中文摘要 這項研究可以引導機器人找到入口或自動門，也可以引導盲人從遠處到自動門附 近。如今，物件檢測模型的效果已經非常強大，高準確度，高幀率，但主要問題是難以 區分便利商店的自動門和玻璃，因此我們使用圖卷積網絡（GCN）來提高準確度。我們 的想法是使用 GCN 模型並透過周遭的物件來辨識自動門。這個系統分為兩個部分，物 件偵測與物件關聯。物件偵測的部分，我們使用高準確度及高幀率的現有模型。YOLOv4 是今年提出最佳的物件偵測模型，但僅使用 YOLOv4 還是會有大量誤判的狀況，因此需 要透過我們提出的方法來改善。物件關聯部分我們提出了結合 GCN 的全連接層。接著， 在開始物件關聯前，我們需要先將物件偵測的結果轉成圖的架構。我們在 GTX 1080 上 進行訓練，並在 AGX 上測試模型。我們的資料集是一個自製的資料集，我們從 Google 街景中收集並錄製位於台灣的街景影片。該資料集由 100 多家便利商店組成，其中包含 內部和外部環境，我們可以透過 GCN 來減少誤偵測的狀況。在我們的測試資料集中獲 得了 86％的準確度。實際以影片或在真實環境測試，我們的模型可以達到 5 FPS 左右， 證明我們提出的模型可以找到自動門。為了證實我們的模型可以解決問題，我們展示了 實驗部分的結果與模型的運作方式;Abstract This research purpose is to navigate robots to find doors and help blind people find entrances and exits even across a room. Nowadays, the object detection model is very powerful, high accuracy, and high frame rate but the main problem is hard to distinguish between glass doors and glass walls in convenience stores. To solve this problem, we use Graph Convolutional Network (GCN) to improve accuracy results. The idea is we use the GCN model to identify the entrance by using surrounding objects. The system consists of two parts, object detector and association part. For the object detector part, we use the advantage of public models for high accuracy and frame rate per second. The YOLOv4 is the new model this year and is state of the art compared with the previous model. But YOLOv4 still has wrong detection so it needs to use our proposed model to fix it. For the association part is our proposed model, Fully Connected Layer (FC) combined with GCN. Then we need to convert the output of the object detector to graph structure first before input to association. We train on GTX 1080 and test the real-time models on the AGX broad. My dataset is custom, collecting from google street view and recording video from Taiwan. The dataset consists of more than 100 convenience stores, including inside and outside environments. We can reduce some wrong detection from the object detector to achieve this. We get test accuracy 86 percent from our test set. However, we are testing on the video to demo and it showed our result gets fps around 5 frames and it can prove our proposed model can find the doors. To confirm our model over the problem, we demonstrate in the experiment part and showed how the model works.
