Deep-learning based recognition on paved road shoulder for the Namibia B2 highway

Abstract

The number of road crash-related deaths worldwide has continued to steadily grow, reaching 1.35 million deaths every year. Low- and middle-income countries (LMIC) bear a disproportionately high number of these deaths in relation to both their population size and the total number of in-use vehicles. One of the daunting challenges facing LMICs is the lack of road safety features and built environment and their required maintenance, which can be attributed to the rising road safety concerns. Namibia, in Sub-Saharan Africa is no exception. Routine road safety audits (RSA) can aid in locating areas of the road network that need maintenance and/or require the installation of safety features. However, constrained by the limited resources for road safety initiatives, RSA are rarely performed in Namibia and LMICs. Therefore, this study demonstrates a low-cost open-source technique that can be fairly used as a supplementary tool to ease the practice of RSA in LMICs. The study presents a Deep-learning approach for classification of the presence of road shoulder and its width on a small dataset from the Highway B2 in Namibia using open access Google Street View images. Results indicate that road shoulder width can clearly be classified with open-source software, readily available models, and open access data. Results from this study have the potential to lower the overall cost of RSA in LMICs and allow for the prudent allocation of limited transportation-related funding that can create a positive impact on road safety problems in these countries.