PROJECT TITLE :

Spatio-temporal matching for Human pose estimation in video - 2016

ABSTRACT:

Detection and tracking humans in videos have been long-standing problems in computer vision. Most successful approaches (e.g., deformable components models) heavily depend on discriminative models to create look detectors for body joints and generative models to constrain potential body configurations (e.g., trees). While these 2D models are successfully applied to pictures (and with less success to videos), a serious challenge is to generalize these models to address camera views. So as to attain read-invariance, these 2D models sometimes require a giant amount of coaching knowledge across views that is difficult to assemble and time-consuming to label. Unlike existing 2D models, this paper formulates the problem of human detection in videos as spatio-temporal matching (STM) between a 3D motion capture model and trajectories in videos. Our algorithm estimates the camera view and selects a subset of tracked trajectories that matches the motion of the 3D model. The STM is efficiently solved with linear programming, and it's strong to tracking mismatches, occlusions and outliers. To the best of our information this is the primary paper that solves the correspondence between video and 3D motion capture information for human create detection. Experiments on the CMU motion capture, Human3.6M, Berkeley MHAD and CMU MAD databases illustrate the benefits of our methodology over state-of-the-art approaches.