JP2008021034A - Image recognition device, image recognition method, pedestrian recognition device and vehicle controller - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve the discrimination precision of an object detected from an image. <P>SOLUTION: When a candidate object as the candidate of discrimination is detected from an input image, rough discrimination using a neural network method (that is, the discrimination of whether or not there is possibility that the detected object is a pedestrian as the object of discrimination) is performed. When there is possibility that the object detected from the input image is a pedestrian as the result of rough discrimination, detail discrimination using the neural network method (that is, the discrimination of whether or not the detected object is a pedestrian as the object of discrimination) is performed. Then, the notification of warning or the control of a vehicle is performed according to the detail discrimination result (for example, the discrimination result showing that the detected object is a pedestrian) or the like. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an image recognition device that discriminates an object that is a candidate for discrimination when an object that is a candidate for discrimination is detected from an image captured by a monocular camera.

  Conventionally, a technique for discriminating an object (for example, a pedestrian) detected from an image picked up by an in-vehicle camera has been proposed for the purpose of avoiding a collision with an object that obstructs vehicle travel. Yes. For example, in Patent Document 1, an object to be discriminated is detected from an image captured by a predetermined camera (for example, a visible light camera), and an object discrimination of a neural network using a pedestrian contour pattern for the object is performed. A technique for discriminating whether or not a detected object is a pedestrian by applying a technique is disclosed.

JP 2004-145660 A

  By the way, the above-described conventional technique has a problem in that the accuracy of discriminating an object detected from an image may be reduced. That is, the above-described conventional technique applies a neural network object discrimination method using a pedestrian contour pattern to an object detected from an image to determine whether the detected object is a pedestrian or the like. Therefore, there is a possibility that an object with a shape similar to the contour of the pedestrian (for example, a tree similar to the shape of the pedestrian) may be identified as a pedestrian, and the object identification accuracy may be reduced. was there.

  Therefore, the present invention has been made to solve the above-described problems of the prior art, and an image recognition device, an image recognition method, and pedestrian recognition that can improve the discrimination accuracy of an object detected from an image. An object is to provide a device and a vehicle control device.

  In order to solve the above-described problems and achieve the object, the invention according to claim 1 is directed to an object that is a candidate for determination when an object that is a candidate for determination is detected from an image captured by a camera. Object determination means for determining a predetermined part constituting the predetermined object to be determined, and determining that the predetermined object is a predetermined object when a determination value obtained as a result exceeds a predetermined threshold It is provided with.

  According to a second aspect of the present invention, in the above invention, the object discriminating unit includes a plurality of predetermined objects constituting a predetermined object to be discriminated for an object that is a discrimination candidate detected from the image. It is characterized in that if the total value obtained by adding the respective discriminant values obtained as a result exceeds a predetermined threshold, it is determined that the object is the predetermined object.

  According to a third aspect of the present invention, in the above invention, the object discriminating unit weights one of the discriminant values and then adds the discriminant values to obtain a predetermined threshold value. If it exceeds, it is determined that the object is the predetermined object.

  According to a fourth aspect of the present invention, in the above invention, the object discriminating unit performs a determination on the operation of the predetermined part for the same object that is a candidate for discrimination detected from a plurality of continuous images. When the discriminant value obtained as a result exceeds a predetermined threshold, it is discriminated that the object is the predetermined object.

  According to a fifth aspect of the present invention, in the above invention, there is provided a possibility determining means for determining a possibility that the object as a determination candidate detected from the image is a predetermined object to be determined. The object discriminating unit further includes the object discriminating unit, when the possibility discriminating unit discriminates that the object that is the candidate for discrimination is the predetermined object, It is characterized in that a determination relating to a predetermined part constituting a predetermined object is performed, and when the determination value obtained as a result exceeds a predetermined threshold, it is determined that the object is the predetermined object.

  According to a sixth aspect of the present invention, in the above invention, the possibility determining means determines the predetermined part according to a position on the image of an object detected from the image as the determination candidate. And determining an object that may be the predetermined object.

  In the invention according to claim 7, in the above invention, the object discriminating unit discriminates an object that is a candidate for discrimination as the predetermined object by using a threshold value according to a predetermined condition. It is characterized by.

  According to an eighth aspect of the present invention, in the above invention, the object determination unit performs a determination on the predetermined part according to a predetermined environmental condition, and determines an object that is a candidate for the determination as the predetermined object. It is characterized by distinguishing.

  According to a ninth aspect of the present invention, in the above invention, the object determination unit is an object that is a candidate for determination according to a position on the image of an object detected from the image as the determination candidate. Is determined as to whether or not it is the predetermined object.

  According to a tenth aspect of the present invention, in the above invention, the object determination unit determines that an object that is a candidate for determination detected from the image is a predetermined object to be determined. In this case, the image processing apparatus further includes an intrusion determination unit that determines whether or not the predetermined object is about to enter the road from the image.

  According to an eleventh aspect of the present invention, in the above invention, the object determination unit performs a determination on the predetermined part that is different from each other for the same object that is a candidate for determination detected from a plurality of continuous images. When the discriminant value obtained as a result exceeds a predetermined threshold value, it is determined that the object is the predetermined object.

  In the invention according to claim 12, in the above invention, when an object that is a candidate for determination is detected from an image captured by a camera, the object that is a candidate for determination is the target of the determination. It includes an object determination step of determining a predetermined part constituting a predetermined object, and determining that the predetermined object is a predetermined object when a determination value obtained as a result exceeds a predetermined threshold value. And

  In the invention according to claim 13, in the above-described invention, when an object that is a candidate for determination is detected from an image captured by a camera, the object that is a candidate for determination is a target of the determination. An object discriminating unit is provided that performs a judgment on a predetermined part constituting the pedestrian and discriminates that the user is a pedestrian when a discrimination value obtained as a result exceeds a predetermined threshold.

  According to the fourteenth aspect of the present invention, when an object that is a candidate for determination is detected from an image captured by a camera, a predetermined object that is a target of the determination is configured for the object that is a candidate for determination. When the determination regarding the predetermined part to be performed is performed and the determination value obtained as a result exceeds a predetermined threshold, the warning notification and And / or a control means for controlling the vehicle.

  According to the first or twelfth aspect of the present invention, with respect to an object that is a candidate for determination detected from an image, a predetermined part that constitutes a predetermined object that is a target of determination (for example, if the target of determination is a pedestrian) (For example, a determination by a neural network method or a pattern matching method) and a discriminant value (for example, an output value) obtained as a result exceeds a predetermined threshold (ie, If the detected object is likely to have a pedestrian's head), it is determined that the object is a predetermined object. For example, determination regarding the pedestrian's part such as the head, torso or limbs By doing this, there is no fear that an object having a shape similar to the overall shape of the pedestrian (for example, a tree similar to the shape of the pedestrian) will be erroneously identified as a pedestrian. To improve the determination accuracy of the detected object from are possible.

