Object Recognition and Positioning with Neural Networks: Single Ultrasonic Sensor Scanning Approach.

Ultrasonic sensing may become a useful technique for distance measurement and object detection when optical visibility is not available. However, the research on detecting multiple target objects and locating their coordinates is limited. This makes it a valuable topic. Reflection signal data obtained from a single ultrasonic sensor may be just enough for the measurements of distance and reflection strength. On the other hand, if extracted properly, a scanned set of signal data by the same sensor holds a significant amount of information about the surrounding geometries. Evaluating this dataset from a single sensor scanning can be a perfect application for convolutional neural networks (CNNs). This study proposes an imaging technique based on a scanned dataset obtained by a single low-cost ultrasonic sensor. So that images are suitable for desired outputs in a CNN, a 3D printer is converted to an ultrasonic image scanner and automated to perform as a data acquisition system for the desired datasets. A deep learning model demonstrated by this work extracts object features using convolutional neural networks (CNNs) and performs coordinate estimation using regression layers. With the proposed solution, by training a reasonable amount of obtained data, 90% accuracy was achieved in the classification and position estimation of multiple objects with the CNN algorithm as a result of converting the signals obtained from ultrasonic sensors into images.