Hyperspectral image dataset of unstructured terrains for UGV perception

Citation Author(s):
Dhanushka
Liyanage
Tallinn University of Technology
Mart
Tamre
Tallinn University of Technology
Robert
Hudjakov
Tallinn University of Technology
Submitted by:
Dhanushka Liyanage
Last updated:
Sat, 02/03/2024 - 16:37
DOI:
10.21227/13bf-pa49
Data Format:
Link to Paper:
Links:
License:
5
1 rating - Please login to submit your rating.

Abstract 

 

Terrain perception for unmanned ground vehicles (UGVs) is one of the challenging tasks in machine vision. Especially when it comes to scene understanding, RGB images carry considerably less information than hyperspectral or multispectral images. Therefore, our approach was to use hyperspectral imaging methods for terrain perception. Even though hyperspectral images contain hundreds of image bands, it can be used to generate multispectral images by band selection using various band selection algorithms. This dataset is a combination of hyperspectral and multispectral images for UGV terrain perception in unstructured terrain (offroad) scenarios. Multispectral images are generated from the original hyperspectral images acquired using Specim IQ mobile hyperspectral camera which offers 400 – 1000 nm spectral range with 204 image bands. With spectral resolution of 7nm, the camera provides 512 x 512 pixels spatial resolution for hyperspectral images while 1280 x 960 pixels spatial resolution of RGB images.

 

The unstructured terrain image data collected in Estonia in the neighbourhood of Tallinn. All the images were captured during daylight hours. Total number of images in the dataset is 137 images including 204 band images of hyperspectral and 9, 16, 25 bands multispectral images. The wavelength details can be found in the accompanying thesis titled “Smart Terrain Perception Using Hyperspectral Imaging”.

 

Instructions: 
  •  Extract each data package separately as they contain images with different number of bands. i.e. HSI 9 dataset contains multispectral images of 9 image bands and its labels.
  • The dataset contains HSI images and labels. Hyperspectral images follow ENVI *.bil image format which can be easily read by using multibandread function on Matlab and SpectralPython libraries. Every image has its raw data which are binary (*.bil band interleaved by line) and header (*.hdr).
  • HSI_raw_dataset contains all the raw images captured from Specim camera with 204 image bands. The folder contains several subfolders which are the images named according to the image captured date. Each subfolder contains “capture”, “metadata”, “results” and RGB image from uncalibrated datacube in *.png format. The “capture” folder contains uncalibrated binary image data. White and dark reference images are also placed in the same folder. All of them follow nband interleaved by line format to store binary data. Since the sensor is a line scanner, white and dark reference images contain raw data for one line. The “results” folder contains RGB converted from hyperspectral datacube in *.png format and its ground truth also in *.png format.  Calibrated hyperspectral datacube file name contains “REFLECTANCE”. The RGB image also converted from hyperspectral included in the same folder in *.png format. Viewfinder image cropped into 645 x 645 pixels spatial resolution which is a *.png file and its ground truth can be find in the same folder. Higher resolution RGB images from the view finder also included.  
  • Refer the associated thesis for ontology.

 

Comments

good job

Submitted by Osman Sonmez on Mon, 02/19/2024 - 00:15