Determination of similarity
Through the Pose Detection Node, the three-dimensional coordinates (x, y, z) of the body part and the accuracy (visibility) of the coordinates are derived. First, array having elements that is composed by body position coordinates is defined as a feature vector. Assuming that n body position data are used, the vector is as follows.
This is used to check the similarity of pose. Data is preprocessed through vector normalization to determine the similarity of pose. And use Cosine Distance and Weighted Distance as Hybrid.
Vector normalization is the operation of matching the scaling of vectors having different sizes. This process prevents data distortion and makes Cosine distance calculation simple. Vector Normalization was normalized based on the L2 Norm of the Euclidean space. The expression is:
Cosine Distance is a very good way to check the similarity of two vectors. Even if the scale difference between the two vectors is large, if the directions are the same, it determine that they are similar. The cosine distance can be easily obtained through the pre-normalized Cosine similarity of two vectors u and v.
Cosine Distance is a good criterion for judging the similarity of two vectors, but an inaccurate vector element is also a criterion for judging the same. To solve this problem, Distance weighted on data accuracy can be used as a criterion for determining similarity.