  According to the second aspect of the present invention, for example, a plurality of predetermined parts constituting a predetermined object to be determined (for example, a determination target is a walking object) If it is a person who has made a judgment on the head, torso, limbs, etc., and the sum of the discriminant values obtained as a result exceeds a predetermined threshold value (that is, in the detected object) Pedestrian's head, torso and limbs are highly likely to be present), it is determined that the person is a pedestrian. Thus, it is possible to comprehensively determine whether or not the object detected from the image is a pedestrian, and it is possible to further improve the determination accuracy of the object detected from the image.

  According to the third aspect of the present invention, the determination as to a plurality of predetermined parts constituting the predetermined object to be determined is performed for each object that is a determination candidate detected from the image, and the result is obtained. Each of the discriminant values is weighted (for example, when the distance is short, sufficient information is obtained and the output value obtained as a result of the judgment on the head with little individual difference is doubled, etc.) Therefore, when the total value obtained by adding the respective discrimination values exceeds a predetermined threshold value, it is determined that the object is a predetermined object, so that the determination accuracy of the object detected from the image can be improved with a simple operation. It is.

  According to the invention of claim 4, the same object that is a candidate for discrimination detected from a plurality of continuous images (for example, the current image and images of several frames input thereafter) is a discrimination target. Discriminations obtained as a result of performing determinations regarding the motion of a plurality of predetermined parts constituting a predetermined object (for example, if the determination target is a pedestrian, determination regarding forward and backward swinging hands), respectively. If each value exceeds a predetermined threshold (that is, if there is a high possibility that a forward hand and a backward hand are present in the detected image), it is determined that the object is a predetermined object. For example, whether or not an object detected from an image is a pedestrian can be determined in consideration of an element of motion, and the determination accuracy of an object detected from an image can be further improved.

  According to the invention of claim 5, it is determined whether an object that is a candidate for determination detected from an image is a predetermined object (for example, a pedestrian) to be determined (for example, a predetermined part). After determining the possibility of being a pedestrian based on whether or not the discriminant value obtained as a result exceeds a predetermined threshold set in advance for simple discrimination. For an object that has been determined to be possible, a determination is made regarding a predetermined part that constitutes the predetermined object, and the determination value obtained as a result is set to a predetermined threshold (for example, high for detailed determination in advance). If it is determined that the object is a predetermined object, for example, after the object detected from the image is determined in advance as an object of pedestrian as a determination target, Head, torso and It performs determination as to the site of pedestrians and feet, by determining whether or not the pedestrian in detail, it is possible to prevent wasteful perform a detailed discrimination process.

  According to the sixth aspect of the present invention, the determination on the predetermined part of the predetermined object (for example, a pedestrian) according to the position on the image of the object detected from the image as the determination candidate is performed, and the result For example, if the obtained discriminant value exceeds a predetermined threshold set in advance for simple discrimination, it is discriminated as an object that may be a predetermined object, so the candidate object is a pedestrian. Use different discrimination patterns to determine whether there is a possibility (for example, use the pedestrian's head pattern when the distance between the detected candidate object grasped from the position on the image and the vehicle is short) If the distance to the candidate object is far, the whole body pattern of the pedestrian is used), and it is possible to accurately determine whether the candidate object may be a pedestrian. Whether or not It is possible to perform a physical efficiently.

  Further, according to the invention of claim 7, since the threshold value is changed according to a predetermined condition, for example, the threshold value can be changed according to conditions such as a travel location, a travel speed, and an object detection position. It is possible to further improve the discrimination accuracy of the object detected from the image.

  Further, according to the invention of claim 8, since the determination regarding the predetermined part constituting the predetermined object to be determined is performed according to the predetermined environmental condition, for example, the difference in pedestrian clothes and the time zone As a result of performing determination according to environmental conditions such as ambient brightness, temperature, weather, and time zone, it is possible to further improve the discrimination accuracy of an object detected from an image.

  According to the ninth aspect of the present invention, whether or not to execute a determination on a predetermined part constituting a predetermined object to be determined according to a position on the image of an object detected from the image as a determination candidate. For example, when an object is detected on the road, a detailed determination process is executed, and when an object is detected outside the road, a detailed determination process is not executed. It is possible to prevent the detailed determination processing from being performed wastefully.

  According to the invention of claim 10, when it is determined that an object that is a candidate for determination detected from an image is a predetermined object to be determined, for example, the position of a pedestrian in the current image When the next image is approaching the road or when a pedestrian is facing the road, it is determined from the image whether or not a predetermined object is about to enter the road. It is possible to determine whether or not the object is going to enter the road.

  According to the eleventh aspect of the present invention, for the same object that is a candidate for discrimination detected from a plurality of continuous images, the determination regarding the predetermined part is performed separately (for example, determination regarding the head in the current image). When the determination is made, the determination regarding the limb is performed in the next image), and when the determination value obtained as a result exceeds a predetermined threshold value, it is determined that the object is the predetermined object. Even when a pedestrian is not clearly imaged, it is possible to improve the discrimination accuracy of an object detected from an image.

  According to the invention of claim 13, a determination regarding a predetermined part constituting a pedestrian to be determined is performed on an object that is a determination candidate detected from an image, and a determination value obtained as a result is obtained. Since it is determined that the person is a pedestrian when the predetermined threshold value is exceeded, an object having a shape similar to the overall shape of the pedestrian ( For example, it is possible to improve discrimination accuracy for discriminating whether or not an object detected from an image is a pedestrian, without erroneously discriminating a pedestrian tree or the like. .

  According to the fourteenth aspect of the present invention, the determination value obtained as a result of performing the determination on the predetermined part constituting the predetermined object to be determined with respect to the object as the determination candidate detected from the image. Is determined to be a predetermined object when it exceeds a predetermined threshold value, and warning notification and / or vehicle control is performed according to the determination result. As a result of accurately determining whether or not there is a warning, it is possible to efficiently perform warning notification and vehicle control.

  Exemplary embodiments of an image recognition device, an image recognition method, a pedestrian recognition device, and a vehicle control device according to the present invention will be described below in detail with reference to the accompanying drawings. In the following, a vehicle equipped with an image recognition device and a vehicle control device according to the present invention will be described as a first embodiment, and then another embodiment included in the present invention will be described.

  In the following first embodiment, the outline and features according to the first embodiment of the present invention, the configuration of the first embodiment, and the flow of processing will be described in order, and finally the effects of the first embodiment will be described.

[Outline and Features (Example 1)]
First, the outline and features of the first embodiment will be described with reference to FIG. FIG. 1 is a diagram for explaining the outline and features of the first embodiment.

  As shown in the figure, in the first embodiment, an object that becomes a candidate for discrimination is detected when an object that becomes a candidate for discrimination is detected from an image captured by a vehicle-mounted camera (for example, a monocular camera or a stereo camera). The outline of performing the predetermined control (notification control and vehicle travel control) is the main feature in that the accuracy of determining the object detected from the image is improved. Here, in order to improve the efficiency of object discrimination, only objects existing in the road on the image are set as discrimination targets.

