Object tracking via adaptive prediction of initial search point on mobile devices

-

Object tracking via adaptive prediction of initial search point on mobile devices




Goal

The goal of this project would be to implement the mapping portion of SLAM, i.e. to develop a mobile application that could create a planar grid given a known reference, then track an object along that grid. In order to constrain the object field, the camera will be fixed, and the object shall be a
cube (although the object location given will be strictly planar). The first step would be to create the grid in the first place, then try and detect the object and track it.

Approach

Tracking an object is difficult because a camera could be easily distracted by local maxima in similarity. Pan et al, however, developed a method for robustly tracking an object by using a Kalman filter to determine the next location of a traveling object [1]. In the test results, it could accurately
detect the location of a car traveling down a highway. From here, the location of the object in the image could be augmented with an estimation of the physical location. The object could be tracked along a planar surface, whose orientation is given with a reference length such as a piece of 8.5” x 11” paper. Assuming the camera matrix is known, the planar surface is not orthogonal to the camera plane, and the dimensions of the reference object, it should be possible to map the entire surface in
the field of view [3] I will use this algorithm with reference to the front face of the cube (always
facing the camera) to determine its location. In order to prove the success of the filter, I will not use a special color for the cube, as that would defeat the purpose of using the Kalman filter to provide a location estimate. However, in order to make the geometry detection simple, I will track a single face, using opencv’s squares.cpp. The performance of the project will be determined by a series of tests,
where the object moves at increasing speed and is then compared to the “dumb” non-filtered detection algorithm. A graph of the estimated position will be compared to the actual position of the object (which can be hand measured on a manual grid or tracked with devices such as the OptiTrack
in the Robotics Lab).

The final project will be a simplified version of the tracking algorithms used in Pilot AI Labs [2]. Given an object to recognize, it will be able to accurately track both the location of the object in the camera field of view and along a 2-D surface grid. I will plan to use my personal Android phone
as the camera for this project. It would be interesting to also implement a physical system that could follow the object, although finding the resources and building the feedback control may be beyond the scope of a three- week project.

Implementation

I do plan to use an android phone for this project, although I have one
myself.

References

http://www.cs.cmu.edu/~jiyanpan/papers/lncs06.pdf
http://pilot.ai
http://szeliski.org/Book/drafts/SzeliskiBook_20100903_draft.pdf
For reading about vision geometry:
http://web.stanford.edu/class/cs231a/resources.html
Multiple View Geometry in Computer Visio

DOWNLOAD CLICK HERE

Comments

Post a Comment

Popular posts from this blog

Light Field Images for Background Removal

-
Image
Light Field Images for Background Removal OUTPUT OVERVIEW Standard edge detection or foreground/background separation techniques, such as Otzu’smethod, require color or intensity differences between the background and regions that need tobe separated. For example, green screens are routinely set up as the background in a scene so thatthere is a clear difference in color between the background and foreground.Light filed images, captured in 4D and passively containing depth information for thescene can be used to approximate this effect. Depth estimation in the scene alone could provide a metric to separate the foreground and background, but more sophisticated methods are available.Considering the edge detection from a single image from a single viewpoint and analyzing the depths from the light field around the edges in that image, occluded edges and intensity or color edges can be distinguished. Along with the rest of the depth information this can allow the foreground and b...
Read more

Using Image Processing to Identify and Score Darts thrown into a Dartboard

-
Image
✔️Using Image Processing to Identify and Score Darts thrown into a Dartboard ✔️Introduction The game of darts is a throwing sport in which participants toss projectiles into a circular target  attached to a vertical surface. The target is divided into many regions which correspond to different point and multiplier values. Darts is commonly played across North America and Europe and is a popular past time for many so an application that allows players to more conveniently keep score would be widely  beneficial. “Darts” is a general term for a targeting game following this basic premise and many game variations exist within this archetype; all utilizing the standard dart projectiles with a regulation dart board. Each game variant may have different objectives for which players to aim but identifying the region in  which the dart has hit in the dartboard is necessary for proper scoring. This project proposes the use of image processing to identify thrown da...
Read more

Face Detection, Extraction, and Swapping on Mobile Devices

-
Image
Face Detection, Extraction, and Swapping on Mobile Devices OUTPUT: DOCUMENTATION: We will use the Viola-Jones face detector in OpenCV as a starting point. Once we have found two faces we will apply some post processing to make sure that we have the entire face without any holes. Once we have a binary mask of the faces’ fully connected components, we plan to use cvBlobsLib (similar to regionprops function in Matlab) to do face region labelling and extraction. Choosing the faces to swap will either be random or done through a simple UI. Initially we will translate the faces to match the new face’s centroid with the centroid of the old face. Then, using the relative locations of features detected using OpenCV (such as the eyes, nose, and mouth) we will determine the orientation of the face in the plane parallel to the camera lens. Then, the relative sizes of these features will allow us to determine the degree of rotation away from the camera. However, if the faces we ar...
Read more