Journal of Biomechanics 33 (2000) 1751}1754 Technical note A method to synchronise cameras using the direct linear transformation technique夽 Philippe Pourcelot *, Fabrice AudigieH , Christophe Degueurce , Didier Geiger, Jean Marie Denoix UMR INRA-ENVA Biome& canique du Cheval, Ecole Nationale Ve& te& rinaire d+Alfort, 7 Av. du Gal de Gaulle, 94704 Maisons-Alfort Cedex, France E.A. CNRS 7052, Laboratoire de Me& canique Physique, Universite& Paris XII 61, Av. du Gal de Gaulle, 94010 Cre& teil Cedex, France Accepted 6 April 2000 Abstract The aim of this paper is to present a method which enables the recordings of cameras that are not equipped with a synchronisation system to be synchronised a posteriori. Using the Direct Linear Transformation technique, this method estimates the phase di!erence between two cameras by minimising the reconstruction errors of a moving point. Once the phase di!erence value is known, one of the recordings is chosen as a reference and the second one is synchronised to the "rst by cubic spline interpolation. 2000 Elsevier Science Ltd. All rights reserved. Keywords: Synchronisation; Camera; DLT; Kinematics; Genlock 1. Introduction The Direct Linear Transformation (DLT) method (Abdel-Aziz and Karara, 1971; Marzan and Karara, 1975) and its re"ned versions (Hatze, 1988; Gazzani, 1993) allow the determination of the 3-D co-ordinates of a point from two or more 2-D views of this point. These methods are commonly used in kinematic analysis of human and animal movement. This is due to the accuracy of the results obtained and the great #exibility in camera set-up (Wood and Marshall, 1986; Hatze, 1988; Challis and Kerwin, 1992; Chen et al., 1994; Hinrichs and McLean, 1995). Another advantage of these methods is that even though only two cameras are required, additional ones can be accommodated (Challis and Kerwin, 1992). Nevertheless, the di!erent cameras need to be perfectly synchronised to preserve the accuracy of the results. Cameras with precise phase-locked synchronisation * Corresponding author. Tel.:#33#1-43-96-70-49; fax: #33#143-96-31-62. E-mail address: [email protected] (P. Pourcelot). 夽 Presented in part at the World Congress on Medical Physics and Biomedical Engineering in Nice, France, 14}19 September 1997. are 10 to 20 times more expensive than home video cameras. It is tempting to use home video cameras to increase the number of cameras at a low cost and to develop a technique allowing the synchronisation of these cameras. In previous studies, cameras' recordings were synchronised by detecting a particular event recorded simultaneously by all the cameras. Miller et al. (1980) synchronised their recordings by "ring a light emitting diode, Degueurce et al. (1996) switched on a light bulb. Such synchronisation methods do not take into account for the relative time o!set between the cameras, which may result in high inaccuracy of 3-D kinematic data. Yeadon (1989, 1990) proposed two di!erent methods to synchronise cameras. The "rst one was based on the digitised data of a ski jumper obtained at take o! and landing and required to know the vertical plane in which the ski jumper's centre of mass lay at take o!. The second one used a timing device in view of each camera. The present study describes a method that enables the synchronisation of multiple cameras by exploiting aspects of the DLT method. Firstly, the cameras are approximately time matched by switching on a light bulb. Secondly, a numerical method, presented in this paper, is used to estimate the time o!set between the cameras. This time o!set is then used to synchronise the recordings of the cameras. 0021-9290/00/$ - see front matter 2000 Elsevier Science Ltd. All rights reserved. PII: S 0 0 2 1 - 9 2 9 0 ( 0 0 ) 0 0 1 3 2 - 9 1752 P. Pourcelot et al. / Journal of Biomechanics 33 (2000) 1751}1754 2. Method The Direct Linear Transformation equations are: ¸ X#¸ >#¸ Z#¸ , x#dx" (1) ¸ X#¸ >#¸ Z#1 ¸ X#¸ >#¸ Z#¸ (2) y#dy" ¸ X#¸ >#¸ Z#1 in which (x, y) are the digitised co-ordinates of a point, (dx,dy) are the errors associated with the co-ordinates, henceforth referred to as DLT errors, (X, >, Z) are the object's space co-ordinates, and (¸ , ¸ ,2, ¸ ) the unknown DLT parameters of each camera. The DLT errors (dx,dy) presented above are the quanti"cation of the inconsistencies of location of a point in two (or more) camera views. Once the object-space coordinates of a point are calculated, these DLT errors can be computed, for each camera, using DLT equations (1) and (2). The use of non-synchronised cameras introduces an additional error into these DLT errors. The phase di!erence between two cameras can be estimated by searching for the phase di!erence value which minimises the mean DLT errors (dx,dy) of a moving point. This mean DLT error per frame and per camera can be calculated as follows: H G, (dx #dy GH GH H G mean DLT error" , 2NbF (3) where dx and dy are the DLT errors calculated at GH GH frame i on camera j along the camera image x- and y-axis, respectively, and NbF is the total number of frames analysed. The phase di!erence value giving the lowest mean DLT error can be determined by successive approximations and then used to synchronise a posteriori both recordings. To test di!erent phase di!erence values, and "nally to synchronise both recordings, interpolated 2-D co-ordinates of the points are calculated for one of the two cameras from their digitised co-ordinates using an interpolation method such as spline functions. 