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

Area of Science:

• Robotics and Automation
• Sensor Technology
• Artificial Intelligence

Background:

• Ultrasonic sensing is valuable for distance measurement and object detection, especially in low-visibility conditions.
• Current research on detecting multiple objects and their coordinates using ultrasonic sensors is limited.
• Scanned ultrasonic signal data contains rich information about surrounding geometries.

Purpose of the Study:

• To propose a novel imaging technique for detecting and locating multiple objects using a single low-cost ultrasonic sensor.
• To adapt a 3D printer as an automated ultrasonic image scanner for data acquisition.
• To apply convolutional neural networks (CNNs) for analyzing scanned ultrasonic data and estimating object coordinates.

Main Methods:

• A single low-cost ultrasonic sensor was utilized to collect scanned reflection signal data.
• A 3D printer was repurposed into an automated system for acquiring ultrasonic datasets.
• A deep learning model, employing CNNs and regression layers, was developed to process the ultrasonic images.

Main Results:

• The proposed technique successfully converts ultrasonic signals into images suitable for CNN analysis.
• The deep learning model achieved 90% accuracy in classifying and estimating the positions of multiple objects.
• The system demonstrates effective feature extraction and coordinate estimation for detected objects.

Conclusions:

• The developed ultrasonic imaging technique, powered by CNNs, offers a viable solution for multi-object detection and localization.
• This approach significantly advances the application of single-sensor ultrasonic systems in complex environments.
• The study highlights the potential of deep learning in extracting valuable geometric information from scanned ultrasonic data.