This paper presents a solution to address the challenging problem of Simultaneous Localization and Mapping (SLAM) in the presence of dynamic objects. SLAM is crucial for the autonomous operation of vehicles and robots, as it enables them to simultaneously determine their own position and create a map of the environment without prior reference knowledge. However, dynamic objects pose a significant challenge in SLAM systems. This paper focuses on improving the performance of SLAM by filtering out points likely to come from dynamic objects using a fusion of camera and LiDAR sensors. The Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) dataset is used to illustrate the proposed solution. Initially, a set of dynamic objects, including cars, pedestrians, bicycles, and trucks, is defined. These objects are detected using the Darknet neural network framework and the pre-trained You Only Look Once Version 3 (YOLOv3) network, utilizing camera data. The paper presents a framework to extend the initial set of dynamic objects, enabling the system to adapt to different environments. The ground plane points, which can introduce noise during registration, are efficiently filtered using a clustering algorithm utilizing Random Sample Consensus. To construct the global map from the filtered point cloud, the Iterative Closest Point algorithm is employed for registration. The paper demonstrates that by filtering out points likely to come from dynamic objects using a fusion of camera and LiDAR sensors, the performance of SLAM can be significantly improved. The effectiveness of the approach is illustrated through a performance comparison, highlighting the advantages of the proposed method. (Bu çalışma birinci sırada yer alan yazarın Yıldız Teknik Üniversitesi, Fen Bilimleri Enstitüsünde yapılan yüksek lisans tezinden üretiliştir. ORCID NO: 0009-0009-1965-4083, 0000-0002-8404-9877)
Anahtar Kelimeler: SLAM, KITTI Dataset, YOLOv3, OpenCV, RANSAC, ICP Registration
|
