Improved Ultrasonic Sensing Using Machine Learning

We present a novel approach for using industrial grade ultrasonic sensors to perform echolocation by detecting ultrasonic echoes in a noisy environment using machine learning. Autonomous driving is expected to become a huge market and among other technical challenges, environmental perception will be the most critical one. For high automation level, classical technologies are limited. On the other hand automotive is cost sensitive. The main part of the lecture starts with state of the art technology and then explains how we have used machine learning approaches to train a net for several classifications tasks: Distinguish whether the ultrasonic echo comes from the sensor or another noise source, distinguish whether the echo is relevant or not and finally a height classification. Results are presented in the form of F1-Score. In addition to this, a method will be presented to use CNN for noise suppression in real time. We demonstrate the potential of using the “bat principle” for perception and prove that by that we also achieve the low cost targets.