Research on Key Technologies of Visual Semantic Segmentation and Localization for Autonomous Vehicles

Addressing the issue of information disparity and intercommunication between visual localization and semantic segmentation, this paper designs an indoor localization system based on the fusion of semantic segmentation and visual localization, mainly applied to indoor localization and environment analysis scenarios. The system acquires images through a monocular camera, utilizes attention convolutional networks in neural architecture search for multi-class pixel segmentation, and calculates the three-dimensional coordinates of target areas through pose estimation and PNP algorithm. Experiments conducted using Nvidia Jetson Nano as the computing platform achieved vehicle tracking and PID motor control. Static environment experiments validated the accuracy of the localization module, with static pixel point localization error generally controlled within 5 cm and vehicle localization error within 7 cm, demonstrating good control of localization errors in both x and y axes. The system exhibits excellent performance in the collaborative ability of semantic segmentation and visual localization, as well as in fusion localization efficiency, providing an effective solution for indoor localization and environment analysis applications.