Object classification in automotive ultrasonic surround sensing using a compact 2 × 2 sensor array and a convolutional neural network

Current park pilot systems are based on ultrasonic surround sensing and, thus, depend on the performance of ultrasonic sensors. Not only capturing the distance to obstacles but also classifying objects is crucial for advanced driver assist systems and ultrasonic perception. However, current single-element sensors are constrained in classification performance due to a lack of directional information that they are able to capture. In this study, we propose replacing the conventional single-element sensor with a small 2 × 2 array sensor to increase object classification accuracy. The array sensor enables the incorporation of directional information, enhancing target discrimination, even in the compact design of 2 × 2 elements. Further, we propose an efficient convolutional neural network (CNN) to classify preprocessed transducer signals based on experimental data. Several feature extraction methods using the delay-and-sum beamformer, minimum variance distortionless response beamformer, acoustic source maps, and an end-to-end approach are evaluated. Promising classification accuracies are achieved for the array sensor when feeding both the preprocessed transducer signals and an acoustic source map into the CNN, significantly outperforming the conventional single-element sensor. Ultimately, this paper demonstrates the potential of enhancing object classification in ultrasonic surround sensing using small aperture array sensors and leveraging directional information.