  Specifically, for example, when one frame of an image captured by a monocular camera is input, a candidate object that is a candidate for determination is detected from the input image by, for example, a saliency calculation method ( (See (1) in FIG. 1). Here, the saliency calculation method is based on a learning base that is generated and stored in advance by a learning process, and features that are prominent in the input image (for example, a general background as an environment around a vehicle such as a road or a building). An object detection method for detecting an object depending on whether or not is detected.

  When a candidate object that is a candidate for discrimination is detected from the input image, a rough discrimination using a neural network technique (that is, whether the detected object is a possibility of a pedestrian to be determined) (See (2) in FIG. 1). Here, the neural network method is to generate and store a learning model learned about an object to be discriminated (for example, a human such as a pedestrian), and input a pixel value of an unknown image to the learning model. This is a technique for discriminating an object according to an output value output in a case.

  Specifically, for example, as illustrated in FIG. 2, the size and shape of the contour of the candidate object detected as a candidate for discrimination are compared with the size and shape of the pattern of the pedestrian's whole body. When the determination output value (for example, A) indicating the degree of coincidence exceeds a predetermined determination threshold value (for example, I) (for example, when A ≧ I), the detected candidate object is determined. It is determined that there is a possibility of a target pedestrian. That is, it is intended to prevent the detailed determination processing from being performed wastefully by making a rough determination on the detected object in advance and determining the possibility of a pedestrian to be determined.

  As a result of rough discrimination, if it is determined that there is a possibility of a pedestrian with respect to an object detected from the input image, then detailed discrimination using the neural network technique (that is, the detected object is It is determined whether or not the subject is a pedestrian (see (3) in FIG. 1).

  Specifically, for example, as illustrated in FIG. 3, the size and shape of an object determined to be a pedestrian by rough determination, and a predetermined part (for example, a head) , Torso, limbs, etc.) (for example, the head) and the determination output value (for example, B) indicating the degree of coincidence exceeds a predetermined determination threshold (for example, G) In this case, it is determined that the person is a pedestrian.

  Subsequently, warning notification and vehicle control are performed according to the determination result of the candidate object (see (4) in FIG. 1). Specifically, when it is determined that the candidate object is a pedestrian, for example, the distance from the position on the image to the pedestrian is calculated, the collision risk with the pedestrian is determined, and the determination result is In response, warning notification and vehicle control are performed. The distance to the pedestrian may be acquired by an on-vehicle radar (for example, millimeter wave radar, optical radar, etc.).

  For this reason, an object having a shape similar to the overall shape of a pedestrian (for example, the shape of a pedestrian, for example, by performing a determination on a pedestrian's part such as the head, torso, or limbs as described above. It is possible to improve the discrimination accuracy of the object detected from the image and whether the object detected from the image is a pedestrian or not. As a result of being able to discriminate with high accuracy, it is possible to perform warning notification and vehicle control efficiently.

[Configuration of Example 1]
Next, the configuration of the image recognition device and the vehicle control device according to the first embodiment will be described with reference to FIG. FIG. 4 is a block diagram illustrating configurations of the image recognition device and the vehicle control device according to the first embodiment. As shown in the figure, the vehicle includes a vehicle control device 10, an imaging device (monocular camera) 11, a navigation unit 12, a radar unit 13, an in-vehicle notification unit 14, and an image recognition device 20. The vehicle control device 10 includes a notification control unit 15 and a vehicle control unit 16, and the image recognition device 20 includes a preprocessing unit 21, a storage unit 22, and a control unit (microcomputer) 23.

  Among them, the navigation unit 12 is a means for setting and guiding a travel route from the specified vehicle position and map data by communicating with a GPS (Global Positioning System) artificial satellite. The navigation unit 12 supplies various information useful for vehicle driving operations such as vehicle position information, road shape, road width, and inclination to the driver via an in-vehicle notification unit 14 described later.

  The radar unit 13 irradiates a radar (for example, a 76.5 GHz millimeter wave radar or an optical radar), and measures a distance, speed, direction, and the like with a vehicle or an obstacle (for example, a pedestrian) ahead. It is. For example, when the radar unit 13 detects the presence of a pedestrian located in front of the vehicle, the radar unit 13 automatically recognizes the pedestrian's feet (ground plane) and acquires the distance from the ground plane to the vehicle by the radar ( After the measurement), the distance information (distance from the vehicle to the pedestrian) is output to the collision risk determination unit 23c.

  The in-vehicle notification unit 14 is means for notifying information from the navigation unit 12 or a notification control unit 15 described later, and includes a monitor, a speaker, and the like. For example, the in-vehicle notification unit 14 receives a command from the notification control unit 15 and outputs an image indicating the presence of a pedestrian to the monitor to alert the driver or to send a message or an alarm sound from a speaker. To warn the driver.

  The notification control unit 15 of the vehicle control device 10 is a processing unit that receives a command from a collision risk determination unit 23c, which will be described in detail later, and performs notification control. For example, when the notification control unit 15 receives a driver's attention or warning command from the collision risk determination unit 23c, the notification control unit 15 uses an image or an alarm sound indicating the presence of a pedestrian to the in-vehicle notification unit 14, A command is output so that the driver is warned and warned. When the collision risk determination unit 23c is configured to determine only the collision risk, the notification control unit 15 performs notification control according to the collision risk received from the collision risk determination unit 23c. The content may be determined and executed by itself, or when the image recognition apparatus 20 is configured not to include the collision risk determination unit 23c, the notification control unit 15 determines the collision risk by itself. The notification control may be executed.

  The vehicle control unit 16 of the vehicle control device 10 is a processing unit that receives a command from the collision risk determination unit 23c and performs vehicle control (vehicle travel control). For example, when the vehicle control unit 16 receives a brake control command from the collision risk determination unit 23c, the vehicle control unit 16 executes brake control to perform speed reduction. When a steering wheel control command is received from the collision risk determination unit 23c, the vehicle control unit 16 performs steering wheel control to avoid collision. When the collision risk determination unit 23c is configured to determine only the collision risk, the vehicle control unit 16 controls the vehicle travel according to the collision risk received from the collision risk determination unit 23c. If the image recognition apparatus 20 is configured not to include the collision risk determination unit 23c, the vehicle control unit 16 determines the collision risk by itself. Then, the vehicle travel control may be executed.

  The preprocessing unit 21 of the image recognition device 20 is a processing unit that performs preprocessing on an image captured by the imaging device 11, and includes a filter unit 21a and a contour extraction unit 21b. Among these, the filter unit 21a is a means for performing preprocessing (for example, sharpness, contrast adjustment, and saturation adjustment) for enhancing the contour of an object projected in an image. The contour extracting unit 21b is means for extracting the contour of an object in the image based on the filtering performed by the filter unit 21a.

