JPH09114977A - Method and device for processing image - Google Patents

Abstract

PROBLEM TO BE SOLVED: To properly extract an object regardlessly of change in the intensity of illumination or the kind of the object. SOLUTION: A television camera 11 is arranged being directed to an observing position and a mobile object, which appears at the observing position, is extracted. A background image memory 2 stores image data previously provided by picking up an image in the state of letting no mobile object exist at the observing position, and image data at the time of extracting processing are inputted to a local correlation circuit 3. The local correlation circuit 3 sets the 5×5 local areas to both the inputted image data and the background image data in the background image memory 2, scans them and executes normalizing correlation arithmetic using the image data at respective scanning positions. At a binarizing circuit 4, a calculated correlation value is binarized by a prescribed threshold value and outputted to a control part or the like.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image processing method and apparatus for imaging a television camera toward a predetermined observation position and extracting a specific object from the obtained image.

[0002]

2. Description of the Related Art As a method of installing a TV camera toward a specific observation position and extracting an object such as a moving object using the input image from the TV camera, conventionally, the input image to be observed and the A method that uses the density difference value with the background image obtained by imaging the state where the object does not exist at the observation position (hereinafter referred to as "background difference method"), or the input image to be observed and the one step before. Method that uses the density difference value from the input image (hereinafter referred to as "time difference method")
Have been proposed.

FIG. 8 shows an example of the configuration of an image processing apparatus to which the above-mentioned background subtraction method is introduced, which is composed of a switch 31, a background image memory 32, a difference circuit 33, a binarization circuit 34 and the like. The switch 31 is for switching the output destination of the grayscale image data (hereinafter simply referred to as “image data”) obtained from a television camera (not shown). When a non-existing state is imaged, it is turned on under the control of a control unit (not shown). As a result, the image data from the television camera is stored in the background image memory 32 as background image data (an example is shown in FIG. 10), after which the switch 31 is turned off.

When the switch 31 is turned off,
The subsequent image data is input to the difference circuit 33. The difference circuit 33 is for calculating the density difference value between the input image data and the background image data in the background image memory 32, and the absolute value of the density difference for each corresponding pixel between both image data. Is calculated, and the calculation result is calculated by the binarization circuit 3
Output to 4. The binarization circuit 34 binarizes each input absolute value with a predetermined threshold value, and outputs the result for each pixel.

FIG. 11 shows an input image obtained when an object (a person in the illustrated example) appears at an observation position, and FIG.
An image (hereinafter referred to as an “extracted image”) after the input image is subjected to the difference processing with the background image in FIG. 10 and the binarization processing by the binarization circuit 34 is shown respectively. Pixel group A corresponding to image A of the object
Only 'is represented as a black pixel.

FIG. 9 shows an example of the configuration of an image processing apparatus to which the time difference method is introduced. A difference circuit 3 similar to the above is used.
In addition to the 3-binarization circuit 34, a frame buffer 35 and the like are included as components.

The image data from the television camera is input to the difference circuit 33 and the frame buffer 35 at the same time. The difference circuit 33 is delayed by one frame by the image data input at the present time and the frame buffer 35. Calculation processing of the density difference value with the input image data is performed. This process calculates the absolute value of the density difference for each corresponding pixel between both image data, as in the case of FIG.
The binarization circuit 34 binarizes the absolute value of each pixel and outputs the result for each pixel to generate an extracted image.

[0008]

However, in the case of the image processing apparatus using the "background subtraction method" of FIG. 8, the illumination intensity that illuminates the imaging area at the time of capturing the background image and the image of the processing target. If there is a large change, the density difference between the portion other than the object in the image to be processed and the corresponding portion of the background image data becomes large, which causes the problem of erroneous extraction as shown in FIG.

In order to prevent erroneous extraction due to such a change in illumination intensity, it is conceivable to increase the threshold value set in the binarization circuit 34. In this case, as shown in FIG. There is a possibility that the corresponding part may not be extracted.

