Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards In-Air Ultrasonic QR Codes: Deep Learning for Classification of Passive Reflector Constellations

Wouter Jansen, Jan Steckel

TL;DR

The paper addresses encoding information in in-air acoustics by moving from single reflectors to constellations to create higher-entropy landmarks for robust perception. It develops a multi-label CNN to identify multiple reflectors from a single cochleogram, and an adaptive beamforming approach with null-steering to isolate reflectors for sequential single-label classification. On a small indoor dataset, the multi-label model achieves high accuracy (F1 = 0.971, Jaccard = 0.928), demonstrating feasibility, while the beamforming method shows substantial practical challenges with accuracy < 0.45. The work points to design, material variations, data augmentation, and exploring alternative signal representations as promising directions to realize high-information acoustic landmarks for robust, real-time recognition.

Abstract

In environments where visual sensors falter, in-air sonar provides a reliable alternative for autonomous systems. While previous research has successfully classified individual acoustic landmarks, this paper takes a step towards increasing information capacity by introducing reflector constellations as encoded tags. Our primary contribution is a multi-label Convolutional Neural Network (CNN) designed to simultaneously identify multiple, closely spaced reflectors from a single in-air 3D sonar measurement. Our initial findings on a small dataset confirm the feasibility of this approach, validating the ability to decode these complex acoustic patterns. Secondly, we investigated using adaptive beamforming with null-steering to isolate individual reflectors for single-label classification. Finally, we discuss the experimental results and limitations, offering key insights and future directions for developing acoustic landmark systems with significantly increased information entropy and their accurate and robust detection and classification.

Towards In-Air Ultrasonic QR Codes: Deep Learning for Classification of Passive Reflector Constellations

TL;DR

The paper addresses encoding information in in-air acoustics by moving from single reflectors to constellations to create higher-entropy landmarks for robust perception. It develops a multi-label CNN to identify multiple reflectors from a single cochleogram, and an adaptive beamforming approach with null-steering to isolate reflectors for sequential single-label classification. On a small indoor dataset, the multi-label model achieves high accuracy (F1 = 0.971, Jaccard = 0.928), demonstrating feasibility, while the beamforming method shows substantial practical challenges with accuracy < 0.45. The work points to design, material variations, data augmentation, and exploring alternative signal representations as promising directions to realize high-information acoustic landmarks for robust, real-time recognition.

Abstract

In environments where visual sensors falter, in-air sonar provides a reliable alternative for autonomous systems. While previous research has successfully classified individual acoustic landmarks, this paper takes a step towards increasing information capacity by introducing reflector constellations as encoded tags. Our primary contribution is a multi-label Convolutional Neural Network (CNN) designed to simultaneously identify multiple, closely spaced reflectors from a single in-air 3D sonar measurement. Our initial findings on a small dataset confirm the feasibility of this approach, validating the ability to decode these complex acoustic patterns. Secondly, we investigated using adaptive beamforming with null-steering to isolate individual reflectors for single-label classification. Finally, we discuss the experimental results and limitations, offering key insights and future directions for developing acoustic landmark systems with significantly increased information entropy and their accurate and robust detection and classification.

Paper Structure

This paper contains 7 sections, 3 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Methodology
  3. In-Air 3D Sonar & Semantic Landmarks
  4. Multi-Label Classification
  5. Adaptive Beamforming with Null-Steering
  6. Experimental Results
  7. Discussion & Conclusions

Figures (3)

  • Figure 1: A multi-reflector made of a constellation of four single reflectors with their hollow hemisphere shape extracted from a sphere with radius $d$ cut off by a percentage factor $c$. The eRTIS sensor is always assumed to be parallel to the origin of the landmark. The top right photograph shows the real measurement setup with the pan-tilt device that controls the azimuth and elevation angles of the 3D-printed reflector constellation.
  • Figure 2: The multi-label classification results for seven different reflector sizes in the form of a confusion matrix. The network reached an overall F1 score of 0.971 and a Jaccard score of 0.928 on the test dataset.
  • Figure 3: An example result of the adaptive-beamforming with null steering on a measurement where a constellation reflector was present. Showing both the time and frequency domain results for the top left and bottom right reflectors respectively with and without null steering applied.