Modeling of ultrasonic range sensors for localization of autonomous mobile robots

This paper presents a probabilistic model of ultrasonic range sensors using backpropagation neural networks trained on experimental data. The sensor model provides the probability of detecting mapped obstacles in the environment, given their position and orientation relative to the transducer. The detection probability can be used to compute the location of an autonomous vehicle from those obstacles that are more likely to be detected. The neural network model is more accurate than other existing approaches, since it captures the typical multilobal detection pattern of ultrasonic transducers. Since the network size is kept small, implementation of the model on a mobile robot can be efficient for real-time navigation. An example that demonstrates how the credence could be incorporated into the extended Kalman filter (EKF) and the numerical values of the final neural network weights are provided in the appendices.