Background Categorization for Automatic Animal Detection in Aerial Videos Using Neural Networks

摘要

This paper addresses the problem of animal detection in natural environment from aerial videos. Since the natural environment is usually composed of several fundamental elements such as trees, grass, streams, etc., it is proposed to distinguish the animal by categorizing the background into several classes. From the manually labeled samples, texture as well as brightness features are extracted to train a feedforward Neural Network. Then the classifier is applied to filter the test frame to locate potential animal regions. Four texture measures calculated from Grey Level Co-occurrence Matrix (GLCM) are used for texture feature description. Instead of obtaining these texture measures from grey level images, it is proposed to carry out calculation for every channel of the RGB image. The implemented results illustrate that this feature extraction method works well and the texture feature is a decisive factor in background categorizing.