Single View Human Action Recognition using Key Pose Matching and Viterbi Path Searching

摘要: 3D human pose recovery is considered as a fundamental step in view-invariant action recognition. However, inferring poses from single view usually slow due to the large number of parameters that need be estimated and recovered are often ambiguous perspective projection. We present an approach does not explicitly infer at each frame. Instead, existing models we search for series actions best match input sequence. In our approach, modeled synthetic 2D rendered wide range viewpoints. The constraints on transition represented by graph model called Action Net. Given input, silhouette matching between frames key performed first using enhanced Pyramid Match Kernel algorithm. matched sequence then tracked Viterbi demonstrate this challenging video sets consisting 15 complex classes.