JPH09282460A - Automatic target recognizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To automatically identify a target even when there are radio interference or radio disturbance by providing a target attribute detector detecting a target attribute from feature values and positional information to unnecessitate a radar device. SOLUTION: A position measuring device 9 measures the position of the device itself and the position of the target. A distance between the device and the target is measured from the position and the attitude angle of an image pickup device 1 and a target position within an image when the target moves at an equal speed. A target attribute detector 5 obtains positional information from the position measuring device 9 and measures the speed of the target, acceleration and a range from the locus of he target within the image. In addition a feature value from a feature extractor 3 and a correction value from an external input device 4 are inputted to detect an aspect and its attribute is inputted to a neural network selector 6. The target attribute detector 5 is provided with a neural network learning by the kind of a target and detects the aspect by it.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an automatic target recognition device for automatically recognizing an object from an image by using a neural network.

[0002]

2. Description of the Related Art FIG. 7 is an overall configuration diagram of a conventional automatic target recognition apparatus. Reference numeral 1 denotes an image pickup apparatus, which is an apparatus for shooting a target and its background, and 2 is processing for adding various image processing from an original image obtained by the image pickup apparatus to easily extract a target feature amount such as a binary image An image processing device for producing an image, a feature extracting device for producing a target feature amount from a processed image,
Reference numeral 14 is a radar device for obtaining target information such as range and aspect, 4 is an external input device for the user to input a correction value of the target information, and 5 is only information necessary for identifying the information obtained from the radar device. A target attribute detecting device for detecting and outputting the target attribute of the neural network, 6 is a neural network for selecting a neural net identifying device from the feature amount extracted by the feature extracting device and the target correction attribute region output by the target attribute detecting device. A selection device, 7 is a neural net identification device that identifies a target from the extracted feature amount, and 8 is an input of the output value and the identification result obtained by each neural net identification device. When the discrimination result is judged and the final target discrimination result is output and one neural net discriminating device is selected, the discrimination of the selected neural net discriminating device is performed. Result identification determination device outputs the final target identification result as an identification result of the ultimate goal, however,
A conventional automatic target recognition device is characterized by using a neural network identification device that obtains target attribute information from a radar device, creates one network for a plurality of targets for each target attribute, and performs learning and recognition. There is.

[0003]

The first problem is the problem of the radar of the conventional automatic target recognition device. In the field of defense, there is a possibility that an enemy may detect by emitting a radio wave by a radar. In addition, a correct radar signal may not be obtained due to radio wave interference.

Further, although the influence of the radar on the human body has not been clarified, it is considered that there is an influence on the person who operates the automatic target recognition apparatus and the object to be photographed.

The second problem is that the conventional automatic target recognition apparatus learns all the targets to be identified, so if there is an automatic identification apparatus for identifying the targets a, b, and c, it is newly called d. In order to identify the target, the already created neural net identification device cannot lead to the result d, so that each neural net identification device becomes invalid and it is necessary to redo all learning.

The present invention was conceived in order to solve such a problem, and the first problem is to change the radar device to a position measuring device, and then learn the neural network identification device. It is an object of the present invention to solve the above-mentioned problem by allocating not to each aspect but to each target and creating a neural network that answers the target aspect.

[0007]

According to a first aspect of the present invention, a position measuring device for measuring a position of a device and a target position, a target attribute detecting device for detecting a target attribute from a feature amount and position information, and a plurality of neural networks which exist. From the net identification device,
The neural network selecting device for selecting the neural network identifying device including the detected target correction attribute region is provided.

According to the second aspect of the present invention, there is provided an aspect detecting device for detecting an aspect and a neural net selecting device for selecting a neural net identifying device including a detected aspect region from a plurality of existing neural net identifying devices. I have it.

According to the third aspect of the invention, a range detecting device for detecting a range to a target and a neural network identifying device including a detected range region are selected from a plurality of existing neural network identifying devices. Equipped with a selection device.

According to the fourth aspect of the invention, a neural network selecting device for detecting a target size and a neural network identifying device for selecting a neural network identifying device including a detected size region from a plurality of existing neural net identifying devices are selected. It is equipped with a device.

