JP2012068965A - Image recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an image recognition device with improved performance in detection of an detection object.SOLUTION: The image recognition device 100 detects an image 50a showing a pedestrian existing in a front area from a front view image 50 generated by an infrared camera 10. The infrared camera 10 generates a front view image of which contrast is adjusted by changing camera gain in response to brightness in the front area. Therefore, the image recognition device 100 includes a database 30 for storing a plurality of previous learning models 31-33 generated according to contrasts of sample images 61a-63a of a plurality of groups, an optimal model selection part 27 for selecting a previous leaning model 31, 32 or 33 corresponding to the camera gain from the plurality of previous leaning models 31-33 stored in the database 30, and a pedestrian recognition part 21 for detecting an image 50a indicating a pedestrian from the front view image 50 based on the selected previous learning model 34.

Description

  The present invention relates to an image recognition device that detects an image representing a detection target from an image generated by an imaging unit.

  2. Description of the Related Art Conventionally, an image recognition technique for detecting an image representing a detection target such as a human or a human face from an image generated by an imaging unit such as a camera is known (for example, Patent Document 1 and Patent Document 2).

  As one type of such technology, for example, Patent Literature 1 discloses an object detection device and a learning database generation device. The object detection device includes an image input device that captures an area around the vehicle, a detection dictionary storage device that stores a detection dictionary, and an image of the object based on the detection dictionary from an image captured by the image input device. And an identification device for detecting. On the other hand, the learning database generation device creates a detection dictionary stored in the above-described detection dictionary storage device using a plurality of images captured in advance. As described above, based on the detection dictionary created using a plurality of images, the identification device of the target object detection apparatus can accurately detect the target object image from the captured image.

  Further, a face recognition device disclosed in Patent Document 2 includes a camera for photographing a person's face, a dictionary information holding unit that stores a plurality of dictionary information, and a dictionary that selects one of the plurality of dictionary information. A switching unit; and a recognition unit that detects a face image based on dictionary information selected from an image captured by the camera. Furthermore, the face recognition device disclosed in Patent Document 2 includes an illumination environment determination unit that determines whether an image captured by a camera is captured in an illumination environment suitable for detecting a face image based on the brightness of the image. It has. The face recognition device adds the determination result by the illumination environment determination unit to the image as attribute information when storing the image from the camera as dictionary information in the dictionary information holding unit.

  With the configuration described above, in the face recognition device disclosed in Patent Document 2, the dictionary switching unit obtains the determination result of the illumination environment when the image is captured from the illumination environment determination unit, and dictionary information similar to the determination result is obtained. Select a plurality of dictionary information. Based on the dictionary information to which similar attribute information is added, the recognition unit can accurately detect the face of a person photographed by the camera.

JP 2007-272420 A JP 2005-84815 A

  Now, unlike the camera included in the face recognition device disclosed in Patent Document 2, an image input device such as a camera used in the object detection device disclosed in Patent Document 1 has an imaging gain according to the brightness of the surrounding area of the vehicle. Can be changed. This is because in the object detection device used in the vehicle, the brightness of the surrounding area easily changes as the vehicle travels. Even when the brightness of the surrounding area changes due to traveling of the vehicle, the camera can generate an image whose contrast is adjusted according to the brightness of the surrounding area by changing the imaging gain.

  When the camera as described above is used as the imaging unit, the contrast between the image representing the detection target in the peripheral image captured by the camera and the background image varies as the imaging gain is changed. In the configuration disclosed in Patent Document 1, when a detection dictionary is created using images with various contrasts, the capacity of the detection dictionary increases in order to maintain detection accuracy. Then, the identification device has to perform detection based on the detection dictionary having an increased capacity. Therefore, there is a possibility that the detection of the object cannot be performed at high speed.

  Therefore, in order to reduce the capacity of the detection dictionary, consider a form in which an image from a camera is classified into a plurality of dictionary information based on the brightness of the image, as in the configuration of Patent Document 2, for example. However, the brightness of the image does not correspond to the contrast between the image representing the detection target and the background image. For example, even if the images have the same brightness as a whole, there are an image obtained by imaging a dark peripheral area with a high imaging gain and an image obtained by imaging a bright peripheral area with a low imaging gain. In the former, the contrast between the image representing the detection object and the background image is high, but in the latter, the contrast is low.

