Condensation makes a comeback! Nice to see that it had an impact on BraMBLe. Anyway, I laughed at the "generalized cylinder" that was used to find human blobs (reminds me of the old internet avatars). However, using such a model would allow great flexability in finding certain blobs. I can image the system can easily track non-humanoids just by changing this model. In certain situations, these blobs might even be used as a rough object recognition if you had multiple cylidrical models (so that of a human and a dog perhaps).

I do wonder about one thing though... In the paper's results, it mentions a case where two people cross in front of a third person (as seen in figure 8). They state that "if a separate foreground model were learned for each object then it might be possible to disambiguate the three people..." It seems to me that the answer could even be simpler than this. If in previous frames you have three independent "blobs" that then become one "blob", it seems that you could make it three "blobs" stacked on top of each other. With the addition of optical flow, a prediction could be made as to where these blobs are. In most cases this should work because people usually walk at a consistent pace. So there is no need to come up with a formal model of each "blob". Of course the REALLY easy solution is to have multiple static cameras surrounding an area (so maybe one 90 degrees of the other camera). Thus if the "blobs" are ambiguous in one, then another camera could disambiguate the result.

BraMBLe actually does rely on a constant velocity assumption to assign identities to blobs through occlusions (p(X_t | X_{t-1}) is a Gaussian centered at the constant velocity prediction). Even using the constant velocity assumption, BraMBLe swaps identities when two people cross in front of a third. A color histogram solution would also make the algorithm more robust, as the constant velocity assumption does not always hold. We are right now looking at extending BraMBLe with the depth-order heuristic from our mouse tracking work (the person in front at the start of the occlusion is the person in front at the end of the occlusion).

We are also trying to apply the BraMBLe algorithm to mouse tracking, using an ellipse instead of a "generalized cylinder".

In section 2 they mention that all filters are fixed at 5 pixels. I'm wondering how scaling would impact the filters and how much this would affect the likelihood algorithm or other higher level parts of the tracker. It seems to me like the graininess of the image filters shouldn't have a large impact, but I could be mistaken.

Is the generalized cylinder object model just a rough way to do object recognition as Matt implied, or is there another reason why it was used as opposed to say an ellipse?

Thanks for the info Kristin. I hadn't thought to use a color histogram but that is a good idea. I look forward to your report on the paper (and potential research).

Seems that a lot of human tracking algorithms have problems with people crossing in front of each other.
When would the combination of color histogram and shape tracking with velocity assumption fail? Under normal circumstances, wouldn't this combination should be sufficient?

Occlusion is really a problem for object tracking. We not only want to know how many objects are overlapped, but also want to get the depth relation of each object. Since there is no single information could completely solve this problem, integrating all the information we can get from the object into a big model may provide a better solution, although it would become time-consuming. That is we should consider the appearance, shape and motion together for a tracking model.

This is an interesting paper, although assuming constant velocity of people in a hallway seems like a stretch. My question is about common data sets for tracking algorithms - are there any? And since surveillance is a commonly mentioned application, why not study some real surveillance videos and try to figure out how often 'hard' sequences occur, and how they relate to the application domain?

I've read a few Segment, Label, and Track algorithms in the recent past. This paper introduced me to a few interesting ideas aside from the 2 principle ideas stated in the paper (appearance model in the form of a likelihood function P(zg|lg), and the bayesian mulitiple-object filter). I list them below.

1. Notion of a "super-pixel" where a patch of 5x5 pixels large is filtered to produce a data vector of 6 values, introduced to remove correlation, i.e. provide a compact representation of images locally.

2. Generalized Cylinders as a method to parameterize volumes of objects for the purpose of recognition.

3. Use of the feet point to determine position on room schematic from a calibrated camera (i know, it's a simple heuristic. But it was new to me).

4. There is a potential of incorporating separate foreground statistics for individual blob labelling to help disambiguate one blob with the next. But i don't see how this can be done without extensive training of many different ppl.

Hi all,

i have found this forum while googling for something related to the bramble algorithm...i don't know if it's still active, but in the case, i would like to pose a question related to an unclear aspect (at least to me!) of the paper:

How the multi-blob likelihood function is actually computed? I suspect that they use training sequences on which they manually label image regions according to algorithm 1, and then the obtained model is used on new sequences, in a supervised framework. Is that right?

Thank you very much.

Walter