Facial Landmarks Tracking Based On Constrained Local Models And SCMS

In recent years,the facial landmarks localization and tracking has became a fundamental but challenging subject in the field of computer vision,and the CLM is one of the most useful method for it.But there are two challenges that limit the furthur development of CLM:the first is the limitation of image information that contained in the image models and the furthur computational complexity issues;the second is to choose a approximation algorithm for the CLM fitting.In this paper we have did some researchs for the issues,proposed a advanced CLM that using ICA and a advanced approximation algorithm for CLM:1. A Advanced CLM based on ICA is proposed.The traditional method for the statistic analysis of CLM is PCA,which is short of higher order statistical information,and has limited potential for the extraction of local features.ICA is based on higher order statistical information,can descripe the local features effectively,and has little limitation for the sample size.In this paper we introduce the ICA method into CLM to extract the shape features,although the computational complexity increased slightly, the fitting accuracy has increased effectively.2. A CLM fitting algorithm based on SCMS is proposed.The common algorithms for CLM fitting usually have some problems,such as apply condition limitations,computational complexity.The SCMS algorithm proposed in this paper use a nonparametric kernel density estimate to estimate the response of the feature point,and use Mean-shift to maximize the response then iterate it to approximate the position of feature point.This algorithm has less computational expense because of the kernel evaluations and reach the local optima at the same time,has a better performance in fitting accuracy and computational complexity.Then we did some facial landmarks localization and tracking experiments through videos that contain single human face to test the permormance of our algorithm,and compared it with algorithm that using ASM.The result shows that the advanced CLM based on SCMS has a much better performance in fitting accuracy compared withASM,and has less posture and angle limitations.The average fitting time for each frame is 30 ms,it’s just one third of the ASM fitting time.The experiments confirmd that the anvanced CLM based on SCMS has a effective increase in fitting accuracy and computational complexity.