  Therefore, even if the dictionary switching means selects dictionary information having attribute information similar to the brightness information of the image, the contrast of the image and the contrast of the image included in the dictionary information selected by the dictionary switching means , Do not correspond to each other. Thus, since the dictionary information corresponding to the contrast of the peripheral image cannot be used, the recognition unit cannot detect the image of the target object from the image with high accuracy.

  The present invention has been made in view of the above problems, and an object thereof is to provide an image recognition apparatus with improved detection performance of a detection target.

  In order to achieve the above object, according to the first aspect of the present invention, the image recognition for detecting an image representing a detection target existing in the peripheral area from the peripheral image generated by the imaging means for capturing the peripheral area of the vehicle. An imaging unit that generates a peripheral image in which contrast is adjusted by changing an imaging gain according to brightness of a peripheral region, and according to contrast of a plurality of groups of sample images serving as a learning reference A storage unit that stores the plurality of pre-learning models that have been created, and an imaging gain when the peripheral image is captured from the imaging unit, and a plurality of pre-learning models that correspond to the imaging gain are stored in the storage unit. Based on the model selection means to select from the pre-learning model and the pre-learning model selected by the model selection means, the peripheral image captured by the imaging means Comprises a detection means for detecting an image representing the object to be detected, the.

  According to the present invention, in the imaging unit, the contrast of the peripheral image is adjusted by changing the imaging gain, so the information on the imaging gain is an index indicating the contrast of the peripheral image. The selection means selects a pre-learning model corresponding to the imaging gain information. Therefore, the contrast of the surrounding image and the contrast of the sample image group used when creating the pre-learning model selected by the model selection unit can correspond to each other. As described above, based on the pre-learning model corresponding to the contrast of the peripheral image, the detection unit can detect the image representing the detection target object from the peripheral image with high accuracy.

  In addition, each pre-learning model is created according to the contrast of the sample images of each sample image group. Since the contrast of the sample images included in each sample image group is approximate, a pre-learning model that can maintain detection accuracy while suppressing the number of sample images for each sample image group can be constructed. Thus, the capacity | capacitance of a prior learning model is suppressed by suppressing the number of sample images. Therefore, the detection means can perform processing for detecting an image representing the detection target at high speed.

  As a result, the image recognition apparatus can detect an image representing the detection target with high accuracy and high speed. Therefore, the detection performance of the detection target by the image recognition device is improved.

  In the invention according to claim 2, the selection unit generates a plurality of pre-learning models created from a sample image group having a high contrast among the plurality of sample image groups as the imaging gain acquired from the imaging unit increases. The detection unit is selected from the learning model, and the detection unit detects an image representing the detection target from the peripheral image based on a pre-learning model created using a group of sample images with high contrast.

  According to the present invention, when the peripheral area is dark, the imaging unit increases the imaging gain. As a result, a peripheral image with high contrast is generated. Therefore, as the imaging gain increases, it is used to create the selected pre-learning model by selecting the pre-learning model created using a group of sample images with high contrast. The contrast of the sample image group can correspond with certainty. Based on a pre-learning model created using a high-contrast sample image group corresponding to the high contrast of the peripheral image, the detection unit can detect the image representing the detection target object from the peripheral image with high accuracy. It becomes like this.

  In the invention according to claim 3, each sample image group includes an assumed object image in which an assumed object assumed as a detection object is captured and an unimagined object image in which an unexpected object other than the assumed object is captured, and each pre-learning The model is constructed by learning features of an image representing an assumed object based on the assumed object image and the non-assumed object image.

  According to the present invention, when the pre-learning model learns the characteristics of the image representing the assumed object assumed as the detection target object, the assumed object image in which the assumed object is captured and the non-assumed image in which the non-assumed object other than the assumed object is captured. An object image is used. Thereby, the prior learning model can learn that the assumed object appears in the assumed object image, and can learn that the non-assumed object reflected in the unexpected object image is not the assumed object. By the above learning, based on the pre-learning model in which the difference between the assumed object and the unforeseen object is learned, the detection unit can reliably detect and detect the image representing the detection object from the surrounding images. .

  According to a fourth aspect of the present invention, the image pickup means includes an infrared camera that generates a peripheral image by detecting infrared light in the peripheral region.

