3-D face acquisition technology has been widely applied in various fields. Its development demands high-precision, rapid and reliable acquisition. Conventional sinusoidal fringe projection profilometry presents an inherent contradiction among measurement reliability, accuracy and speed, making it difficult to achieve collaborative optimization of the three aspects. However, the existing related improved methods still suffer from limitations in practical application scenarios such as complex poses, occlusions and dynamic acquisition. Aiming to resolve the above contradictions, this study combines relevant technologies and experience to enhance the comprehensive performance of 3-D face acquisition so as to meet the requirements of practical applications.

High-precision 3-D face data were obtained using the phase of sinusoidal fringes. A reference disparity map based on facial feature points was constructed to assist phase matching. An adaptive search window construction method based on phase difference was also proposed. The main steps were as follows. First, a face detection algorithm was used to segment the region of each face from the background image of the deformed fringes, and facial feature points were detected. Second, the disparity was calculated based on the correspondence between feature points in the left and right views. Then the feature points were triangulated, and the coarse disparity of the face was obtained through interpolation in the triangular mesh, which served as the reference disparity. Finally, under the constraint of reference disparity, high-precision disparity information was obtained by stereo matching using wrapped phase as a feature, and stereo reconstruction was performed to achieve high-precision reconstruction of 3-D facial shape.

First, a head model was used to verify the proposed method, and the 3-D face obtained by stereo matching using absolute phase from the three-frequency heterodyne method was taken as a reference to compare the reconstruction result obtained by reliability-guided phase unwrapping and that obtained by the proposed method. The models reconstructed by the three methods were shown that the results of the proposed method exhibited the smallest deviation from the reference morphology. Then, the reconstruction time of the three methods was compared, as shown in Table 1, and the three-frequency heterodyne method took the longest time in the phase unwrapping stage because three wrapped phase maps needed to be calculated. The time for obtaining coarse disparity with the proposed method was close to that of reliability-guided phase unwrapping, but in the subsequent phase matching stage, the time of the proposed method was only 61% of that of the reliability-guided algorithm because the search range was smaller. In addition, the reliability-guided algorithm required manual specification of the starting point of phase unwrapping, while the proposed method had the fastest overall reconstruction speed. Finally, an experiment of capturing multiple faces simultaneously in a complex background was carried out and the results were shown in Fig.6. Fig.6a showed one deformed fringe captured by the right camera. Fig.6b showed the background image of the fringe, and the detected faces were in the red boxes. Fig.6c showed the positions of the detected facial feature points. Fig.6d showed two reconstructed 3-D faces and their relative spatial positions. The experimental results showed that the proposed method accurately reconstructed the 3-D shape of multiple faces.

This paper proposes a reference disparity assisted stereo matching method based on facial feature points, which avoids the phase unwrapping operation. Compared with the temporal phase unwrapping method, this method reduces the number of projected patterns and improves the acquisition speed and the 3-D reconstruction speed. Compared with the spatial phase unwrapping method, this method improves the reliability of acquisition results and the 3-D reconstruction speed. The experimental results show that even if there are multiple faces in the image, the proposed method ensures the reconstruction accuracy and reliability of 3-D facial shape, and the capture speed is over 70 fps. The proposed method has broad application prospects in 3-D face acquisition in crowded scenes.