  The storage unit 22 of the image recognition apparatus 20 is a storage unit (storage unit) that stores data and programs necessary for various processes performed by the control unit 23. In particular, as closely related to the present invention, a learning base storage unit 22a, a learning model storage unit 22b, and a collision risk determination table 22c are provided.

  Among these, the learning base storage unit 22a is a storage unit that stores a learning base necessary for object detection processing by the object detection unit 23a described in detail later. The generation of the learning base will be specifically described. First, a plurality of learning images obtained by shooting the same object (background including roads and the like) in which there is no object (for example, a pedestrian) to be discriminated. Is input to cut out the learning area (for example, 20 × 20 pixels). Subsequently, this cut-out learning area is treated as an N-dimensional vector (N is a pixel value, for example, 400 dimensions if 20 × 20 pixels), and a principal component analysis is performed on the learning area consisting of the N-dimensional vector. Then, a feature vector (a principal component of an N-dimensional vector) is extracted. Then, feature vectors are extracted from all input learning images, and learning bases are generated and stored from feature vector components common to the same object.

  The learning model storage unit 22b is a storage unit that stores a learning model necessary for object discrimination processing by the object discrimination unit 23b described in detail later. Specifically, patterns for pedestrians that are objects to be discriminated (patterns of the pedestrian's whole body and patterns of heads, limbs, and torso that are each part of the pedestrian) are set under various conditions (for example, , Learning in advance according to the distance from the vehicle, the weather in which clothing changes are considered, and the time zone when pedestrians are difficult to see).

  The collision risk determination table 22c is a storage unit that stores a determination table necessary for the collision risk determination process by the collision risk determination unit 23c described in detail later. Specifically, as illustrated in FIG. 5, when an object (candidate object) that is detected from an image and is a candidate for determination is “not a pedestrian”, regardless of the distance to the candidate object (that is, “ALL”). )), The risk level is “0”, the vehicle control / notification control is “none”. On the other hand, if the candidate object is “pedestrian”, the distance to the candidate object is “30 m or more” and the risk level is “level”. 1 ”, vehicle control / informing control“ output image indicating presence of pedestrian ”, distance“ 10-30 m ”, risk“ level 2 ”, vehicle control / informing control“ speed deceleration by brake control ”, distance“ “Less than 10 m”, the risk level is “level 3”, and the vehicle control / notification control is “prevented from collision by steering wheel control”.

  Note that the configuration of the collision risk determination table 22c described above is merely an example. For example, even when the candidate object is not a pedestrian, risk setting, vehicle control, or notification control may be performed. It is good to update the distance to the candidate object automatically according to the traveling speed of the vehicle (for example, when the speed of the vehicle is high, the setting distance is automatically updated to be longer so that a collision can be avoided). Alternatively, the contents of the vehicle control or notification control may be different (for example, “pre-crash” as vehicle control when the risk level is high, or “notice control” when the risk level is medium. You may make it employ | adopt by prescribing | regulating.

  The control unit (microcomputer) 23 of the image processing apparatus 20 has an internal memory for storing predetermined control programs, programs defining various processing procedures, and necessary data, and processes for executing various processes using these programs. Part. In particular, an object detection unit 23a, an object determination unit 23b, and a collision risk determination unit 23c are closely related to the present invention.

  Among these, the object detection unit 23a is a processing unit that detects candidate objects that are candidates for discrimination from the input image by the saliency calculation method. Specifically, the object detection unit 23a reads the learning base from the learning base storage unit 22a and is not seen in the learning base in the input image as compared with the input image currently input from the imaging device 11. When a prominent feature (for example, a peculiar vector) appears, the region is detected from the image as a candidate object that is a candidate for discrimination. Note that the object detection unit 23a is not limited to detecting an object that is a candidate for discrimination from the input image by the saliency calculation method. For example, the object detection unit 23a is detected by image processing or edge extraction processing by the background difference method. You may do it. The object detection unit 23a performs a candidate object detection process for each predetermined image frame (for example, one frame or several frames) input from the preprocessing unit 21.

  The object discriminating unit 23b is a processing unit that discriminates in detail by a neural network method whether or not a candidate object that is a discrimination candidate detected from the input image by the object detection unit 23a is a pedestrian.

  Specifically, first, the object determination unit 23b configures the pedestrian stored in the learning model storage unit 22b and the contours and sizes of candidate objects in the input image currently input from the imaging device 11. A determination (region determination) is performed by comparing the size and shape of any pattern of a predetermined region (eg, head, torso, limbs, etc.), and a determination output value indicating the degree of coincidence (degree of coincidence) is obtained. Ask.

  Then, if the determination output value exceeds a predetermined determination threshold value (for example, 70% with 100% being a perfect match), the object determination unit 23b determines a predetermined part (for example, the head) that the candidate object constitutes the pedestrian. , Body, limb, etc.), the candidate object is determined to be a pedestrian, and the determination result (for example, the determination result of “being a pedestrian”) is output to the collision risk determination unit 23c. To do.

  Further, the object determination unit 23b determines, for example, the size and shape of the pattern of the pedestrian's whole body and the size and shape of the contour of the candidate object for the candidate object detected by the object detection unit 23a. (Whole body determination) is performed to roughly determine whether the candidate object is likely to be a pedestrian, and if the candidate object is likely to be a pedestrian, further region determination is performed. You may make it discriminate | determine in detail. When there are a plurality of candidate objects detected by the object detection unit 23a, the object to be determined in detail is narrowed down by determining whether or not the object is a pedestrian.

  Specifically, the object determination unit 23b compares the size and shape of the pedestrian's whole body pattern stored in the learning model storage unit 22b with the size and shape of the contour of the candidate object. If the determination output value indicating the degree of coincidence is equal to or greater than a predetermined determination threshold (for example, if 100% is 30% or more with complete coincidence), it is determined that the candidate object may be a pedestrian. Subsequently, in order to obtain a final conclusion as to whether or not the candidate object is a pedestrian, a predetermined part constituting the pedestrian stored in the learning model storage unit 22b (for example, head, torso, limbs) Etc.) is determined in detail by performing part determination on the degree of matching with each pattern.

  If a candidate object is determined based on a comparison with only the whole body pattern, even if the candidate object is actually a pedestrian, the determination threshold for determining whether the candidate object is a pedestrian is increased. If it is set, there is a risk that the candidate object detected in a slightly ambiguous state (shape) from the input image may be “not a pedestrian”, and if the discrimination threshold is lowered, the shape of the pedestrian There is a risk of discriminating objects such as trees similar to pedestrians. On the other hand, if a candidate object is determined based on a comparison with only one of predetermined patterns (eg, head, torso, limbs, etc.) constituting a pedestrian, for example, an object similar to the shape of the head is walked. When there are a plurality of candidate objects detected from the input image, a predetermined part constituting the pedestrian (for example, head, torso, limbs, etc.) It is inefficient to perform the discrimination processing for any of the patterns. In other words, the purpose of performing detailed determination on a candidate object that may be a pedestrian after determining whether the candidate object is a pedestrian through rough determination is based on such a situation. It is. Note that whole body determination may be performed after narrowing down candidate objects in the region determination.

