Efficient and Lightweight Convolutional Neural Network for Lane Mark and Road Segmentation

Abstract

Semantic segmentation is one of an important task in computer vision that takes a great part in the perception needs of intelligent autonomous vehicles. ConvNets excel at this task, as they can be adaptively trained end-to-end to yield a set of robust hierarchies of features. The proposed key method is to reduce the unnecessary weights to build an efficient and lightweight network to acquire high accuracy on lane mark and road segmentation at pixel level. The proposed fully convolutional neural network achieves 360x480@28 fps and 97.6% accuracy on our in-house pixel-based hand-annotated lane mark and road datasets. All our models and results are trained and evaluated on an NVIDIA GTX 1080 GPU device.