Modality of Adaptive Neuro-Fuzzy Classifier for Acoustic Signal-Based Traffic Density State Estimation Employing Linguistic Hedges for Feature Selection

Abstract

Modality of adaptive neuro-fuzzy classifier (ANFC) for vehicular traffic density estimation using linguistic hedges (LH) and feature selection (FS) approach is proposed in this work. Cumulative vehicular acoustic signal is collected from roadside installed Omni-directional microphone followed by acoustic feature extraction using Mel frequency cepstral coefficients for varying combination of frame size and shift size. ANFC is modeled to classify traffic density states as low, medium, and heavy. Classification performance is further improved through modeling of ANFC with LH. Feature selection criteria are considered for varying combinations of frame size and shift, where linguistic hedges are employed for FS followed by ANFC to model traffic density states. We are getting better classification performance as compared to state of art literature for lower combinations of frame and shift size and even for consideration of single frame feature vector. Consideration of multiple contiguous frames will definitely increase the accuracy but with cost of computational time.