  In addition, the object discrimination processing executed in the object discrimination unit 23b is not limited to the case where a candidate object that is a discrimination candidate detected from the input image is discriminated by a neural network method, but a pattern (for example, a pedestrian) The present invention can be similarly applied to a case where object discrimination is performed by a pattern matching method using a template such as a predetermined part or a whole body constituting a pedestrian such as a head, a torso, and a limb.

  The collision risk determination unit 23c is a processing unit that determines the collision risk with the object detected from the input image based on the determination result by the object determination unit 23b, and outputs a command for vehicle control or notification control. . Specifically, upon receiving the determination result from the object determination unit 23b, the collision risk determination unit 23c reads the collision risk determination table (see FIG. 5) from the collision risk determination table 22c. Then, the collision risk determination unit 23c determines the determination result received from the object determination unit 23b and the distance to the pedestrian ascertained from the position on the image (if it is located at the upper part of the image, it is located at the lower part of the image. If the collision risk is determined according to the distance received from the radar unit 13 or the predetermined distance, the vehicle control command is output to the vehicle control unit 16 or Then, a notification control command is output to the notification control unit 15.

  That is to say, with an example, the collision risk determination unit 23c has no collision risk (risk level) except when the determination result “is a pedestrian” is received from the object determination unit 23b. “Level 0”) is determined, and no vehicle control or notification control command is output. On the other hand, when the collision risk degree determination unit 23c receives the determination result “is a pedestrian” from the object determination unit 23b, the collision risk determination unit 23c, for example, according to the distance from the position on the image to the pedestrian, for example, When the distance to the pedestrian is long (for example, 30 m or more), it is determined that the risk of collision is low (risk level “level 1”) and driving based on an image showing the presence of the pedestrian A command is output to the notification control unit 15 so as to give a warning or warning to a person, and when the distance to the pedestrian is close to a medium level (for example, 10 m or more and less than 30 m), a collision occurs. Is determined to be moderate (level 2), a command is output to the vehicle control unit 16 to perform speed reduction by brake control, and the distance to the pedestrian is extremely close (for example, 10 m) Less than Risk of outputting a command to the vehicle control unit 16 so as to determine high and (Level 3) performs collision avoidance by steering wheel control. In addition, you may make it grasp | ascertain the distance to a pedestrian based on the distance information acquired from the radar part 13 (for example, a millimeter wave radar, an optical radar, etc.).

  Further, the collision risk determination unit 23c may perform only the collision risk determination and output the collision risk determination result to the notification control unit 15 or the vehicle control unit 16. In this case, the notification control unit 15 and the vehicle control unit 16 each determine and execute the content of the notification control and the content of the vehicle control according to the collision risk received from the collision risk determination unit 23c.

[Processing by Image Recognition Device and Vehicle Control Device (Example 1)]
Next, processing by the microcomputer (control unit) and the vehicle control device of the image control device according to the first embodiment will be described with reference to FIG. FIG. 6 is a flowchart illustrating a flow of processing by the microcomputer (control unit) and the vehicle control device of the image control device according to the first embodiment. The process performed by the vehicle control device described below is repeatedly executed every time one frame of an image is input.

  As shown in the figure, when one frame of an image is input (Yes at step S601), the preprocessing unit 21 performs preprocessing on the input image (step S602). Next, the object detection unit 23a detects candidate objects that are candidates for discrimination from the input image by the saliency calculation method (step S603).

  Specifically, the object detection unit 23a reads the learning base from the learning base storage unit 22a and is not seen in the learning base in the input image as compared with the input image currently input from the imaging device 11. When a prominent feature (for example, a peculiar vector) appears, the region is detected from the image as a candidate object that is a candidate for discrimination.

  Then, the object determination unit 23b determines the degree of coincidence between the size and shape of the candidate object that is a determination candidate detected from the input image by the object detection unit 23a and the size and shape of the pattern of the pedestrian's whole body. By performing (whole body determination) and roughly determining, it is determined whether or not there is a possibility that the candidate object is a pedestrian (step S604).

  Specifically, the object determination unit 23b compares the size and shape of the pedestrian's whole body pattern stored in the learning model storage unit 22b with the size and shape of the contour of the candidate object, If the determination output value indicating the degree of coincidence is equal to or greater than a predetermined determination threshold (for example, if 100% is 30% or more with complete coincidence), it is determined that the candidate object may be a pedestrian.

  As a result, when it is determined that the candidate object has a possibility of a pedestrian (Yes in step S604), the object determination unit 23b obtains a final conclusion as to whether or not the candidate object is a pedestrian. A part determination regarding the degree of coincidence with each pattern of a predetermined part (for example, head, torso, limbs, etc.) constituting the pedestrian stored in the learning model storage unit 22b is performed to determine in detail (step S605). . On the other hand, when the candidate object is not determined to be a pedestrian (No in step S604), the object determination unit 23b waits for input of an image of the next frame.

  Specifically, the object discriminating unit 23b uses the patterns of predetermined parts (for example, the head, the torso, and the limbs) constituting the pedestrian stored in the learning model storage unit 22b for the contours of the candidate objects. And if there is a pattern that roughly matches, compare the size and shape of the contour of the candidate object with the size and shape of the pattern that approximately matches, A determination output value indicating the degree of coincidence is obtained, and if the determination output value exceeds a predetermined determination threshold (for example, 70% with 100% being the complete match), the candidate object is a predetermined portion (for example, a pedestrian) The candidate object is determined to be a pedestrian.

  If it is determined in detail that the candidate object is determined to be a pedestrian (Yes at step S605), the object determination unit 23b uses the determination result (for example, a determination result of “being a pedestrian”) as a collision risk. It outputs to the degree determination part 23c (step S606). On the other hand, if the candidate object is not determined to be a pedestrian (No at step S605), the object determination unit 23b does not output the determination result and waits for the input of the next frame image.

  Subsequently, the collision risk determination unit 23c determines the collision risk with the candidate object detected from the input image based on the determination result received from the object determination unit 23b (step S607), and performs vehicle control and notification. Outputs a control command. Specifically, when the collision risk determination unit 23c receives the determination result from the object determination unit 23b, the collision risk determination unit 23c reads the collision risk determination table (see FIG. 5) from the collision risk determination table 22c, and receives the determination result from the object determination unit 23b. Determine the collision risk according to the received determination result and the distance to the pedestrian grasped from the position on the image (distant if it is located at the top of the image, close if it is located at the bottom of the image), A vehicle control command is output to the vehicle control unit 16, or a notification control command is output to the notification control unit 15.

