Acoustic Scene Classification for Personal Commuting Mode: Detecting Polluting vs. Non Polluting Vehicles

In the non-lane driven and chaotic roads of the developing countries analyzing the moving vehicles is a great challenge to Intelligent Transport System (ITS) as the vehicles are not homogenous. Inspecting or analyzing the number and type of moving vehicles on the road can generate a significant amount of information for traffic management as the roads are crowded with either Public vehicles or Private vehicles. Further, the information about how many commuters use public vehicle could be used to frame various traffic management policies. In this paper, we have proposed a technique wherein we collect the commuter's Smartphone acoustic data to perform multilevel classification. The first step involves determining the mode of transportation, i.e., either pedestrian, or public vehicle such as Auto-rickshaws, followed by identifying the type of public vehicle, i.e., Auto-rickshaw, either as polluting, i.e., Diesel or as non-polluting, i.e., CNG. Since Diesel autos cause more pollution as compared to CNG vehicles, so the information collected in the second step will come in handy while framing certain environment management policies. The data collected is subjected to classification process involving framing, windowing, feature extraction, normalization, and classification. In addition to these, various factors which effect the classification rate, are experimented, like frame size, overlapping size, and windowing along with combinations of different feature sets. Various Temporal and Spectral features being analyzed are: ZCR, Spectral flux, Spectral Centroid and MFCC. Results have shown that MFCC outperforms other features set and has been passed to SVM classifier for further classification. Classification accuracy for Level 1 is found to be 95%, and Level 2 is 78.95%.