Thermal Face Recognition over time is a difficult challenge due to faces varies with different factor such as metabolism or ambient conditions. Thus, the aim of this work is to improve recognition rates of thermal faces acquired in time lapse mode, since the results available in other articles are not entirely satisfactory in this modality, which is mainly due to the large variation in the thermal characteristics of the faces in time lapses. To improve the recognition rates the approach called "Sparse Representation" was chosen. This method represents an input image as a linear combination of a dictionary composed of images of different subjects and a vector of sparse coefficients. The results are obtained using the two sets of UCHThermalFace database. The method shows high performance in the time lapse for thermal images.