A Framework for Object Classification via Camera-Radar Fusion with Automated Labeling

This paper introduces a robust deep learning-based framework designed to annotate Radar and camera datasets with the goal of enhancing classification and object detection performance in intricate driving environments. The realm of autonomous driving frequently encounters obstacles when it comes to interpreting complex scenes, as relying solely on camera data may not suffice to capture the full scope of information. In many cases, Radar sensors are deployed alongside cameras in commercial vehicles, providing additional and complementary data streams that can enrich the understanding of the environment. The framework put forth in this study makes use of deep learning approach to annotate a radar data that lacks elevation information, also has proposed a new neural architecture to annotate Radar and camera data, thereby facil-itating precise classification tasks and ultimately improving the performance of autonomous systems. Moreover the annotated files thus generated have been tested with a newly proposed mid level fusion architecture. An evaluation conducted on a benchmark dataset showcases an impressive overall classification accuracy rate of 99%, underscoring the efficacy of the proposed framework. Moreover, testing carried out on a dataset gathered onsite yielded a commendable overall accuracy of 90%, further validating the framework's robustness and real-world applicability. Through the utilization of this framework, researchers are empowered to leverage the combined strengths of Radar and camera fusion, thereby bolstering the capabilities of autonomous systems across a wide array of driving scenarios.