Hand Tracking via Efficient Database Indexing

Tracking the pose of a moving hand from a monocular perspective is a difficult problem. In the current study the tracking problem is recast as efficient database indexing. Recognition is hastened by approximating chamfer distances between hand poses which have been embedded into a common vector space. Using this embedding, a small number of candidate matches for each query can be quickly identified using simple vector comparisons. These candidates are then processed with an expensive metric (chamfer distance) to determine the optimal match. This approach does not suffer from the traditional weakness of tracking methods, and lends itself naturally to real time solutions.