Also, as another method for preventing erroneous extraction due to a change in illumination intensity, it has been proposed to update the contents of the background image memory at predetermined time intervals so as to respond to changes in the environment of the image pickup area. However, in this method, if the background image update rate is slowed down, it becomes impossible to cope with a sudden change in illumination intensity, and the same result as in FIG. 13 is brought about. On the contrary, if the update speed of the background image is increased, there is a problem that a moving object having a slow motion cannot be extracted.

Next, when the "time difference method" of FIG. 9 is used, if the object is a moving body with slow movement, a density difference exceeding the threshold value does not occur between the image data of one frame before and the image data of one frame before, It becomes impossible to extract the movement of the object. Also,
Even with this method, it is impossible to cope with a rapid change in illumination intensity.

The present invention has been made in view of the above problems, and provides an image processing method and apparatus capable of correctly extracting an object regardless of the change in illumination intensity or the type of the object. The purpose is to

[0013]

According to a first aspect of the present invention, a television camera is imaged at a predetermined observation position, image data obtained is input, and a specific object is extracted from the input image. In the image processing method, a background image in which the object does not exist at the observation position is stored in advance, and image data in a predetermined area on the input image and image data on the background image corresponding to this image data are stored. After performing the normalized correlation calculation of 1), it is determined whether the image data corresponding to the object exists in the area on the input image by using the calculation result.

According to a second aspect of the present invention, there is provided an image processing device for picking up an image with a television camera directed to a predetermined observation position, inputting the obtained image data, and extracting a specific object from the input image. , Background image storage means for storing background image data in which the object does not exist at the observation position, image data in a predetermined area on the input image, and image data on a background image corresponding to this image data Image data extraction means, calculation means for executing a normalized correlation calculation on both image data extracted by the image data extraction means, and a correlation value calculated by the calculation means for targeting in the area on the input image. And a discriminating means for discriminating whether or not image data corresponding to the object exists.

According to a third aspect of the present invention, the image data extracting means sets a local area having a predetermined size on the input image and scans the input image data, and the input image data in the local area and the image at each scanning position. The image data on the background image corresponding to the data is sequentially extracted.

According to a fourth aspect of the present invention, there is provided an image processing apparatus, wherein a television camera is imaged at a predetermined observation position, image data obtained is input, and a specific object is extracted from the input image. Background image storage means for storing background image data in which the object does not exist at the observation position, and feature point extraction means for extracting a plurality of feature points from the input image,
For each feature point extracted by the feature point extraction means, image data extraction means for extracting image data in a predetermined local area including the feature point and image data on a background image corresponding to this image data, An arithmetic means for executing a normalized correlation operation on both image data extracted by the image data extracting means for each characteristic point, and a characteristic point corresponding to an object by using each correlation value calculated by the arithmetic means. And a discriminating means for discriminating the object.

According to the invention of claim 5, the feature point extracting means extracts pixels constituting an edge in the input image as feature points.

According to a sixth aspect of the present invention, an image in which a plurality of television cameras are respectively imaged toward a predetermined observation position, and the obtained image data of each television camera is simultaneously input to perform an object recognition process. In the processing device, a background image storage unit that stores background image data in which the object does not exist at the observation position for each TV camera, and a plurality of feature points are extracted for each input image from each TV camera. Image data on a background image corresponding to the image data in a predetermined local area including the feature points for each feature point of each input image extracted by the feature point extraction means and the feature point extraction means Image data extracting means for extracting data, and arithmetic means for executing a normalized correlation operation between both image data extracted by the image data extracting means for each feature point, Correspondence between each input image with respect to the characteristic point specified by the specifying section and the specifying point for specifying the characteristic point corresponding to the object for each input image using each correlation value calculated by the calculation section The associating means for attaching, the coordinate extracting means for three-dimensionally measuring each feature point associated by the associating means and extracting the three-dimensional coordinates of each feature point, and the coordinate extracting means for extracting. And a determination unit that determines the three-dimensional shape of the object using the three-dimensional coordinates of each feature point.

