| Summary: | 碩士 === 國立成功大學 === 電腦與通信工程研究所 === 96 === Human walking pose estimation plays an important role in modern computer vision. In this paper, we propose a method to estimate 3D human walking pose by a learning architecture. First, establishing the database by synthesize the human walking sequence. Then, projecting the 3D human model to 2D image plane obtains the 2D position from different views. By Applying the skeleton vector, which subtracts the nearest the joint position and previous orientation in image sequence to represent the direction of human body part, we obtain the spatial and temporal of human pose information. Due to the dimensionality of features are less, we use the fast training method called classification and regression trees combined with multi-gradient treeboost to improve the disadvantage of long time training and get the fast testing tree. The testing phase, we do an auto-initialization of human body part( head, chest, and lower limbs)on first image, then applying particle filter tracking on body parts obtains the following 2d position in image sequence. Then, we use skeleton approximating to get the other body part position. Finally, using the output of skeleton as the input feature of decision tree infers the 3D human walking pose. Our quantitative analysis of experimental results showed a good performance by using our proposed method.
|
|---|
