Deep Learning Based Vehicle Localization Using Angular Characteristics of Millimeter Wave
Yaser Awadh Bakhuraisa, Azlan Bin Abd Aziz, Tan Kim Geok, Azwan Mahmud, Mohamad Yusoff Alias, Mohd Khanapiah Bin Nor · 2024
Accurate vehicle localization is essential for connected and automated Vehicles (CAVs). However, there are significant deficiencies in the performance of Machine Learning-based localization techniques that utilize received signal strength (RSS) from millimeter-wave (mm-Wave) communications, primarily due to instability in dynamic scenarios such as those involving vehicles. In this work, we propose a vehicle localization method based on deep neural networks (DNNs) with mm-Wave propagation characteristics, including RSS and angular characteristics. Specifically, three types of artificial neural networks (ANNs), namely feed-forward, function fitting, and cascade-forward, are developed. A commercial ray tracing simulation is utilized to construct a real vehicle-to-infrastructure (V2I) communication scenario and model mm-Wave propagation characteristics. The performance of the proposed method is evaluated by testing the accuracy of the proposed ANN using different datasets. Moreover, the accuracy of the proposed method is tested by using a corrupted dataset to study its robustness against characteristics uncertainties. The numerical results show that the feed-forward model outperforms the other NN models. The constructed empirical cumulative distribution function (CDF) demonstrates that 90% of vehicles have a localization error of below 2.5m. However, the results of the corrupted dataset indicate that the accuracy degraded significantly. This aspect will be further investigated and addressed in future work.