JP2011090681A - Image recognition device and method, and navigation apparatus - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an image recognition device that can execute image processing with a rational processing load if a travel road includes a point where an event as an obstacle for the image processing is likely to occur relatively frequently. <P>SOLUTION: The image recognition device, which is configured to be mountable on a traveling body traveling on travel roads R and which includes an image recognition means for image-recognizing objects 3 to be recognized existing on the travel roads R from an imaged image i1 picked up by an imaging means according to predetermined recognition logic, includes a recognition processing changing means for changing the recognition processing by the image recognition means between recognition obstacle points O that are points where an obstacle making image recognition difficult is likely to occur and points N other than the recognition obstacle points. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

The present invention is configured to be equipped with a traveling body traveling on a traveling road,
The present invention relates to an image recognition apparatus provided with an image recognition means for recognizing a recognition target existing on the travel path from a captured image picked up by an image pickup means in accordance with a predetermined recognition logic, and to an image recognition method used in the apparatus.

Conventionally, image recognition devices having various configurations have been proposed as image recognition devices mounted on navigation devices. In this type of image recognition device, for example, the recognition of a white line lane mark drawn along the traveling path is an important theme.
Japanese Patent Application Laid-Open No. 2004-133260 discloses performing an edge extraction process, a lane mark candidate extraction process by pairing, a Hough conversion process, and an edge point extraction process in the vicinity of a lane mark as image processing on a captured image.

  In the technique disclosed in this document, the edge point extracted by the above method is converted from the image coordinate system to the overhead coordinate system, and the edge point converted to the overhead coordinate system is approximated to a cubic curve using the least square method. By outputting the white line lane mark as a function of a cubic polynomial, the far white line lane mark can be estimated by a curve function. As a result, it is possible to accurately detect a partition line (white line lane mark) farther than the straight line portion on the image.

  This type of image processing is generally devised to perform the processing satisfactorily in a state where there is no obstacle to the image processing on the traveling road.

JP-A-9-218937 (Claims, FIG. 2)

However, it is rare that there are no obstacles in image processing on the road.
Taking the above white line lane mark as an example, illegal parking frequently occurs at points subject to image processing, and in fact, when the white line lane mark is imaged from the in-vehicle camera, it is captured only as a piece There is. On the other hand, when there are few parked vehicles on the roadside like farm roads, there are also cases where almost no parked vehicles exist like highways.

  As described above, the image recognition apparatus needs to perform appropriate image processing for various situations, but conventionally, regarding image processing in a situation where illegal parking is often performed and the white line lane mark can only be captured as a tear line, in practice, It is processed as image processing impossible.

  In the image processing method, for example, when an extraction process of a white line lane mark is performed on a captured image of a point where illegal parking is permitted, as shown in FIGS. 8 (A), (B), and (C), first, Extracting the image portion related to the car, deleting the image portion from the captured image, sequentially extracting the white line lane mark portion appearing as a tear line, and recognizing that the portion is an image portion related to the white line lane mark At the same time, it is necessary to connect the broken portion and recognize it as a single white line lane mark.

  However, such processing is far more expensive than, for example, white line lane mark recognition processing on a highway, and such a processing method cannot always be employed.

  An object of the present invention is to execute image processing with a reasonable processing load when a point where an event that causes processing problems occurs relatively frequently is included in the traveling road. An image recognition apparatus capable of performing the processing and an image processing method capable of performing the processing are obtained.

In order to achieve the above-mentioned purpose, it is configured so that it can be equipped on a traveling body traveling on a traveling road,
The characteristic configuration of the image recognition apparatus provided with the image recognition means for recognizing the recognition target existing on the travel path from the captured image captured by the imaging means according to a predetermined recognition logic is as follows:
It is provided with a recognition process changing means for making the recognition process in the image recognition means different between a recognition failure point that is a point where the possibility of occurrence of a failure that makes image recognition difficult is high and a point other than the recognition failure point. .

In this image recognition device, for a traveling body traveling on a traveling road,
Characteristic means of the image recognition method for recognizing a recognition target existing on the travel path from a captured image captured by the imaging means according to a predetermined recognition logic,
Image recognition is performed by different recognition processing in the image recognition means at a recognition failure point that is a highly likely failure occurrence point that makes the image recognition difficult, a recognition failure point, and a point other than the recognition failure point. It is supposed to be done.

  In the image recognition method according to the present application, the recognition process is made different between the recognition failure point and the other points. In other words, an image processing method suitable for the point is adopted at a point where the possibility of failure is high and image recognition is difficult, and a relatively easy image recognition method is adopted at other points, for example. . This configuration will be described with respect to the case shown in the example of the problem earlier. White line lane mark recognition processing at a point where illegal parking occurs relatively frequently and a point where there is almost no illegal parking such as an expressway. Will be different. As a result, it is possible to cope with the point and the recognition load of the point, and as a result, the performance required for the image recognition apparatus can be satisfied with a reasonable processing load.

Now, the recognition failure point and other points can be identified according to the following method.
That is, the image recognition apparatus includes a recognition difficulty determination unit that determines difficulty of image recognition by the image recognition unit at a point on the travel path.
A recognition failure point determination unit that determines whether or not a point on the travel route corresponds to the recognition failure point based on a statistical result of recognition difficulty determination by the recognition difficulty determination unit at a point on the travel route. Is provided.
In this apparatus, the recognition difficulty determination means determines the difficulty of image recognition at a point on the travel path. Then, the recognition failure point determination unit determines whether or not the point corresponds to a recognition failure point based on the statistical result of the recognition difficulty determination by the recognition difficulty determination unit.

