No two humans are alike. We all look, and move, a bit differently. In fact, there have been some sci-fi movies that show that gait is as unique as a person’s fingerprint (or ears, which are as unique as your fingerprints as well), and I can tell you that they are right. I, for one, can pick out my friends by their gait, even if they are so far away that I can’t see their faces.
Now, if we are talking about a normal, healthy population, though each person is still slightly different, they generally fall within standard categories, so it’s not surprising that our markerless system is pretty robust in tracking them; while they are different, they are also, more or less, the same.
One thing that we regularly encounter are questions regarding the suitability of our solution to track non “normal” individuals. For example, can we track babies, or people who are missing a fourth toe? How about different walking patterns like those found in clinical biomechanics, such as children with cerebral palsy? If they don’t behave normally, how is it possible that a system can track both normal and non-normal patterns?
One big difference is that the data used to train our algorithms is not from normal healthy adults walking in a lab. It consists of many millions of images, with hundreds of thousands of different individuals doing all sorts of different activities. There is such a massive variety in the appearance of the individuals in this set, whether it be the clothing they wear, the environment they are in, or the individual's age, to name a few, that this allows our algorithms to generalize.
Now what does this term mean? It effectively indicates that because our data set is so diverse, we don’t really care what you look like! We collaborate with many institutions that collect video data of children with cerebral palsy, and though their kinematics are very distinct, when you look at the anatomy of the knee for instance, and in particular how it looks on an image, it looks like a knee. And because our data sets have seen so many different types of knees, it knows what to look for. Now, if you get into more significant disabilities this can prove problematic; however, in general the statement holds true.
The second aspect of our algorithms that enables us to track those who aren’t “normal” is that for the most part, the kinematic constraints are very similar in most types of movement. Our data team continually adds new features to track on the body, which allows us to open up joint constraints that were historically not possible with markerless systems. For instance, our feet can be tracked with 6 degrees of freedom, which allows us to measure all sorts of non-normal patterns.
The final, and probably most important, aspect is that we do not have an underlying assumption of the movement that we are measuring. What this means is that we don’t tell the system “this is normal walking” or “this is running” that fits an underlying model to the data. We just measure the key points and apply constraints. So the output is not defined by our underlying assumptions, but by the movement itself.
Getting back to the original question of, how can the same system be used to measure babies as adults? I like to finish the conversations by saying that we measure humanoids, and that babies, and those with skeletal abnormalities, very much fit into those categories!
Comments