Further, in the fifth invention, a position measuring device for measuring the position of the device itself and the position of the target, a target attribute detecting device for detecting the target attribute, a size detecting device for detecting the size of the target, and a target And a neural net selection device that selects a neural net identification device based on the attribute.

Further, in the sixth invention, a position measuring device for measuring the position of the device itself and the position of the target, a target attribute detecting device for detecting the target attribute from the characteristic amount and position information, and the size of the target are detected. A size detection device, a neural network selection device that selects a neural network identification device based on a target correction attribute region, a neural net aspect identification device that has learned by target, and an aspect detection device that uses a neural network are provided.

[0013]

BEST MODE FOR CARRYING OUT THE INVENTION

First Embodiment FIG. 1 is a first embodiment of an automatic target recognition apparatus according to the present invention.
FIG. 1 to 8 are the same as the conventional one, and 9 is a position measuring device for measuring the position of the device itself and the target position. The automatic target recognition device according to the present invention has
9.

The following method can be used as a method for detecting the target distance, velocity, and acceleration when the position measuring device 9 outputs the position and posture angle of the device. First, if the target is moving at a constant velocity, the distance from the target can be measured from the position and posture angle of the imaging device and the target position in the image using a passive ranging method. Further, when stereoscopic vision is used, the distance to the target can be measured although the accuracy depends on the distance between the two viewpoints, and the target velocity and acceleration can be measured from the target trajectory in the image.

The target attribute detecting device 5 obtains position information from the position measuring device 8 to measure the target speed, acceleration, and range, and further, the feature amount is inputted from the feature extracting device 3 and the correction value is inputted from the external input device. By doing so, the aspect is detected, and its attribute is input to the neural network selection device 6. As the aspect detection method, the change in the feature amount of the target depending on the aspect is used, and if the target type is the same such as the target is a vehicle or an aircraft, the target shapes in the same aspect region are similar, Aspect regions can be detected by learning the features for each target type using a neural network. Therefore, the aspect can be detected by providing the target attribute detection device with the neural network learned for each target type.
It is also possible to estimate the aspect from the target traveling direction.

As described above, by using the position detecting device and the target attribute detecting device, the same result as when the radar device is provided can be obtained.

Embodiment 2 FIG. 2 shows Embodiment 2 of the automatic target recognition apparatus according to the present invention.
FIG. 1 to 4 and 6 to 8 are the same as the conventional one,
An aspect detection device 10 is a device that detects an aspect region from the feature amount obtained from the feature extraction device 3 and the target type input from the external input device. The automatic target recognition apparatus according to the present invention is constituted by the above 1 to 7.

The operation of the above automatic target identifying device will be described below. A target type such as an aircraft or a vehicle is input as a correction value by the external input device 4, and the corrected aspect area is detected by the aspect detection device as in the first embodiment. FIG. 8 is a diagram in which the entire aspect direction of the target is divided into several regions, or the neural net identification device of FIG. 7 learns the target within one aspect amount region shown in FIG. Create one neural network,
One neural net identification device is constructed.

Third Embodiment FIG. 3 shows an automatic target recognition apparatus according to a third embodiment of the present invention.
FIG. 1 to 4 and 6 to 8 are the same as the conventional one,
Reference numeral 4 is an external input device, which outputs a correction value, and 11 is a range detection device, which is a device for detecting a range region by a neural network from the feature amount and the correction value obtained from the feature extraction device 3. The automatic target recognition device according to the present invention has
4, 6 to 8 and 11.

FIG. 9 is a diagram in which the target range is divided.
Input the target type as a correction value with an external input device,
A neural network that outputs the range area shown in the figure for each target type is created in the range detection device.

FIG. 11 shows an example of the characteristic amount for detecting the range. Taking the same goal with different distances as an example,
At 21 the image is binarized by the image processing device and is input to the feature extraction device, at 22 the target contour and center of gravity are detected, and at 23 the conversion into logarithmic polar coordinates is performed. It is possible to create a neural network that outputs the range region by learning the feature amount in 23 stages, where the difference in distance becomes a shift in the logarithmic direction. When the shift is made in the logarithmic axis direction like 24, the feature amount is invariant to the size in the screen.

