自駕車的議題在研究領域產生的論文絡繹不絕，在實際的應用上也必須克服種種的困難來達到更進一程度的自動化。而所有的應用都仰賴於地圖的建置，尤其是在新環境的地圖建置，近年來SLAM的發展越來越多樣化，支援的感測器也越來越多元。也有些結合機器學習的版本，但是相對所需的運算資源也較多。如何在有限的資源上達到最大效益，降低誤差值以影響到後續的應用。本論文使用Raspberry Pi 3 model B在上面實現ORB-SLAM2改良的算法。利用更換更有效率的描述符，以得到匹配率高的立體匹配特徵點。另外也在迴路檢測中調整參數，以修正在建圖中所累積的誤差。搭配前端與後端的修正，在Raspberry Pi 3（CPU ARM A53）的平台上也能達到可接受的精準度。;The topic of self-driving cars has produced an endless stream of papers in the research field, and various difficulties must be overcome to achieve a higher degree of automation in practical applications. All applications rely on the establishment of maps, especially those in new environments. In recent years, the development of SLAM has become more and more diversified, and the supported sensors have become more and more diverse. There are also versions that combine machine learning, but relatively require more computing resources. How to achieve maximum benefit on limited resources and reduce the error value to affect subsequent applications. This paper uses Raspberry Pi 3 as the platform to implement the improved algorithm of ORB-SLAM2 on it. Replace the more efficient descriptors to obtain the stereo matching feature points with high matching rate. In addition, the parameters are adjusted in the loop detection to correct the errors accumulated in the construction drawing. With front-end and back-end corrections, acceptable accuracy can also be achieved on the CPU ARM A53 platform.
