Investigation of SLAMs applicable on computational-resource-constrained cars in urban environments

To achieve autonomous driving, Simultaneous Localization and Mapping (SLAM) is an indispensable system enabling cars to localize and navigate urban areas. SLAMs suitable for cars should not require too much computational resources and high-end hardwares, while being able to identify locations with similar textures and rely primarily on vision cameras. Several Vision SLAMS are therefore reviewed and further investigation with other studies trying to tackle the issues in previous studies. This study mainly focuses on SLAMs that utilize 2D visual data to match 3D point cloud data. Later studies try to reduce the storage space required for 3D maps and make the matching process robust to lighting and weather changes. This paper also investigates SLAMs that use only 2D views and 2D image datasets. Later studies improve the localization precision and efficiency of matching real-time views and images from the dataset. Finally, a comparison between the 2 methods are presented, showing that 2D-oriented system can run more quickly, while being less precise, and 2D-3D combined systems tend to generate more precise localization results but requires longer run time.