Embodiment 4 FIG. 4 shows an overall configuration of an embodiment of the present invention. 1 to 4 and 6 to 8 are the same as the conventional ones, and 12 is a size detection device, which is a device for detecting a target size from a binarized image obtained by the image processing device 2. The automatic target recognition apparatus according to the present invention is the above-mentioned 1-4, 6-8, 1
2.

FIG. 10 shows an example of shooting the target in the same range and a case of shooting the target in the same size. FIG.
(A) is a target photographed from the same distance, and the sizes of the targets are different on the screen. FIG. 10B shows that by changing the distance between the two targets, the sizes of the targets are almost the same in the screen. By limiting the targets to be recognized by the neural network identification device to the same size, it is possible to improve the recognition result by coping with the reduction of the identification targets due to the size.

Embodiment 5 FIG. 5 shows Embodiment 5 of the automatic target recognition apparatus according to the present invention.
FIG. 1 to 8 are the same as the conventional device, and 9
Is the same as that of the first embodiment, and 12 is the same device as that of the fourth embodiment. The automatic target recognition device according to the present invention is constituted by the above 1 to 9 and 12.

Sixth Embodiment FIG. 6 shows a sixth embodiment of the automatic target recognition apparatus according to the present invention.
FIG. 1 to 8 are the same as the conventional device, and 9
Is the same as the first embodiment, 11 is the same as the third embodiment, 12 is the same device as the fourth embodiment, and 13
Is a neural network and aspect identification device that learns by target and answers the target aspect. The automatic target recognition device according to the present invention is constituted by the above 1-9, 11-13.

Here, a neural network / aspect identification device is selected by the neural network selection device based on the target correction attribute region, and a neural network that answers the aspect region for each target is created for learning and recognition. The aspect area output from the selected neural network identification device and the output from the aspect detection device are compared. If there is one neural network identification device that outputs the same aspect area, the target indicated by the neural network identification device is the final target identification. If there are multiple results, the one with the largest output is the final target identification result.

[0027]

As described above, according to the present invention, since the radar device is unnecessary, it is possible to automatically identify the target even if there is a radio interference or an interference image.

Further, if the entire aspect of the target is divided into some aspect regions and learning is performed in each region, the difference in the target becomes clear and the recognition can be improved.

In addition, the target feature quantity may not actually be invariable even by a method that is theoretically invariant depending on the range. Therefore, by dividing the region into regions and learning, It is possible to perform corresponding learning and improve recognition.

Further, as in the case of the range, the target feature quantity may not be actually changed even by a method which is theoretically invariable depending on the size. Therefore, the area is divided according to the size and learning is performed. Therefore, learning corresponding to each size can be performed, and recognition can be improved.

Further, by combining the target correction attribute area and the correction size area, it is possible to select the minimum necessary neural net identification device and improve the recognition rate.

Finally, the neural network aspect ratio discriminator selected by the neural net selector selects one neural network for each target in the target attribute region to improve the convergence of the network, and When learning a new goal, the already learned network can be used as it is.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an automatic target recognition apparatus showing a first embodiment of the present invention.

FIG. 2 is a configuration diagram of an automatic target recognition device showing a second embodiment of the present invention.

FIG. 3 is a configuration diagram of an automatic target recognition device showing a third embodiment of the present invention.

FIG. 4 is a configuration diagram of an automatic target recognition device showing a fourth embodiment of the present invention.

FIG. 5 is a configuration diagram of an automatic target recognition device showing a fifth embodiment of the present invention.

FIG. 6 is a configuration diagram of an automatic target recognition device showing a sixth embodiment of the present invention.

FIG. 7 is a block diagram showing a conventional automatic target recognition device.

FIG. 8 is a diagram in which all target aspect directions are divided into several regions.

FIG. 9 is a diagram in which a target range is divided into several areas.

FIG. 10 is a diagram of an image of a vehicle having the same size as an image of two vehicles having the same range.

FIG. 11 is a configuration diagram showing a feature extraction method.

