Real-time detection system for elephant and dangerous wildlife intrusion along the forest-outskirt villages
Loading...
Date
2017-06
Authors
Law Hoi Chian
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Over the years, human have to exploit the nature for the sake of globalization growth. These activitiesare forcing the wildlife to move their habitat to a safer and more resourceful place; i.e., the habitat of human beings. Dangerous animals may intrude into the habitat of human beings for food or shelter. Elephant herds, tiger, sun bear and wild boar contributes the most for human-wildlife conflict in peninsular Malaysia. Hence, this project aims to build a real-time image detection of the intrusion of animalsmoving towards human population, especially at forest-outskirt areas. In this project, computer vision based detection is used, which an artificial intelligent system is developed to detect the intrusion of dangerous wildlife before they approach and destroy to the fences of the village. The target animals to be detected are elephant, sun bear, tiger and wild boar. The recognition system is modelled by two different methods; the Bag of Words (BoW) and Artificial Neural Network (ANN). Feature extraction algorithm to extract the features of the training images are Speeded-up Robust Features (SURF) and Color-based features. The learning algorithm to develop the recognition model is Support Vector Machines and feed forward back propagation of Artificial Neural Networks. Different testing procedures are done to test the performance of the recognition model and perform fine tuning of the training parameter of both the recognition models. The parameters to be tested to the BoW model develop the ideal recognition are: the type of training images, image preprocessing methods to the training and testing data set, train to validation ratio of the training image set and threshold value. For ANN recognition model, the parameters to be tested are type of training images, image preprocessing methods to the training and testing data set, number of neurons in the hidden layer, threshold value and number of retrain. The performance of two recognition models with ideal parameters are compared in terms of accuracy, sensitivity and specificity. The BoW model is selected to develop the recognition model, with accuracy of 69.29%, 76.32% overall sensitivity and 61.02% specificity, higher than the ANN model of 55.93% accuracy, 66.89% overall sensitivity and 41.38% specificity.