Geo-referencing and mosaicing agricultural field images from a close-range sensing platform

Jiang, Yanshui

Permalink

https://hdl.handle.net/2142/15981 Copy

Description

Title

Geo-referencing and mosaicing agricultural field images from a close-range sensing platform

Author(s)

Jiang, Yanshui

Issue Date

2010-05-18T18:53:23Z

Director of Research (if dissertation) or Advisor (if thesis)

Tian, Lei

Department of Study

Engineering Administration

Discipline

Agricultural & Biological Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Geo-referencing
• Image Mosaic

Abstract

Field images are becoming more frequently used for sensing the crucial properties of the crops in Precision Agriculture. Some of them contain distortions that need to be removed before further analysis. Data from images, such as the coordinates of the crops, also need specific algorithm to be extracted. For these purposes, two computer vision algorithms were developed for pre-processing the field images from two monocular vision systems. One algorithm was used for a tower remote sensing system data pre-processing for image distortion removal and the mosaic to generate geo-referenced images. The other was for the image data interpretation used for the vision system of a field robot. Satellite and aerial remote sensing systems are the two major platforms for collecting remote sensing images for agriculture. However, due to the critical drawbacks of these systems, such as low spatial and temporal resolution, a tower remote sensing system with a 360-degree rotatable camera on the top has been established in the experiment field to obtain the multi-spectra images for monitoring the status of the plants. In this research, the geo-reference and image mosaic algorithms were developed for data acquisition. While taking remote sensing images of the field, the camera will turn 360o horizontally and 90o vertically. This creates the difficulty of geo-reference because different images have different distortions. Therefore, traditional ways of geo-reference, such as using Ground Control Points (GCP), are no longer appropriate. A three-axis digital compass was used to provide the absolute orientation of the camera, which can be used to geo-reference a single image. The calibration of both camera and compass was introduced, and necessary parameters for geo-reference were estimated. Based on the angles, positions and optical parameters of the camera, a transformation from the image coordinate system to the ground coordinate system was introduced. After the transformation, the performance of the geo-reference method was evaluated with data from a Real-Time Kinematic Global Positioning System (RTK-GPS) to assure the accuracy. Since there are not enough features in the field, the algorithm of the image mosaic for the tower system is based on geographical information rather than features. Moreover, comparing to aerial and satellite systems, the images from a tower remote sensing system have usually 10 to 15 times larger pitch angles that result in large geometric distortion. Thus further processing is needed to remove distortion. The algorithm firstly used coordinate transformation to compute top-view coordinates of all pixels in the image. The new coordinates were used for reorganizing the pixels. Due to the large geo-metric distortion, the Pixel Combination was applied. After computing the top-view image, global alignment was applied to generate an initial image mosaic. This global alignment method was based on the geographical information of each image, which allows a pixel-level mosaic, without limitation of detection of feature points. The mosaic image was then improved by local alignment. In local alignment, the movement of each pixel was computed by an optical flow method. For covering the entire field, seventy-one images were taken for the Energy Farm near the campus of the University of Illinois at Urbana-Champaign. The accuracy of the image mosaic was tested using several markers in the field. The other algorithm of computer vision is the image data interpretation of a vision system. A monocular vision system for a field robot was developed to replace a former binocular stereo vision system. With this system, the 3D coordinates of plants can be geometrically estimated. Many approaches of monocular stereovision use mirrors and prisms to convert monocular images to binocular ones. However, the high frequency encoder on the robot makes it possible that without any optical accessories, one camera can still recover the 3D coordinates of the object, due to the availability of real time velocity measurement. The camera installed on the robot functions similarly to the tower camera, which can transform the coordinates in the image frame to the ground frame. Therefore, the monocular vision system is able to estimate the relative position of a plant in front of the robot. Furthermore, with the velocity of the robot detected, the vision system can even recover the height of the plant. The robot has an encoder installed that can measure the real-time speed of the robot. With this information, the displacement between two images can be obtained, thus the height of the plant can be estimated. Therefore, with this method, the 3-D geometric information of a plant can be obtained. Two types of experiments--laboratory test and field test--were conducted for evaluating accuracy. After the experiments, the results from both tests were compared to address the possible sources of errors. Compared to the ideal environment in a lab, the outdoor conditions could decrease the system’s performance. However, the final results show that the method has the ability to provide relatively accurate measurement.

Graduation Semester

2010-5

Permalink

http://hdl.handle.net/2142/15981

Copyright and License Information

Copyright 2010 Yanshui Jiang

Owning Collections