  That is, to explain with an example, the collision risk determination unit 23c has no collision risk (level 0) except when the determination result “is a pedestrian” is received from the object determination unit 23b. In particular, the vehicle control and notification control commands are not output. On the other hand, when the collision risk degree determination unit 23c receives the determination result “is a pedestrian” from the object determination unit 23b, the collision risk determination unit 23c, for example, according to the distance from the position on the image to the pedestrian, for example, When the distance to the pedestrian is long (for example, 30 m or more), it is determined that the risk of collision is low (risk level “level 1”) and driving based on an image showing the presence of the pedestrian A command is output to the notification control unit 15 so as to give a warning or warning to a person, and when the distance to the pedestrian is close to a medium level (for example, 10 m or more and less than 30 m), a collision occurs. When the risk level of the vehicle is determined to be medium (risk level “level 2”), a command is output to the vehicle control unit 16 to reduce the speed by brake control, and the distance to the pedestrian is extremely close (For example, when it is less than 10m , The risk of collision and outputs a command to the vehicle control unit 16 to perform collision avoidance by steering wheel control is determined high and (risk "level 3").

  Then, the vehicle control unit 16 that has received a vehicle control command from the collision risk determination unit 23c, or the notification control unit 15 that has received a notification control command performs vehicle control (vehicle travel control) or warning notification, respectively. (Step S608). When the collision risk determination unit 23c is configured to determine only the collision risk, the notification control unit 15 performs notification control according to the collision risk received from the collision risk determination unit 23c. The content may be determined and executed by itself, or the notification control unit 15 may determine the collision risk by itself and execute the notification control. Similarly, the vehicle control unit 16 may determine and execute the content of the vehicle travel control by itself according to the collision risk received from the collision risk determination unit 23c, or the vehicle control unit 16 may You may make it determine vehicle collision risk and perform vehicle travel control.

[Effect of Example 1]
As described above, according to the first embodiment, with respect to an object that is a candidate for discrimination detected from an image, a predetermined part constituting a predetermined object that is a discrimination target (for example, a discrimination target is a pedestrian). If there is a decision on the head, torso, limbs, etc. (for example, a decision by a neural network method or pattern matching method), and the resulting discriminant value (for example, output value) exceeds a predetermined threshold Since it is determined that the object is a predetermined object (that is, when there is a high possibility that a detected object is likely to be a pedestrian's head), for example, a pedestrian part such as a head, a torso or a limb By making a determination on the object, there is no risk that an object having a shape similar to the overall shape of the pedestrian (for example, a tree similar to the shape of the pedestrian) will be mistakenly identified as a pedestrian. It is possible to improve the determination accuracy of the detected object from an image.

  Further, according to the first embodiment, an object that is a candidate for discrimination detected from an image is discriminated as to whether or not it is a predetermined object (for example, a pedestrian) to be determined (for example, determination regarding a predetermined part). To determine whether or not the object is a pedestrian based on whether or not the discriminant value obtained as a result exceeds a predetermined threshold set in advance for simple discrimination. A determination is made regarding a predetermined part that constitutes a predetermined object for an object that has been determined to have a characteristic, and a determination value obtained as a result is a predetermined threshold (for example, a threshold that is set high in advance for detailed determination) ), It is determined that the object is a predetermined object. For example, after the object detected from the image is previously determined as an object that may be a pedestrian to be determined, Torso and limbs It performs determination as to the site of the pedestrian, by determining whether or not the pedestrian in detail, it is possible to prevent wasteful perform a detailed discrimination process.

  Further, according to the first embodiment, a determination regarding a predetermined part that constitutes a predetermined object to be determined is performed on an object that is a determination candidate detected from an image, and a determination value obtained as a result is predetermined. If the threshold is exceeded, it is determined that the object is a predetermined object, and a warning is issued and / or the vehicle is controlled according to the determination result, so whether the object detected from the image is a pedestrian As a result, it is possible to efficiently perform warning notification and vehicle control.

  Although the embodiments of the present invention have been described so far, the present invention may be implemented in various different forms other than the embodiments described above. Therefore, another embodiment included in the present invention will be described below.

(1) After a rough discrimination, a detailed discrimination is made at a plurality of parts. In the first embodiment, a predetermined part (for example, a pedestrian) (for example, a candidate object that is determined to be a pedestrian by a rough discrimination) However, the present invention is not limited to this, and a plurality of predetermined parts constituting the pedestrian ( For example, it may be determined in detail whether the person is a pedestrian by performing determinations on the head, limbs, and torso.

  Specifically, for example, as illustrated in FIG. 7, the object determination unit 23b may determine the possibility of a pedestrian by performing rough determination using a pedestrian's whole body pattern stored in the learning model storage unit 22b. The contours and sizes of candidate objects determined to be present, and the sizes of all patterns of a plurality of predetermined parts (for example, the head, limbs, and torso) constituting the pedestrian stored in the learning model storage unit 22b The total value obtained by adding the respective determination output values (for example, C for the head, D for the torso, E for the hand, and F for the foot) obtained by performing the determination of comparing the height and the shape with the comparison is predetermined. If it exceeds the determination threshold value (for example, H), it is determined that the person is a pedestrian.

  Further, the object discriminating unit 23b substantially matches the patterns of predetermined parts (for example, the head, the torso, the limbs, etc.) constituting the pedestrian stored in the learning model storage unit 22b with respect to the contour of the candidate object. If there is a pattern that almost matches, the size and shape of the contour of the candidate object are compared with the size and shape of the pattern that matches, and the degree of match A determination output value indicating the pedestrian may be determined if the determination output value exceeds a predetermined determination threshold value.

  Thereby, for example, it is possible to comprehensively determine whether or not the object detected from the image is a pedestrian, by making a determination on each part of the pedestrian such as the head, torso and limbs, The discrimination process can be performed efficiently, and the discrimination accuracy of the object detected from the image can be further improved.

(2) Detailed Discrimination Without Rough Discrimination In the above-described first embodiment, candidate objects that may be pedestrians are roughly discriminated using the pedestrian's whole body pattern, etc., and then narrowed down. Although the case where it is determined whether or not the candidate object is a pedestrian has been described, the present invention is not limited to this, and the contour and shape of the candidate object that is a candidate for determination detected from the input image, and learning A determination (part determination) is performed by comparing the size and shape of a pattern of a predetermined part (eg, head, torso, limbs, etc.) constituting the pedestrian stored in the model storage unit 22b. A determination output value indicating a degree (a degree of coincidence) may be obtained, and it may be determined whether or not the person is a pedestrian from the beginning without roughly determining and narrowing down candidate objects that may be pedestrians.

  This makes it possible to walk on an object having a shape similar to the overall shape of the pedestrian (for example, a tree similar to the shape of the pedestrian) by making a determination on the pedestrian's part, such as the head, torso or limbs. Therefore, it is possible to quickly determine the object detected from the image.

