SURF is a fast and
performant scale and rotation invariant point of interest descriptor as well as
detector 25. The employed of integral images reduce the number of
operations drastically for simple box convolutions and independent of the
chosen scale. This results in the high gain of speed as compared with SIFT.
of the weakness of SIFT is that the high dimensionality of the descriptor at
the matching step. In SURF, the interest point detector is based on the Hessian
approximation which is better than the state-of-the-art interest point
detector. SURF display a nearly real-time computation without any loss in
performance. Descriptor in SURF refers to the sums of Haar wavelet components.
As compared with histogram approach, description of the nature of the
underlying image intensity pattern is much more distinctive. Due to its
simplicity and the uses of integral images, the descriptor is said to be faster
than SIFT. Last but not least, the Laplacian-based indexing strategy makes the
matching step in SURF is faster than in SIFT. The faster matching step have no
any loss in terms of its performance.
26 applied SURF in generating the panoramic street view
images of four perspective images which totally covered the horizontal angle of
360 degree and vertical angle of 270 degree. SURF-based matching algorithm is
employed between each consecutive pair of images along the video sequence. In
order to estimate the depth map for each image sequentially, Multiview
super-pixel stereo method is applied.
27 indicated that, when multiple images of same scene
taken in different times, from different viewpoints, by different sensors, the
process to overlay all the images called image registration. The reference and
sensed images are geometrically aligned. In image registration, through the
combination of different data sources such as image fusion, change detection,
and multichannel image restoration, the final information of the images can be
obtained. Image registration is widely applied in the field of remote sensing,
medicine and computer vision.
24 applied image registration to register a frame in an
image sequence to map. With that, two approach of registration are employed
iteratively which are frame-to-frame registration and frame-to-map
registration. Frame-to-frame registration register a frame to its previous
frame which been registered to the previous frame in the past iteration.
Frame-to-map registration warp the frame to the map by the estimated
transformation. This compensate for scale and difference in the rotation and
finally perform an area according matching by applying Mutual Information to
obtain the correspondence between the warped frame and the map.
28 proposed to apply image registration to sequentially
compute the correspondences between road regions in the vehicle camera image
and the aerial image. Assumption of low resolution of aerial image is made in
the proposed method. Hence, image registration is applied to match images
consistently by using the entire image. To generate a more accurate road image,
the result of previous frame is used to stable the initial parameter estimation
for the image registration. Image registration is performed between the vehicle
camera image It and the aerial image J, as shown in Figure 2.7.