Application of deep learning methods for obstacle classification in ultrasonic surround sensing
Jona Eisele,André Gerlach,Marcus Maeder,Andreas Koch,Steffen Marburg
DOI: https://doi.org/10.1121/10.0027488
2024-03-01
The Journal of the Acoustical Society of America
Abstract:Ultrasonic sensors are widely used in driver assist systems for close-range surround sensing. Based on the pulse-echo method, the distance to obstacles is calculated, particularly in scenarios such as parking and maneuvering. For advanced ultrasonic perception systems, however, the classification of obstacles should be considered, too. For instance, it is relevant to distinguish between traversable and non-traversable objects, as well as to detect pedestrians. Classifying obstacles by ultrasonic echoes remains an area of ongoing research, and, in contrast to ultrasonic ranging, implementation in current systems is rudimentary at best. This work examines the application of deep learning methods for ultrasonic-based object classification in driver assist systems. Features based on raw time signals, amplitude- and phase-based echo features as well as time-frequency images are compared. The preprocessing and feature extraction methods are evaluated based on experimental data captured in low and high clutter environments. Promising classification results are achieved using convolutional neural networks, demonstrating a significant improvement over traditional methods. Finally, we discuss the benefits of employing small aperture ultrasonic sensor arrays to increase classification robustness using beamforming methods.
acoustics,audiology & speech-language pathology