Detection of one-horned rhino using multispectral images

Abstract

Animal detection and surveillance is an important field of research to address the needs for protection of endangered species among others. The challenges in animal detection include low-contrast and poor image quality which is commonly observed during night time. Researchers have mostly worked on day-light, low-contrast and thermal images. To handle the challenges of detection during night time, multispectral images in combination with deep architectures may be used for better detection performance. In the present work, one-horned rhino is considered for detection because they are getting endangered for reasons like poaching, natural calamities and diseases. A novel multispectral one-horned rhino dataset is introduced and the multispectral data is obtained by combining the channels of color images and the corresponding thermal images. Instance segmentation based techniques YOLACT and YOLACT++ are used here to detect rhinos with the above multispectral dataset. The performances of the detectors are studied in terms of mAP and FPS.