  When an infrared camera is provided as the imaging means as in the present invention, the imaging gain of the infrared camera changes remarkably due to infrared rays emitted from street lights and headlights of oncoming vehicles. Therefore, in the peripheral image, the contrast between the image representing the detection target and the background image varies significantly. However, in the configuration described above, based on the pre-learning model corresponding to the contrast of the peripheral image, the detection unit can detect the image representing the detection target object from the peripheral image with high accuracy. Thus, the effect of improving the detection accuracy by being based on a plurality of pre-learning models created according to the contrast of a plurality of sample image groups is remarkably exhibited in the form in which the peripheral image is generated by the infrared camera. It is.

1 is a configuration diagram schematically illustrating a configuration of an image recognition apparatus according to an embodiment of the present invention. It is a figure for demonstrating the process for an image recognition circuit to detect the image showing a pedestrian from a front image. It is a figure for demonstrating a front image in case a camera gain is high, Comprising: (a) is a figure in which a front image is shown, (b) And (c) is an image showing the pedestrian reflected in a front image. It is a figure shown. It is a figure for demonstrating a front image in case a camera gain is medium, Comprising: (a) is a figure in which a front image is shown, (b) is an enlarged image showing the pedestrian reflected in a front image. FIG. It is a figure for demonstrating a front image when a camera gain is low, Comprising: (a) is a figure in which a front image is shown, (b) is expanded and the image showing the pedestrian reflected in a front image is shown. FIG. It is a figure for demonstrating a prior learning model. It is a flowchart in which the process of the image recognition circuit which detects the image showing a pedestrian from a front image is shown.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

  FIG. 1 shows a configuration of an image recognition apparatus 100 according to an embodiment of the present invention. The image recognition apparatus 100 detects an image 50a representing a detection target such as a pedestrian existing in the front area from the front image 50 obtained by imaging the front area of the vehicle as shown in FIG.

  As shown in FIG. 1, the image recognition apparatus 100 includes an infrared camera 10, an image recognition circuit 20, and the like. The image recognition device 100 is connected to the display device 40 and outputs a detection result to the display device 40.

  The infrared camera 10 can pick up an image even in an environment with little visible light by detecting near infrared rays. The infrared camera 10 is installed so that the traveling direction of the vehicle is directed, and can capture particularly the front region of the vehicle in the peripheral region of the vehicle. The infrared camera 10 captures a front area of the vehicle to generate a front image 50 as a peripheral image. Then, the infrared camera 10 sequentially outputs the generated front image 50 to the image recognition circuit 20 and the display device 40.

  The image recognition circuit 20 includes a processor that performs various arithmetic processes, a flash memory that stores programs used for the arithmetic processes, a RAM that functions as a calculation work area, and the like. In addition, the image recognition circuit 20 has an interface that is connected to the infrared camera 10 and acquires information from the infrared camera 10.

  The image recognition circuit 20 includes a camera gain acquisition unit 25, a database 30, an optimal model selection unit 27, and a pedestrian recognition unit 21. These elements 25, 30, 27, and 21 are functional blocks of the image recognition circuit 20. The functions of these elements 25, 30, 27, and 21 are performed by the image recognition circuit 20 when a predetermined program is executed by the processor of the image recognition circuit 20.

  The camera gain acquisition unit 25 acquires a camera gain from the infrared camera 10. The camera gain acquisition unit 25 acquires, from the infrared camera 10, information on camera gain when the front image 50 output from the infrared camera 10 to the image recognition circuit 20 is captured by the infrared camera 10. The acquired camera gain information is acquired by the optimal model selection unit 27 through the camera gain acquisition unit 25.

  The database 30 stores a plurality of prior learning models 31 to 33. These pre-learning models 31 to 33 are obtained by learning the characteristics of images of pedestrians and the like, and are created according to the contrast of a plurality of groups of sample images 61a to 63a serving as a learning reference by a method described later. ing. The database 30 is connected to the optimum model selection unit 27, and in response to a request from the optimum model selection unit 27, one of the stored prior learning models 31 to 33 is selected as the optimum model selection unit. 27.

