Convolutional gated recurrent networks for video semantic segmentation in automated driving

Abstract

Semantic segmentation is an important visual perception module of automated driving. Most of the progress has been focused on single-frame image segmentation with an optional temporal post-processing. In this paper, we propose a novel algorithm to utilize the temporal information in the semantic segmentation model using convolutional gated recurrent networks. The main motivation is to design a spatio-temporal network which can leverage motion cues for aiding segmentation and providing temporally consistent results. The proposed algorithm makes use of a fully convolutional network (FCN) that is embedded into a gated recurrent architecture. We use FCN because of its simplicity and ease of extension and the embedding can extend to other architectures. We also chose an FCN model with reasonable computational complexity suitable for real-time applications. Experimental results show consistent accuracy improvements over the baseline FCN in several datasets and it is also visually evident in our test videos shared on YouTube. The accuracy improvements for binary segmentation using F-measure were 5% and 3% in SegTrack2 and Davis respectively and the improvements for semantic segmentation in mean IoU were 5.7% and 1.7% in Synthia and Camvid respectively. To our knowledge, no prior work has been done for CNN based spatio-temporal video segmentation for automated driving and we hope that our results encourage further research in this area.