Detection of Animals in Thermal Imagery for Surveillance using GAN and Object Detection Framework

Wild animals have been a challenge to farmers worldwide as they are very active during the nighttime. Animals like elephants, deer, monkeys, cows, rats, peacocks, and many cause severe damage to crops by trampling. It is easier to protect crops in daylight, but it is tough for farmers to protect the field at night. Even in the forest, it is hard for zoologists to understand the activity pattern of animals at night. To tackle the challenge of detecting and tracking the animals at night, we propose a model that focuses on animal detection on thermal images. Although object detection is an advanced problem in computer vision, they mainly focus on color images rather than thermal images. Hence, a powerful object detection technique is required to detect and recognize the objects in thermal images. In addition, plenty of datasets are available for normal objects. However, there is a dearth of the thermal for animals to carry out the research. The work aims to create the dataset by collecting thermal images from FLIR videos. In addition, the dataset lacks the training data required for deep learning methods. Hence, the ThermalGAN framework uses color images to convert into thermal images. After that, YOLOv4 is trained to estimate the position of the animal. The proposed model predicts the location of animals with an average precision of 84.77% and an F1-score of 94%.