In this paper, we addressed the particularly challenging problem of fall detection using very low resolution thermal images. We proposed a new method for fall detection only based on the matches and a determined threshold. By classifying a pair of matched points on the ground or not on the ground, we could easily determine how many percent of the shape of a person is on the ground. Thus, we could determine if there is a fall or not. The experiments show that the method is able to classify features of human silhouette as one the ground or not on the ground.