  The optimal model selection unit 27 selects a pre-learning model corresponding to the camera gain acquired through the camera gain acquisition unit 25 from the plurality of pre-learning models 31 to 33 stored in the database 30. The optimum model selection unit 27 selects a pre-learning model created from a plurality of sample image groups having a high contrast from the plurality of pre-learning models 31 to 33 as the acquired camera gain increases. Specifically, in the optimum model selection unit 27, two camera gain reference values for selecting a pre-learning model are set in advance. For convenience, of the two reference values, the higher one is the first reference value and the lower one is the second reference value. The optimal model selection unit 27 selects the pre-learning model 31 when the acquired camera gain is higher than the first reference value. The optimal model selection unit 27 selects the pre-learning model 32 when the acquired camera gain is equal to or lower than the first reference value and higher than the second reference value. The optimal model selection unit 27 selects the pre-learning model 33 when the acquired camera gain is equal to or less than the second reference value. The prior learning model 34 selected by the optimal model selection unit 27 is read from the database 30 and output to the pedestrian recognition unit 21.

  The pedestrian recognition unit 21 is connected to the infrared camera 10, the optimum model selection unit 27, and the display device 40. The pedestrian recognition unit 21 acquires a front image 50 captured by the infrared camera 10. The pedestrian recognition unit 21 acquires the pre-learning model 34 selected by the optimal model selection unit 27. The pedestrian recognition unit 21 outputs a detection result of a detection target such as a pedestrian to the display device 40.

  The pedestrian recognition unit 21 includes a feature amount extraction unit 22 and a collation unit 23.

  The feature amount extraction unit 22 scans the front image 50 and extracts feature amount data of an object shown in the front image 50. Specifically, as illustrated in FIG. 2, the feature amount extraction unit 22 scans the front image 50 acquired from the infrared camera 10 using a search window 55 having a predetermined size. The feature amount extraction unit 22 moves the search window 55 in the horizontal direction and the vertical direction of the front image 50 (see the dashed arrows shown in FIG. 2), and the image 56 in the range surrounded by the search window 55. (Hereinafter, window image 56) is cut out in order. Further, the feature amount extraction unit 22 converts the cut window image 56 into feature amount data suitable for detection processing. The feature amount extraction unit 22 sequentially outputs the converted feature amount data to the collation unit 23.

  The matching unit 23 detects an image 50 a representing a pedestrian from the front image 50 based on the pre-learning model 34 selected by the optimal model selection unit 27. Specifically, the collation unit 23 collates the feature amount data extracted by the feature amount extraction unit 22 with the selected pre-learning model 34, and the feature amount data is learned by the selected pre-learning model 34. It is determined whether or not it matches the feature amount data of the image 50a representing the pedestrian. When an object that matches the feature amount data learned by the selected pre-learning model 34 is detected, the collation unit 23 displays the position information of the object in the front image 50 as a detection result, as shown in FIG. Output to the device 40.

  The display device 40 provides various information to the driver by displaying various images on a display screen of a liquid crystal display disposed at the center of the instrument panel in the vehicle compartment. The display device 40 of the present embodiment is connected to the infrared camera 10 and can acquire the front image 50 from the infrared camera 10. The display device 40 displays the front image 50 on the display screen based on an operation of a user such as a driver. In addition, the display device 40 superimposes a frame-shaped image on the front image 50 so as to surround the image 50 a representing the pedestrian in the front image 50 based on the position information as the detection result acquired from the pedestrian recognition unit 21. Are displayed (see FIG. 3A, etc.).

  Next, the front image 50 captured by the infrared camera 10 and output to the pedestrian recognition unit 21 will be described in more detail based on FIGS.

  The infrared camera 10 (see FIG. 1) can change the camera gain according to the brightness of the front area. The camera gain indicates the amount of amplification when the intensity of light is converted into an electrical signal in the infrared camera 10. The camera gain indicates the sensitivity of the infrared camera 10.

  FIG. 3 shows a front image 51 that is captured when the front area is dark, for example, when a vehicle is traveling on a streetlight and a road with a small amount of traffic. When the front area is dark, the infrared camera 10 adjusts the contrast of the front image 51 by increasing the camera gain. In this case, the contrast between the image 51a representing the pedestrian in the front image 51 (see FIGS. 3B and 3C) and the background image is high.

