Motion detection method, motion detection program, storage medium in which motion detection program is stored, and motion detection apparatus

1. A method of detecting motion of a particular part from a sequence of successive input images, comprising the steps of:

detecting motion vectors of a plurality of parts of the input images by tracking feature points in the successive input images; and

detecting the motion of the particular part by processing the motion vectors, the motion detection step including

producing a histogram of the motion vectors,

processing the histogram such that a mountain-like distribution segment each having a peak at which a frequency distribution of motion vectors has a maximum in the histogram are classified, and a mountain-like distribution segment corresponding to the particular part is detected,

defining class information for each feature point to indicate whether each feature point belongs or does not belong to the mountain-like distribution segment detected in the histogram processing step, and

calculating motion of the particular part by analyzing the motion vectors belonging to the mountain-like distribution segment detected in the histogram processing step, the histogram processing step including,

detecting the mountain-like distribution segment corresponding to the particular part on the basis of the class information defined for feature points.

2. The motion detection method according to claim 1 , wherein the histogram processing step includes segmenting the histogram into a plurality of mountain-like distribution segments by an image region segmentation method and detecting the mountain-like distribution segment corresponding to the particular part.

3. The motion detection method according to claim 1 , wherein

the histogram producing, step includes producing a histogram with high resolution and a histogram with low resolution, and

the histogram processing step includes detecting a low-resolution mountain-like distribution segment corresponding to the particular part from the low-resolution histogram, and detecting a mountain-like distribution segment corresponding to the low-resolution mountain-like distribution segment from the high-resolution histogram, thereby detecting the mountain-like distribution segment corresponding to the particular part.

4. The motion detection method according to claim 1 , wherein the histogram producing step includes

calculating a confidence level indicating reliability of the class information, and

and performing weighting in accordance with the confidence level in producing of the histogram.

5. The motion detection method according to claim 4 , wherein the confidence level is given by the number of times that success has been successively achieved in tracking the feature point of interest.

6. The motion detection method according to claim 4 , wherein the confidence level is given by the degree of change in distance between the feature point of interest and an adjacent feature point.

7. The motion detection method according to claim 1 , wherein the histogram processing step includes

for an initial frame, detecting a mountain-like distribution segment corresponding to the particular part on the basis of the height of mountain-like distribution segments on the histogram, and

for frames following the first frame, detecting a mountain-like distribution segment corresponding to the particular part by evaluating a plurality of mountain-like distribution segments on the histogram in accordance with class information.

8. The motion detection method according to claim 1 , wherein

the particular part is a background, and

the histogram processing step includes selecting a mountain-like distribution segment having a greatest height from mountain-like distribution segments in the histogram and setting the selected mountain-like distribution segment as the mountain-like distribution segment corresponding to the particular part for an initial frame.

9. The motion detection method according to claim 1 , wherein

the particular part is a background, and

the histogram processing step includes

for an initial frame, selecting a mountain-like distribution segment corresponding to feature points set in a peripheral part of the input image and setting the selected mountain-like distribution segment as the mountain-like distribution segment corresponding to the particular part.

10. The motion detection method according to claim 1 , wherein the histogram processing step includes

calculating a peak likelihood for each mountain-like distribution segment on the basis of the class information, the peak likelihood indicating likelihood of being a mountain-like distribution segment corresponding to the particular part, and

selecting a mountain-like distribution segment having greatest peak likelihood as the mountain-like distribution segment corresponding to the particular part.

11. The motion detection method according to claim 10 , wherein the peak likelihood is given by the ratio of the number of feature points classified, by the class information, as belonging to the mountain-like distribution segment corresponding to the particular part.

12. The motion detection method according to claim 10 , wherein the peak likelihood is given by the product of a value indicating the ratio of the number of feature points classified, by the class information, as belonging to the mountain-like distribution segment corresponding to the particular part and a value obtained by quantizing the height of the mountain-like distribution segment.

13. The motion detection method according to claim 10 , wherein the histogram processing step includes

calculating confidence level indicating reliability of the class information, and

weighting the peak likelihoods in accordance with the confidence level.

14. The motion detection method according to claim 13 , wherein the confidence level is given by the number of times that success has been successively achieved in tracking the feature point of interest.

15. The motion detection method according to claim 13 , wherein the confidence level is given by the degree of change in distance between the feature point of interest and an adjacent feature point.

16. The motion detection method according to claim 1 , wherein the motion calculation step includes

evaluating the distance from the center of the mountain-like distribution segment detected in the histogram processing step for each motion vector in the mountain-like distribution segment with respect to a predetermined threshold value, and, if a motion vector whose distance from the center of the mountain-like distribution segment is greater than the predetermined threshold value is detected, removing the detected motion vector so that it will not be processed, and

calculating the average of motion vectors belonging to the mountain-like distribution segment and remaining after the evaluation step, and detecting motion of the particular part from the calculated average of the motion vectors.