3. Experimental evaluation A routine implementing this synchronisation method was developed and added to a more general computer program that tracks the markers on the 2-D video images, calibrates the cameras and calculates the 3-D coordinates of the markers. 3.1. Apparatus and recording procedure In order to test the accuracy of the results obtained with this method, the rotations (1.16 rps) of a 64 cm Fig. 1. The motion of a wheel "tted with one marker was recorded by four Hi 8 mm video cameras placed in such a way that each side of the wheel was recorded by 2 cameras. To avoid temporary concealment of the marker, the wheel's axis was suspended to a rigid body with nylon yarns. diameter wheel "tted with one marker was recorded in outdoor conditions by four Hi 8 mm video cameras (Sony FX 700, 50 Hz) (Fig. 1). These cameras imaged a 6 m long "eld of view. Their shutter speed was set at 1/1000 s. While the wheel was recorded, a light bulb, placed in the "eld of the 4 cameras, was switched on every 10 s. For each camera, the frame on which the "rst light of the bulb appeared was selected to synchronise approximately the four recordings. 3.2. Film digitisation and marker detection The image processing was conducted o!-line. The recordings were analysed for sequences of about 3 s (1.5 s before and after the "ring of the light bulb). These sequences were digitised using a video card (VideoVision Studio, Radius) with a resolution of 768 columns by 576 lines in 256 grey scale. The computer program determined automatically the positions of the marker. 3.3. Camera synchronisation The numerical synchronisation of the cameras was made for each digitised sequence. This synchronisation was conducted as follows. One camera was chosen as a reference, and the others were synchronised with it. To synchronise one camera with the reference one (called slave and master cameras, respectively), di!erent values of phase di!erence between both cameras were considered and tested. The tests were performed by calculating, for the slave camera, interpolated 2-D locations of the moving marker using cubic spline functions. For each phase di!erence value considered, the mean DLT error was computed for one entire rotation of the moving marker using formula (3). The value of phase di!erence corresponding to the lowest mean DLT error was P. Pourcelot et al. / Journal of Biomechanics 33 (2000) 1751}1754 1753 Fig. 2. Mean DLT error versus phase di!erence between two cameras. The phase di!erence value corresponding to the minimal value gives an estimation of the phase di!erence value between the cameras (arrow). determined using an iterative method (Fig. 2). This phase di!erence value was then used to synchronise both recordings by calculating, for the slave camera, interpolated 2-D co-ordinates of the marker using cubic spline functions. 3.4. 3-D reconstruction of the marker To evaluate the e!ects of this synchronisation method on the kinematic data, four di!erent reconstructions of the marker were carried out: two left reconstructions using cameras 1 and 2, and two right reconstructions using cameras 3 and 4. In both cases, the "rst reconstruction was performed with numerical synchronisation of the cameras and the second without. For each of these two cases, the distance between the left and right reconstructions of the marker was calculated. The two distances were then compared. Moreover, the consistency of the phase di!erences between cameras 1 and 2, 1 and 3, 1 and 4, 2 and 3, 2 and 4, and 3 and 4 were studied. 4. Results 4.1. Distances between the left and right reconstructions of the marker The distances calculated with and without synchronisation between the left and right reconstructions of the marker are shown in Fig. 3. This "gure was obtained by superimposition of 3 rotations of the wheel. The start of the rotation (t"0 s) corresponded to a position of the marker at 3 o'clock. The phase di!erence values calculated between the left (1 and 2) and right (3 and 4) cameras were !36 and !5% of the time interval between two frames, respectively. In other words, as the recording frequency of the cameras used in this study was Fig. 3. Distances calculated with and without synchronisation between the left and right reconstructions of a marker (superimposition of 3 rotations). The highest distances were observed with non-synchronised cameras when the marker was at 6 and 12 o'clock. 50 frames per second, the time interval between two consecutive frames was 0.02 s and consequently, the position of the marker was recorded by the second camera 0.0072 s before the "rst one (as indicated by the negative value of the phase di!erence). The mean distances obtained with and without synchronisation were 0.68$0.20 and 1.81$0.75 cm, respectively. 4.2. Consistency of the phase diwerences between cameras The phase di!erences values between cameras 1 and 2, 1 and 3, 1 and 4, 2 and 3, 2 and 4, and 3 and 4 were !36, !23, !27, #15, #6 and !5% of the time interval between two frames, respectively. The consistency of these phase di!erence values was high but not perfect. As an example, the phase di!erence value between cameras 1 and 2 (!36%) was not perfectly equal to the sum of the phase di!erences between cameras 1 and 3, and cameras 3 and 2 (!23 !15"!38%). The highest inconsistencies were less than 5% of the time interval between two frames. 