  FIG. 4 shows a front image 52 that is captured when the front region is brighter than the state shown in FIG. 3, such as when a vehicle is traveling on a streetlight and a road with a relatively large traffic volume. When the front area becomes brighter, the infrared camera 10 adjusts so that strong halation does not occur in the front image 52 due to light emitted from the streetlight and light emitted from the headlight of the oncoming vehicle by suppressing the camera gain. To do. In this case, the contrast between the image 52a representing the pedestrian shown in the front image 52 (see FIG. 4B) and the background image is lower than the image 51a representing the pedestrian when the front area is dark. Become.

  FIG. 5 shows a front image 53 in a case where the front area is brighter than the state shown in FIG. 4, such as when a vehicle is traveling on a streetlight and a road with a large amount of traffic. When the front area is bright, the infrared camera 10 reduces the camera gain so that strong halation does not occur in the front image 53 due to light emitted from the streetlight and light emitted from the headlight of the oncoming vehicle. Adjust. In this case, the contrast between the image 53a representing the pedestrian in the front image 53 (see FIG. 5B) and the background image is lower than that of the image 52a representing the pedestrian described above.

  Next, the pre-learning models 31 to 33 stored in the database 30 will be described in detail with reference to FIG.

  As shown in FIG. 6, the pre-learning models 31 to 33 are created by the learning model creation device 70. In order to create the pre-learning models 31 to 33, a large number of sample images 61a to 63a are used. Many sample images 61a-63a are classified into a plurality of sample image groups 61-63 according to the contrast of each sample image. In the present embodiment, the sample images 61a to 63a correspond to the number of pre-learning models, and the sample image group 61 having a high contrast, the sample image group 62 having a medium contrast, and the sample image group 63 having a low contrast. It is classified into one image group.

  In each of the sample image groups 61 to 63, an assumed object that is assumed as a detection target, that is, an assumed object image 64 in which a pedestrian is photographed, and an unimagined object image in which only a supposed sad object other than a pedestrian is photographed (not shown). It is included. Specifically, non-assumed objects are vehicles other than pedestrians, traffic lights, signs, and the like. Each of the prior learning models 31 to 33 is constructed by learning the characteristics of the image 50a representing the pedestrian that is the detection target based on the assumed object image and the unforeseen object image. The images shown in FIGS. 6A to 6C are a part of the assumed object image.

  The learning model creation device 70 reads the sample images and scans each sample image by the search window 55 (see FIG. 2) in the same manner as the pedestrian recognition unit 21. Thereby, the learning model creation apparatus 70 cuts out the window image 56 (see FIG. 2) from each sample image. And the feature-value data of the cut-out window image 56 are learned. In the learning model creation apparatus 70 according to the present embodiment, the Adaboost algorithm is used as an algorithm for learning feature data.

  The learning model creation device 70 causes the pre-learning model 31 to learn the features of the image 50a representing the pedestrian by weighting the feature amount data. Specifically, the learning model creation device 70 increases the weighting of the feature amount data extracted from the assumed object image 64. On the other hand, the learning model creation apparatus 70 reduces the weighting of the feature amount data that is extracted from the unexpected image. In this way, by reading a large number of assumed object images 64 and non-assumed object images, the learning model creation apparatus 70 shows a strong reaction with respect to an image in which a pedestrian is shown, and an image in which no pedestrian is shown For the discriminator, pre-learning models 31 to 33 are created that show no response or only a weak response.

  When these pre-learning models 31 to 33 are created, target recognition rates and erroneous recognition rates are set in advance. The recognition rate is a probability that the prior learning model can correctly recognize the assumed object image 64 as the assumed object image 64. The misrecognition rate is a probability that an unexpected image is erroneously recognized as an assumed image 64. Features of the image 50a representing a pedestrian by causing the learning model creation device 70 to read the sample images 61a to 63a until the recognition rate set as the target exceeds the recognition rate set as the target and falls below the false recognition rate set as the target. The pre-learning models 31 to 33 that learn the above are created.

  The pre-learning models 31 to 33 are created by causing the learning model creation device 70 described so far to read the classified sample image groups 61 to 63. The pre-learning model 31 selected when the contrast between the image 50a representing the pedestrian and the background image is high due to the high camera gain is generated from the sample image group 61 with high contrast. From the sample image group 62 having a medium contrast, a pre-learning model 32 that is selected when the camera gain is medium is generated. From the low-contrast sample image group 63, a pre-learning model 33 that is selected when the contrast between the image 50a representing the pedestrian and the background image is low due to low camera gain is generated. As described above, when the camera gain is high, the collation unit 23 can detect the image 50a representing the pedestrian from the front image 50 based on the pre-learning model 31 created using the sample image group 61 with high contrast. it can. On the other hand, when the camera gain is low, the collation unit 23 can detect the image 50a representing the pedestrian from the front image 50 based on the pre-learning model 33 created using the sample image group 63 with low contrast.