17. The motion detection method according to claim 16 , wherein the average calculation step includes selecting motion vectors corresponding to the particular part from the motion vectors belonging to the mountain-like distribution segment and remaining after the evaluation step on the basis of the class information, and calculating the average of the selected motion vectors.

18. The motion detection method according to claim 16 , wherein the threshold value with respect to which to evaluate the distance is determined in accordance with the standard deviation of motion vectors belonging to the mountain-like distribution segment detected in the histogram processing step.

19. The motion detection method according to claim 16 , wherein threshold value with respect to which to evaluate the distance is determined in accordance with a change in the distribution of motion vectors in the mountain-like distribution segment detected in the histogram processing step.

20. The motion detection method according to claim 19 , wherein the motion calculation step includes

calculating confidence level indicating reliability of the class information, and

producing the threshold value weighted by the confidence level.

21. The motion detection method according to claim 20 , wherein the confidence level is given by the number of times that success has been successively achieved in tracking the feature point of interest.

22. The motion detection method according to claim 20 , wherein the confidence level is given by the degree of change in distance between the feature point of interest and an adjacent feature point.

23. The motion detection method according to claim 1 , further comprising the steps of:

dividing each frame of the input image into a plurality of blocks;

initially setting a predetermined number of feature points in each block in an initial frame of the input image;

in successive frames following the initial frame of the input image, tracking the feature points to detect the motion vectors; and

correcting unevenness of the distribution of the feature points in the successive frames of the input image.

24. The motion detection method according to claim 1 , wherein

the particular part is a background, and

the method further comprising the step of compensating for camera motion due to hand shake on the basis of the motion vectors.

25. A processor which executes a program to implement a method of detecting motion of a particular part from a sequence of successive input images, the program comprising the steps of:

detecting motion vectors of a plurality of parts of the input images by tracking feature points in the successive input images; and

detecting the motion of the particular part by processing the motion vectors

the motion detection step including

producing a histogram of the motion vectors,

processing the histogram such that a mountain-like distribution segment each having a peak at which a frequency distribution of motion vectors has a maximum in the histogram are classified, and a mountain-like distribution segment corresponding to the particular part is detected,

defining class information for each feature point to indicate whether each feature point belongs or does not belong to the mountain-like distribution segment detected in the histogram processing step, and

calculating motion of the particular part by analyzing the motion vectors belonging to the mountain-like distribution segment detected in the histogram processing step, the histogram processing step including

detecting the mountain-like distribution segment corresponding to the particular part on the basis of the class information defined for feature points.

26. A non-transitory computer-readable storage medium in which a program is stored, the program being executable by processing means to implement a method of detecting motion of a particular part from a sequence of successive input images, the program comprising the steps of:

detecting motion vectors of a plurality of parts of the input images by tracking feature points in the successive input images; and

detecting the motion of the particular part by processing the motion vectors, the motion detection step including

producing a histogram of the motion vectors,

processing the histogram such that a mountain-like distribution segment each having a peak at which a frequency distribution of motion vectors has a maximum in the histogram are classified, and a mountain-like distribution segment corresponding to the particular part is detected,

defining class information for each feature point to indicate whether each feature point belongs or does not belong to the mountain-like distribution segment detected in the histogram processing step, and

calculating motion of the particular part by analyzing the motion vectors belonging to the mountain-like distribution segment detected in the histogram processing step, the histogram processing step including

detecting the mountain-like distribution segment corresponding to the particular part on the basis of the class information defined for feature points.

27. A motion detection apparatus configured to detect motion of a particular part from a sequence of successive input images, the apparatus comprising:

a motion vector detection unit configured to detect motion vectors of a plurality of parts of the input images by tracking feature points in the successive input images; and

a motion detection unit configured to detect the motion of the particular part by processing the motion vectors, the motion detection unit including

a histogram producing unit configured to produce a histogram of the motion vectors,

a histogram processing unit configured to process the histogram such that a mountain-like distribution segment each having a peak at which a frequency distribution of motion vectors has a maximum in the histogram are classified, and a mountain-like distribution segment corresponding to the particular part is detected,

a feature point classifying unit configured to classify feature points into a class of feature points belonging to the mountain-like distribution segment detected by the histogram processing unit and a class of the other feature points, and

a motion calculating unit configured to calculate motion of the particular part by analyzing the motion vectors belonging to the mountain-like distribution segment detected by the histogram processing unit, the histogram processing unit configured to detect the mountain-like distribution segment corresponding to the particular part on the basis of the class information defined for feature points.