使用AI晶片電梯樓層面板辨識之自主移動機器人

Abstract

本論文主要透過AI開發板做電梯樓層面板的辨識，為了減少人力成本，一個成本低且搭載AI開發板的自主跨樓層導航車孕育而生，可以辨識電梯按鈕，並且按壓電梯按鈕，不論是辦公大樓的公文傳遞或是工廠貨物的搬運，可以因地制宜，此任務被區分成三個部份，室內的導航、電梯按鈕的辨識、與機械手臂的控制。首先透過導航車的雷射測距儀進行區域定位，而後透過所設計的定位方式與手臂控制方式完成按壓電梯按鈕的任務，其中電梯樓層按鈕的辨識可以藉由少量的原始照片透過視芯有限公司的影像處理軟體自動生成出不同亮度、模糊度、銳利度與不同的平移量的樣本，進而訓練出一種包含上、下與樓層數字按鈕的卷積神經網路模型，而機械手臂控制的部分則使用兩種不同的機械手臂進行按壓電梯按鈕，一支機械手臂為自行設計的四軸機械手臂，另一支則是六軸的myCobot機械手臂，經由逆向運動學與影像的像素差值做回授控制，最後透過實際場景驗證，本論文所提出的系統架構是有效的。This thesis mainly recognizes the elevator button using an AI evaluation board. In order to reduce labor costs, a cheap autonomous mobile robot is built with an AI evaluation board. It can recognize the elevator button and press the elevator buttons. Whether it is document delivery goods, it can adapt to the local conditions. This task is separated into three parts: indoor navigation, elevator button recognition, and robotic arm control. This thesis proposed a local localization method and a way to control the robotic arm to press the elevator button. Besides, many photos can be utilized to create datasets with different brightness, blurriness, sharpness, and shift automatically from taking photos by ourselves with the image process software developed by a technology company"AVSdsp". The convolutional neural network (CNN) model can identify the upstairs, downstairs, and floor number signs. This thesis uses two different robotic arms. Both of them are controlled by inverse kinematics and image pixel differences. Finally, the experimental results showed the effectiveness of the proposed system.