Hyper-spectral imaging for target detection and tracking

Abstract

Nowadays, there is a growing interest in the defense and security industry to track pedestrians in urban environments. In order to detect abnormal behavior, monitor crowded and sensitive landmarks, and study day to day patterns. The current state-of-the-art methods that are utilized to detect and track pedestrians tend to face many difficulties when the person is stationary, partially occluded, or if the lighting conditions of the environment are poor and continually changing. Hyperspectral cameras in recent years have been developing at a fast rate, which has opened the doors to many applications such as real-time tracking. Hyperspectral Images (HSI) in Near-infrared (NIR) are capable of detecting human skin, based on the reflection at skin tissue layers. Other human-related materials can also be detected, such as natural fabric and artificial fabric. In this thesis, a new tracking system that uses a hyperspectral camera is presented. The proposed tracking system works by analyzing the different spectral signatures of each pixel in an HSI, and classifying them into pre-trained classes. Spectral Angle Mapper (SAM) classifier is used to categorize pixels into different classes such as human skin, hair, and clothing. Information from the classifier enables us to detect and track pedestrians based on their spectral signatures. The proposed system is validated through experiments on a newly developed dataset. Additionally, the tracking system is compared to current state-of-the-art tracking systems. The experiments show that the proposed tracking system is capable of outperforming the current tracking system by tracking stationary and partially occluded pedestrians, in addition to being illuminance invariant.

Date of Award	Nov 2019
Original language	American English