[0019]

According to the first and second aspects of the invention, the image data in the corresponding areas of the input image and the background image are
The object is discriminated using the result of the normalized correlation calculation. Since the normalized correlation calculation is not affected by the illumination variation, the calculation result reflects whether or not the object exists in the area.

According to the third aspect of the invention, the local area is scanned over the entire input image, and the above processing is performed for each scanning position.

According to the fourth aspect of the invention, a plurality of feature points are extracted from the input image, and image data in a predetermined local area including the feature points and background image data corresponding thereto are extracted for each feature point. Then, the normalized correlation calculation is executed, and the feature point corresponding to the object is specified using the result.

According to the fifth aspect of the invention, the contour portion of the object is extracted by extracting the pixels forming the edge in the input image as the feature points.

According to the invention of claim 6, the input images obtained by a plurality of television cameras are respectively defined by claim 4.
After performing the same process as above, the feature points identified as points corresponding to the object are associated between the input images, and the three-dimensional coordinates of each feature point are extracted using the result of the association. Then, the three-dimensional shape of the object is determined based on the extraction result.

[0024]

BEST MODE FOR CARRYING OUT THE INVENTION The image processing apparatus of FIGS. 1 to 3 is an example of the configuration for implementing the inventions of claims 1 to 3, and the image of FIG. 4 is an example of the configuration for implementing claims 4 and 5. As the processing apparatus, the image processing apparatus of FIG. 5 is given as a configuration example for carrying out claim 6, and a detailed description of each image processing apparatus will be developed below.

[0025]

1 is a block diagram showing the arrangement of an image processing apparatus according to the first embodiment of the present invention. This image processing device, in advance, produces grayscale image data (hereinafter referred to as “target image data”) obtained by imaging a television camera 11 (hereinafter simply referred to as “camera 11”) toward a predetermined observation position, Whether or not a moving object such as a human (hereinafter referred to as "moving object") appears at the observation target position by using the background image data obtained by imaging the state where the extraction target does not exist at the observation position. It is for recognition, and includes a switch 1, a background image memory 2, a local correlation circuit 3, a binarization circuit 4 and the like as a configuration.

The switch 1 is turned on / off by a control unit (not shown), and is shown in FIG.
Similarly to the above, the background image data is stored in the background image memory 2 by being turned on when the background image data is captured prior to the extraction processing. After that, the switch 1 is turned off, and the target image data is input to the local correlation circuit 3.

The local correlation circuit 3 sets a local area of a predetermined size in the target image data obtained from the camera 11 and the background image data in the background image memory 2 and scans them. The correlation value R calculated at each scanning position is sequentially input to the binarization circuit 4 in order to execute the normalized correlation calculation between the image data. Binarization circuit 4
Is for binarizing the input correlation value R based on a predetermined threshold value, and a control unit (not shown) uses the result of this processing to find data corresponding to the object in the target image data. It is determined whether or not it exists.

As shown in FIG. 2, the local correlation circuit 3 in this embodiment vertically and horizontally relates to the target image data G _I input from the camera and the background image data G _M input from the background image memory 2. Local regions R _I and R _M of 5 pixels are set, and these local regions R _I and R _M are horizontally scanned from the upper left of each image data G _I and G _{M at} the same timing, and local regions at each scanning position are set. Region R _I ,
Using the image data in R _M , the normalized cross-correlation operation between both local regions is executed.