  The target values of the recognition rate and the misrecognition rate set when generating the above-described pre-learning models 31 to 33 may be different for each of the pre-learning models 31 to 33. The recognition rate may be set to a higher value as the contrast of the sample image for creating the pre-learning model becomes higher. On the other hand, the false recognition rate may be set to a lower value as the contrast of the sample image for creating the pre-learning model becomes lower.

  If the contrast of the sample image is high, the outline between the image 50a representing the pedestrian and the background image becomes clear. Therefore, even if the target value of the recognition rate is set high or the target value of the erroneous recognition rate is set low, the pre-learning model 31 that achieves these target values can be created.

  On the other hand, when the contrast of the sample image is low, the outline between the image 50a representing the pedestrian and the background image becomes ambiguous. Therefore, if the target value of the recognition rate is too high or the target value of the erroneous recognition rate is too low, it may be difficult to create a pre-learning model. In addition, even when a pre-learning model is created, the capacity of the pre-learning model 33 increases. Therefore, when the contrast of the sample image is low, the target values of the recognition rate and the erroneous recognition rate should be relaxed as compared with the case where the contrast of the sample image is high.

  Next, the process in which the image recognition apparatus 100 detects the image 50a representing the pedestrian from the front image 50 will be described in detail with reference to FIG. The process described below is performed by the image recognition circuit 20 when the user performs an operation for displaying the front image 50 on the display device 40, for example. The processing by the image recognition circuit 20 is repeated until the user performs an operation for stopping the display of the front image 50 on the display device 40.

  In S101, the pedestrian recognition unit 21 acquires the front image 50 from the infrared camera 10, and the process proceeds to S102. In S102, the camera gain acquisition unit 25 acquires the camera gain from the infrared camera 10, and the process proceeds to S103.

  In S103, the pre-learning model corresponding to the camera gain acquired in S102 is selected by the optimum model selection unit 27, and the process proceeds to S104. The prior learning model 34 selected by the optimal model selection unit 27 is output to the collation unit 23 of the pedestrian recognition unit 21.

  In S <b> 104, the window image 56 is cut out in order from the front image 50 by the search window 55. Then, the feature amount extraction unit 22 extracts feature amount data from each clipped window image 56, and the process proceeds to S105. The extracted feature amount data is output from the feature amount extraction unit 22 to the collation unit 23.

  In S105, the feature amount data extracted in S104 is collated with the feature amount data of the image 50a representing the pedestrian learned by the pre-learning model 34 selected in S103, and the process proceeds to S106. In S105, when the window image 56 suitable for the feature amount data of the image 50a representing the pedestrian is detected, the position information of the window image 56 in the front image 50 is stored.

  In S <b> 106, it is determined whether or not an image 50 a representing a pedestrian that is a detection target is detected from the front image 50 based on the collation result in S <b> 105. In S106, when it is determined that the image 50a representing the pedestrian is not detected, the process returns to S101. On the other hand, when it determines with detecting the image 50a showing the pedestrian from the front image 50 in S106, it progresses to S107.

  In S107, the positional information of the window image 56 memorize | stored in S105 is output to the display apparatus 40 as a detection result, and it returns to S101. Based on the position information acquired from the collation unit 23, the display device 40 displays a display in which the image 50a representing the pedestrian is surrounded by a frame-shaped image on the display screen.

  In the present embodiment described so far, the camera gain information is an index indicating the contrast of the front image 50. Therefore, the contrast of the front image 50 and the contrast of the sample image group used when creating the pre-learning model 34 selected according to the camera gain can correspond to each other. Thus, based on the pre-learning model 34 corresponding to the contrast of the front image 50, the pedestrian recognition unit 21 can detect the image 50a representing the pedestrian as the detection target from the front image 50 with high accuracy. become.

