JP2875292B2 - Adaptive head position detection device - Google Patents

Description

Description: TECHNICAL FIELD [0001] The present invention relates to an adaptive head position detecting device that appropriately switches an arithmetic expression according to the motion of a head of a person having a complicated shape.

(Prior Art) Information is exchanged between a machine and a person manually through a keyboard, a mouse, or the like, or the amount of movement of the head is detected to instruct the machine.

That is, conventionally, means for detecting and processing the motion of a person includes means for performing image processing.

FIG. 6 is an explanatory view of a conventional means for calculating the amount of movement of the head of a person. An example of an image 2 and an upward image 3 is shown. Each image mainly contains hair areas 1H, 2H, 3H
Area information and the area information of the face area 1F, 2F, 3F. The image processing unit 5 calculates the area of the hair H and the face F from each image, calculates the center of gravity of each area, obtains the image-processed head image 6, and calculates the moving amount calculation processing unit 7. To perform the movement amount calculation processing. In this case, the present invention is applied to an arithmetic expression for calculating a rotation amount (a rotation amount about an arbitrary axis in a space) and a translation amount (a straight movement amount of an object in a three-dimensional space) previously derived from a human head model. .

However, since an arithmetic expression using a single person model has been conventionally applied to all head images of a person having various hair and face shapes, there is a considerable problem in detection accuracy. Was. That is, in the same person, the pattern of the hair part and the face part has a complicated appearance in a front image, a sideways image and the like, and is asymmetric even in the front image.

As described above, in the conventional head movement amount calculation processing means, the rotation amount and the translation amount are obtained using a single person model.
There is a problem that the detection error increases as the rotation amount of the head increases.

(Object of the Invention) The present invention solves the problems of the above-mentioned conventional means, and calculates a rotation amount and a translation amount of a head of a person moving with a complicated three-dimensional shape in a non-contact manner. It is another object of the present invention to obtain a device with a small detection error by adaptively switching arithmetic expressions.

(Structure of the Invention) (Differences between Features of the Invention and the Prior Art) In order to achieve the above object, the present invention provides a method of processing a head image of a person captured from a camera by applying image processing to the head image. After extracting the outline of the head as an approximation process, the first
, And then, as a second approximation process, after extracting the outline of the hair part in the head, the head image model including the head part outline is applied. Various parameters determined by the first approximation model and the second approximation model are used for selecting an arithmetic expression. That is, since model approximation is adaptively performed on a head image of a person captured in time series, a considerable improvement in detection accuracy is achieved compared with the related art.

The difference from the prior art is that, in the means for detecting the moving amount of the head image, a plurality of models instead of a single model is adaptively selected for the same human image, and the arithmetic expression is appropriately switched.

(Embodiment) The basic configuration for implementing the present invention is as follows. In order to detect the rotation amount and the translation amount of the head of a person, only the head of the person is taken from the camera, and the direction of the person to the camera is obtained by image processing. Accordingly, the model is adaptively switched, and the arithmetic expression is switched according to the model, thereby dynamically detecting the moving amount of the head of the person.

The appearance of a person's head is generally not only asymmetric. It does not have a uniform shape even at that angle. Therefore, when the model is applied to the head of the person, improvement in detection accuracy is not expected unless the model and its arithmetic expression are switched in each direction of the head.

FIG. 1 is a flowchart of a process for detecting a rotation direction of a person's head according to an embodiment of the present invention. The background image is input to the head image input from the camera of the input unit 8, and the difference processing is performed on the head image sequentially captured. In the first approximation processing unit 9, the contour line of the head is extracted from the difference image by the first head approximation processing unit 91, and the model approximation fitting to the head contour figure is performed by the first head approximation model adaptation unit 92.

As a second stage, in the second approximation processing unit 10, the shape of the boundary line between the hair region and the face region inside the head region is approximated by the second approximation processing unit 101 in the head, and a head image is formed. On the other hand, a second head approximation model adaptation unit 102 performs fitting to a head model including a boundary line.

In this manner, the model for each direction of the head and the arithmetic expression for calculating the amount of movement are adaptively obtained by the adaptive calculation expression processing unit 11, and the rotation amount and the translation amount of the head are detected by the head position detection. It is determined by the processing unit 12 and is determined by repeating this sequentially as shown by a route 13.

FIG. 2 is a detailed explanatory diagram of the first approximation processing unit 9 in FIG. In order to cut out the head image from the background, the head image (1) input from the camera and the background image (2) are subtracted to obtain a silhouette image (3).

In FIG. 2, a midpoint 18 of a line 17 connecting the vertex 14 of the head and the midpoint 16 of the neck 15 is obtained as shown in (4). As shown in (3), a plurality of representative points 20 are found on the contour line of the head from the outside of the head image using the same threshold value as shown in (3) toward the midpoint 18 thus obtained. Multiple representative points
A spline approximation is performed between 20 and a closed curve 21 is obtained as shown in (5). In this way, an approximate figure for the head contour is determined.

