Formant-based acoustic features for cow´s estrus detection in audio surveillance system

In this paper, we developed an optimal formant feature subset algorithm for the detection of cow´s estrus vocalizations and introduced a prototype system to distinguish the differences between estrus and normal sounds from pattern recognition perspectives. Primarily, we found that there exist 19 formants in a spectrogram of Korean native cow vocalization, and this important finding initiated us to introduce a formant-based feature subset selection algorithm. We obtained the optimal formant feature subset {F1, F2, F4, F7, F14, F19} for the detection of Korean native cow´s estrus. Finally, performance evaluation was conducted using real vocalizations in a commercial loose barn, in which the average detection accuracy reached 97.5%, with false positive rate and false negative rate on average approaching 5.0% and 2.5%, respectively, when AdaBoost.M1 was used as a detector.