Designing Machine Learning Algorithm for Autonomous Driving in the Snow

Abstract

Autonomous vehicles have evolved to make decisions with great confidence and accuracy. Systems like the Waymo self-driving clearly have the ability to control a vehicle with level 4 autonomy under good conditions where lane markings and other driving guides like road signs are clearly visible [2]. Elon Musk has announced that Tesla autopilot should reach level 4 in 2023 [3]. However, many such systems fail to perform at a similar level in non-perfect weather conditions. The most difficult such weather appears to be a snowy climate as the vehicle significantly underperforms and the autonomous functionality is often not applicable [1]. This paper aims to train a model capable of navigating snowy weather by collecting data pertaining specifically to this weather, including unorthodox visual cues to stand in place of road markings. Such cues include tire tread marks and surrounding snow piles. These serve as a viable substitute for lane markings, allowing the vehicle to operate safely while mimicking human driving tendencies. Such a model is able to effectively drive in both clear and snowy weather when tested in simulation software. We consider a model effective for the purposes of this study in the snow when it is able to execute a full turn before failure.