Smartphone based traffic state detection using acoustic analysis and crowdsourcing

Typically an Intelligent Transportation System (ITS) employs various types of infrastructure-based technologies into vehicles and roadways for monitoring traffic. But these solutions have high installation, maintenance, and operational costs. Further, most of these solutions are based on the assumptions of lane-based organized and homogeneous traffic, due to which these are ineffective in less organized traffic conditions which are common in developing countries. In this paper, we have proposed a cost-effective approach to infer the traffic state of the road by analyzing the cumulative acoustic signal collected from the microphone sensor of the user's smartphone. To capture the distinctive characteristics of various traffic scenes, we explored two different types of features: Mel Frequency Cepstral Coefficients (MFCCs) and Wavelet Packet Transform (WPT). Based on the understanding of acoustic signals of different traffic scenes, various parameters of these features such as window size and MFCC dimensions are tuned for better detection accuracy and robustness. To validate the approach, field experiments were conducted in varied conditions on the roads of City X. Experimental results revealed that for binary traffic scene classification ('busy-street' vs. 'quiet-street'), MFCC features are sufficient to get an overall accuracy of 100%. However, for 'congestion' vs. 'medium-flow' vs. 'free-flow' traffic scene classification, MFCC features yield a bit lower accuracy of 77.64%. In this scenario, it was observed that WPT features can be used to reduce the false positive rate, thereby providing an absolute gain of 11.38% in the classification accuracy over the MFCC baseline. Also, it has been observed that by crowdsourcing the traffic state information from multiple users' smartphones, an effective accuracy improvement can be achieved for each traffic scene.