Comparing evolutionary algorithms and particle filters for Markerless Human Motion Capture

Abstract

Markerless Human Motion Capture is the problem of determining the joints' angles of a three-dimensional articulated body model that best matches current and past observations acquired by video cameras. The problem of Markerless Human Motion Capture is high-dimensional and requires the use of models with a considerable number of degrees of freedom to appropriately adapt to the human anatomy.Particle filters have become the most popular approach for Markerless Human Motion Capture, despite their difficulty to cope with high-dimensional problems. Although several solutions have been proposed to improve their performance, they still suffer from the curse of dimensionality. As a consequence, it is normally required to impose mobility limitations in the body models employed, or to exploit the hierarchical nature of the human skeleton by partitioning the problem into smaller ones.Evolutionary algorithms, though, are powerful methods for solving continuous optimization problems, specially the highdimensional ones. Yet, few works have tackled Markerless Human Motion Capture using them. This paper evaluates the performance of three of the most competitive algorithms in continuous optimization -Covariance Matrix Adaptation Evolutionary Strategy, Differential Evolution and Particle Swarm Optimization -with two of the most relevant particle filters proposed in the literature namely, the Annealed Particle Filter, and the Partitioned Sampling Annealed Particle Filter.The algorithms have been experimentally compared in the public dataset HumanEva-I by employing two body models with different complexities. Our work also analyzes the performance of the algorithms in hierarchical and holistic approaches, i.e., with and without partitioning the search space. Non-parametric tests run on the results have shown that: (i) the evolutionary algorithms employed outperform their particle filter counterparts in all the cases tested; (ii) they can deal with high-dimensional models thus leading to better accuracy; and (iii) that the hierarchical strategy surpasses the holistic one.

Parallelization strategies for markerless human motion capture

“…Thus, we would recommend the one having 1500 evaluations and 27393 vertices because it provides both accurate results and relatively fast runtime. This configuration agrees with the proposed in [44] as the best performance solution.…”

“…It is employed to calculate the tracking error of a complete sequence as the average of all its frames. Finally, we have employed the CMAES algorithm, which has recently been reported to obtain the best results [44] for the MMOCAP problem.…”

“…Zhao and Liu [47] proposed a Hierarchical Annealed Genetic Algorithm to infer the three-dimensional pose from a single monocular camera. Yeguas-Bolivar et al [44] perform an experimental comparison of three relevant evolutionary algorithms namely Covariance Matrix Adaptation Evolutionary Strategy (CMAES) [19], Differential Evolution (DE) [35], and PSO [23], with two particle filters, namely APF [13] and PSAPF [1]. The results obtained show that the evolutionary algorithms evaluated performed significantly better than particle filters.…”

“…In this work, we propose f (x) as an optimized version of the silhouette matching function employed in most related works [1,4,13,16,22,44]. In short, given a model configuration x, its projection (silhouette) in all the cameras is matched against the foreground information obtained by background subtraction.…”

“…The CMAES algorithm has been run using the parameters indicated in [44]. Table 1 shows the average runtime (in milliseconds) of each strategy in evaluating a video frame, i.e., the time employed in evaluating all the solutions plus the time employed by the CMAES algorithm.…”