  In addition, each pre-learning model 31-33 is created from each sample image group 61-63 classified according to contrast. Since the contrast of the sample images 61a to 63a included in each of the sample image groups 61 to 63 is approximate, the prior learning model 31 that can maintain the detection accuracy while suppressing the number of sample images for each of the sample image groups 61 to 63. ~ 33 can be constructed. Thus, the capacity | capacitance of the prior learning models 31-33 is suppressed by suppressing the number of sample images 61a-63a. Therefore, the collation part 23 can perform the process for detecting the image 50a showing a pedestrian at high speed.

  Accordingly, the image recognition apparatus 100 can detect the image 50a representing the pedestrian with high accuracy and high speed. Therefore, the detection performance of the detection target by the image recognition apparatus 100 is improved.

  In addition, in the present embodiment, the optimum model selection unit 27 selects a pre-learning model created using the sample image group 61 having a high contrast as the camera gain increases. Since the contrast of the front image 50 increases as the camera gain increases, the contrast of the front image 50 and the contrast of the sample image group 61 used to create the selected pre-learning model 34 can reliably correspond. And when the contrast of the front image 50 is high, it is based on the pre-learning model 31 created using the sample image group 61 having a high contrast. Based on the pre-learning model 33 created using the image group 63, the pedestrian recognition unit 21 can detect the image 50a representing the pedestrian from the front image 50 with high accuracy.

  Further, in the present embodiment, when the pre-learning models 31 to 33 learn the characteristics of the image 50a representing the pedestrian assumed as the detection target object, the assumed object image 64 and the unexpected object image are used. Thus, each of the prior learning models 31 to 33 can learn that a pedestrian is reflected in the assumed object image 64 and can learn that what is reflected in the non-imaginary object image is not a pedestrian. Based on each learning model 31-33 which learned the difference between a pedestrian and a thing other than a pedestrian by the above learning, the pedestrian recognition part 21 displays the image 50a showing a pedestrian from the front image 50, It will be possible to reliably select and detect.

  Furthermore, in the present embodiment, the pre-learning model 31 created using the sample image group 61 having a high contrast is recognized at the time of model creation than the pre-learning model 33 created using the sample image group 63 having a low contrast. Target values for rate and false recognition rate are strictly set. Therefore, when the camera gain is high, the pedestrian recognition unit 21 can reliably detect the image 51a representing the pedestrian based on the pre-learning model 31 that has achieved a strict target value. On the other hand, when the camera gain is low, the pedestrian recognition unit 21 can detect the image 53a representing the pedestrian at high speed based on the pre-learning model 33 whose capacity is suppressed by the relaxation of the target value. As described above, the image recognition apparatus 100 can achieve both the accuracy of detection and the speed of processing.

  Furthermore, the camera gain of the infrared camera 10 used in the present embodiment changes significantly due to infrared rays radiated from street lights and headlights of oncoming vehicles. Therefore, in the front image 50, the contrast between the image 50a representing the pedestrian and the background image also varies significantly. However, in the configuration described above, the pedestrian recognition unit 21 can detect the image 50a representing the pedestrian from the front image 50 with high accuracy based on the pre-learning models 31 to 33 corresponding to the contrast of the front image 50. become. As described above, the action of improving the detection accuracy based on the plurality of pre-learning models 31 to 33 created according to the contrast of the plurality of sample image groups 61 to 63 generates the front image 50 by the infrared camera 10. In the form, it is exhibited remarkably.

  In the present embodiment, the infrared camera 10 corresponds to the “imaging means” recited in the claims, the pedestrian recognition unit 21 corresponds to the “detection means” recited in the claims, and obtains the camera gain. The unit 25 and the optimum model selection unit 27 correspond to “model selection means” recited in the claims, the database 30 corresponds to “storage means” recited in the claims, and the front images 50 to 53 represent patents. The image 50a to 53a representing a pedestrian corresponds to the “image representing the detection target” described in the claims, and the camera gain corresponds to the “peripheral image” recited in the claims. This corresponds to the “imaging gain” described.

(Other embodiments)
As mentioned above, although one embodiment by the present invention was described, the present invention is not interpreted limited to the above-mentioned embodiment, and can be applied to various embodiments within the range which does not deviate from the gist.

  In the said embodiment, the optimal model selection part 27 selected one from the three prior learning models 31-33 produced according to the contrast of the sample images 61a-63a. However, the number of pre-learning models stored in the database 30 is not limited to three. For example, the number of pre-learning models stored in the database 30 may be two, or four or more.

