Road-scene segmentation is a part of general image segmentation problems which tries to characterize the road-scene and divides it into labeled area/objects, such as road, building, car, pedestrian, pavement, etc. This problem has a huge range of applications, e.g. Intelligent Vehicle and Advanced Driver Assistance System (ADAS). We develop RCC-Net, a deep learning-based algorithm for achieving a real-time road-scene segmentation for Advanced Driver Assistance System (ADAS). We have successfully built the RCC-Net on a low-cost embedded system, NVIDIA Jetson TK1, opening the possibilities for in-car deployment.
The system achieves 5-7 FPS for running on a Jetson TK1 (forward inference only). It runs over 30 FPS on a workstation PC with GTX 1080.
I. Ardiyanto, T.B. Adji, “Deep Residual Coalesced Convolutional Network for Efficient Semantic Road Segmentation”, IPSJ Transactions on Computer Vision and Applications (IPSJ-CVA), vol. 9:6, Springer, 2017. (Invited paper of MVA 2017, ISSN: 1882-6695)