Object Recognition with Radar Signal Features in Autonomous Driving Scenes

Abstract

The integration of advanced sensor suites is fundamentally critical to the safe and reliable operation of autonomous vehicles in complex dynamic environments. Among the various sensing modalities, automotive radar systems provide unparalleled robustness against adverse weather and lighting conditions. However, the sparsity and noise inherent in radar signals pose significant challenges for high-fidelity object recognition compared to high-resolution optical and laser-based sensors. This paper presents a comprehensive investigation into the relationship between granular radar signal features and the efficacy of object recognition algorithms, utilizing extensive network simulations modeled after real-world autonomous driving scenarios. By isolating specific signal characteristics such as radar cross-section, micro-Doppler signatures, and range-azimuth profiles, we evaluate their individual and synergistic impacts on the classification accuracy of deep neural networks. The research methodology relies on a high-fidelity synthetic environment to generate annotated radar datasets across a variety of traffic configurations and environmental states. Through rigorous evaluation, the study identifies the optimal combinations of signal features that maximize recognition precision while minimizing computational overhead. The findings provide critical insights into the design of next-generation perception architectures, demonstrating that simulated environments can effectively bridge the data scarcity gap in radar-based deep learning. These results carry significant implications for the development of resilient, all-weather autonomous driving systems.