JP2013093639A - Vehicle periphery monitoring device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a vehicle periphery monitoring device capable of improving recognition accuracy without repetitive erroneous detection and non-detection, by learning data on an occurrence factor of the erroneous detection or non-detection occurring in image recognition.SOLUTION: In the case of the erroneous detection or non-detection of an object by an object detection processor 3 with respect to an image inputted by imaging means 1, user input means 6 designates the corresponding region on output means 7. An image recognition DB correction unit 4 corrects dictionary data on an erroneous detection region and stores the corrected data in dictionary data in an image recognition DB 5.

Description

  The present invention relates to a vehicle surrounding monitoring apparatus that detects a person or other obstacle existing around a host vehicle with an in-vehicle camera, and prevents erroneous detection or non-detection of the same target object.

  In order to avoid a collision with a person while the vehicle is running, a device has been developed in which a camera is installed on the vehicle and a person around the vehicle is recognized by image recognition. For example, there is one using pattern matching. In this method, various characteristics of a person are prepared in advance as a database. In the human detection process, if the area of the input image cut out from the captured image is similar to the image stored in the database, there is a person in that area, otherwise there is no person. The method of judging the presence or absence of a person with the idea of, is disclosed. (See Patent Document 1).

  Also disclosed is a method for detecting a person by combining an object detected using a radar and information on an image obtained by a camera for the object, and discriminating a person or an object other than a person. (See Patent Document 2).

JP 2009-070344 A JP 2005-157765 A

  However, in Patent Document 1, the method for increasing the accuracy of human recognition is to collect as many types of human image data as possible. However, in this method, the database becomes enormous and the cost increases. Since problems such as a large amount of processing occur, data registration in the database is limited.

  It is difficult to distinguish people from all backgrounds (structures, road textures, fences, trees, signs, etc.) and people against a database with a finite amount of data that can be registered. There is a case where a non-region is erroneously detected as a person, or a person who has melted into the background cannot be detected and is not detected.

  Further, Patent Document 2 is a method for more reliably detecting a person by using a plurality of sensing devices such as a camera and a radar. Even false detection cannot be eliminated. For example, when the shape of a tree is visible to a person, it is detected by a radar and also detected by a camera as a person. Furthermore, there is a problem that the number of sensing devices is increased and the cost is increased.

  In addition, since such erroneous detection performs pattern matching on a sensing result with a certain threshold value process, there is a problem that a pattern that has been erroneously detected once is repeatedly erroneously detected.

The present invention has been made to solve the conventional problems. For a target area that has been erroneously detected or not detected, the user specifies the target area on the screen and corrects and registers the image database. And it aims at providing the vehicle periphery monitoring apparatus which can improve detection accuracy so that a false detection and a non-detection may not be repeated.

  The present invention provides an input means for inputting a false detection region or a non-detection region when the user views the output means and determines that it is false detection or non-detection, and a false detection region input by the input means. Alternatively, an image recognition database correction unit for correcting dictionary data of the image recognition database in the non-detection area is stored, and the dictionary data corrected by the image recognition database correction unit is stored in the image recognition database.

  In the vehicle periphery monitoring apparatus of the present invention, when the output result of the object detection processing unit is erroneously detected or not detected, the input means inputs the erroneously detected region or the undetected region, and the image recognition database correcting unit Corrects the false detection area or non-detection area image recognition database dictionary data input by the input means and stores it in the image recognition database, so that the same target object can be erroneously detected or not detected. Repeat can be prevented.

The block diagram which shows the structure of the vehicle periphery monitoring apparatus in one embodiment of this invention The figure which shows the example of camera installation in the same device The figure which shows the picture which is photographed with the camera in the same device The figure which shows the detection result of person detection in the same device The figure which shows the false detection result of person detection in the same device Flow chart of operation explanation in the same device The figure which shows the outline of human detection in the same device The figure which shows the designation method of the false detection area in the same device The figure which shows dictionary data in the same device The figure which shows the example of correction of the dictionary data at the time of erroneous detection in the same apparatus The figure which shows the designation method of the non-detection area in the same device The figure which shows the example of correction of the dictionary data at the time of non-detection in the same apparatus

  Hereinafter, a vehicle periphery monitoring device according to an embodiment of the present invention will be described with reference to the drawings.

  FIG. 1 is a block diagram showing a configuration of a vehicle periphery monitoring device according to an embodiment of the present invention.

  In the present embodiment, an example in which the detection target is assumed to be a person will be described below.

  The imaging means 1 images the situation around the vehicle. The image pickup means 1 has a CCD camera or a CMOS camera as a device, and is installed in the vicinity of the license plate or emblem at the rear of the vehicle, or at the top to pick up the rear of the vehicle, or on the side mirror and on the side of the vehicle. There is a side camera that images the vehicle, or a front camera that images the front of the vehicle from the vicinity of an emblem in front of the vehicle or from the rear of the rear-view mirror inside the front window.

  The signal processing means 2 includes an object detection processing unit 3 and an image recognition DB (database) correction unit 4.

