lassi onk: a system framework to annotate and classify vehicular honk from road traffic.

Area of Science:

• Environmental Science
• Acoustics
• Artificial Intelligence

Background:

• Vehicular traffic and honking are major contributors to urban noise pollution, impacting health and road safety.
• Existing honk detection models struggle with classification in real-world, noisy environments.
• Classifying honks by vehicle type can offer valuable contextual information about urban areas.

Purpose of the Study:

• To develop a novel framework for sensing, labeling, and classifying vehicular honks based on vehicle type.
• To address the limitations of current models in classifying honks amidst ambient noise.
• To infer contextual information about urban locations using classified honk signatures.

Main Methods:

• A novel framework, 'lassionk', was developed for raw vehicular honk sensing and data labeling.
• Spectrogram images were generated from spatio-temporally collected audio samples.
• A deep learning-based Multi-label Autoencoder (MAE) model was used for automated data labeling (97.64% accuracy).
• An Ensembled Transfer Learning (EnTL) model, integrating Inception V3, ResNet50, MobileNet, and ShuffleNet, was proposed for honk classification.

Main Results:

• The MAE model achieved 97.64% accuracy in automated labeling of unlabeled honk data.
• The EnTL model demonstrated superior performance in classifying vehicle honks, achieving 97.64% accuracy.
• The EnTL model outperformed individual pre-trained models like Inception V3, ResNet50, MobileNet, and ShuffleNet.
• Classified honk signatures were used to identify the context of urban locations.

Conclusions:

• The proposed 'lassionk' framework effectively detects and classifies vehicular honks by type, even in noisy conditions.
• The EnTL model offers a robust solution for accurate vehicle honk classification.
• This technology can enhance environmental noise pollution monitoring and provide insights into urban dynamics.