IntroductionAugmented, virtual or mixed reality? Still are some confusion due to the semantics in this area. But what it matter to me are the myriad of applications that are currently under development. Decided to explore the field, I have done my own application (simple enough). From my opinion, I consider this application that falls into the Augmented/Mixed reality approach, it overlays an object on top of scenario making the object to appear par of the scene. How does it work? While the approach can be describe very simple, getting it to work seems quite hard. For this application I have used a state of the art detector/tracker called CMT (Clustering of Static-Adaptive Correspondences for Deformable Object Tracking), while I could go for hours trying to explain the overall approach. I have focused in how I could use it. However, there is no restriction in what type of algorithm or approach is selected, but usually is nice to be able to have an algorithm that gives features such as plane rotation, scale change, and handles some deformation, most of the point-based algorithms are able to handle this up to some point. Descriptors such as FAST, ORB, BRISK and some others can be used to accomplish this task (in fact, CMT uses BRISK in a consensus manner. Hint!). Now we are able to have a model of the target and calculate plane rotations, scale changes, and probably some homography of the object for out-of-plane rotation. Another set of algorithms that could be used are those that are template matching based such as the cross-correlation, some extra implementation would be necessary to account for scale change. However, this type of approaches do not handle well out-plane-rotations. ResultsI have placed a Kalman Filter that estimates the center pixel coordinates, the height and width of the target (for more about the Kalman Filter, you can visit my older post in the following link) with an outlier rejection mechanism (the famous Mahalanobis distance) to be able to remain in the last position if the tracker/dectetor is lost. The overall approach works well with targets with a lot of features at somewhere around the 25FPS.
0 Comments
Leave a Reply. |
About meEvery venture in an unknown territory is exciting to me. I ended up working with autonomous robots using knowledge from fields such as; computer vision, Bayesian estimation, control theory, neural networks, and SLAM. I have always been fascinated by aerial and ground mobile vehicles. Thankfully I had the chance to work on algorithms that bring them to life. Archives |