A hybrid approach for acoustic signal segmentation by computing similarity matrix, novelty score and peak detection for vehicular classification in wireless sensor networks

Vehicle acoustic signals have long been considered as significant source in sensor networks for classification. In this research acoustic signals generated by each vehicle will be used to detect its presence and classify the type. Vehicle acoustic signal segmentation is important for continuous signal recognition because it reduces the search space effectively in vehicle's signal recognition. However, for Vehicle classification, it is difficult to segment the signal input reliably into useful sub-units because i) vehicle sound units can often be located roughly via intensity changes ii) energy changes in signal spectrum or amplitude help to estimate unit boundaries, but these cues are often unreliable. In this paper the series of steps proposed are signal segmentation is presented. includes decomposition of signals into successive frames of 50 ms without overlap. The computations of the spectrum representation (FFT) of the frames are carried out. The similarity matrix that shows the similarity between the spectrums of different frames is computed. Estimation of the novelty score related to the similarity matrix is done. The detection of the peaks in the novelty score is made and finally segmenting the vehicle acoustic signals using the peaks as position is done. These segmented signals are further used for feature extraction and classification.