Designs and Evaluation of Raccoon Detector for Anti-Vermin Surveillance Cameras

Abstract

Aiming for anti-vermin surveillance cameras, this paper presents evaluation and trial-implementations of classifiers based on machine learning and construction of database to improve the accuracy for actual images. Based on a database composed of images taken in a zoo, we construct classifiers with HOG features and a support vector machine (HOG + SVM), HOG features and a neural network (HOG + NN), and a convolutional neural network (CNN). The results show that the HOG + SVM classifier is the fastest and the CNN classifier achieves the best accuracy. Trial implementations of the two classifiers on an embedded processor show that the HOG+NN achieves 88.5% accuracy and takes 8.7sec processing time for a VGA sized image frame while the CNN achieves 94.5% accuracy and takes 45.8sec processing time. For images actually taken in shrines and temples, an experiment reveals that the CNN achieves merely 20-30% of accuracy but that training with a database composed of differently characterized image sets improves the accuracy to around 60%. An improvement of the program with background subtraction technique is given and the processing time is reduced to few seconds or less. The results indicate that the CNN trained with the carefully constructed image database shows good accuracy to detect raccoons on low-computing-performance surveillance cameras.