents. 2. Low cost: Owing to very dense features of the use of monitor systems on urban roads, use this system to replace manual enforcement, which reduces a lot of labor costs. 3. Trending: With the advancement of technology, in the Internet of Things and the generation of artificial intelligence, we will combine monitoring systems with automated identification systems to achieve technological enforcement.