Physics-based Person Tracking Using the Anthropomorphic Walker

Contribution:
Evaluation:
Reproducibility:
Improvements:

Contribution : Physics based Model for tracking people,as an alternative to other methods, for instance kinematic models, is suggested. Instead of modeling full body dynamics, a biomechanics based lower body tracking model is implemented.

Evaluation: The model was evaluated for 4 different types of walking and environmental variations. Three of these were generated specifically to test the model performance in changes in speed, simulated occlusion and turning. The model then was applied to an external dataset (HumanEva).

Reproducibility: The components of the model have been explained considerably well, though lots of the have been referred to previous work and weren't revisited. THe parameters of most of these components appeared to have been set manually or empirically while evaluation. So, there might stay some room for ambiguity while reproducing the results completely.

Improvements: Overall the paper was well written and explained. Model parameter selection could have been emphasized more.

-- Main.sumanm - 01 Dec 2011

Contribution: A new method for tracking is proposed, where the underlying model is not kinematic but rather physics-based. The method function without learning and is real time, but focuses mainly on walking, and the tracking is based on monocular video sequences.

Evaluation: I think this part was actually very good. The showed a very good visualization of their results vs reality by an overlaid image, and the carried out a variety of experiments (various speeds, etc) to show robustness. Some numerical values would have been perfection, but those are in general difficult to obtain.

Improvements: performance would probably be boosted by using more than one point of view. As it is, it seems they can only track movements that lie perfectly on a plane perpendicular to the direction of the camera, which while a good first step, is far from optimal.

Reproducibility: I think this is the part where the paper is most laking. While explicit formulas and algorithms are given and spelled out, parameters are not specified, nor are the conditions of the recording and so on. This paper is fairly complex and hard to describe with perfect accuracy, however, so the authors had to compromise between length and precision.

-- Main.ginestra - 01 Dec 2011

Contribution: The paper presents a method for tracking the locomotion of a human from single-camera video sequences. This is accomplished by first contructing a 2-dimensional physics-based model of lower-body locomotion called an Anthropomorphic Walker. This is used to form a generative model of a simplified 3-dimensional kinematic human body. These two components are then used along with a likelihood function that describes the relationship between poses and images captured in the video. Through Bayesian inference, the most likely pose at each frame is computed.

Evaluation: The system is evaluated in four different experiments. The first three experiments use monocular video data of different nontrivial locomotion scenarios: variable speed walking, occlusion (walking behind obstacles), and turning out of the camera plane. The fourth experiment involves testing on a pre-existing benchmark dataset. The fourth experiment allows for explicit quantitative analysis of the accuracy of the results.

Reproducible: I believe this paper would be very difficult to reproduce. Many details of the implementation are left out or incomplete, such as the methods used to extract data from the video frames. It shouldn't be impossible, however, since most of the gaps can likely be filled in with other papers and tools from related fields like computer vision, inverse kinematics and Bayesian inference. By itself, though, I don't think this paper provides sufficient detail to reproduce its results.

Improvement: The paper is well-organized and the results are presented in a way that's easy to understand and appreciate. I found the formulas to be confusing however, usually because it wasn't entirely obvious what each symbol being used meant. Referring back to the reproducibility of the paper, I also feel the paper would benefit from additional detail on some of the algorithms and tools used to implement their system.

-- Main.cdoran - 01 Dec 2011

Contribution:
The study proposes a physics-based model for 3D tracking of a person from monocular video sequences. The presented model applies a probabilistic approach and it exploits basic physical principles in the design of a prior density model over human motion. It also uses a low-dimensional, biomechanically inspired models. As a result, it achieves crucial characteristics of bipedal locomotion such as balance and ground contact. Moreover, the model aims for reducing the dependence of models on mocap data which in turn enable models to generalize to different styles, such as varying walking speed, step length, and mass, as well as avoiding some of the problems of other approaches, e.g. footstake.

Evaluation:
The study includes four experiments for evaluating the accuracy and the generalization capabilities of the proposed model. Experiment 1 and 3 demonstrates the genralization of the model across different walking speeds, and motions styles, such as turning. Experiment 2 shows the robustness of the method as it can still track a person even when there are occlusions. Lastly, fourth experiment supports the results by using the HumanEva dataset, as it quantitatively compares and assesses the tracking results with the ground truth. It turns out that estimating the depth information is, unsurprisingly, the hardest part of such a problem.

Reproducibility:
Even though the theoretical explanations are detailed out, most of the implementation details, as well as the hand-tuned parameters are left out from the paper. One has to look at the given reference papers for the implementation. On the other hand, they have released a version of their implementation, which would definitely be helpful to reproduce the results.

Improvements:
A possible improvement for that study would be to investigate possible controllers for generalizing over different morphologies, or even moods of characters. The paper is well-written, but it's not very easy to follow as it incorporates ideas/methods from different areas.

---

Contribution: The paper presents a physics-based model to track walking people in monocular video while most current methods rely on kinematic models.

Evaluation: The result is evaluated by 4 experiments. The first one tests the changes in speed from 7 to 3 km/h. The second one tests occlusion by blacking out an image region. Third one tests the track of 3D turning. The last one tests against humanEva dataset.

Reproducibility: This paper presents a complex topic. While the key ideas, formulas, and general ideas of experiments are presented with details, the actual implementation still misses a lot of explanation. It may not be easy to reproduce exactly what the paper does.

Improvement: Overall the paper is well organized and presents the key ideas very well. However I still find it is hard to follow because it requires so much background knowledge from different areas.

