| dc.description.abstract |
For computers to understand human activity or behavior in a variety of scenarios, reli
able 3D human posture estimation is a prerequisite. A number of difficulties have made
such work more complex as it is influenced by various factors, including image quality,
background, garment texture and diversity, body shape, and the presence of other objects
alongside persons in the image which has depicted the necessity of adopting the technique
of computer vision. While much work has been done on 2D human pose estimation, show
ing state-of-the-art performance, the objective of this research is to estimate 3D human
pose from 2D joint positions. We have investigated deep neural networks comprising of
linear layers with residual blocks and proposed a hybrid deep learning framework in order
to achieve this objective. We experimented the proposed by raising the number of residual
blocks to anlaysis the performance. The final proposed architecture (HEpose) comprises
of three parallel models, one model is base one only the linear layers concept, second one
is based on the residul connection without normalization, and third model gathers the in
formation of connection among the joints. We combined ouputs of the three model and
f
inally used a fully connected linear layer to estimate 3D pose. We also showed compar
ative training results. Finally, the proposed architecture was evaluated on H3WB dataset
and presented the evaluation results considering the evaluation metrics of the mean per
joint position error (MPJPE) and the percentage of correct keypoints (PCK). The proposed
architecture performed about 50% better in terms of MPJPE and PCK@150mm for three
residual block. We had also compared the performance of HEpose with other state-of-the
art methods of 3D pose estimator and achieved inevitable performance. |
en_US |
