Compact Object Detection with MetaFormers from Millimeter-Wave FMCW Radar Signals

Abstract

With the development of sensor technologies, frequency-modulated continuous wave radars have seen increasing applications in the automotive industry. Operating on the millimeter-wave frequency bands, these radars offer stronger penetration capabilities through adverse weather conditions than optical sensors such as cameras and lidars, and provide accurate range and velocity measurements, making them a suitable choice for object detection tasks in autonomous driving systems. These embedded systems possess stringent requirements on the compactness and computational efficiency of the employed object detection models, which have not been adequately addressed by existing works in the literature. Motivated by the recent advancements in vision Transformers and the MetaFormer architecture, this thesis explores the design of well performing yet compact radar object detection models based on these novel architectures. A U-net shaped 3D Swin Transformer model is first developed to effectively capture spatio-temporal radar features for improved detection performance. Quantitative and qualitative analysis on the CRUW ROD2021 dataset demonstrate the effectiveness of the proposed architecture in modeling radar features and aggregating temporal information. Building upon this, this thesis proposes mRadNet, a more compact radar object detection model inspired by the MetaFormer architecture. In addition to the flexibility of leveraging both convolution-based and attention-based token mixers, mRadNet incorporates efficient token encoding and decoding strategies to further reduce model size and inference time, improving state-of-the-art performance on the CRUW ROD2021 dataset by a considerable margin with 20% smaller model size and 60% lower inference latency.