. This is indeed an open research problem - ``How likely is the current
configuration given the image observation?'' There are couple of things
to be considered to approach the problem. First, the image features should
be determined. As in the current implementation, edges can be looked for
image features. Other candidates for image features are chromatic cues,
texture and so on. Whatever feature the observation module looks for, it
should be something that allows the algorithm to efficiently compute the
above density function. Tradeoff between global features and local features
should be considered also. The whole edge map can be used for the image
feature, but it is claimed that the edges lying on the normals of the previous
estimate are sufficient to compute the density function. However, the accuracy
of the observation density in that case does not seem to be guaranteed.
The current observation model in the implementation is really for tutorial
purpose, since it can not generate the multiple hypothesis. Hybrid observation
model combining the above image features might be an alternative approach.
The dynamic model parameters are also key factors in the performance of the system. The heuristically chosen parameters can sometimes cause the slow or wrong response of the tracker, and this was shown in the figure 5. The more accurate parameters should be chosen through learning from the data, and this procedure is well described in [2]. There are some other issues of the implementation. The number of random samples chosen from the effective prior at each iteration has some tradeoff between the real-time performance and the accurate state probability estimation. It also depends heavily on the shape of the target for obvious reason. Also, the camera used in the experiment turns out to have very weird procedure of obtaining colors, and the ordinary RGB-YIQ conversion does not produces an acceptable grey-level images. The ``xv'' software available on the leland machines are used instead to convert the color images into the grey images.