Parameter Tuning for Speed Changes Detection in On-Road Audio Recordings of Single Drives

Abstract

AbstractEconomical driving not only saves fuel, but also reduces the carbon dioxide emissions from cars. Apart from environmental benefits, road safety is also increased when drivers avoid speeding and sudden changes of speeds. However, speed measurements usually do not reflect speed changes. In this paper, we address automatic detection of speed changes, based on audio on-road recordings, which can be taken at night and at low-vision conditions. In our approach, the extraction of information on speed changes is based on spectrogram data, converted to black-and-white representation. Next, the parameters of lines reflecting speed changes are extracted, and these parameters become a basis for distinguishing between three classes: accelerating, decelerating, and maintaining stable speed. Theoretical discussion of the thresholds for these classes are followed by experiments with automatic search for these thresholds. In this paper, we also discuss how the choice of the representation model parameters influences the correctness of classification of the audio data into one of three classes, i.e. acceleration, deceleration, and stable speed. Moreover, for 12-element feature vector we achieved accuracy comparable with the accuracy achieved for 575-element feature vector, applied in our previous work. The obtained results are presented in the paper.KeywordsDriver behaviorHough transformIntelligent transportation systems