Baoxuan- 01 Dec 2011

-- Main.ooguz - 01 Dec 2011

Contribution: The paper presents a human walk video-tracking method based on physics. The key component of the paper is usage of Anthropomorphic Walker dynamical model and sequentially tracking the walk using monte-carlo methods by evaluating MAP estimation of the position on the next frame based on the previous one. The advantage is the same as the one for any physics-based simulation methods: possible generalizations.

Evaluation: The proposed method was evaluated on 4 different experiments. I think the most relevant is the HumanEva experiment, when they actually compare their tracking to the corresponding mocap data.

Repoducibility: The method looks to be quite straightforward to implement.

Possible improvements: The model is very simple, and probably this is the key to success, it's easier to track. In reality they would need a more complex model - this is a good direction of improvement.

-- MikhailBessmeltsev - 01 Dec 2011

---

** Contribution? It uses physics-based model that generalizes naturally to variations in style of different speed, step-length, and mass, and avoids common problems such as footskate.

** How are the results evaluated? They evaluated the results by doing four experiments: change in speed, occlusion, turning and using the HumanEva mocap data. To measure the quality of tracking, they plot the spped of subject along with the posterior probability of which leg is the stance leg. They also watch the 3D animation to see how realistic the walking motion is.

** Reproducible? Looks to be reproducible. The methods and equations are well documented and values of parameters are also given.

** Possible improvements? This paper crosses many different fields. There are many concepts that are not explained clearly to me. It feels very difficult to read this paper. Otherwise, it's a well written paper except figure 8 comes after figure 11.

-- Main.shuoshen - 01 Dec 2011

-- MichielVanDePanne - 27 Nov 2011

---

Contribution: The authors present a model for people tracking, the Anthropomorphic Walker, that uses a low-dimensional dynamic controller to inform a physically realistic prediction of a person's pose and motion from source video. The predicted motions avoid several errors associated with purely kinematic solutions, such as footskate, implausible whole-body rotations, and discontinuous accelerations of the person.

Evaluation: The paper includes the trajectory tracking results for four different experimental clips. The experiments were made to demonstrate performance in spite of complications such as changing walking speed, occlusion of the walker, and a changing walking direction. They additionally determined the error of their tracker on part of the HumanEva data set, which includes ground truth motion capture data alongside the test video.

Reproducibility: The authors provide source code and results that demonstrate their chosen optimal control values given a step length and walking speed. Additionally, they provide explicit parameter values used in their system. As such, the paper's methodology should be reproducible.

Improvements:

• What justification exists for using a very simple 2D dynamic system with low-dimensional control to inform the 3D kinematic model? Would adding more control parameters to the dynamic model increase tracking accuracy?
• Section 5 (Sequential Monte Carlo Tracking) delves fairly deeply into symbolic math and can be confusing, especially for somebody not familiar with optical flow and this technique. It would have been appreciated to have an occasionally surface-for-air that gave motivation for each equation in terms of the final desired result.

-- BenHumberston - 01 Dec 2011

Contribution: The authors use physics-based dynamics to develop natural parameterizations of human motion. Physical interactions with the environment are also possible. The scheme generalizes to different speeds, masses, etc. of walking models. By modeling the underlying dynamics, more realistic motions and accurate tracking are achieved in comparison to modeling based solely on kinematics.

Evaluation: The authors evaluate their system quite vigorously. Comparison of the performance of the system to mocap data is much appreciated. A statistical analysis of how well the model generalizes to variability such as changes in speed, introducing occlusions and turning. The use of test data (from an independent project, HumanEva) not contained in the training set is essential and, although results are not exceptional, gives a much better gauge the generalization capabilities of the system.

Reproducibility: Though details in terms of formulas and theory of the underlying system are well presented, there seems to be a lack in practical advise on how to reimplement their solution in software. The authors make up for this deficiency by publishing the matlab source code of the project.

Improvements: Being a journal article, the work is high quality. If I had to complain about anything, it would be that complexity of the system is quite high. How they managed to squeeze a discussion of particle filters on top of all that is beyond me.

-- DanielTroniak - 02 Dec 2011

---

What is the contribution of the paper?

This paper focused on the physically-based dynamics for human motion tracking. This approach can produce naturally tracking reaction for the changing of environment and motion properties, such as mass distribution, speed. In this paper, the author design a generative model for human tracking with a stochastic controller to generate muscle forces, and a higher-dimensional kinematic model conditioned on the low-dimensional dynamics.

How are the results evaluated?

This paper evaluates the result by doing four experiments. The first three experiments use the same set of parameters for the kinematic evolution and the same prior over the control parameters for the dynamics. The parameters for the fourth experiment were set to similar values, but adjusted to account for a difference in frame rate. The first experiment tested the when changing the speed. The second one focused on dealing with occlusion. The third test was for turning in the 3D world. And the last experiment report results on the HumanEva benchmark dataset.

Is the paper reproducible?

Except for no source code, this paper is reproducible. The dataset used here is available and the tracking parameter is shown in the paper.

How could the paper research or paper writing be improved?

If the examples in the paper are on different human models, it will be better. I mean like on male/female, the young/elder, short/high, thin/fat.

-- Main.chuanzhu - 02 Dec 2011

---

Contribution:

The physics-based models offer significant benefits in terms of accuracy, stability, and generality for person tracking. They even try to solve some problems like occlusion.

How are the results evaluated:

In my opinion, the result is good evaluated. The author tried to test the algorithm in several ways, such as changes in speed, simulated occlusion and turning. The last experiment use HumanEva dataset and it is the quantitative comparison to the truth ground.

Reproducability:

The method looks really straightforward, but some details are still not been seen from the paper.

Improvement:

The paper writing could be better understandable if more pictures appear in the algorithm part to illustrate them in a visual way.

-- Main.Jingxian Li - 02 Dec 2011