깊이 정보를 활용한 자율주행을 위한 실시간 의미론적 영상 분할 네트워크

Abstract

Semantic segmentation is one of the most fundamental perception tasks for Autonomous Electric vehicle (AEV). It provides an overall understanding of the driving environment, including road and pedestrians. Its high computational cost with high-resolution images makes real-time implementation difficult in time-critical and resource-constrained AEV. To resolve this issue, this paper proposes a Depth-fused Trilateral Network (DTN) with dilated convolution and depthwise separable convolution that reduces 90% of the overall computation of baseline network[1] and achieves 94.73% MaxF on KITTI Road dataset and 58.67% mIOU on Cityscape 7 dataset.