Figure 1.
Components of lightweight human pose estimation
Figure 2.
COCO (left) and OpenPose (right) coding diagrams
Figure 3.
The process of image inspection
Figure 4.
Open Pose’s network structure diagram
Figure 5.
Diagram of the forecast internal structure
TABLE I.
Lightweight backbone performance comparison
| AP, % | |
|---|---|
| MobileNet v1 (cut to conv4_1) | 37.9 |
| Dilated MobileNet v1 (cut to conv5_5) | 42.8 |
| Dilated MobileNet v1 (cut to conv5_6) | 43.2 |
| Dilated MobileNet v2 (cut to conv6_3) | 39.6 |
Figure 6.
Merge branching diagram
Figure 7.
Improved convolution kernel structure
Figure 8.
OpenPose model key point confidence distribution chart
Figure 9.
Lightweight OpenPose model key point confidence distribution chart
TABLE III.
Comparison of Average Confidence
| Average confidence of person i | 1 | 2 | 3 |
|---|---|---|---|
| OpenPose method | 0.572 | 0.473 | 0.473 |
| Lightweight OpenPose method | 0.625 | 0.718 | 0.779 |
TABLE IV.
Comparison before and after network structure improvement
| Type | FPS |
|---|---|
| OpenPose | 0.035 |
| Lightweight OpenPose | 20.889 |
Figure 10.
Multi-person pose recognition results