FIG. 3 shows a detailed structure of the local correlation circuit 3, which is a target image scanning unit 5, a background image scanning unit 6, and a background image scanning unit 6.
It includes a normalized correlation calculator 7 and the like. The target image scanning unit 5 and the background image scanning unit 6 each include four line buffers LB and 20 flip-flops FF arranged in 4 rows × 5 columns. , From each column, density data I _{0 to} I ₂₄ of each pixel in the local area R _I , from each row and column of the background image scanning unit 6, density data M ₀ of each pixel in the local area R _M. ~ M ₂₄ are extracted respectively. Each of the extracted density data is input to the normalized correlation calculation unit 7, and the correlation value R between the local regions R _I and R _M is obtained by executing the calculation based on the following equation (1). In the following equation, n is 25.

[0030]

(Equation 1)

When the target image data shown in FIG. 11 is input to the above image processing apparatus and collated with the background image data shown in FIG. 10, a correlation value R in a portion corresponding to the background in the target image data is obtained. However, the correlation value R in the image A portion of the moving body (person) becomes low, and as a result of the processing by the binarization circuit 4, only the pixels corresponding to the object are converted into black pixels and the extracted image data is obtained. (Shown in FIG. 12) is obtained. Since the normalized correlation calculation can always obtain stable results without being affected by the fluctuation of illumination, even if the illumination at the time of inputting the target image data fluctuates greatly from the time of inputting the background image data, Only the image can be accurately extracted.

In the above embodiment, the matching processing with the background image data is performed with all the pixels of the target image data as the processing target. Feature points of
You may make it perform the matching process with a background image about the local area | region containing this extracted feature point.

FIG. 4 shows an example of a device configuration for realizing the second method, in which the CPU 17 is the main control unit,
Switch 12, edge extraction unit 13, edge image memory 1
4, a background image memory 15, an image memory 16, an edge coordinate memory 18, a feature point coordinate memory 19 and the like are included as components.

The switch 12 switches the connection to either the background image memory 15 or the image memory 16 under the control of the control unit mainly composed of the CPU 17, whereby the background image data is stored in the background image memory 15. The target image data is stored in the image memory 16, respectively.

The edge extracting unit 13 is for inputting the target image data from the camera, extracting the edge components in the image, and generating an edge image in which only the edge constituent points are represented by black pixels. The created edge image is stored in the edge image memory 14, and the coordinates of the edge constituent points in the edge image are further stored in the edge coordinate memory 18.

In this embodiment, a program for realizing the functions of the normalized correlation calculator 7 and the binarization circuit 4 of the first embodiment is stored in a ROM (not shown) or the like, and C
The PU 17 refers to the edge coordinate memory 18 and, with respect to the target image data in the image memory 16 and the background image data in the background image memory 15, 5 × 5 centered on a pixel corresponding to each edge constituent point. Local areas are sequentially set, and the normalized correlation calculation processing and the binarization processing based on the program are performed on the image data in both local areas. Further, the CPU 17 compares the correlation value R obtained for each edge composing point with a predetermined threshold value, and only the edge composing points for which the correlation value R is below the threshold value are contoured to the moving body. It is determined that the feature point is a constituent point (hereinafter referred to as a “moving body feature point”), and the coordinates thereof are stored in the feature point coordinate memory 19 and used for a moving body recognition process or the like.

FIG. 5 shows an image processing apparatus according to the third embodiment of the present invention. This image processing device is capable of recognizing the three-dimensional shape of the moving body at the observation position by taking in the image data from the two cameras 11a and 11b. In addition to the configuration in which the two configurations of the second embodiment are combined, , A model memory 20 and a three-dimensional coordinate memory 21. It should be noted that, in the drawing, of the configurations common to the second embodiment, the CPU 17
The edge coordinate memory 18 is commonly used for processing both image data from the cameras 11a and 11b, but other configurations are provided for each of the cameras 11a and 11b.

Therefore, in the figure, the same reference numerals are given to the respective constituents common to the second embodiment, and the constituents related to the processing of the image data from the camera 11a have a at the end of the reference numeral and the camera 1
In the configuration related to the processing of the image data from 1b, b is added to the end of the code, and the description thereof is omitted. The cameras 11a and 11b are both fixedly arranged toward the observation position and their positions are adjusted so that their optical axes are parallel to each other.

