Real-time road marking recognition algorithm for embedded systems

Abstract

Introduction. The system of lane exit warning and keeping the car in the lane are among the main ones for driver assistance systems (ADAS) and autonomous driving systems. The traffic lane both on the highway and in urban conditions is usually limited by road marking lines on both sides, which are obligatoraly to be considered while implementing these systems. The purpose of the study was to develop a primary road marking recognition algorithm by means of video in real-time execution on a low-power Jetson TX2 compact computer to provide the further use of the algorithm while developing and testing Russian transport vehicles (TV) in Russian road conditions. Methodology and research methods. The task set was purposed to achieve high detection quality, resistance to environmental conditions and real-time computation on a low-power device. The analysis of open publications testified of the limitations in the use of neural network detectors with a complex architecture. As a result, it was decided to develop an algorithm that would be a set of preliminary high-performance analytical search to be refined further by the neural network classifier. The result of the study was an algorithm for recognizing road markings by video image in real-time on a Jetson TX2 with an accuracy of 0.97 and a response of 0.84 on a sample of videos taken in Moscow. Scientific novelty of the work is the ability to integrate into the Russian transport vehicles design. Practical significance. The development of the primary algorithm on the selected hardware platform, the real vehicles neural network classifier training and testing developed as part of the research work have been completed at the given stage of the research. This algorithm can be further improved in the future, integrated into Russian TV designs and tested in the conditions in which it will be used.