Automatic Annotation of Hyperspectral Images and Spectral Signal Classification of People and Vehicles in Areas of Dense Vegetation with Deep Learning

• Main Authors: Adam Papp, Julian Pegoraro, Daniel Bauer, Philip Taupe, Christoph Wiesmeyr, Andreas Kriechbaum-Zabini
• Format: Article
• Language: English
• Published: MDPI AG 2020-07-01
• Series: Remote Sensing
• Subjects: hyperspectral imaging; deep learning; computer vision; automatic annotation
• Online Access: https://www.mdpi.com/2072-4292/12/13/2111

Despite recent advances in image and video processing, the detection of people or cars in areas of dense vegetation is still challenging due to landscape, illumination changes and strong occlusion. In this paper, we address this problem with the use of a hyperspectral camera—installed on the ground or possibly a drone—and detection based on spectral signatures. We introduce a novel automatic method for annotating spectral signatures based on a combination of state-of-the-art deep learning methods. After we collected millions of samples with our method, we used a deep learning approach to train a classifier to detect people and cars. Our results show that, based only on spectral signature classification, we can achieve an Matthews Correlation Coefficient of 0.83. We evaluate our classification method in areas with varying vegetation and discuss the limitations and constraints that the current hyperspectral imaging technology has. We conclude that spectral signature classification is possible with high accuracy in uncontrolled outdoor environments. Nevertheless, even with state-of-the-art compact passive hyperspectral imaging technology, high dynamic range of illumination and relatively low image resolution continue to pose major challenges when developing object detection algorithms for areas of dense vegetation.