The three-dimensional coordinate memory 21 is for storing the three-dimensional coordinates of each moving body feature point obtained by the moving body feature point associating process and the three-dimensional measuring process described later, and is stored in the model memory 20. Stores a plurality of types of three-dimensional models of a moving object that is a recognition target of this image processing apparatus. The CPU 17 collates the shape of the object specified by each feature point in the three-dimensional coordinate memory 21 with each three-dimensional model in the model memory 20, and recognizes the type of the moving object in the observation target based on the collation result.

FIG. 6 shows the processing procedure of the image processing apparatus of the second embodiment. First, when the background image data from the camera 11 is stored in the background image memory 15 in step 1 (indicated by "ST1" in the figure), the switch 12 is switched, and the subsequent target image data is stored in the image memory 16. (Step 2).

This target image data is also given to the edge extraction unit 13 at the same time, and the edge constituent points are extracted to generate an edge image. FIG. 15 shows an edge image generated for the target image data of FIG. 11, in which the contour portion of the image A of the person and the contour portion of the object placed on the background are extracted.

When the target image data for one frame is stored in the image memory 16 and simultaneously the edge image data corresponding to this image data is also stored in the edge image memory 14, the CPU 17 executes the following step 3. The coordinates of each edge constituent point in this edge image are extracted and stored in the edge coordinate memory 18.

In the next step 4, the CPU 17 sets, for the image memory 16 and the background image memory 15, the predetermined coordinate data stored in the edge coordinate memory 18 as the center of the pixel specified by the coordinate data. After extracting the image data in the 5 × 5 local region, the normalized correlation calculation between the two local regions is performed in the following step 5 by executing the equation (1) for both the extracted image data. Run.

Next, the CPU 17 compares the correlation value R calculated in step 5 with a predetermined threshold value Th, and when the correlation value R is below the threshold value Th, the edge constituent point to be processed moves. It is determined that it is a body feature point, and its coordinates are stored in the feature point memory 19 (step 7).

The processes of steps 4 to 7 are carried out for all the edge composing points for which the coordinate data is stored in the edge coordinate memory 18. When this process is completed, the judgment of step 8 is "YES". Become Step 9
Move to.

In step 9, the CPU 17 recognizes from the coordinate data stored in the feature point coordinate memory 19 whether or not the target moving body exists at the observation target position. FIG. 16 shows the result of performing the correlation calculation process and the binarization process on the edge image of FIG. In the case of the illustrated example, the correlation value in the local region set on each edge composing point of the image portion of the person is low, but the correlation value in the local region set on another edge composing point is high, and as a result, The contour portion of the image A of the person is extracted. Based on this extraction result, the CPU 17 recognizes that a moving object exists at the observation target position, and stores the recognition result in a RAM (not shown) or the like.

The above processing is performed every time one frame of the target image data is input, and the CPU 17 recognizes the movement and the moving direction of the moving body from the temporal transition of the recognition result at the time of inputting each image. To do.

FIG. 7 shows a processing procedure of the image processing apparatus according to the third embodiment. The processing from steps 1 to 3
The image data from the camera 11a and the image data from the camera 11b are respectively processed in parallel, and in step 1, the background image data from the cameras 11a and 11b are stored in the background image memories 15a and 15b, respectively. Then, in subsequent step 2, each switch 12a, 1
2b is switched, and the target image data is input to the image memories 16a and 16b.

In the next step 3, the same processing as in steps 3 to 8 of FIG. 6 is performed for each target image data, and as a result, it is determined as a moving body feature point in each target image data. Only the coordinates of the edge constituent points are stored in the feature point coordinate memories 19a and 19b.

