부산대학교 도서관

A dense pixel-level prediction problem, semantic segmentation of road scenes for autonomous driving requires real-time processing. While sizeable deep learning networks have high accuracy, they are unsuitable for resource-constrained environments such as embedded or edge devices. Lightweight networks suffer from accuracy loss due to limited feature representation capability at a particular scale. To solve this problem, we present a novel convolutional neural network-based lightweight architecture which jointly learns features at multiple scales in a hierarchical fashion. The proposed Hierarchical Context Integration Network (HCINet) follows a multi-pathway design with feature fusion from adjacent branches for effective segmentation. The decoder uses attention and context guidance from a higher-resolution branch to further improve accuracy while being fast and lightweight. Extensive experiments on Cityscapes and CamVid datasets show the effectiveness of the proposed method. Specifically, with only 1.65 Million parameters, our network can achieve 74.5% mean IoU on Cityscapes at 82.4 frames-per-second. At the same time, it achieves 67.9% mIoU at 70 fps on the CamVid dataset.