Steering Angle Prediction From a Camera Image as a Backup Service

In this letter, we describe a method to predict the steering angle of the front wheels of a small-sized vehicle from a rear-view camera image. Our approach can be applied to

give backup feedback for controlling units of autonomous vehicles or to detect anomalies in the running characteristics of the wheels. We determine how the emitted control (pulsewidth modulation, PWM) signal correlates with the position of the wheel edges in the camera image. Thus, we can use the PWM values in a supervised learning scenario to train an appropriate convolutional neural network-based architecture to predict the steering angle. In our experiments, we found an approximately 95% prediction accuracy for this task; we also made our data and source codes publicly available.