Robust Gunshot Features and Its Classification Using Support Vector Machine for Wildlife Protection

摘要

In the present study, automatic gunshot sound event detection in forest areas for wildlife protection has been considered. The feature extraction method used is robust to any length of gunshot sound events, which also takes care of the burst of multiple gunshots. First, low-level DWT-based features of large dimension were extracted which projected onto high-level histogram feature vector of small dimension using the bag-of-words approach. Support vector machine (SVM) classifier was considered to classify input audio signals into gunshot or forest ambience. The obtained results are highly reliable with a classification accuracy of 96.04% and area under ROC curve of 0.9866 indicating low false alarming rate. Automatic audio event detection can effectively expand the overall consistency of forest surveillance systems.