Intelligent method for moving object identification based on feature clustering in the mel-frequency cepstral domain

Abstract

The implementation of intelligent monitoring systems in modern industrial automation and advanced transportation networks requires the development of highly reliable methods for moving object identification. The relevance of this study is driven by the necessity to create systems capable of stable operation under conditions of intense acoustic noise, limited optical visibility, and dynamic environmental parameters. Traditional amplitude-based sound processing methods prove to be ineffective due to their high sensitivity to signal non-stationarity and the Doppler effect, which prevents accurate segmentation of an object’s dynamic states as the distance to the sensor changes.

The objective of this work is to develop and scientifically substantiate a method for identifying motion phases based on the complex application of Mel-frequency cepstral analysis and intelligent clustering. The proposed approach is based on the synthesis of Short-Time Fourier Transform) algorithms using the Hamming window to minimize spectral leakage, psychoacoustic filtering based on Mel-scale filters, and Discrete Cosine Transform for feature disintegration and decorrelation in the cepstral domain.

The research methodology is based on converting a non-stationary acoustic signal into a time-frequency representation followed by the calculation of Mel-Frequency Cepstral Coefficients. This allowed for the formation of a multidimensional feature matrix that acts as a unique "digital fingerprint" of the object. A significant advantage of this method is its invariance to amplitude (loudness) fluctuations, achieved through logarithmic energy scaling within each critical frequency band. For automatic state recognition, an unsupervised machine learning method – K-means clustering – was applied, allowing for the grouping of feature vectors into compact centroids corresponding to specific motion phases.

The scientific novelty of the results lies in the improvement of the dynamic state segmentation process for moving objects by transitioning from time-domain characteristic analysis to analysis within the cepstral feature space. It has been experimentally confirmed that distinct differentiation is possible between the phases of approach, transit (point of closest proximity), and departure of the object. The results demonstrate high computational stability and speed, ensuring the possibility of integrating the method into local microprocessor-based automated control systems, PLCs, and real-time technical diagnostic systems. The practical value of the work lies in the potential to create autonomous acoustic detectors for monitoring industrial equipment and traffic flows without the need for expensive video surveillance systems.