Millimetre Wave Radar in Residual Neural Network with Multiple Large Model Parameters

In recent years, millimetre-wave radar has shown important application value in automatic target detection, autonomous driving perception, and multimodal data fusion tasks. After combining it with deep learning, especially with residual neural network (ResNet) and its derivative structures, millimetre-wave radar has further improved its performance in object detection. This study systematically compares the experimental performance of four deep learning models (ResNet50, ResNet101, Res2Net101 and Swin Transformer, referred to as Swin-T) on KITTI, nuScenes and BDD datasets. We evaluated key performance indicators including mAP, mAP0.5, mAP0.75, mAPS, mAPM, mAPL, mAR, mARS, mARM, mARL, FPS, and inference speed (FPS) in order to provide guidance for model selection for mmWave radar combined with deep learning.