(3) Rough discrimination using a part corresponding to the position on the image In the first embodiment, the detected candidate object can be a pedestrian by making a rough discrimination using the pedestrian's whole body pattern or the like. However, the present invention is not limited to this, and depending on the position (distance from the vehicle) on the image where the candidate object is detected, The contour and size of the candidate object are compared with the size and shape of any pattern of a predetermined part (for example, head, torso, limb, etc.) constituting the pedestrian stored in the learning model storage unit 22b. A determination output value indicating the degree of coincidence (coincidence degree) is obtained, and it is possible for a pedestrian to determine whether or not the determination output value exceeds a predetermined discrimination threshold. Whether there is sex You may make it discriminate | determine.

  Specifically, for example, as illustrated in FIG. 8, the vehicle control device is configured such that the distance between the object detected as a discrimination candidate and the vehicle (distant when the image is located at the upper part of the image and located at the lower part of the image). Is determined from the position on the image, and for example, if the distance is close, the head pattern size and shape are compared with the contour and size of the candidate object. If the distance is a long distance, the size and shape of the whole body pattern is compared with the contour and size of the candidate object, and when each of the resulting determination output values exceeds a predetermined threshold, The candidate object is determined to be a pedestrian as a pedestrian's head or a pedestrian's head.

  The rough determination for determining whether or not the candidate object may be a pedestrian and the detailed determination for determining whether or not the candidate object is a pedestrian include a predetermined part ( For example, even if the head) pattern is the same, the same processing is not performed because the discrimination threshold value is different.

  Thus, depending on the distance to the candidate object, different determination patterns are used for determining whether the candidate object may be a pedestrian (for example, the detected candidate object grasped from the position on the image) If the distance between the vehicle and the vehicle is short, the pedestrian's head pattern can be used, and if the distance to the candidate object is far, the pedestrian's whole body pattern can be used), and the candidate object may be a pedestrian As a result of accurately determining whether or not there is a pedestrian, it is possible to efficiently execute the process of determining whether or not the user is a pedestrian.

(4) Detailed determination by weighting the decision output value, majority decision, etc. In addition, in the first embodiment, each decision output value output as a result of the determination on a plurality of predetermined parts constituting the pedestrian is weighted. (For example, when the distance is close, sufficient information is obtained and a determination output value obtained as a result of determination using a head pattern with a small individual difference is doubled, etc.) Whether the candidate object is detected as a pedestrian or not may be determined in detail. In addition, for example, if the determination result is determined as a pedestrian, ○, if not determined as a pedestrian, is output as x, etc., and the object detected from the input image as a determination candidate by the majority of the determination result, You may make it discriminate | determine in detail whether it is a pedestrian.

  For this reason, it is possible to improve the discrimination accuracy of the object detected from the image with a simple operation.

(5) Discrimination taking into account elements of operation In addition, in the first embodiment, the discriminant is a candidate for discrimination detected from a plurality of continuous images (for example, the current image and images of several frames input thereafter). For the same candidate object, the contour and size of the candidate object, and the pedestrian to be determined, for example, the shape and size of the forward hand pattern and the shape and size of the backward hand pattern A comparison may be made, and if the determination output value obtained as a result exceeds a predetermined determination threshold value, it may be determined that the object is a predetermined object.

  For this reason, for example, it is possible to determine whether an object detected from an image is a pedestrian by taking into account an element of motion, and further improve the accuracy of determining the object detected from the image Is possible.

(6) Change of threshold value according to predetermined conditions In the first embodiment, the determination threshold value used for determining the object is set as the travel location (the downtown area or the suburbs), the travel speed and the object detection position (the road position). You may make it change according to conditions, such as inside and outside. For example, if the travel location is a busy street or the vehicle travel speed is high, the determination threshold is changed to a low value.On the other hand, if the travel location is a suburb or the vehicle travel speed is slow, Change the discrimination threshold value higher.

  As a result, the threshold value can be changed according to a predetermined condition, and as a result, the discrimination accuracy of the object detected from the image can be further improved.

(7) Discrimination according to predetermined environmental conditions In the first embodiment, for example, conditions such as temperature, weather, and time zone such as a difference in clothes of pedestrians and ambient brightness in a time zone. A learning model learned in accordance with this may be prepared in advance, and the object determination unit 23b may execute the object determination process using an optimal learning model according to environmental conditions such as temperature, weather, and time zone.

  For this reason, the determination according to the environmental conditions such as the temperature, the weather, and the time zone can be executed. As a result, it is possible to further improve the determination accuracy of the object detected from the image.

(8) Determining whether or not to perform detailed determination according to the object detection position In the first embodiment, the position of the candidate object detected from the image as a determination candidate (for example, the road Whether or not to execute a detailed determination process for determining whether or not the candidate object is a pedestrian may be determined according to whether the candidate object is a pedestrian or not. For example, detailed discrimination processing is executed when an object is detected on the road, and detailed discrimination processing is not executed when an object is detected outside the road.

  For this reason, it is possible to prevent the detailed determination process from being performed wastefully.

(9) Determination of intrusion In addition, in the first embodiment, when it is determined that a candidate object that is a candidate for determination detected from the image is a pedestrian, is the pedestrian trying to enter the road from the image? It may be determined whether or not. For example, as shown in FIG. 9, when the position of the pedestrian in the current image is approaching the road in the next image, it is determined that the pedestrian is about to enter the road.

  For this reason, it is possible not only to discriminate the candidate object detected from the image as a pedestrian but also to determine whether or not it is about to enter the road.

(10) Detailed Discrimination is Performed with Image Input After Rough Discrimination In the first embodiment, candidate objects that are candidates for discrimination detected from within the input image are pedestrians within the input image of one frame. Although the case where it is determined whether or not there has been described, the present invention is not limited to this. For example, the determination of a detection detected from a plurality of continuous images (for example, the current image and the next image to be input next). For the same candidate object as a candidate, make a determination on a predetermined part constituting each separate pedestrian (for example, if a determination is made regarding the head in the current image, a determination is made regarding the limb in the next image) You may make it discriminate | determine whether it is a pedestrian.

  Specifically, as illustrated in FIG. 10, when an object that is a candidate for discrimination is detected from the current input image (see (1) in FIG. 5), the object discriminating unit 23b A determination (part determination) that compares the contour and size of the object with the size and shape of the pattern of the head that constitutes the pedestrian, for example, is performed to determine in detail whether or not the user is a pedestrian.

  Further, when the object determination unit 23b newly detects a candidate object in the vicinity of the object detection position of the current image that has been input up to the next image input next to the current image ((3 in FIG. 5). )), And a determination (part) that compares the contour and size of the newly detected candidate object with the size and shape of a pattern such as a limb that constitutes the pedestrian, for example, different from that used in the current image It is determined in detail whether or not it is a pedestrian (see (4) in FIG. 5). As a result, when it is determined that the person is a pedestrian in both the current image and the next image, the object detected from the image is determined to be a pedestrian.

  As a result, it is possible to further improve the discrimination accuracy of the object detected from the image as a result of judging whether or not the pedestrian is a pedestrian by making a judgment on each pedestrian's separate walking parts in a plurality of images. It is.

