We present a method for estimating pose information from a single depth image given an arbitrary kinematic structure without prior training. For an arbitrary skeleton and depth image, an evolutionary algorithm is used to find the optimal kinematic configuration to explain the observed image. Results show that our approach can correctly estimate poses of 39 and 78 degree-of-freedom models from a single depth image, even in cases of significant self-occlusion.
Being able to estimate three-dimensional pose of an articulated articated object, such as a robot or human, is important for a variety of applications (eg. [8]). While recent technological advances have made capturing depth images both convenient and affordable, extracting pose information from these images remains a challenge-even when the kinematic structure of the target is provided. Popular approaches often rely domain specific knowledge and extensive training, thus providing little generality to arbitrary skeletons where little or no training data exists.
This paper presents results on estimating poses of an arbitrary kinematic skeleton from a single depth image without prior training. The pose estimation is defined as a modelbased estimation problem and an evolutionary algorithm is applied to find the optimal pose. Rather than using a priori beliefs or pre-trained models, this algorithm extracts the most likely configuration based solely on the kinematic structure to explain the observed depth image (Fig. 1).
The vast majority of pose estimation research focused specifically on the human kinematic skeleton. Recent surveys [5,6] describe two primary directions: pose assembly via probabilistic detection of body parts and example-based methods. For example, Shotton et al. [7] described a particularly successful approach to human pose recognition that builds a probabilistic decision tree to first find an approximate pose of body parts, followed by a local optimization step. While this technique is fast and reliable, it relies on significant training: 24000 core hours of training on 1 million randomized poses. A primary limitation of these techniques is their reliance on domain specific information regarding human kinematics which does not generalize to arbitrary skeletons without explicit and additional training. In comparison, Gall et al. [3] used motion capture with markerless camera systems to find poses of complex models, such as those generated from animals and non-rigid garments. However, this approach required laser scanned visual-hulls which were mapped, by human experts, to an underlying kinematic structure.
In an alternative approach, Katz et al. [4] inferred relational representations of articulated objects by tracking visual features, but is limited to planar objects and requires interactions to infer the underlying structure.
The pose estimation is defined as an optimization problem:
where s(θ) is skeleton model with parameters θ, p n is a point from the observed depth image, p * is the closest point on the model to the point p n given the parameters θ, and σ is the standard deviation of point distances in the depth image. This optimization problem is challenging as it is non-convex with numerous local optima. Therefore, we use an evolutionary algorithm to find the pose parameters. The evolutionary algorithm is a population-based, heuristic algorithm that iteratively selects and combines solutions to produce increasingly better models [2]. The skeleton is represented as an acyclic graph of links, with parameterizable joint angles and length. Traditional evolutionary operators are used: random mutations are applied to the parameters and recombination swaps branches of links between parents to produce offspring.
We captured a data set of an articulated robot using a Kinect camera’s depth sensor [1]. Images of two drastically distinct subjects were captured: the first is a spider model is a based on a quadruped robot with 8 links resulting in 39 degrees-of-freedom, while the second is a humanoid model consisting of 17 links amounting to 78 degrees-of-freedom. For the spider model, we arranged a robot in four distinct poses, and collected five images ranging in inclination angles was taken per poses; resulting in a total of twenty depth images. There were multiple examples of self-occlusion in the data set. For the humanoid model, eight images were taken of four subjects, totaling to 32 images. The images in both data sets were pre-processed with background subtraction.
We ran the learning algorithm for 10 9 objective function evaluations, which is approximately 10000 iterations. On a single core 2.8GHz Intel processor, this required approximately 30 and 70 minutes of computational effort per image for the spider and humanoid models, respectively.
Results of the evolutionary algorithm indicate successful pose estimation for both the spider and humanoid models, even with significant self-occlusion (Fig. 2). Quantitatively, our model placed links in the correct position with an accuracy of 99% and 84% for the spider and humanoid model, respectively. Qualitatively, on a scoring survey with a scale of 5, spiders scored 4.9, while the humanoid model achieved a 4.1 score.
Compared results to a hill-climbing baseline, our learning algorithm produced mariginally superior results for the lowdimensional spider model. However, for the high-dimensional humanoid model, there is a sharp contrast in pe
This content is AI-processed based on open access ArXiv data.