Since most of the obtained graphic (6) is elliptical from the appearance of the head, the maximum lengths m and n in the vertical direction 22 and the horizontal direction 23 of the graphic are obtained and shown in FIG. In the processing circuit, a model is selected in advance by template matching with each of the models 25, 26, and 27 in the memory 24, and equations 25A, 26A, and 27A for calculating the movement amount for each model are simultaneously selected. Here, since each function is a function of the vertical and horizontal lengths m and n of the figure, the movement amount calculation result 25
B, 26B, 27B can be reserved. In this way, the moving amount calculation time can be significantly reduced.

FIG. 3 is a flowchart relating to the second approximation processing unit 10 of FIG. After the approximation process is performed on the head contour by the first approximation processing unit 9, division by the in-head small area dividing unit 103 is performed in the same manner as the first approximation process. Subsequently, the first and second differential filter processing sections 104 and 105 respectively perform the first and second differentiation processes on the divided small regions.

In the primary differential filter processing unit 104, a rough distribution area can be roughly regarded as a head area and a smooth distribution area can be regarded as a face area from the calculated gradient value distribution. 2
In the next differential filter processing unit 105, the boundary between the hair region and the face region is roughly determined from the calculated distribution of the zero crossings.

The boundaries between the hair region and the face region are determined by the two filter processing units 104 and 105, and the sampling point 105 samples a plurality of representative points 20 from the boundary line. Then, the sampled points are connected by the linear connecting unit 106. Smoothing processing is performed by the smoothing processing unit 107 on the polygonal lines generated by the connection.

FIG. 4 is a second approximation processing unit shown in FIGS. 1 and 3.
FIG. 10 is a detailed explanatory view of FIG. 10, showing a typical rotation state of the head. In each of the front-facing (1), right-facing (2), downward (3), and left-facing (4) surface images, the image processing unit 108 calculates the area 40 of the hair region and the area 41 of the face region. In order to calculate the rotation direction of the head, a plurality of models 28 to 31 are previously stored in a memory as shown in FIG.
Is stored. Then, arithmetic expressions 28A, 29A, 30A, 31A for the respective models 28 to 31 are prepared. In this case, each function defines the area 40 of the hair region and the area of the face region 41 as variables S ₁ and S _{2, respectively.}
luggage. The results according to the combination of these two variables S ₁ and S ₂ can be stored in advance as Tables 28B, 29B, 30B and 31B.

In this way, the arithmetic expression selected from the first and second approximations
The rotation amount and the translation amount of the head are obtained using 25A to 31A.

In the present embodiment, since the switching of the model and the selection of the arithmetic expression are performed successively for the same person, even from a head having a complicated and asymmetrical shape, as shown in FIG. Regarding the detection accuracy in each direction of the section, comparison of the actual measurement value A, the detection value B by the conventional device, and the detection value C by the present device shows that the detection accuracy is improved. The figure shows each direction, upward (1), downward (2), left (3),
Regarding the rightward direction (4), the horizontal axis of the graph indicates the number of frames in which the head image is serially captured from the camera, and the vertical axis indicates the detected rotation amount of the head. In general, in the case of B using the conventional apparatus, since the detection is based on a single model as the actual amount of rotation of the head increases, the accuracy drops rapidly from around 15 degrees, The detected angle no longer shows a change.

As the actual amount of rotation increases, a detection accuracy of 84% or more can be achieved by the device of the present invention as compared with a detection device based on a single model of the conventional device.

(Effect of the Invention) As described above, according to the present invention, when calculating a rotation amount and a translation amount of a person having a complicated shape in front of a camera, an arithmetic expression based on a plurality of head models is used. Since the switching is performed in accordance with the shape pattern, data of the movement amount with higher accuracy can be detected.

[Brief description of the drawings]

FIG. 1 is a flowchart of a process for detecting the rotation direction of the head of a person according to one embodiment of the present invention, FIG. 2 is a detailed explanatory diagram of the first approximation processing unit 9 in FIG. 1, and FIG. Second approximation processing unit in FIG.
10, FIG. 4 is a detailed explanatory diagram of the second approximation processing unit shown in FIG. 1 and FIG. 3, FIG. 5 is a diagram showing a comparison between the measured values of the present invention and the conventional device, and FIG. FIG. 6 is an explanatory view of means for calculating the amount of movement of the head of a person in the conventional apparatus. 8 ... head image input unit, 9 ... first approximation processing unit, 10 ...
2nd approximation processing unit, 91 ... head 1st approximation processing unit, 92 ... head 1st approximation model adaptation unit, 101 ... 2nd approximation processing unit in head, 102 ... 2nd approximation in head Model adaptation unit, 11 ... Adaptive calculation formula processing unit, 12 ... Head position detection processing unit.

Claims (1)

(57) [Claims] An input means for capturing a head image of a person, and an approximation processing section performs an approximation process on a contour of the head image captured by the input means, and a model adaptation section adaptively applies the contour image to the contour image. First approximation processing means for fitting a model in a memory;
A second approximation processing means for approximating the boundary line between the hair region in the head and the head region in the approximation processing unit and adaptively fitting the model in the memory in the model adaptation unit, and corresponding to the sequentially selected model An adaptive calculation formula processing means for calculating the amount of movement of the head to be moved, and head position detection processing means for obtaining the amount of movement according to the shape of the head from the means. Part position detection device.