Automatic segmentation of echocardiography videos is critical for assessing various cardiac functions and improving the diagnosis of cardiac diseases. Convolutional Neural Networks (CNNs) have recently demonstrated their ability to segment 2D cardiac ultrasound images, but they suffer from temporal inconsistency. The main approaches for more reliable spatio-temporal analysis include 3D deep learning methods, recurrent segmentation algorithms, and post-processing of 2D segmentations. This paper presents a new efficient method for temporally regularized segmentation of cardiac ultrasound images. To address this problem, we propose a post-processing procedure based on the description of left ventricular, endocardial and epicardial, boundaries by their central distance signatures, which emerges as a powerful representation with excellent temporal coherence. Principal component analysis of the boundary signature is shown to provide a concise model for its representation. In particular, the coefficient of the first component has an interesting functional interpretation. Temporal smoothing of the obtained parametric representation enforces the temporal consistency of the segmentation. Results are given on the TED dataset [11] to illustrate the regularized segmentation and to measure some anatomical cardiac features.