Limits...
Articulated Non-Rigid Point Set Registration for Human Pose Estimation from 3D Sensors.

Ge S, Fan G - Sensors (Basel) (2015)

Bottom Line: We introduce a visible point extraction method to initialize a new template for the current frame from the previous frame, which effectively reduces the ambiguity and uncertainty during registration.Third, to support robust and stable pose tracking, we develop a segment volume validation technique to detect tracking failures and to re-initialize pose registration if needed.The experimental results on both benchmark 3D laser scan and depth datasets demonstrate the effectiveness of the proposed framework when compared with state-of-the-art algorithms.

View Article: PubMed Central - PubMed

Affiliation: School of Electrical and Computer Engineering, Oklahoma State University, Stillwater, OK 74078, USA. song.ge@okstate.edu.

ABSTRACT
We propose a generative framework for 3D human pose estimation that is able to operate on both individual point sets and sequential depth data. We formulate human pose estimation as a point set registration problem, where we propose three new approaches to address several major technical challenges in this research. First, we integrate two registration techniques that have a complementary nature to cope with non-rigid and articulated deformations of the human body under a variety of poses. This unique combination allows us to handle point sets of complex body motion and large pose variation without any initial conditions, as required by most existing approaches. Second, we introduce an efficient pose tracking strategy to deal with sequential depth data, where the major challenge is the incomplete data due to self-occlusions and view changes. We introduce a visible point extraction method to initialize a new template for the current frame from the previous frame, which effectively reduces the ambiguity and uncertainty during registration. Third, to support robust and stable pose tracking, we develop a segment volume validation technique to detect tracking failures and to re-initialize pose registration if needed. The experimental results on both benchmark 3D laser scan and depth datasets demonstrate the effectiveness of the proposed framework when compared with state-of-the-art algorithms.

No MeSH data available.


Validation and re-initialization results for a passing case (first row) and three failed cases (second to fourth row). Columns (a–e) are the point set in the current frame, that in the previous frame, correspondence estimation results by GLTP (with body segment labels), segment volume validation and pose estimation/re-initialization results, respectively.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4541828&req=5

f19-sensors-15-15218: Validation and re-initialization results for a passing case (first row) and three failed cases (second to fourth row). Columns (a–e) are the point set in the current frame, that in the previous frame, correspondence estimation results by GLTP (with body segment labels), segment volume validation and pose estimation/re-initialization results, respectively.

Mentions: In practice, we found that both M1 and M2 have very distinct values in the passing and failing cases, indicating their sensitivity for volume validation. In this work, we chose M1 and M2 to be 0.3 and 10, respectively. The threshold of the torso's M2 is 1.4 to reflect the maximum allowable height change. In all 28 testing sequences, the total frame-wise pass rate is over 98%, and there are 1.89% of frames that require re-initialization (Case I or II). Twenty one out of 28 sequences have a 100% passing rate, and Case III is only detected for a few frames in Sequence 25. Some validation examples are given in Figure 19, which shows a passed case (the first row) and three failed cases: (1) In the second row (Case I), the right arm is visible in the previous frame (red points in column (b)), but invisible in the current frame (column (a)). The right arm has invalid M1 (column (d)). The re-initialization result (re-do GLTP with a template where the right arm is set as invisible) is shown in column (e). (2) In the third row (Case II), the left arm is trapped in the torso, and the right arm has an enlarged volume to cover the points from both arms (column (c)). The left arm has invalid M1, and the right arm has invalid M2 (column (d)). Column (e) shows the re-initialization result with the recovered left arm after GLTP registration using the “T-pose” template. (3) In the fourth row (Case III), both left and right arms and part of the torso are missing, caused by large self-occlusions. Correspondence estimation results are invalid (column (c)), leading to invalid M1 and M2 for most segments (column (d)). Column (e) shows the pose estimation result by using pose continuity and physical constraints.


Articulated Non-Rigid Point Set Registration for Human Pose Estimation from 3D Sensors.

Ge S, Fan G - Sensors (Basel) (2015)

Validation and re-initialization results for a passing case (first row) and three failed cases (second to fourth row). Columns (a–e) are the point set in the current frame, that in the previous frame, correspondence estimation results by GLTP (with body segment labels), segment volume validation and pose estimation/re-initialization results, respectively.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4541828&req=5

f19-sensors-15-15218: Validation and re-initialization results for a passing case (first row) and three failed cases (second to fourth row). Columns (a–e) are the point set in the current frame, that in the previous frame, correspondence estimation results by GLTP (with body segment labels), segment volume validation and pose estimation/re-initialization results, respectively.
Mentions: In practice, we found that both M1 and M2 have very distinct values in the passing and failing cases, indicating their sensitivity for volume validation. In this work, we chose M1 and M2 to be 0.3 and 10, respectively. The threshold of the torso's M2 is 1.4 to reflect the maximum allowable height change. In all 28 testing sequences, the total frame-wise pass rate is over 98%, and there are 1.89% of frames that require re-initialization (Case I or II). Twenty one out of 28 sequences have a 100% passing rate, and Case III is only detected for a few frames in Sequence 25. Some validation examples are given in Figure 19, which shows a passed case (the first row) and three failed cases: (1) In the second row (Case I), the right arm is visible in the previous frame (red points in column (b)), but invisible in the current frame (column (a)). The right arm has invalid M1 (column (d)). The re-initialization result (re-do GLTP with a template where the right arm is set as invisible) is shown in column (e). (2) In the third row (Case II), the left arm is trapped in the torso, and the right arm has an enlarged volume to cover the points from both arms (column (c)). The left arm has invalid M1, and the right arm has invalid M2 (column (d)). Column (e) shows the re-initialization result with the recovered left arm after GLTP registration using the “T-pose” template. (3) In the fourth row (Case III), both left and right arms and part of the torso are missing, caused by large self-occlusions. Correspondence estimation results are invalid (column (c)), leading to invalid M1 and M2 for most segments (column (d)). Column (e) shows the pose estimation result by using pose continuity and physical constraints.

Bottom Line: We introduce a visible point extraction method to initialize a new template for the current frame from the previous frame, which effectively reduces the ambiguity and uncertainty during registration.Third, to support robust and stable pose tracking, we develop a segment volume validation technique to detect tracking failures and to re-initialize pose registration if needed.The experimental results on both benchmark 3D laser scan and depth datasets demonstrate the effectiveness of the proposed framework when compared with state-of-the-art algorithms.

View Article: PubMed Central - PubMed

Affiliation: School of Electrical and Computer Engineering, Oklahoma State University, Stillwater, OK 74078, USA. song.ge@okstate.edu.

ABSTRACT
We propose a generative framework for 3D human pose estimation that is able to operate on both individual point sets and sequential depth data. We formulate human pose estimation as a point set registration problem, where we propose three new approaches to address several major technical challenges in this research. First, we integrate two registration techniques that have a complementary nature to cope with non-rigid and articulated deformations of the human body under a variety of poses. This unique combination allows us to handle point sets of complex body motion and large pose variation without any initial conditions, as required by most existing approaches. Second, we introduce an efficient pose tracking strategy to deal with sequential depth data, where the major challenge is the incomplete data due to self-occlusions and view changes. We introduce a visible point extraction method to initialize a new template for the current frame from the previous frame, which effectively reduces the ambiguity and uncertainty during registration. Third, to support robust and stable pose tracking, we develop a segment volume validation technique to detect tracking failures and to re-initialize pose registration if needed. The experimental results on both benchmark 3D laser scan and depth datasets demonstrate the effectiveness of the proposed framework when compared with state-of-the-art algorithms.

No MeSH data available.