Trials of Automated Detection of Raccoons for Surveillance Cameras

Abstract

Raccoons, that are alien species in Japan, are increasing significantly and recently cause serious damage to farm products and historical buildings. Our objective is to contribute the analysis of the behavior and distribution of raccoons and the prompt and effective reaction by developing more intelligent surveillance cameras. This manuscript reports trials of automated vermin detection for embedded systems including surveillance cameras. Following popular methods of image recognition, we adopt HOG (histograms of oriented gradients) as feature extraction scheme and SVM (support vector machine) or NN (Neural Network) as categorization scheme. We also adopt a variation of deep learning designated as CNN (convolutional neural network). We implement the image recognition methods of HOG+SVM, HOG+NN, and CNN as a software programs running on a PC and evaluated with training and test images of raccoons and raccoon dogs taken in a zoo. The experiment results in the recognition rates of 89.8%, 88.5%, and 94.5%, respectively, and the execution time of 111ms, 0.681ms, and 11.9ms, respectively.