Next, the CPU 17 executes a process of associating the moving body feature points extracted between the first and second target image data with one-to-one correspondence (step 4). In this process, for example, attention is paid to a predetermined moving body feature point of the target image data from the first camera 11a (for convenience of description, Pi),
After extracting a plurality of moving body feature points (for convenience of description, q1, q2 ...) Located on the same epipolar line as the moving body feature point Pi in the second target image data, The mobile body feature point qi having the highest degree of similarity to the mobile body feature point Pi is determined as a corresponding point by performing template matching on the density data of the points and several pixels in the vicinity thereof.

When the above associating process is completed, the CPU
17 applies the principle of three-dimensional measurement to each associated moving body feature point, calculates the three-dimensional coordinates of each moving body feature point, and stores them in the three-dimensional coordinate memory (step 5).

When the three-dimensional coordinates of all the moving body feature points are obtained, the CPU 17 collates the shape of the object specified by these three-dimensional coordinates with each three-dimensional model in the model memory 20, and extracts the extraction result. The model closest to is adopted as the recognition result (step 6). As the matching process, for example, for each three-dimensional model, the Euclidean distance obtained for each corresponding point between the model constituent points and each mobile body feature point in the three-dimensional coordinate memory 21 is averaged,
A method of comparing the average values with each other is used.

Similar to the second embodiment, the above process is repeated every time one frame of the target image data is input, and the CPU 17 moves based on the transition of the recognition result at the time of inputting each image. Recognize body shape and movement direction.

In each of the second and third embodiments, the edge composing points are used as the characteristic points of the target image, but the present invention is not limited to this. For example, pixels having density data within a predetermined range in the image may be used. It may be the target. Further, in the first to third embodiments, the moving body is the extraction target, but the extraction target is not limited to this, and an event such as a certain stationary object disappearing from the observation position may be the extraction target.

[0055]

According to the first and second aspects of the invention, the object discrimination processing is performed using the result of the normalized correlation calculation for the image data in the corresponding areas of the input image and the background image, respectively. Since the calculation result is used to determine whether or not an object exists at the observation position, it is possible to accurately determine whether the object exists at the observation position without being affected by lighting fluctuations or the moving speed of the object. Therefore, the accuracy of image processing can be improved.

According to the third aspect of the invention, the local area is scanned over the entire input image, the image data in the local area at each scanning position and the corresponding background image data are extracted, and the above processing is performed. It is possible to extract the detailed structure of an object.

According to the invention of claim 4, a plurality of feature points are extracted from the input image, and image data in a predetermined local area including the feature points and background image data corresponding thereto are extracted for each feature point. Then, the normalized correlation operation is executed, and the feature point corresponding to the object is specified using the result, so that the object can be extracted at high speed.

According to the fifth aspect of the present invention, the pixels forming the edge in the input image are extracted as the feature points, so that the contour portion of the object can be accurately extracted.

According to the invention of claim 6, the input images obtained by a plurality of television cameras are respectively defined by claim 4.
After performing the same process as above, the feature points identified as points corresponding to the object are associated between the input images, and the three-dimensional coordinates of each feature point are extracted using the result of the association. Moreover, since the three-dimensional shape of the object is determined based on the extraction result, the shape and orientation of the object can be accurately extracted.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an image processing apparatus according to a first embodiment of the present invention.

FIG. 2 is an explanatory diagram showing a method of matching a target image with a background image.

FIG. 3 is a block diagram showing a detailed configuration of a local correlation circuit.

FIG. 4 is a block diagram showing a configuration of an image processing apparatus according to a second embodiment of the present invention.

FIG. 5 is a block diagram showing a configuration of an image processing apparatus according to a third embodiment of the present invention.

FIG. 6 is a flowchart showing a processing flow of the image processing apparatus of the second embodiment.

FIG. 7 is a flowchart showing a processing flow of the image processing apparatus of the third embodiment.

FIG. 8 is a block diagram showing a configuration example of a conventional image processing apparatus.

