During my MSc Eng. in Computer Vision i worked on various vision projects. The focus on these projects has been object tracking utilizing both 2D and 3D data sources. The application for thesis project has been for autonomous remote satellite capture.
Seen here are three different classes of object trackers:
This Iterative Closest Point tracker is capable of tracking rigid object at over 200fps utilizing a FPGA platform. Because the FPGA plateform allow for massivly parallel processing the focus of the project was to improve the performance of the nearest neighbor processing step
Some results of the system in operation can be seen here:
A 2D variant of Efficient Tracking with the Bounded Hough Transform is introduced to track objects in a sequence of images acquired from a digital camera. This 2D variant is able to track an object’s X and Y camera coordinates, X-axis rotation and its current scale factor.
By using the Bounded Hough Transform, we limit the possible inter-frame transformation space to only 9, 27 or 81 possible states. This depends on whether translation alone (9); translation and X-axis rotation (27); or translation, Z-axis rotation and scale (81) are tracked.
In the 3D approach, the acquisition of range data is the bottleneck of the system. Therefore, by tracking in 2D, we will effectively eliminate this acquisition bottleneck that is caused by the capturing and the processing of range data. Consequently, this method will need only to use an inexpensive digital camera rather then expensive range sensors.
Tests results show that tracking can be performed at a high rate of 180 frames per second when only translation and X-axis rotation are considered. This is due to the algorithm efficiently running at O(NM) for N data pixels sampled from a frame and M possible states. Further tests show good robustness to slight changes in lighting, occlusion, outliers and even camera generated noise.
To extend this tracker so that all possible rotations in three dimensions can be tracked, a view vertex sphere (geodesic sphere) can be used to build to compute a three dimensional model of the object. On this sphere each vertex (or viewing vertex) contains the corresponding image from its view point. See next video with cube object surrounded by a view vertex sphere.
To track the view point be now only need to compare the new data with the current and neighbour vertices (assuming that there are bounds on the objects interframe motion).
While with the two dimensional variant previously mentioned it was only required to perform a total of 81 comparisons to determine the new pose, this new three dimensional extension will only increase the number of comparisons by approximately seven (one for the current view and six for the neighboring view vertices.
This are the results of tracking the objects rotations:
About | Contact Michael Belshaw | ©2008
