Enhanced ERFNet Decoder for Road Segmentation Model

This study presents a road segmentation model with high accuracy and real-time performance. Road segmentation is a crucial problem in the application of autonomous vehicles, which requires real-time performance and accuracy. Nowadays, current models require high-resolution images to realize high accuracy performance, while the low-resolution input losses many details. However, high-resolution input creates much inference time and can’t meet the real-time requirement. To handle this problem, based on Efficient Residual Factorized ConvNet (ERFNet), we insert the feature fusion and enhance its upsampling blocks, which capture more of the roads edge information and achieve higher accuracy. This enhanced ERFNet decoder is designed with three main components: multi-scale feature fusion, revised factorized layers and dense upsampling convolutions. The proposed model shows impressing results on both inference time and segmentation accuracy, and this efficient model can be applied to autonomous vehicle systems.