  The target object detection processing unit 3 refers to a feature amount representing a person registered in the image recognition DB (database) 5, and in the image data captured by the imaging unit 1, there is an area close to the feature amount of the person. If there is, that portion is detected as a human area.

  The image recognition DB correction unit 4 extracts the feature amount of the image for the image region designated by the user input means 6 and updates the content of the image recognition DB 5.

  The image recognition DB 5 is dictionary data obtained by extracting feature values from various types of human image data and creating a database.

  The user input means 6 is detected by the object detection processing unit 3 and output to the output means 7 in the case of erroneous detection in which an object other than a person is extracted as a person, or a person's unique posture is the background. In the case of non-detection that cannot be detected as a person, the user designates an erroneous detection / non-detection area on the screen in order to acquire data for updating the image recognition DB 5.

  The output means 7 superimposes and displays the result detected by the object detection processing unit 3 on the image data input from the imaging means 1. As the device, for example, a navigation monitor, a monitor in the instrument panel, This is a monitor built in the room mirror.

  The case where the vehicle periphery monitoring device configured as described above correctly detects a person (FIG. 4) and the case of erroneous detection (FIG. 5) will be described below.

  First, the installation position of the imaging means on the vehicle will be described. In FIG. 2, the imaging means is attached to the height h20 and depression angle θ21 from the road surface. An image picked up by the image pickup means attached in this way is shown in FIG. When a person detection function is added to the image pickup means for projecting such an image, images as shown in FIGS. 4 and 5 are displayed on the output means. In Rect1 and Rect2 of FIG. 4, the existence areas of the detected persons obj1 and obj2 are correctly displayed. However, as shown in FIG. 5, when an object other than the person to be detected is similar to the feature amount of the person, it may be mistakenly recognized as a person. In FIG. 5, a region (obj3) where a sign and a tree branch overlap is erroneously detected as a person and displayed.

  Next, the operation | movement is demonstrated below about the vehicle periphery monitoring apparatus of this Embodiment.

  FIG. 6 shows a flowchart of signal processing in the vehicle periphery monitoring apparatus of the present embodiment.

  At the start of the signal processing, first, in step S60, the signal processing means 2 reads dictionary data (data in which human feature values are recorded in the present embodiment) from the image recognition DB 5, and inputs it to the object detection processing unit 3. save. The dictionary data is stored in, for example, a rewritable FlashROM or the like, and is loaded from the object detection processing unit 3 therefrom. Dictionary data is a collection of data representing human features, such as head shape, shoulder shape, and arm shape. Similar to the input image feature values, the dictionary area has an arbitrary threshold value. If it is above, it is the data used for detecting as a person.

  In step S <b> 61, the signal processing unit 2 reads new image data from the imaging unit 1 and stores it in the object detection processing unit 3.

In step S62, the object detection processing unit 3 of the signal processing means 2 detects a person using the image data read in step S61 and the dictionary data read in step S60.
In the detection of a person, a search window R1 for determining whether or not the image is the person area is set for the image data read as shown in FIG. Scan window R1. The object detection processing unit 23 compares the scanned image data of the search window R1 region with a feature amount representing a person in the dictionary data of the image recognition DB 5 if there is a certain correlation or more (FIG. 7). In this example, P1) is detected as a person area. (Pattern matching method).

  In step S63, the detection result of the person in the object detection processing unit 3 in step S62 is output to the output means 7.

  For example, as shown in FIG. 4, an output result is obtained by superimposing rectangles Rect1 and Rect2 representing a person area on image data captured by the imaging unit 1.

  In step S64, if the user finds an erroneous detection or non-detection of a person in the detection result output in step S63, it is determined whether or not the detection result is to be corrected (correction mode). The correction mode is selected by pressing a correction key set on the screen. If the correction mode is selected, the process proceeds to step S65. If the correction mode is not selected, the process returns to step S61.

  In step S65, when the correction mode is selected in step S64, it is selected whether to correct the false detection or to correct the non-detection.

  If erroneous detection is selected, the process proceeds to step S66. If non-detection is selected, the process proceeds to step S68.

  In step S66, when the correction mode for erroneous detection is selected in step S65, the correction target area is designated by the user input means 6. As shown in FIG. 8 (b), when there is only an erroneously detected area in the output result, it is not necessary to specify the area again, and press the misrecognition registration key set in the screen. To set. Further, as shown in FIG. 8C, the position of the extracted area d1 is directly designated on the touch panel, or the ID (d1 to d3 in the figure) indicating the area is shown in the remote control as shown in FIG. 8D. There are a method of designating using a numeric keypad, or a method of selecting a corresponding area by sequentially changing a selection area with a key such as a remote controller in a recognition area in the screen.

  In step S67, the feature amount is extracted for the erroneous detection region designated in step S66, and erroneous detection information (not a person in the present embodiment) is added to the dictionary data of the image recognition DB 5. As described above, the feature amount is, for example, a combination of edge luminance gradient information or a collection of contour (shape) information. As addition of dictionary data, for example, for feature amount data that is a source of erroneous detection, Thus, the difference between the feature amounts of the designated area is updated as non-detection additional information.