FIG. 9 is a block diagram showing another configuration example of a conventional image processing apparatus.

FIG. 10 is a diagram showing an example of a background image.

FIG. 11 is a diagram showing an example of a target image.

FIG. 12 is a diagram showing an example of an extracted image.

FIG. 13 is a diagram showing an example of an extracted image.

FIG. 14 is a diagram showing an example of an extracted image.

FIG. 15 is a diagram showing an edge image obtained by subjecting the target image of FIG. 11 to edge extraction processing.

FIG. 16 is a diagram showing an example of extraction of mobile body feature points.

[Explanation of symbols]

2, 15, 15a, 15b background image memory 3 local correlation circuit 4 binarization circuit 7 normalized correlation operation circuit 11, 11a, 11b camera 13, 13a, 13b edge extraction unit R _I , R _M local region

Claims (6)

[Claims] 1. An image processing method in which a television camera is imaged at a predetermined observation position, image data obtained is input, and a specific object is extracted from the input image. A background image in which an object does not exist is stored in advance, and after performing a normalized correlation calculation between the image data in a predetermined area on the input image and the image data on the background image corresponding to this image data, An image processing method characterized by determining whether or not image data corresponding to an object exists in the area on the input image using the calculation result. 2. An image processing device for imaging a television camera toward a predetermined observation position, inputting the obtained image data, and extracting a specific object from the input image. Background image storage means for storing background image data in which an object does not exist, and image data extraction means for extracting image data in a predetermined area on the input image and image data on a background image corresponding to this image data An operation means for executing a normalized correlation operation on both image data extracted by the image data extraction means; and an image corresponding to an object in the area on the input image based on the correlation value calculated by the operation means. An image processing apparatus comprising: a determination unit that determines whether or not data exists. 3. The image data extracting means sets and scans a local area having a predetermined size on an input image, and at each scanning position, the input image data in the local area and a background corresponding to this image data. The image processing apparatus according to claim 2, which is a unit that sequentially extracts image data on an image. 4. An image processing apparatus for imaging a television camera to a predetermined observation position, inputting the obtained image data, and extracting a specific object from the input image. Background image storage means for storing background image data in which the object does not exist, feature point extraction means for extracting a plurality of feature points from the input image, and each feature point extracted by the feature point extraction means, respectively. Image data extracting means for extracting image data in a predetermined local area including a feature point and image data on a background image corresponding to this image data, and both extracted by the image data extracting means for each feature point. Calculation means for performing a normalized correlation calculation on the image data, and a discriminant for discriminating the object by specifying feature points corresponding to the object using the correlation values calculated by the calculation means. The image processing apparatus comprising a means. 5. The image processing apparatus according to claim 4, wherein the feature point extracting means extracts pixels forming an edge in the input image as feature points. 6. An image processing device for recognizing an object by simultaneously imaging a plurality of television cameras toward a predetermined observation position and simultaneously inputting the obtained image data of the television cameras. A background image storage unit that stores background image data in which the object does not exist at the observation position for each TV camera, and a feature point extraction unit that extracts a plurality of feature points for each input image from each TV camera. For each feature point of each input image extracted by the feature point extracting means, image data in a predetermined local area including the feature point and image data on a background image corresponding to this image data are extracted. Image data extraction means, operation means for executing a normalized correlation operation between both image data extracted by the image data extraction means for each feature point, and the operation means Using the correlation values calculated by the above, the specifying means for specifying the feature points corresponding to the object for each input image, and the feature points specified by the specifying means are associated with each other between the input images. Associating means, coordinate extracting means for performing three-dimensional measurement on each feature point associated by the associating means, and extracting three-dimensional coordinates of each feature point, and each feature extracted by the coordinate extracting means. An image processing apparatus comprising: a discriminating means for discriminating a three-dimensional shape of an object using three-dimensional coordinates of points.