抄録
In recent years, damage to agricultural products, livestock, and power transmission systems by crows is regarded as a serious problem; countermeasures against crow damage are urgently necessary. This paper proposed a method for detecting crow calls in various environmental sounds. If detection and discernment of crow calls were possible, various actions could be undertaken to prevent the damage. Wildlife call detection, not only that for crows, should be executed in extremely noisy environments. We then introduced both a spectrograph estimation technique with AR modeling in which AR coefficients are temporally smoothed and interpolated and a background noise elimination technique to obtain higher-quality crow call templates. The input sounds are compared with these templates by DP matching in the metric vector space of a logarithmic cepstrum. Every input sound whose minimal distance to the template database is less than the specified threshold value is detected as a crow call. The maximal detection performance can be obtained when five call templates in the template database are utilized; the implication is that carrion crows have five distinguishable call patterns. It is shown that the proposed method achieves 95% detection rate when 1.66% misdetection rate is allowed.
