Modeling Lane and Road Characteristics for Drive Quality Rating

India has one of the largest road networks. Countries such as India, South East Asia, Africa etc., consists of roads that have one or two lanes and not always well-maintained. In fact, many of the roads have obstacles, occluded lane markings, broken dividers, cracks, obstacles, potholes, etc. that puts the drivers at significant risk while driving. In this regard, this work proposes to focus on lane and road quality assessment in order to improve the drivability of the roads. Notably, this work proposes a new framework, LaneRoadNet (LRNet), to address this problem with an integrated mechanism considering lane and road parameters using deep learning. Subsequently, road quality is analyzed with help of a modular scoring function. Significantly, the final score provides the drivability assistance, given the particular instance of the road quality on the abovementioned factors. A detailed comparative analysis for Convolution Neural Network (CNN) based and Involution Neural Networks (INN) is given. An accuracy of up to 68.69% and 71.3% was observed for CNN and EfficientNet, respectively for road material classification. Importantly, an accuracy of 95.6 % was observed for segmentation-based lane detection. this is a very significant result; and would the basis for real-world lane and road assessment.