A Rapid Video Frame Correspondence Algorithm for Agricultural Video Field Surveying

Dev S. Shrestha, Brian L. Steward, Kelly R. Thorp and Bo Li
2004 2004, Ottawa, Canada August 1 - 4, 2004   unpublished
Video frame correspondence is a key operation in video processing for image-based video field surveying for precision agriculture applications. Video frame correspondence takes significant processing time and can be a major constraint in real time video processing applications. A Kalman filter was used to predict future shifts from previously measured shifts, and a gradient ascent method was developed to search for the maximum normalized cross correlation in the vicinity of predicted shifts.
more » ... redicted shifts. Compared with the results from the previous minimum error method developed by authors, the time required to compute the shift for a 30 by 30 pixel image patch with a 90 by 90 pixel search region was approximately ten times less than searching for a match over the entire search region. In a Matlab® script implementation, only 9.5 seconds were required to find the correspondence between 500 video frames of a corn field using the new algorithm whereas with the minimum error method, 114.6 seconds were required. The gradient ascent method with Kalman shift prediction can be used to find the image shift for real time applications. However, the success of the process depends on the closeness of the predicted shift to the actual shift and the characteristics of the correlation surface. Abstract. Video frame correspondence is a key operation in video processing for image-based video field surveying for precision agriculture applications. Video frame correspondence takes significant processing time and can be a major constraint in real time video processing applications. A Kalman filter was used to predict future shifts from previously measured shifts, and a gradient ascent method was developed to search for the maximum normalized cross correlation in the vicinity of predicted shifts. Compared with the results from the previous minimum error method developed by authors, the time required to compute the shift for a 30 by 30 pixel image patch with a 90 by 90 pixel search region was approximately ten times less than searching for a match over the entire search region. In a Matlab ® script implementation, only 9.5 seconds were required to find the correspondence between 500 video frames of a corn field using the new algorithm whereas with the minimum error method, 114.6 seconds were required. The gradient ascent method with Kalman shift prediction can be used to find the image shift for real time applications. However, the success of the process depends on the closeness of the predicted shift to the actual shift and the characteristics of the correlation surface.
doi:10.13031/2013.17086 fatcat:ityllpcscvhsrbrwiivshfbnj4