Audible Sonar Research Articles

Artificial neural network classification of foliage targets from spectrograms of sequential echoes using a biomimetic audible sonar

Classifying foliage targets using echolocation is important for recognizing landmarks by bats using ultrasonic emissions and blind human echolocators (BEs) using palatal clicks. Previous attempts to classify foliage used ultrasonic frequencies and single sensor (monaural) detection. Motivated by the echolocation capabilities of BEs, a biomimetic sonar emitting audible clicks acquired 5600 binaural echoes from five sequential emissions that probed two foliage targets at aspect angles separated by 18°. Echo spectrograms formed feature vector inputs to artificial neural networks (ANNs) for classifying two targets, Ficus benjamina and Schefflera arboricola, with leaf areas that differ by a factor of four. Classification performances of ANNs without and with hidden layers were analyzed using tenfold cross-validation. Performance improved with input feature size, with binaural echo classification outperforming that using monaural echoes for the same number of emissions and for the same number of echoes. Linear classification accuracy was comparable to that using nonlinear classification with both achieving fewer than 1% errors with binaural spectrogram features from five sequential emissions. This result was better by a factor of 20 compared to previous classification of these targets using only the time envelopes of the same echoes.

Artificial neural network classification of surface reflectors and volume scatterers using sequential echoes acquired with a biomimetic audible sonar.

This paper investigates classifying two target groups, surface reflectors (SR) and volume scatterers (VS), using echo envelope features. SR targets have convex surface patches that exhibit echo persistence over aspect angle, while VS targets are composed of random range-distributed and oriented reflectors producing echoes that become uncorrelated with small changes in aspect angle. The SR target group contains single-post (P1) and multiple-post (PM) types and the VS group contains Ficus benjamina (F) and Schefflera arboricola (S) foliage types with leaf areas that differ by a factor of 4. A biomimetic sonar emitting audible clicks acquired sequences of up to three binaural echoes from target views separated by 18°. Two artificial neural networks performing linear and nonlinear classification first differentiated SR/VS target groups and then P1/PM and F/S types. Classification performance improved with echo number, from a single monaural echo to three pairs of binaural echoes, demonstrating the benefit of sequential echoes. Linear and nonlinear classification of SR/VS targets achieved a minimum generalization error probability PEG = 0.003. Nonlinear P1/PM classification achieved PEG = 0.009 that was four times smaller than linear classification. Nonlinear F/S classification achieved PEG = 0.220, indicating that envelope features by themselves are inadequate to accurately differentiate foliage targets.

Generating cognitive maps using echo features from a biomimetic audible sonar.

A sonar cognitive map displays target components that are specified by signal features extracted from a single binaural echo pair. A biomimetic audible sonar probes targets configured using posts connected by tangential planes. Echo envelopes are processed to extract values of eight parameters that govern the mapping process. Being tuned to recognize posts and planes, a cognitive map is composed of these two components using the posts' centers and radii as landmarks. A platform with translational and rotational degrees of freedom implements a landmark-centric scanning trajectory whose step size adaptively changes with echo information. The sonar tracks the target surface by maintaining a constant first-echo arrival time and by equalizing binaural echo times to form singular echoes that identify landmarks. The mapping process employs five states from detection to termination that pass through the singular echo state. Separate states process echo interference caused by two posts and echoes from planar surfaces. Sonar scanning stops when the current landmark parameters match those of the first landmark. Two targets configured with three posts and an added plane illustrate the procedure. Cognitive maps exhibit landmark locations that are accurate to ±5% with post radius estimates accurate to ±20%.

Forming maps of targets having multiple reflectors with a biomimetic audible sonar.

A biomimetic audible sonar mimics human echolocation by emitting clicks and sensing echoes binaurally to investigate the limitations in acoustic mapping of 2.5 dimensional targets. A monaural sonar that provides only echo time-of-flight values produces biased maps that lie outside the target surfaces. Reflector bearing estimates derived from the first echoes detected by a binaural sonar are employed to form unbiased maps. Multiple echoes from a target introduce phantom-reflector artifacts into its map because later echoes are produced by reflectors at bearings different from those determined from the first echoes. In addition, overlapping echoes interfere to produce bearing errors. Addressing the causes of these bearing errors motivates a processing approach that employs template matching to extract valid echoes. Interfering echoes can mimic a valid echo and also form PR artifacts. These artifacts are eliminated by recognizing the bearing fluctuations that characterize echo interference. Removing PR artifacts produces a map that resembles the physical target shape to within the resolution capabilities of the sonar. The remaining differences between the target shape and the final map are void artifacts caused by invalid or missing echoes.

Audible sonar images generated with proprioception for target analysis

Some blind humans have demonstrated the ability to detect and classify objects with echolocation using palatal clicks. An audible-sonar robot mimics human click emissions, binaural hearing, and head movements to extract interaural time and level differences from target echoes. Targets of various complexity are examined by transverse displacements of the sonar and by target pose rotations that model movements performed by the blind. Controlled sonar movements executed by the robot provide data that model proprioception information available to blind humans for examining targets from various aspects. The audible sonar uses this sonar location and orientation information to form two-dimensional target images that are similar to medical diagnostic ultrasound tomograms. Simple targets, such as single round and square posts, produce distinguishable and recognizable images. More complex targets configured with several simple objects generate diffraction effects and multiple reflections that produce image artifacts. The presentation illustrates the capabilities and limitations of target classification from audible sonar images.

Experimental audible sonar to model echolocation by the blind

Some blind humans echolocate by emitting palatal clicks and processing echoes, even when there is temporal emission/echo overlap (EEO). Our previous work indicates that high frequencies in the emission that travel directly to the ears are attenuated and, when high frequencies are heard, they come from target echoes. Binaural processing of these high frequency components in the EEO signals provide target range and bearing. Our experiments with 3D-printed parabolic pinnae and a speaker emitter indicate that while the high frequency components are important for object classification, the low frequencies in the emission provide more robust bearing localization because pinna diffraction effects are reduced. Classifying targets with EEO signals requires analysis in the power spectral domain because of the phase insensitivity of hearing. The power spectrum Fourier inverse estimates the autocorrelation function of the target reflector sequence. Robust autocorrelation estimates occur when the emission and echo contain comparable energies. Our experiments demonstrate that the echo energy varies with target type (e.g., planar or cylindrical) and the optimum range for classification depends on the target and itself forms a target classification feature.

Modeling human echolocation of near-range targets with an audible sonar.

Blind humans echolocate nearby targets by emitting palatal clicks and perceiving echoes that the auditory system is not able to resolve temporally. The mechanism for perceiving near-range echoes is not known. This paper models the direct mouth-to-ear signal (MES) and the echo to show that the echo enhances the high-frequency components in the composite MES/echo signal with features that allow echolocation. The mouth emission beam narrows with increasing frequency and exhibits frequency-dependent transmission notches in the backward direction toward the ears as predicted by the piston-in-sphere model. The ears positioned behind the mouth detect a MES that contains predominantly the low frequencies contained in the emission. Hence the high-frequency components in the emission that are perceived by the ears are enhanced by the echoes. A pulse/echo audible sonar verifies this model by echolocating targets from 5 cm range, where the MES and echo overlap significantly, to 55 cm. The model predicts that unambiguous ranging occurs over a limited range and that there is an optimal range that produces the highest range resolution.