[Explanation of symbols]

 1 Imaging Device 2 Image Processing Device 3 Feature Extraction Device 4 External Input Device 5 Target Attribute Detection Device 6 Neural Net Selection Device 7 Neural Net Discrimination Device 8 Discrimination Judgment Device 9 Position Measurement Device 10 Aspect Detection Device 11 Range Detection Device 12 Size Detection Device 13 Neural Net Aspect Identification Device 14 Radar Device 15 Aircraft 16 Part of Region Divided All Aspects 17 Part of Divided Entire Range 18 Imaging Device 21 Binary Image 22 Target Contour 23 Polar Coordinate Map 24 Shifted Polar Coordinate Map

Claims (6)

[Claims] 1. An image pickup apparatus for photographing a target and outputting an original image, and an image processing apparatus for inputting an original image created by the image pickup apparatus and for processing the original image to output a processed image created. And a processed image processed by the image processing device is input, a target feature amount is extracted from the processed image and is output, and a position measurement that outputs position information of the target imaged with itself. A device, an external input device for inputting a correction value by which a user corrects a target attribute, and outputting the correction value, a target feature amount extracted by the feature extraction device, and position information output by the position measuring device. And a target attribute detection device that outputs a target correction attribute region such as a target range or aspect from a correction value by the external input device,
The target correction attribute region detected by the target attribute detection device and the target feature amount extracted by the feature extraction device are input, and the network is configured to learn the feature amount including the attribute in the target correction attribute region. A neural network selecting device that outputs a target feature amount to the selected neural net identifying device, and learning using this input target feature amount.
And a plurality of neural net identification devices that perform identification and output identification results and output values, and when a plurality of the neural net identification devices are selected, the output values and identification results obtained by each neural net identification device are The final target identification result is judged from those input results, the final target identification result is output, and when one neural net identification device is selected, the identification of the selected neural network identification device is performed. An automatic target recognition device, comprising: an identification determination device that outputs a final target identification result as a final target identification result. 2. An aspect detection device which receives a correction value obtained by an external input device and a target feature amount extracted by the feature extraction device and outputs a corrected aspect region, and an aspect detection device which detects the aspect ratio region. The correction aspect region and the target feature amount extracted by the feature extraction device are input, and the neural network identification device configured by a network that learns the feature amount including the attribute in the correction aspect region is selected, The automatic target recognition device according to claim 1, further comprising a neural network selection device that outputs a target feature amount to the selected neural network identification device. 3. A correction value obtained by an external input device and a target feature amount extracted by the feature extraction device are input, and a range detection device that outputs a correction range region and an output by the range detection device. The correction range region and the target feature amount extracted by the feature extraction device are input, and the neural network identification device configured by a network that learns the feature amount including the attributes in the correction range region is selected, The automatic target recognition device according to claim 1, further comprising a neural network selection device that outputs a target feature amount to the selected neural network identification device. 4. A size detection device for outputting a target correction size region from a correction value obtained by an external input device and an image binarized by the image processing device, and a correction size output by the size detection device. The region and the target feature amount extracted by the feature extraction device are input, and the neural network identification device configured by a network that learns the feature amount including the attribute in the correction size region is selected,
The automatic target recognition device according to claim 1, further comprising a neural network selection device that outputs a target feature amount to the selected neural network identification device. 5. The position measuring device, the neural network attribute detecting device, the size detecting device, an external input device for inputting a correction value for a user to correct a target attribute, and outputting the correction value, The neural network identification device, which is configured by a network in which the target correction attribute region detected by the target attribute detection device and the correction size region detected by the size detection device are input, and which learns a feature amount including the input attribute 2. The automatic target recognition device according to claim 1, further comprising a neural net selection device that selects a target feature amount and outputs a target feature amount to the selected neural net identification device. 6. The position measurement device, the neural network attribute detection device, the target correction attribute region detected by the size detection device and the target attribute detection device, and the correction size region detected by the size detection device. A neural network selection device that selects the above-mentioned neural net identification device that is configured by a network that learns a feature amount that is input and that includes the input attribute, and that outputs a target feature amount to the selected neural net identification device. , A plurality of neural network / aspect identification devices that perform learning and recognition for each target by using the feature amount of the target input by the neural net selection device, identify the target aspect, and output the identification result and the output value And the outputs of the aspect detection device and the neural network aspect identification device. 2. The automatic target recognition device according to claim 1, further comprising an identification determination device which inputs the identification result and the corrected aspect area output from the aspect detection device and outputs the final target identification result.