Achieving Efficient and Accurate Segmentation for Autonomous Driving

Autonomous driving technology, as a mainstream trend in today’s technological development, holds significant commercial value. Semantic segmentation, a core technology in this field, faces challenges with current mainstream models due to their complex structures, large numbers of parameters, and the need for high-performance hardware. These issues hinder their application in mobile devices. Therefore, this paper proposes a road scene semantic segmentation algorithm based on dynamic asymmetric convolution and capsule networks, named Asymmetric Capsule Convolutional Networks (ACCN). Specifically, the algorithm uses dynamic asymmetric convolution to construct a baseline network that combines dynamic convolution and capsule dynamic routing algorithms to enhance the model’s representational capabilities. Additionally, a dispersed attention module is built to strengthen the dependency between features to achieve more adaptable network weights. Experiments demonstrate that ACCN achieves a good balance between segmentation accuracy and detection speed. On the CamVid and Cityscapes test datasets, the proposed method in this paper achieved an average Intersection over Union (mIoU) of $\mathbf{8 0. 8\%}$ at 55.2 fps, showcasing better trade-offs in segmentation accuracy and efficiency. The experimental results indicate the effectiveness of our method.