Embeddable Temporal Road-User Detection from Radar: A Hybrid CNN-MetaFormer Approach

Abstract

Thanks to significant breakthroughs in millimeter-wave radar technology, deep learning architectures, and edge computing capabilities, the pursuit of robust all-weather perception systems for autonomous vehicles has intensified. With various environmental challenges and safety-critical scenarios demanding reliable object detection, researchers are addressing fundamental limitations in sensor-based perception systems. One of the most pressing challenges is achieving accurate road-user detection using automotive radar while maintaining computational efficiency for embedded deployment. Given that modern vehicles require real-time processing to operate on limited computational resources, this thesis presents a hybrid deep learning framework that leverages temporal radar data through a novel CNN-MetaFormer architecture to perform efficient detection and classification of dynamic road users. We provide a comprehensive analysis of traditional radar processing methods and their evolution toward deep learning approaches, examining both convolutional-based and transformer-based architectures for radar object detection. We also thoroughly investigate temporal modeling strategies and sensor-aware design principles specific to radar data characteristics. Furthermore, we present detailed development of our proposed CompactRADNet architecture that processes sequences of range-azimuth radar frames, introducing the Adaptive Quadratic ReLU (AQR) activation function and radar aware, multipart loss function . Our extensive experiments on the CRUW dataset demonstrate superior performance over state-of-the-art methods. The real-world deployment demonstrates the framework's implementation feasibility, highlighting the impact of hybrid architectural design, temporal sequence optimization, radar-specific adaptations, and the critical balance between detection accuracy and computational efficiency in automotive radar perception systems.