Object Understanding using Ultrasounds and Neural Networks
The application of ultrasound signals in the domain of object identification and recognition has received much interest by research scholars in recent years. The use of ultrasonic sensors providing beneficial applications to a range of industries has been of particular interested, such as in the medical field, where ultrasonic sensing can be used to generate snapshot representations of the internal body structure of a patient which can then be utilized to performed faster and more accurate diagnostics.
Ultrasound sensors have numerous advantages over other types of sensors; notably their ability to operate in the absence of light. However, the generated beams of a single element have a directivity pattern which can provide only a very low resolution – or fuzzy impression – of an object being sensed, which is no way as accurate as visual imaging of the object.
This research study focuses on studying the characteristics of objects using ultrasound waves. Beam-steering is applied to a sensor arrays to detect and compute the times of flight (TOFs) of the echoed beams. The data obtained is then used to train three convolution neural networks to predict the shape, the position and proximity of objects detected.
The proposed approach has been evaluated against two experiments, the one experiment using beam-steering techniques to transmit the incident signals and the other experiment without using these methods. The inclusion of the beam-steering techniques, although having a higher computation cost, provided better results in predicting the characteristics of the target object with identification accuracies of 93%, 90 and 77% respectively for the shape, the proximity and the position of an object.
Dr Simon Winberg
Tel: +27 (0)21 650-2793
Menzies Building, Library Road, Upper Campus
University of Cape Town