LaneScanNET: A deep-learning approach for simultaneous detection of obstacle-lane states for autonomous driving systems

Abstract

Autonomous driving is the future of the automotive industry across the globe. Many challenges must be resolved in designing and developing successful Autonomous Driving Systems (ADS), especially in developing countries with poor road infrastructure. Detecting obstacles in lanes is essential for robust ADS to decide whether to change lanes, slow down, or even stop. Developing countries like India present challenges, including poor or no lane line demarcation, poor traffic management, and diverse driving behaviors. That brings many challenges to developing robust lane-keeping or lane-changing decision systems in an ADS. Publicly available driving datasets have an inherent bias towards the road infrastructure of developed countries and hence cannot cater to the needs of the Indian requirements. That makes the existing obstacle-lane segment detection models unsuitable for road infrastructures in developing countries. Due to these factors, the rate of advancement of ADS and Advanced Driver Assistance Systems (ADAS) in developing countries is meager compared to other developed countries. This makes it even more critical for future research to focus on these developing countries. This paper proposes a deep-learning-based novel decision-making network named LaneScanNET to assist the ADS in lane-changing or lane-keeping decision-making. The proposed system assists ADS in detecting obstacles, localization of the Ego Vehicle (EGV) on roads, and estimating lane status in its Field of View (FOV). The proposed LaneScanNET uses a parallel pipeline with an Obstacle Detection Network (ODN) and a Lane Detection Network (LDN) to simultaneously process the incoming image frames for detecting obstacles and segmenting lane lines, respectively. Further, the Obstacle-Lane Fusion Network (OLFN) fuses these results to predict the status of the obstacle lane in the FOV of autonomous vehicles. Vellore Institute of Technology’s (VIT’s) real-time driving dataset on indigenous roads has been collected to train the proposed LaneScanNET with 2464 obstacle-lane images obtained by driving over 60 km. The dataset generated for obstacle detection and its corresponding lane detection has also been made publicly available to promote research work for these developing countries. The proposed system outperforms all existing networks on Indian roads with an accuracy of 75.28% in obstacle detection and 91.36% in lane detection. Additionally, the proposed system can analyze the lane status with an accuracy of 92.54%. The proposed network performs exceptionally well in unforeseen circumstances like shadows, fog, dust, and occlusions making LaneScanNET a robust network that can be an add-on for ADS to make lane-keeping or lane-changing decisions. The LaneScanNET can be integrated into real-time vehicles to assist drivers or ADS.