Deep Learning Enhances Smartphone Navigation in GPS-Denied Environments

Navigating through environments where GPS signals are unavailable, such as tunnels and underground parking structures, has been a significant challenge for smartphone-based navigation systems. A collaborative team from Wuhan University and Chongqing University has introduced a novel solution, the Data- and Model-Driven Vehicle Dead Reckoning (DMDVDR) framework, which leverages deep learning to estimate a vehicle's position accurately without relying on GPS signals. This innovation is detailed in a study published in Satellite Navigation in June 2025.

The DMDVDR framework utilizes a custom-designed deep neural network, AVNet, to process data from a smartphone's Inertial Measurement Unit (IMU) and estimate the vehicle's orientation and velocity. These estimates are then integrated into an Invariant Extended Kalman Filter (InEKF) to compensate for sensor inaccuracies. The system's ability to adapt to various driving conditions through a data-driven filter parameter adapter further enhances its accuracy and robustness. Test results have demonstrated the framework's effectiveness, with minimal positional drift in GPS-denied environments.

This advancement represents a significant leap forward in smartphone-based navigation, offering a scalable and cost-effective alternative to traditional in-vehicle navigation systems. The potential applications of the DMDVDR framework extend to autonomous parking assistance, fleet management in covered facilities, and improved navigation in urban canyons and tunnels. By combining deep learning with classical control theory, the researchers have paved the way for more reliable and intelligent navigation solutions in challenging environments.