Evaluation of Real-Time Vehicle Localization Accuracy Using Vision-Based Deep Learning and Sensor Fusion in GPS-Denied Scenarios

This study presents a real-time vehicle localization framework tailored for GPS-denied environments. The proposed dynamic vehicle system integrates deep learning-based panoramic image stitching with localization computation from bird’s-eye view imagery and error correction using inertial measurement unit (IMU) data. Kalman filtering is applied to enhance system stability. A deep convolutional model is trained to accurately match real-time visual scenes with a high-resolution aerial parking map generated by an unmanned aerial vehicle (UAV), ensuring reliable positioning even under occlusions or sensor drift. Experimental results in structured parking environments demonstrate the high localization accuracy and adaptability of the system to diverse operational conditions. This lightweight and scalable design supports real-time deployment in intelligent parking and transportation systems. The proposed method offers a robust and cost-effective localization solution for autonomous vehicles in GPS-denied scenarios.