Fig. 1. Line segment L and vector .

Fig. 2. Let s = constant, line segment [−0.25, 0.25] convolved with Cauchy function, in first row changing s can adjust the shape of the convolution surface; in second row, keep s = constant, adjusting (q₀, q₁) can change the shape of the convolution surface; in third row, adjusting (q₀, q₁) can change the shape of the convolution curve.

Fig. 3. a–c and d–f show human body with two gestures in different viewpoints, h is the correspondence convolution curve of model g.

Fig. 4. First two images show the skeleton and convolution surface model without image information. Last four images show the initialization process with a specified gesture.

Fig. 5. First row is the gray video sequences, second row is image after foreground segmentation.

Fig. 6. The figure with red line is the rough contour and the figure with blue line is the contour after smoothing. (For interpretation of the references in colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 7. The sketch map of the joint and human skeleton likelihood.

Fig. 8. Continuous 10 frame image and the recovered 3D model of a walking man from CMU video datasets: a row shows the gray image sequence. The b–k row show the estimation of a walking man. The 1th column shows the human contour and convolution curve, the cross-mark means the projection of 3D joint point; the 2th column shows the skeleton after image skeletonization; the 3th column shows the estimated 3D skeleton; the 4th column shows the reconstruction 3D human model.

Fig. 9. First image shows us the optimization with joint and skeleton constrain. Second image shows us the optimization without joint and skeleton constrain.

Fig. 10. The sketch map of convolution surface projection under orthogonal projection, the broken line on convolution surface is its contour generator and the real line on image plane is convolution curve.

Fig. 11. Human model under orthogonal projection.

The DOF numbers of main model joints