A dataset for environmental sound recognition in embedded systems for autonomous vehicles

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Autonomous vehicles require advanced sensory systems beyond vision and touch.
• Environmental sound recognition can mimic human auditory perception for enhanced situational awareness.
• Existing datasets may not be optimized for autonomous vehicle sound recognition needs.

Purpose of the Study:

• To adapt the UrbanSound8K dataset for autonomous vehicle applications.
• To create a specialized dataset for training sound recognition models for autonomous systems.
• To improve the decision-making capabilities of autonomous vehicles through auditory cues.

Main Methods:

• Extended the UrbanSound8K dataset by creating new 'background' and 'silence' classes.
• Incorporated audio samples from Freesound.org for the 'silence' class.
• Annotated 4,908 WAV files totaling 4.94 hours across 6 classes for the US8K_AV dataset.

Main Results:

• Developed the UrbanSound8K for Autonomous Vehicles (US8K_AV) dataset.
• The US8K_AV dataset contains 4.94 hours of annotated audio data.
• The dataset is suitable for training predictive models for embedded systems like Raspberry Pi.

Conclusions:

• The tailored US8K_AV dataset is crucial for advancing environmental sound recognition in autonomous vehicles.
• This dataset facilitates the development of more robust and aware autonomous driving systems.
• The research supports the integration of auditory sensing for safer and more efficient autonomous navigation.