Track classifiction based on shape/texture

We have been investigating ways of classifying objects by their shape or texture that can be used in addition to the current test based on fitting a rectangular model. At this point we are primarily interested in recognizing objects such as foliage or ground returns which are hard to track succesfully, causing false motion detections, however we plan to develop a more general classification framework that should also be helpful for classifications of interest to the clients of the tracker data such as pedestrian, car, building, etc.

We first investigated classifying based on the spectrum of the object appearance, treating the mismatch between the measured surface and a line fit as a 1D signal. The results are somewhat discouraging.

Although there is clear difference between the spectra, there is not much new information not given by the RMS line match error and object size. In particular, though the amount of energy below the 0.5 meter scale varies considerably, the general shape of the spectra show similar high frequency rolloffs. This means that there is not much information not already present in the line match error amplitude.

We are currently investigating aligning multiple scans of the same object and examining the degree of similarity to characterize whether the object has stable structure beyond the first-order linear fit. Here is the result of aligning 30 scans of a hummer (approx 1 second of data.)

We can see that there is stable structure, and this conclusion is borne out by statistical analysis. This graph displays the mean of the aligned scans and the standard deviation. When the mean amplitude exceeds one standard deviation from zero there is good evidence of stable structure at that position.

We can summarize this graph by two numbers: the total RMS match error and the RMS amplitude of the mean (stable structure) signal. For the hummer, the numbers are 0.016 m and 0.045 m. The ratio of amplitude/match error can be taken as a general index of structure stability, with values > 1 indicating probable stable structure. In this case, the value is 2.8.

Compare this to the same graphics for the bush:

We can see that the scans are much more dissimilar, and there is no significant stable structure. The match error is 0.02 and the structure amplitude is 0.001, giving a stable structure index of 0.05. For these two cases, the stable structure index varies by a factor of 56, strongly distinguishing the two cases.

We plan to explore this method of shape characterization further, then to see how it can be applied in a particular classifier framework, and finally to integrate this work back into the real-time tracker.