Convolutional neural network with data augmentation for object classification in automotive ultrasonic sensing.

Area of Science:

• Robotics and Autonomous Systems
• Sensor Signal Processing
• Machine Learning for Automotive Applications

Background:

• Current automotive ultrasonic sensors have limited range and classification capabilities.
• Future autonomous driving and parking assist systems require enhanced sensor performance for object recognition and traversability assessment.
• Processing raw ultrasonic sensor data is crucial for extracting meaningful information about the environment.

Purpose of the Study:

• To investigate the classification of object classes and traversability using a single ultrasonic sensor.
• To develop and evaluate a novel signal processing chain and machine learning model for ultrasonic sensor data.
• To explore the impact of data augmentation techniques on model robustness and accuracy.

Main Methods:

• Acquisition of raw time signals from ultrasonic sensors.
• Transformation of time signals into time-frequency images (scalograms).
• Classification using a convolutional neural network (CNN) and comparison with a LeNet-5 baseline.
• Implementation of offline and online data augmentation strategies.
• Conducting stationary and dynamic measurements in laboratory and outdoor settings.

Main Results:

• The proposed scalogram-based CNN model significantly outperforms the LeNet-5 baseline.
• Object classification achieved 90.1% accuracy in the lab and 66.4% outdoors for seven classes.
• Traversability classification reached 96.4% accuracy in the lab and 91.5% outdoors.
• Data augmentation techniques were shown to improve model robustness.

Conclusions:

• A single ultrasonic sensor, coupled with advanced signal processing and machine learning, can effectively classify objects and assess traversability.
• The developed method offers a promising approach for enhancing perception in autonomous driving systems.
• Data augmentation is a critical component for building robust models for real-world automotive applications.