Steering Angle Prediction and Algorithm Performance Testing in Autonomous Driving Simulators

The majority of traffic accidents are caused by human error. To counteract that, Advanced Driver-Assistance Systems (ADAS) are being developed and installed in vehicles. One of those systems is the Lane Keeping Assist System (LKAS), which helps to keep the vehicle centered in the corresponding lane. The problem of calculating a steering angle that is to be applied to the steering wheel to keep the vehicle in the corresponding lane is called steering angle prediction (SAP). In this paper, a novel SAP neural network (NN) architecture is presented. It is based on Shifted Windows (Swin) Transformer architecture and uses an image from the vehicle's front-view camera to predict the steering angle. Two new datasets for the training and testing of the NN were created. The datasets consist of the vehicle's front-view camera images and the corresponding steering wheel angles from two different autonomous driving simulators, CARLA and MetaDrive. A variety of weather and road conditions are present in the datasets, including clear, rainy, and cloudy weather and two-, four-, and six-lane roads with various road markings (white and yellow, solid and dashed lines). The presented SAP solution has shown promising results on created test sets and in simulated test environments.