/CV_PoseEstimation

Single Human Pose Estimation Using SVM

Primary LanguagePython

Single Human Pose Estimation Using SVM

Introduction

The problem of human pose estimation, defined as the problem of localization of human joints. Human pose estimation in this work provide the method that not using Deep Learning approach as much as possible.

Approach

Foreground Segmentation

  • Background Subtraction, Morphology, Canny Edge Detection, Convex Hull

Depth Estimation (Deep Learning)

Superpixel Generation

  • Reduce time used for the classification

Superpixel Classification (using SVM)

Joint Estimation

Classification Feature

Superpixel Relative Position

  • X
  • Y

Superpixel Depth

  • Alpha

Result

From the implementation has showed that trained SVM classifier will have a good accuracy only when the image have a clear depth estimation value.
In this work estimate 16 key points: Head, Neck, Spine, Pelvis, r_Ankle, l_Ankle, r_Knee, l_Knee, r_Hip, l_Hip, r_Shoulder, l_Shoulder, r_Elbow, l_Elbow, r_Wrist, l_Wrist

Head l_Shoulder r_Shoulder l_Elbow r_Elbow l_Wrist r_Wrist Neck
All 10 10 10 10 10 8 9 9
Correct 7 4 2 2 0 1 1 4
% 70.00% 40.00% 20.00% 20.00% 0.00% 12.50% 11.11% 44.44%
Spine Pelvis l_Hip r_Hip l_Knee r_Knee l_Ankle r_Ankle
All 9 8 10 10 10 10 9 8
Correct 8 3 5 4 0 0 2 2
% 88.88% 38.00% 50.00% 40.00% 0.00% 0.00% 22.22% 25.00%

Result Example