Facial Feature Tracking Using Efficient Particle Filter and Active Appearance Model

Abstract

For natural human-robot interaction, the location and shape of facial features in a real environment must be identified. One robust method to track facial features is by using a particle filter and the active appearance model. However, the processing speed of this method is too slow for utilization in practice. In order to improve the efficiency of the method, we propose two ideas: (1) changing the number of particles situationally, and (2) switching the prediction model depending upon the degree of the importance of each particle using a combination strategy and a clustering strategy. Experimental results show that the proposed method is about four times faster than the conventional method using a particle filter and the active appearance model, without any loss of performance.

Estimating Parameters of Nonlinear Systems Using the Elitist Particle Filter Based on Evolutionary Strategies

“…As evaluation measure, we choose the time-dependent echo return loss enhancement (ERLE n ) ERLE n = 10 log 10 E{d 2 n }/E{e 2 n } (39) and stop the filter adaptation after 9 seconds. This allows to evaluate the online performance during the first half of the signal and should illustrate how well the instantaneous solution generalizes during the second half, which is an indicator for the actual system identification performance.…”

“…and setâ n = [0, 0, 0] T during the first 0.1 seconds to facilitate an initial convergence of the NLMS algorithm. Afterwards, we employ the ERLE n defined in (39) to investigate the impact of the preprocessor estimation (by applying the EPFES) for the identification of the Hammerstein system. The following evaluation is split into three parts, where each subsection considers a different EPFES configuration dependent on the length M z of the latent state vector z n and the number of L particles: For each of these EPFES configurations C1, C2 and C3, we compare three EPFES parametrizations P1, P2 and P3 dependent on the threshold ω th,n and the smoothing factor λ:…”

“…Pr1 -High computational load: It is well known in the particle-filtering community that both estimation accuracy and computational cost increase with the number of L particles [27], [28]. One approach to avoid a dissatisfying tradeoff is the selection of elitist particles by dropping all particles with weights smaller than a specific threshold [39].…”

“…th,n = 1 (GPF) P2: ω th,n in(14), λ = 0.5 P3: ω th,n in(14), λ = 0.Simulation results for the ERLEn defined in(39) for three different choices of the parameters ω th,n and λ, where ω th,n = 1 represents the Gaussian particle filter (GPF), Mz = 44 and L = 100 (Configuration C2).…”

Facial/License Plate Detection Using a Two-level Cascade Classifier and a Single Convolutional Feature Map

“…For example, the performance of face recognition was dramatically affected by the results of face detection [27]. The first step of facial expression recognition or facial feature tracking was also the application of successful face detection [28].…”