  In the above embodiment, the optimal model selection unit 27 selects a pre-learning model corresponding to the acquired camera gain. However, the optimal model selection unit 27 may select the pre-learning model based on other information together with the camera gain. For example, the way in which the image 50a representing the pedestrian in the front image 50 is reflected may change depending on the weather conditions around the vehicle such as rain and snow. Therefore, the image recognition apparatus is provided with weather information acquisition means that is connected to a communication device that can communicate with the outside of the vehicle, such as a mobile communication network, or a rain sensor and acquires weather information. With this weather information acquisition means, the image recognition device can acquire rainfall information, snow cover information, and the like. Further, the learning model creation device 70 creates a pre-learning model corresponding to the weather information as well as the camera gain, and stores it in the database 30. With the above configuration, the image recognition apparatus can accurately detect an image 50a representing a pedestrian from the front image 50 even under various weather conditions.

  In the above embodiment, the Adaboost algorithm is used as an algorithm for causing each of the pre-learning models 31 to 33 to learn the feature amount data of the image 50a representing the pedestrian. However, various methods may be used to cause each of the prior learning models 31 to 33 to learn feature amount data. For example, a feature amount learning method using a neural network may be used to create the prior learning model.

  In the above embodiment, the target values of the recognition rate and the misrecognition rate when creating each pre-learning model 31 are changed according to the contrast of each of the sample image groups 61 to 63. However, these recognition rates and erroneous recognition rates may be constant values regardless of the contrast of the sample image groups 61 to 63.

  In the above embodiment, the infrared camera 10 is used as an imaging unit for capturing the front image 50. However, the imaging means is not limited to the infrared camera 10 as long as the front image 50 can be captured. For example, a visible light camera that captures an image using visible light may be an imaging unit. Alternatively, the imaging means may have both an infrared camera and a visible light camera.

  In the above embodiment, the image recognition device 100 detects a pedestrian as a detection target. However, the detection target detected by the image recognition device is not limited to a pedestrian. For example, a person who rides on a vehicle, a bicycle or a two-wheeled vehicle, an animal such as a deer or a frog may be set as the detection target.

DESCRIPTION OF SYMBOLS 10 Infrared camera (imaging means), 20 Image recognition circuit, 21 Pedestrian recognition part (detection means), 22 Feature amount extraction part, 23 Collation part, 25 Camera gain acquisition part (model selection means), 27 Optimal model selection part ( Model selection means), 30 database (storage means), 31-34 pre-learning model, 40 display device, 50-53 forward image (peripheral image), 50a-53a image representing a pedestrian, 55 search window, 56 window image, 61 to 63 sample image group, 61a to 63a sample image, 64 assumed object image, 70 learning model creation device, 100 image recognition device

Claims (4)

An image recognition device that detects an image representing a detection target existing in the peripheral area from a peripheral image generated by an imaging unit that captures the peripheral area of the vehicle,
The imaging means for generating the peripheral image in which the contrast is adjusted by changing the imaging gain according to the brightness of the peripheral area;
Storage means for storing a plurality of pre-learning models created according to the contrast of a plurality of groups of sample images serving as a learning reference;
A model for acquiring the imaging gain when the peripheral image is captured from the imaging unit, and selecting the pre-learning model corresponding to the imaging gain from the plurality of the pre-learning models stored in the storage unit A selection means;
Detection means for detecting an image representing the detection target object from the peripheral image captured by the imaging means based on the prior learning model selected by the model selection means;
An image recognition apparatus comprising: The selection unit is configured to convert the pre-learning model created from the sample image group having a high contrast among the plurality of sample image groups to a plurality of the advance learning models as the imaging gain acquired from the imaging unit increases. Select from the learning model,
The detection unit detects an image representing the detection target from the peripheral image based on the pre-learning model created using the sample image group having a high contrast. The image recognition apparatus described. Each of the sample image groups includes an assumed object image in which an assumed object assumed as the detection object is captured and an unimagined object image in which an unimagined object other than the assumed object is captured,
Each of the pre-learning models is constructed by learning a feature of an image representing the assumed object based on the assumed object image and the non-imaginary object image. The image recognition apparatus described.   The image recognition apparatus according to claim 1, wherein the imaging unit includes an infrared camera that generates the peripheral image by detecting infrared light in the peripheral region.