Multimedia Systems Компьютерные сети и коммуникации
High-fidelity semantic segmentation for dynamic small objects in autonomous driving scenes
Аннотация
Autonomous driving perception relies on accurate semantic segmentation, especially for small, high-dynamic objects such as pedestrians and traffic cones. Existing methods suffer from severe information loss during downsampling, detail smoothing, and noise interference in feature fusion. This paper presents a high-fidelity segmentation framework named RSCD-SegMAN to tackle these issues. The model introduces a newly designed Rectangular Self-Calibration Downsampling module to preserve geometric structures, Wavelet Convolution to enhance high-frequency details, and Efficient Local Attention to refine shallow features. Extensive experiments on UrbanDrive-Seg and Cityscapes datasets show that RSCD-SegMAN achieves 71.57% mIoU and 75.55% mIoU respectively, with low computational complexity of 32.93 GFLOPs, demonstrating a favorable balance between segmentation accuracy and computational efficiency.