Track points in video using Kanade-Lucas-Tomasi (KLT) algorithm
The point tracker object tracks a set of points using the Kanade-Lucas-Tomasi (KLT), feature-tracking algorithm. You can use the point tracker for video stabilization, camera motion estimation, and object tracking. It works particularly well for tracking objects that do not change shape and for those that exhibit visual texture. The point tracker is often used for short-term tracking as part of a larger tracking framework.
As the point tracker algorithm progresses over time, points can be lost due to lighting variation, out of plane rotation, or articulated motion. To track an object over a long period of time, you may need to reacquire points periodically.
Starting in R2016b, instead of using the
to perform the operation defined by the System
object™, you can
call the object with arguments, as if it were a function. For example,
= step(obj,x) and
y = obj(x) perform
pointTracker = vision.PointTracker returns
a System object,
pointTracker, that tracks a set
of points in a video.
pointTracker = vision.PointTracker( configures
the tracker object properties, specified as one or more name-value
pair arguments. Unspecified properties have default values.
To initialize the tracking process, you must use the
to specify the initial locations of the points and the initial video
points to track and sets the initial video frame. The initial locations
must be an M-by-2 array of [x y] coordinates. The
initial video frame,
I, must be a 2-D grayscale
or RGB image and must be the same size and data type as the video
frames passed to the
Define and set up your point tracker object using the constructor.
step method with the input
I, and the point tracker object,
See the following syntax for using the
After initializing the tracking process, use the
to track the points in subsequent video frames. You can also reset
the points at any time by using the
[points,point_validity] = step(pointTracker,I) tracks
the points in the input frame,
I using the point
pointTracker. The output
an M-by-2 array of [x y] coordinates that correspond
to the new locations of the points in the input frame,
point_validity provides an M-by-1
logical array, indicating whether or not each point has been reliably
A point can be invalid for several reasons. The point can become
invalid if it falls outside of the image. Also, it can become invalid
if the spatial gradient matrix computed in its neighborhood is singular.
If the bidirectional error is greater than the
this condition can also make the point invalid.
[points,point_validity,scores] = step(pointTracker,I) additionally
returns the confidence score for each point. The M-by-1
scores, contains values between
These values correspond to the degree of similarity between the neighborhood
around the previous location and new location of each point. These
values are computed as a function of the sum of squared differences
between the previous and new neighborhoods. The greatest tracking
confidence corresponds to a perfect match score of
setPoints(pointTracker, points) sets the
points for tracking. The method sets the M-by-2
of [x y] coordinates with the
points to track. You can use this method if the points need to be
redetected because too many of them have been lost during tracking.
lets you mark points as either valid or invalid. The input logical
point_validity of length M,
contains the true or false value corresponding to the validity of
the point to be tracked. The length M corresponds
to the number of points. A false value indicates an invalid point
that should not be tracked. For example, you can use this method with
to determine the transformation between the point locations in the
previous and current frames. You can mark the outliers as invalid.
Number of pyramid levels
Specify an integer scalar number of pyramid levels. The point
tracker implementation of the KLT algorithm uses image pyramids. The
object generates an image pyramid, where each level is reduced in
resolution by a factor of two compared to the previous level. Selecting
a pyramid level greater than 1, enables the algorithm to track the
points at multiple levels of resolution, starting at the lowest level.
Increasing the number of pyramid levels allows the algorithm to handle
larger displacements of points between frames. However, computation
cost also increases. Recommended values are between
Each pyramid level is formed by down-sampling the previous level by a factor of two in width and height. The point tracker begins tracking each point in the lowest resolution level, and continues tracking until convergence. The object propagates the result of that level to the next level as the initial guess of the point locations. In this way, the tracking is refined with each level, up to the original image. Using the pyramid levels allows the point tracker to handle large pixel motions, which can comprise distances greater than the neighborhood size.
Forward-backward error threshold
Specify a numeric scalar for the maximum bidirectional error.
If the value is less than
Using the bidirectional error is an effective way to eliminate
points that could not be reliably tracked. However, the bidirectional
error requires additional computation. When you set the
Size of neighborhood
Specify a two-element vector, [height, width] to represent the
neighborhood around each point being tracked. The height and width
must be odd integers. This neighborhood defines the area for the spatial
gradient matrix computation. The minimum value for
Maximum number of search iterations
Specify a positive integer scalar for the maximum number of
search iterations for each point. The KLT algorithm performs an iterative
search for the new location of each point until convergence. Typically,
the algorithm converges within 10 iterations. This property sets the
limit on the number of search iterations. Recommended values are between
|initialize||Initialize video frame and points to track|
|setPoints||Set points to track|
|step||Track points in video using Kanade-Lucas-Tomasi (KLT) algorithm|
Create System objects for reading and displaying video and for drawing a bounding box of the object.
videoFileReader = vision.VideoFileReader('visionface.avi'); videoPlayer = vision.VideoPlayer('Position',[100,100,680,520]);
Read the first video frame, which contains the object, define the region.
objectFrame = step(videoFileReader); objectRegion = [264,122,93,93];
As an alternative, you can use the following commands to select the object region using a mouse. The object must occupy the majority of the region. figure; imshow(objectFrame); objectRegion=round(getPosition(imrect))
Show initial frame with a red bounding box.
objectImage = insertShape(objectFrame,'Rectangle',objectRegion,'Color','red'); figure; imshow(objectImage); title('Red box shows object region');
Detect interest points in the object region.
points = detectMinEigenFeatures(rgb2gray(objectFrame),'ROI',objectRegion);
Display the detected points.
pointImage = insertMarker(objectFrame,points.Location,'+','Color','white'); figure; imshow(pointImage); title('Detected interest points');
Create a tracker object.
tracker = vision.PointTracker('MaxBidirectionalError',1);
Initialize the tracker.
Read, track, display points, and results in each video frame.
while ~isDone(videoFileReader) frame = step(videoFileReader); [points, validity] = step(tracker,frame); out = insertMarker(frame,points(validity, :),'+'); step(videoPlayer,out); end
Release the video reader and player.
Lucas, Bruce D. and Takeo Kanade. “An Iterative Image Registration Technique with an Application to Stereo Vision,”Proceedings of the 7th International Joint Conference on Artificial Intelligence, April, 1981, pp. 674–679.
Tomasi, Carlo and Takeo Kanade. Detection and Tracking of Point Features, Computer Science Department, Carnegie Mellon University, April, 1991.
Shi, Jianbo and Carlo Tomasi. “Good Features to Track,” IEEE Conference on Computer Vision and Pattern Recognition, 1994, pp. 593–600.
Kalal, Zdenek, Krystian Mikolajczyk, and Jiri Matas. “Forward-Backward Error: Automatic Detection of Tracking Failures,” Proceedings of the 20th International Conference on Pattern Recognition, 2010, pages 2756–2759, 2010.
Usage notes and limitations:
Generates portable C code using a C++ compiler that links to OpenCV (Version 3.1.0) libraries. See Portable C Code Generation for Functions That Use OpenCV Library.
See System Objects in MATLAB Code Generation (MATLAB Coder).