5. Discussion The distance between the left and right reconstructions of the marker obtained for non-synchronised cameras presented two maxima for each rotation of the wheel (Fig. 3). The "rst one occurred approximately when the marker was at 6 o'clock and the second one at 12. These high distances were mainly due to di!erences between the transversal co-ordinates (Y-axis) of the left and right 1754 P. Pourcelot et al. / Journal of Biomechanics 33 (2000) 1751}1754 make it impossible the cabling between cameras. In practice this synchronisation method can be performed using the digitised data of landmarks or markers on the body of interest. Acknowledgements This study was supported by the Institut National de la Recherche Agronomique and the Service des Haras et de L'Equitation. References Fig. 4. In this example, the marker has been recorded at time t by the camera 1 and at time t#dt by the camera 2. At 12 o'clock the 3-D co-ordinates of the marker were greatly translated along Y-axis but the DLT errors calculated were very small because of the almost perfect intersection of the marker's rays. At 9 o'clock, opposite results were found, no signi"cant translation along the Y-axis occurred but large DLT errors were observed due to the high distances which separated the marker's rays at the point where they were the closest. reconstructions. This phenomenon is due to the phase di!erences between the cameras and to the position of these cameras with respect to the wheel. At 6 and 12 o'clock, the marker's location (the point for which the marker's rays are the closest) obtained with the left reconstruction was not in the main plane of the wheel but greatly translated along the Y-axis (Fig. 4). This phenomenon occurred to a lesser degree for the right reconstruction of the marker because of the small phase di!erence value between both right cameras. In contrast, the DLT errors calculated at 6 and 12 o'clock were very small because of the quasi perfect intersection of the marker's rays. At 3 and 9 o'clock, opposite results were found. No signi"cant Y translation occurred and large DLT errors were observed. This was due to the fact that where the marker's rays were the closest, the distance in between was very high. With synchronised cameras, the high distances observed at 6 and 12 o'clock disappeared so that the mean distance between the left and right reconstructions of the marker was equal to the one calculated when the marker was static. This synchronisation technique allows the use of cameras which cannot be synchronised either because they are devoid of a genlock or the recording conditions Abdel-Aziz, Y.I., Karara, H.M., 1971. Direct linear transformation from comparator coordinates into object-space coordinates in closerange photogrammetry. In: Proceedings of the ASP/UI Symposium on Close-Range Photogrammetry. American Society of Photogrammetry, Falls Church, VA, pp. 1}18. Challis, J.H., Kerwin, D.G., 1992. Accuracy assessment and control point con"guration when using the DLT for photogrammetry. Journal of Biomechanics 25, 1053}1058. Chen, L., Armstrong, C.W., Raftopoulos, D.D., 1994. An investigation on the accuracy of three-dimensional space reconstruction using the direct linear transformation technique. Journal of Biomechanics 27, 493}500. Degueurce, C., Dietrich, G., Pourcelot, P., Denoix, J., Geiger, M., 1996. Three-dimensional kinematic technique for evaluation of horse locomotion in outdoor conditions. Medical and Biological Engineering and Computation 34, 249}252. Gazzani, D., 1993. Comparative assessment of two algorithms for calibrating stereophotogrammetric systems. Journal of Biomechanics 26, 1449}1454. Hatze, F., 1988. High-precision three-dimensional photogrammetric calibration and object space reconstruction using a modi"ed DLTapproach. Journal of Biomechanics 21, 533}538. Hinrichs, H., McLean, R.N., 1995. NLT and extrapolated DLT: 3-D cinematography alternatives for enlarging the volume of calibration. Journal of Biomechanics 28, 1219}1223. Marzan, G.T., Karara, H.M., 1975. A computer program for direct linear transformation solution of collinearity condition, and some applications of it. In: Proceedings of the Symposium on CloseRange Photogrammetric Systems, Champaign. American Society of Photogrammetry, Falls Church, VA, pp. 420}476. Miller, S.P., Shapiro, N.R., McLaughlin, R., 1980. A technique for obtaining spatial kinematic parameters of segments of biomechanical systems from cinematographic data. Journal of Biomechanics 13, 535}547. Wood, T.M., Marshall, G.A., 1986. The accuracy of DLT extrapolation in three-dimensional "lm analysis. Journal of Biomechanics 19, 781}785. Yeadon, R.N., 1989. A method for obtaining three-dimensional data on ski jumping using pan and tilt cameras. International Journal of Sport Biomechanics 5, 238}247. Yeadon, M.R., 1990. The simulation of aerial movement * I. The determination of orientation angles from "lm data. Journal of Biomechanics 23, 59}66.
© Copyright 2026 Paperzz