(11) Device Configuration The components of the vehicle control device 10 and the image recognition device 20 shown in FIG. 2 are functionally conceptual and need not be physically configured as illustrated. . That is, the specific form of dispersion / integration of the vehicle control device 10 and the image recognition device 20 is not limited to the illustrated one. For example, the notification control unit 15 and the vehicle control unit 16 are integrated, and the object determination unit 23b is roughly determined. Disperse all or part of the functions, such as processing functions and detailed discrimination processing functions, functionally or physically in arbitrary units according to various loads and usage conditions. Can be configured. Furthermore, the processing functions performed by the vehicle control device 10 and the image recognition device 20 (the vehicle control function of the vehicle control device 10, the object detection function and the pedestrian discrimination function of the image recognition device 20) are all or any one of them. The unit may be realized by a CPU and a program that is analyzed and executed by the CPU, or may be realized as hardware by wired logic.

  Each processing method described in the present embodiment (see FIG. 6) can be realized by executing a prepared program on a computer such as a personal computer or a workstation. This program can be distributed via a network such as the Internet. The program can also be executed by being recorded on a computer-readable recording medium such as a hard disk, a flexible disk (FD), a CD-ROM, an MO, and a DVD and being read from the recording medium by the computer.

  As described above, the image recognition device, the image recognition method, the pedestrian recognition device, and the vehicle control device according to the present invention perform the determination when an object that is a determination candidate is detected from an image captured by the camera. This is useful when discriminating an object as a candidate, and is particularly suitable for improving the discrimination accuracy of an object detected from an image.

FIG. 3 is a diagram for explaining an overview and features of the first embodiment. FIG. 3 is a diagram for explaining an overview and features of the first embodiment. FIG. 3 is a diagram for explaining an overview and features of the first embodiment. 1 is a block diagram illustrating configurations of an image recognition device and a vehicle control device according to a first embodiment. It is a figure which shows the structural example of a collision risk determination table. 4 is a flowchart illustrating a flow of processing of the image recognition device and the vehicle control device according to the first embodiment. It is a figure for demonstrating the outline | summary of the pedestrian site | part determination which concerns on Example 2. FIG. It is a figure for demonstrating the outline | summary of the pedestrian discrimination | determination based on Example 2. FIG. It is a figure for demonstrating the outline | summary of the pedestrian discrimination | determination based on Example 2. FIG. It is a figure for demonstrating the outline | summary of the pedestrian whole body determination which concerns on Example 2. FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Vehicle control apparatus 11 Imaging device 12 Navigation part 13 Radar part 14 In-vehicle notification part 15 Notification control part 16 Vehicle control part 17 Brake 18 Handle 20 Image recognition apparatus 21 Pre-processing part 21a Filter part 21b Contour extraction part 22 Storage part 22a Learning rule Storage unit 22b Learning model storage unit 22c Collision risk determination table 23 Control unit 23a Object detection unit 23b Object determination unit 23c Collision risk determination unit

Claims (14)

  When an object that is a candidate for determination is detected from an image captured by the camera, for the object that is a candidate for determination, a determination is made regarding a predetermined part that constitutes the predetermined object that is the target of determination, An image recognition apparatus comprising: an object discriminating unit that discriminates a predetermined object when a discrimination value obtained as a result exceeds a predetermined threshold value.   The object discriminating unit performs a judgment on a plurality of predetermined parts constituting a predetermined object to be discriminated for an object that is a discrimination candidate detected from the image, and each discrimination obtained as a result thereof The image recognition apparatus according to claim 1, wherein when the total value obtained by adding the values exceeds a predetermined threshold value, the image recognition apparatus determines that the object is the predetermined object.   The object discriminating unit weights one of the discriminant values, and discriminates that the object is the predetermined object when a total value obtained by adding the discriminant values exceeds a predetermined threshold value. The image recognition apparatus according to claim 2.   The object discriminating unit performs a judgment on the operation of the predetermined part for the same object that is a candidate for discrimination detected from a plurality of continuous images, and a discrimination value obtained as a result exceeds a predetermined threshold value. The image recognition apparatus according to claim 1, wherein the image recognition apparatus determines that the object is the predetermined object. The object that is a candidate for determination detected from the image further includes a possibility determination unit that determines a possibility that the object is a predetermined object to be determined.
The object determining means, when the possibility determining means determines that the object that is a candidate for determination is a possibility of the predetermined object, for the object that is a candidate for determination A determination is made regarding a predetermined part constituting the object, and when the determination value obtained as a result exceeds a predetermined threshold, it is determined that the object is the predetermined object. The image recognition apparatus according to 1.   The possibility determining means determines an object that may be the predetermined object by performing a determination on the predetermined part according to a position on the image of an object detected from the image as a candidate for the determination. The image recognition apparatus according to claim 5, wherein:   The said object discrimination | determination means discriminate | determines that the said object used as the candidate for a discrimination | determination is the said predetermined object using the threshold value according to predetermined conditions. The image recognition apparatus described.   The object discriminating unit performs a judgment on the predetermined part according to a predetermined environmental condition, and discriminates an object that is a candidate for the discrimination as the predetermined object. The image recognition apparatus as described in any one of.   The object discriminating unit determines whether or not to discriminate an object that is a candidate for discrimination as the predetermined object according to a position on the image of an object detected from the image as the discrimination candidate. The image recognition apparatus according to claim 1, wherein the image recognition apparatus is an image recognition apparatus.   When the object discriminating unit discriminates that an object that is a candidate for discrimination detected from the image is a predetermined object to be discriminated, the predetermined object will enter the road from the image. The image recognition apparatus according to claim 1, further comprising an intrusion determination unit that determines whether or not   The object determination means performs a determination regarding the predetermined part separately for the same object that is a candidate for determination detected from a plurality of continuous images, and a determination value obtained as a result exceeds a predetermined threshold value. The image recognition apparatus according to claim 1, wherein the image recognition apparatus determines that the object is the predetermined object.   When an object that is a candidate for determination is detected from an image captured by the camera, for the object that is a candidate for determination, a determination is made regarding a predetermined part that constitutes the predetermined object that is the target of determination, An image recognition method, comprising: an object discrimination step of discriminating that a given object is a predetermined object when a discrimination value obtained as a result exceeds a predetermined threshold.   When an object that is a candidate for discrimination is detected from an image captured by a camera, the object that is a candidate for discrimination is determined for a predetermined part that constitutes a pedestrian that is the target of the discrimination. A pedestrian recognition apparatus comprising object discrimination means for discriminating that a person is a pedestrian when a discrimination value obtained as a result exceeds a predetermined threshold.   When an object that is a candidate for determination is detected from an image captured by the camera, for the object that is a candidate for determination, a determination is made regarding a predetermined part that constitutes the predetermined object that is the target of determination, Control means for notifying a warning and / or controlling the vehicle according to the result of discrimination by the object discriminating means for discriminating that it is the predetermined object when the discriminant value obtained as a result exceeds a predetermined threshold value. A vehicle control device comprising the vehicle control device.

Images