Effective deep learning aided vehicle classification approach using Seismic Data.

Area of Science:

• Transportation Engineering
• Machine Learning
• Signal Processing

Background:

• Intelligent transportation systems (ITSs) rely on vehicle classification (VC) for traffic management.
• Traditional VC methods using visual or sensor data are limited by environmental factors and privacy concerns.
• There is a need for robust VC techniques resilient to adverse conditions.

Purpose of the Study:

• To develop a novel vehicle classification technique using seismic data.
• To address the limitations of traditional VC methods in ITSs.
• To leverage self-supervised learning for seismic signal classification.

Main Methods:

• A self-supervised contrastive learning approach was employed for seismic signal classification.
• Specialized data augmentation techniques were used to create positive and negative sample pairs.
• An encoder network and projection head were utilized for feature extraction and representation refinement.
• Contrastive loss function was applied to align similar seismic features and separate dissimilar ones.

Main Results:

• The proposed method achieved state-of-the-art performance in seismic signal classification.
• An accuracy of 99.8% was attained, demonstrating high precision in vehicle classification.
• The approach proved effective even with limited training data, highlighting its utility in data-scarce scenarios.
• The method showed reduced susceptibility to environmental conditions compared to traditional techniques.

Conclusions:

• Seismic data offers a robust alternative for vehicle classification in ITSs.
• Self-supervised contrastive learning is effective for analyzing seismic signals for VC.
• The developed technique provides a promising solution for accurate and reliable vehicle classification, especially in challenging environments.