  FIG. 9 shows an example of dictionary data in the image recognition DB 5. In the present embodiment, for example, extraction of a person is taken as an example, but even if a person is referred to, various orientations (front, sideways, diagonal, etc.), postures (upright, walking, bending, sitting, etc.), belongings ( Umbrellas, bags, etc.) are diverse, and are composed of a collection of a plurality of data obtained by quantifying these feature quantities.

  FIGS. 10A to 10D show the concept until the feature amount of the erroneous detection region is added to the dictionary data of the image recognition DB 5. FIG. 10A is dictionary data in which an object other than a person is erroneously determined to be a person. On the other hand, FIG. 10B shows the feature amount of a region mistakenly detected as a person from the dictionary data of FIG. In FIG. 10D, a feature amount for non-detection determination that is a difference between the original dictionary data in FIG. 10A and the feature amount of the erroneously detected region is registered. FIG. 10C shows the address (100 address) of the additional data added to the original dictionary data, and is added to the dictionary data in the image recognition DB 5 together with the additional feature amount data of FIG. Thus, the original dictionary data in FIG. 10A is updated.

  By updating the dictionary data in the image recognition DB 5 by adding / correcting the dictionary data in this way, when the dictionary data in FIG. 10C is referred to the same erroneous detection area, the additional data (address 100) 10 (d) in FIG. 10D), it is not a person if the correlation is high compared to the non-detection feature quantity of FIG. 10 (d), or a person if the correlation is low. In addition, it is possible to determine whether or not the person is a person in more detail.

  In step S68, when the undetected correction mode is selected in step S65, the correction target area is designated by the user input means 6.

  For example, as shown in FIG. 11, the area can be designated by a method in which the user designates a rectangular area (P1 to P2) on the screen through a remote controller or a touch panel.

  In step S69, the image recognition DB correction unit 4 extracts the feature amount for the non-detection area designated in step S68, and adds non-detection information (in this embodiment, a person) to the dictionary data of the image recognition DB 5. Is done.

  As the update of the dictionary data in the image recognition DB 5, for example, the dictionary data having the highest correlation in the image recognition DB 5 (data that is not recognized as a person and is not detected because the correlation does not exceed the threshold value) The feature amount that is the difference from the feature amount of the specified area is added as detection additional information and updated.

  12A to 12D show the concept until the feature amount of the non-detection area is added to the dictionary data. FIG. 12A assumes that the dictionary data has the highest correlation with the feature amount of the target area that has not been detected. On the other hand, FIG. 12B shows the feature amount of the target area that has not been detected. In FIG. 12D, a feature value for detection determination, which is a difference between the original dictionary data of FIG. 12A and the feature value of the non-detected area, is registered. FIG. 12C shows an additional data address (address 200) added to the original dictionary data, and is added to the dictionary data in the image recognition DB 5 together with the additional feature amount data shown in FIG. Thus, the original dictionary data in FIG. 12A is updated.

  In this way, by updating the image recognition DB 5 by adding / modifying dictionary data, when the dictionary data in FIG. 12C is referred to the same non-detection area, the additional data (address 200) is present. With reference to the additional feature amount in FIG. 12D, the details are such that, if the correlation is high compared to the detection feature amount in FIG. It is determined whether the person is a person.

  According to the vehicle periphery monitoring apparatus of the present embodiment as described above, it is easy to specify a region that has been erroneously detected / not detected by the user input means 6 and to make a feature amount that is a factor in dictionary data for person recognition in the image recognition DB 5. Therefore, it is possible to eliminate the same erroneous detection / non-detection of the object in the next and subsequent operations.

  In the present embodiment, an example of detecting a person has been given, but the same object can be obtained by preparing dictionary data in the image recognition DB for other objects (light vehicles, vehicles, signs, traffic lights). An effect is obtained.

  In the present embodiment, brightness gradients and contour features are given as examples of dictionary feature values. However, by adding color information or the like, it may be configured to further reduce false detection / non-detection effects. .

  As described above, according to the present invention, when the output result of the object detection processing unit is erroneously detected or not detected, the input unit inputs the erroneously detected region or the undetected region, and the image recognition database correcting unit Corrects erroneous detection or non-detection of the same target object by correcting the dictionary data in the image database of the erroneous detection area or non-detection area input by the input means and storing it in the image recognition database. It is useful as a vehicle periphery monitoring device that can prevent the above.

DESCRIPTION OF SYMBOLS 1 Imaging means 3 Object detection processing part 4 Image recognition DB correction part 5 Image recognition DB
6 User input means 7 Output means

Claims (1)

  An image capturing unit that captures an image of the surroundings of the vehicle, an image recognition database that extracts feature values from various types of human image data, and is stored as dictionary data; and the image recognition for the captured image captured by the image capturing unit An object detection processing unit for recognizing an object with reference to a database, an output unit for displaying an output result of the object detection processing unit, and a user seeing the output unit for erroneous detection or no detection Image recognition for correcting the dictionary data in the image recognition database of the erroneous detection area or the non-detection area input by the input means and the input means for inputting the erroneous detection area or the non-detection area And a database correction unit, wherein the dictionary data corrected by the image recognition database correction unit is stored in the image recognition database. Apparatus.