Improved Self-Localization for Intelligent Vehicles in Underground Parking Lots: Fusing Visual Features and Semantic Topology

Accurate self-localization is increasingly required for Intelligent Vehicles (IVs) in underground parking lots. Among various self-localization methods, the vision-based approach is considered an effective scheme due to its low cost and capability to provide rich information. In this paper, a novel method is proposed to enhance the localization performance. A fusion method, named Dynamic K-KNN, is developed to improve scene recognition by fusing handcrafted features, convolutional features, and semantic topology. For localization computation, a Maximum Likelihood Estimation based on semantic topology constraints is proposed to improve localization accuracy. The proposed method has been tested in two scenarios. Experimental results demonstrate that the proposed method achieves improved localization accuracy, reaching sub-meter level performance.