Real-time road surface and semantic lane estimation using deep features

In this article, we present a robust real-time road surface and semantic lane marker estimation algorithm using the deconvolution neural network and extra trees-based decision forest. Our proposed algorithm simultaneously performs three environment perception tasks on colour and depth images, even under challenging conditions, namely road surface estimation, lane marker localization, and lane marker semantic information estimation. The lane marker semantic information implies the lane marker type such as dotted lane marker or continuous lane marker. The task of road surface estimation is performed with a trained deconvolution neural network. For the lane marker localization task, a scene-based extra trees regression framework is used to localize the lane markers in the given road. To account for the variations in the number and characteristics of the lane markers in the road scene, multiple regression models indexed with scene labels are used. The pre-defined scene labels correspond to the lane marker variations in a given scene, and an extra trees-based classification model is trained to estimate them from the road features. The road features, given as an input to the extra trees frameworks, are extracted from the road image using the trained filters of the deconvolution network. The proposed algorithm is validated using multiple acquired datasets. A comparative analysis is also conducted with baseline algorithms, and an improved accuracy is reported. Moreover, a detailed parameter evaluation is also performed. We report a computational time of 90 ms per frame.