Chamfer Matching, Tracking and Occlusion
Uploaded by
lukerichman29Chamfer Matching, Tracking and Occlusion
Uploaded by
lukerichman29Chamfer Matching, Tracking and
Occlusion
Chamfer Matching
A Chamfer is a transitional edge between two surfaces
of an object which is flattened or levelled to a certain angle.
Chamfer matching is an image registration method in which objects are matched with
templates using their boundaries. It basically matches a drawing with an image.
In this context, the word chamfer is used to denote an edge or a groove which forms the
boundaries of an object.
Chamfer Matching
The chamfer matching algorithm is mainly based on the distance transform function called
chamfer distance. Before that, we need to carry out the following steps:
● Image Segmentation: Any suitable segmentation algorithm can be applied to crop
out the required image and the necessary template with which it needs to be
compared.
● Binary Thresholding: If the input images are colored, they are binarized using a
thresholding technique for accurate results.
● Contour Detection: Contours or the outlines of both the image and the template are
detected and the contour/edge points are obtained.
Chamfer Distance
Chamfer distance between a template T and image I is the average of all the distances
between the edge points of template T and the nearest edge point in the image I. This is
given by the equation,
where, NT is the number of edge points in template T and DFI is the distance between the
edge point of T and nearest edge point of I.
Chamfer Distance
t1
i1
Fig. 1 Fig. 2 Fig. 3
If we consider Fig.1 to be the image and Fig. 2 to be the template, the distance between an
edge point of Fig. 2 denoted by t1 and nearest edge point in Fig. 1 denoted by i1 is DFI(1).
So, the average of all such distances from every edge point in the template is the chamfer
distance.
Chamfer Distance
If the two images are a perfect match the chamfer distance is zero. On the contrary, the
more the distortion or irregularity, the higher will be the chamfer distance.
The distance is computed in only N steps, where N is the number of points in template.
The method works best when the point set is sparse, so that the accurate location is
achieved, and computation is reduced.
Reducing the number of points too far will produce inaccurate results.
To obtain the best results, N may be on the order of a few hundred points for 2D to a few
thousand points for 3D image registration.
Improvements over original method
The original version of chamfer distance is limited because any outliers (caused by
occlusion or segmentation errors) will lead to matching problems. To limit this problem a
truncated distance can be used for matching:
where d is an empirical value which can be varied until the errors are rectified as best as
possible.
Improvements over original method
Another interesting approach is using the Hausdorff distance. The Hausdorff distance is
the longest distance from a point in one of the two sets, to the other set. In other words, it
is the greatest of all the distances from a point in one set to the closest point in the other
set. The two sets are of edge points of the image and the template.
The lesser the distance, better the match.
Tracking
● Object tracking is an essential problem in the field of video and image
processing.
● Tracking problem includes two main parts:-
● The first is to extract features of the image at time t and t+1 which can
distinguish the object and the background effectively.
● The second is to compare the extracted features with a template and estimate
the position of the object at t+1 based on its status at t.
Using Chamfer Matching algorithm for tracking
● The chamfer matching algorithm introduced a distance transform to find the
best fit of edge points from different images.
● Since computing exact Euclidean distance directly is resource demanding,
Gunilla Borgefors proposed a simplified method, 3-4 DT, to estimate the
distance.
● Borgefors proposed a hierarchical chamfer matching algorithm to find the best
match by a coarse-to-fine search.
● Chen proposed an improved version which can obtain an optimal fit through
simulated annealing.
● Thayananthan used chamfer matching to track hand shapes in cluttered image.
Improved Chamfer Matching
● Rotation Invariance: The original chamfer matching algorithm is sensitive to
rotation. The result can be affected by rotation distortion. Eigen
decomposition was employed on binary images to implement rotation
invariance.
● Orientation Distance transformation (ODT):The traditional distance
transform generates a distance map which can be used to estimate the
likelihood between two binary edge images. Where the object lies may be
concerned more than how much an image is similar with the template. ODT
assigns a complex number to each pixel. The magnitude and the phase angle
of a pixel’s value represent the distance and the direction to the nearest edge
pixel, respectively.
● Likelihood measurement:Classical chamfer matching uses root mean
square average as follow, to measure the likelihood.
The likelihood value is defined by:
● Shift Estimation: When the mask is superimposed over the orientation
distance map, the values of hit pixels in the map can be used to not only
measure a likelihood value, but also estimate a shift vector that implies where
the object may lie. The map was superimposed onto its orientation distance
transform result after moving it along an orientation for several pixels.
Results
Tracking objects in motion
Occlusion
“is the effect of one
object blocking another
object from view”
Handling Occlusion
Detecting Occlusion Occlusion Correction
Detection
Step 1: Object segmentation and colour tracking
Detection
Step 2: Object specification
ID:{
Area:” ”,
Width:” “,
Height:” “,
Co-ordinates: “ “,
}
Detection
Step 3: Repeat the process to all the frames and compute.
● Max_Area
● Max_Width
● Max_Height
● Co-Ordinates of Biggest Bounding Box
Step 4: Compare these results with all frames and detect the occlusion.
Correction
Case I: when two or more areas of with the same mean colour is detected
occlusion issue will be solved by replacing the 2
bounding boxes of two different areas with a
reconstructed bounding box to track the
middle-occluded object from behind.
Correction
Case II: when the change in width or height of an object is detected
solved by reconstructing the smaller bounding box that
only tracks the exposed part of an object, with a new
bounding box that tracks the whole object including its
occluded part
Advantages of Occlusion
In Immersion Technology
It makes the virtuality look realistic
Thank You
