SCAAT certainly takes a different approach than most trackers by providing incomplete tracking information. To be honest, I'm not sure that the commercial market will pick up on this right away. Many people who use commercial trackers want "final" and accurate results rather than partial information. Although the incremental improvements might produce pseudo-final results, if there is a measurement that is way off (either because of noise or because the tracked object moved really fast), it is hard to know what actually happened just from one measurement. Thus, the result at that instant might produce unwanted results.

In their defense however, this is an idea that deserves some study. SCAAT is a lot faster and more responsive than most, if not all, existing tracking systems. This alone can be a huge advantage if the application demands this. I also feel that the use of previous measurements is important to reduce errors in measurements. So, I think that some of the ideas SCAAT proposes will be important for tracking system in the future. It will be interesting to see what develops out of it.

To me, the difference between their tracking system and existing tracking system is similar to that of synchronous/asynchronous communication. Each system has advantages and disadvantages. If you want "final" and accurate results, an existing tracking system is probably best. However, if you want a system that is fast and responsive, then the SCAAT system may be the best.

Matt, as I understand it, a SCAAT system uses incomplete information, but its pose estimate is "complete" at the end of each measurement update. Visualization in state space might be more appropriate here - standard tracking systems try to estimate closed-form solutions for points in state space, while SCAAT pushes the current state estimate along the "track" most consistent with the most recent observation.

Though, one needs to ensure that the measurements can complementarily retrieve pose.

It's true that the paper is largely silent on the effect of measurement errors due to significant target dynamics, in particular whether such errors would be more than those in standard EKF implementations.

But the possibility of such errors is less in a SCAAT system. SCAAT allows for autocalibration in each iteration, so the device (source or sensor) that is most frequently used is also the most well-calibrated. This has an effect on accuracy akin to Amdahl's law on speed, consequently, one-off measurement errors (or random noises) actually may not have a very large effect.

I am intringued by the use of Kalman Filters in this example. Here the kalman filter is being used to track as well as calibrate using data that comes in sequentially.

I have questions about this paper this time: What is meant by "improve" when the so called isolation enforced by the SCAAT approach improves on situations when individual elements of the constraints are corrupted by indepedent noise over a batch or ensemble estimation scheme? (sec 2.2) I've always learned that the more data points one's got, the better the estimate.

For the 6 states that describes the incremental angle and rate of incremental angle change, we are really looking at the angular velocity of the point and tne angular acceleration. Why do we consider the acceleration for the angles and not the acceleration of the displacements? And this zeroing out the orientaiton elements of the state vector step...what's that for? Shouldn't it's changed be accounted for in the process noise or even the process transition matrix?

It is amazing how SCAAT works: its ability to use incomlete information to generate a reasonable estimate. Amazing that it works at all.

Wish the presentation would have "real results" mentioned in the paper to show. Watching video demonstration is always better than reading simulation results. =)

To answer the questions - in a SCAAT implementation, we are using as many data points as we'd use in a conventional implementation. Only, they are incorporated sequentially and not assumed to be simultaneous.

Extraordinary measurement noise is easily attributed to the source or sensor responsible for it, since there is a single source-sensor pair per estimate. In practice, extremely high estimation rates makes it possible to simply discard the tracker data for such noisy measurements, at the same time, retaining the device autocalibration refinement.

The incremental orientation elements are maintained internal to the EKF, but the derivatives are those of angles themselves and not the incremental angles. This is because the angle derivatives behave as orthogonal vectors and do not introduce non-linearities in pose computation as the orientation elements do.

SCAAT looks to be a really great algorithm not just for tracking but for dealing with unconstrained systems in general. Its efficient, it seems to be more accurate than other existing algorithms, and heck even the rate at which estimates are generated is improved... Now the paper doesn't mention any short comings, but are there any that anyone can see?

Matt mentions the case where a single outlier might throw off the whole system. Now, wouldn't this single outlier have affected the accuracies of commercial systems also?

Yeah, it is amazing. I was highly impressed by the rate and accuracy.

BTW, I recall the problem that KF has when the inputs are steady for a long time then suddenly change. Do you know if SCAAT has any improvement in this situation?

In section 3.4, the authors mention that it might be possible to do something smarter than the simple round-robin approach to measurement ordering. This seems to be an acknowledgment that some error could be introduced for some kinds of motion by always listening to the measurements in the same order - this problem is also present in traditional systems, and it's interesting that there is an opportunity to correct for it here, but they don't do it. It seems like even random choices for the next observation to use would be better than round-robin.