Weakly Aligned Multi-part Bag-of-Poses for Action Recognition from Depth Cameras

摘要: In this work, we propose an efficient and effective method to recognize human actions based on the estimated 3D positions of skeletal joints in temporal sequences depth maps. First, body skeleton is decomposed a set kinematic chains, position each joint expressed locally defined reference system, which makes coordinates invariant translations rotations. A multi-part bag-of-poses approach then defined, permits separate alignment parts through nearest-neighbor classification. Experiments conducted MSR Daily Activity dataset show promising results.