Fig. 1. Pin-hole camera model of a video camera of focal length (f). The location of the projected point p′ on the virtual image plane is given by the perspective projection (x_i′=fX_P/Z_P,y_i′=fY_P/Z_P).

Fig. 2. Capturing a scene for structure-from-motion analysis using either camera or object motion. In this example, either method will produce the same perceived temporal image sequence (right).

Fig. 3. A SfM system translating in the direction V with rotation Ω. The FOE can be seen in the middle of the virtual image plane.

Fig. 4. A typical stereo camera configuration. The distance between the optical centres of the cameras is the baseline length, b. Two corresponding epipolar lines are also shown.

Fig. 5. The processing stages used in the FSC SV algorithm.

Fig. 6. Combined motion-stereo matching using initial stereo matching (1) followed by stereo matching (4) via motion correspondences (2,3).

Fig. 7. Example of how the MS-TCS algorithm can use variable correlation block sizes (right) to match the scene shown in the left hand image.

Fig. 8. The left start frame from the synthetic moving cubes sequence. The arrows show how two of the cubes move during the sequence. Each image is 512×512 pixels.

Fig. 9. Sample images from the synthetic moving room sequence. The arrows show the direction taken by the vision system through the scene. Each image is 512×512 pixels.

Fig. 10. A typical frame from the translating train sequence. The arrow shows the direction in which the train moves during the sequence. Each image is 748×560 pixels.

Fig. 11. The re-projected 3D results from the start, middle, and end frames of the cubes sequence as processed by the CCS algorithm.

Fig. 12. The re-projected 3D results from the start, middle, and end frames of the cubes sequence as processed by the TCS algorithm.

Fig. 13. The re-projected 3D results from the start, middle, and end frames of the cubes sequence as processed by the MS-TCS algorithm.

Fig. 14. The number of matches and outliers per frame produced by the CCS, TCS, and MS-TCS algorithms for the cubes sequence.

Fig. 15. The number of multiplies and additions per frame required by the CCS, TCS, and MS-TCS algorithms for the cubes sequence.

Fig. 16. The re-projected 3D results from the start, middle, and end frames of the room sequence as processed by the CCS algorithm.

Fig. 17. The re-projected 3D results from the start, middle, and end frames of the room sequence as processed by the MS-TCS algorithm.

Fig. 18. The number of matches and outliers per frame produced by the CCS and MS-TCS algorithms for the room sequence. Due to the large number of frames in this sequence, the results are grouped into bins of 50 frames showing the mean value for those frames together with 10 and 90% percentile markers.

Fig. 19. The number of multiplies and additions per frame required by the CCS and MS-TCS algorithms for the room sequence. Due to the large number of frames in this sequence, the results are grouped into bins of 50 frames showing the mean value for those frames together with 10 and 90% percentile markers.

Fig. 20. The re-projected 3D results from the start, middle, and end frames of the train sequence as processed by the CCS algorithm.

Fig. 21. The re-projected 3D results from the start, middle, and end frames of the train sequence as processed by the MS-TCS algorithm.

Fig. 22. The number of matches per frame produced by the CCS and MS-TCS algorithms for the train sequence.

Fig. 23. The number of multiplies and additions per frame required by the CCS and MS-TCS algorithms for the train sequence.