  That is, regarding a point on the traveling road, a point where the situation where image processing is determined to be difficult is statistically high is set as a recognition failure point. As a result, in this image recognition apparatus, the difficulty related to image processing is learned at each point, and a point where a trouble event is likely to occur and relatively difficult image processing is required is determined at that point. By executing the necessary image processing, it is possible to realize an appropriate processing load.

  In the image recognition apparatus having this configuration, the image processing method is based on the result of running on the road and executing image recognition in the past. Will be identified.

Now, in the image recognition apparatus described so far,
Whether or not the recognition difficulty determination means can recognize an image;
It is preferable that the statistical result of the recognition difficulty determination is a statistical probability result regarding possibility or impossibility in the determination of availability.

In this configuration, the recognition difficulty determination unit determines whether or not the recognition is possible, and obtains a statistical probability of the determination at each point. That is, every time the traveling body passes a specific point, the recognition difficulty determination result at that point is determined in the simplest availability state, and the statistics are collected. Then, for example, a point with a high statistical probability of being disabled is set as a recognition failure point.
By doing in this way, the objective information of this application can be achieved effectively by the simplest and rational method of the experience information regarding recognition difficulty.

Furthermore, as an apparatus configuration of the image recognition apparatus described so far,
In order to make the recognition process in the image recognition means different between the recognition failure point and a point other than the recognition failure point, the same recognition logic is adopted, and the recognition level is at least the first recognition level, the first recognition level, It is preferable to make the recognition process different from the second recognition level having a recognition ability higher than the recognition level.

  A common recognition logic is used in the recognition process, and the first recognition level and the second recognition level, which differ in recognition ability, are prepared, and the recognition process follows the same recognition logic. However, the processing of the present application can be performed only by changing the recognition level, and the object of the present application can be achieved while effectively using the limited recognition logic.

  In this case, recognition processing at the second recognition level is executed at the recognition failure point, and recognition processing at the first recognition level is executed at points other than the recognition failure point.

In the image recognition apparatus described so far, the lane mark on the road is the target of recognition on the road. Examples of this type of lane mark include a white line lane mark, a broken line lane mark, a center line, a double line, and a zebra zone.
Further, the obstacles that occur are the obstacles temporarily located on the road. Such obstacles include vehicles parked on the roadside, fallen leaves from roadside trees in the fall or winter, and notice signs.

Now, the image recognition apparatus is configured to include a recognition failure point determination unit that determines a recognition failure point on the map based on the travel point information from the travel point determination unit that determines the travel point of the traveling body. preferable.
With this configuration, it is possible to identify a recognition failure point and other points on a map basis, and to obtain identification information regarding an arbitrary point on the map, which can be used for subsequent image processing.

  Now, the image recognition apparatus described so far is used by being incorporated in a navigation apparatus, and obtains information on a map-based recognition failure point and other points to recognize a travel path required by the navigation apparatus. (For example, whether the travel path is a straight path or a curve, whether the travel path has reached an intersection or a branch point, etc.), travel state recognition (for example, whether the vehicle is traveling along the lane of the travel path) Or which driving lane the vehicle is traveling on).

The figure which shows the picked-up image of the vehicle-mounted camera with which the formation condition of the recognition failure point which concerns on this invention and the traveling body which drive | works the said point are equipped Explanatory diagram showing the contents of the recognition failure point database The block diagram which shows the outline of a structure of the navigation apparatus provided with the image recognition apparatus which concerns on this invention. The figure which shows the processing flow of recognition failure point registration The figure which shows the processing flow of image recognition processing The figure which shows the input image when there is no recognition obstacle The figure which shows the image processing procedure in the point other than the recognition obstacle point The figure which shows the image processing procedure in the recognition obstacle point

  An image recognition apparatus according to the present application will be described by taking a navigation apparatus 100 equipped with the image recognition apparatus as an example.

  FIG. 1 shows a road network in which the image recognition apparatus 101 according to the present application is preferably used, and shows a state of a recognition failure point O which is a problem of the present application. In the figure, in the balloon, a captured image i1 obtained by capturing the front of the host vehicle from the in-vehicle camera 2 mounted on the host vehicle as the traveling body is shown. In the figure, it is shown that illegal parking occurs relatively frequently at the recognition failure point O, and the image recognition of the white line lane mark 3 is relatively difficult.

  In the present application, the point where the occurrence rate of the obstacles that cause the image recognition as described above is relatively high (the recognition obstacle point O shown in FIG. 1 is an example thereof), and the other points N and The image recognition apparatus 101 is configured to execute a quick image process with a low image recognition load by changing the recognition process.

  Further, as will be described later, the device of the present application determines whether or not image recognition is possible within the image recognition device 101 by running while continuously using the image recognition device 101, and the image recognition can be performed. It is possible to generate a recognition failure at each point by storing the points that did not exist one after another and determining whether image recognition is possible each time the point is passed and statistically processing the image recognition result at that point. It is configured so that sex can be evaluated.

  In addition, the evaluation is continuously performed, and the identification / registration of the point that should actually be recognized as the recognition failure point O is automatically advanced, and finally, at least once at the point where the vehicle has traveled. A recognition failure point database DBo that can identify whether or not it is a recognition failure point O as referred to in the present application is generated for each point where recognition is not possible. A configuration example of this recognition failure point database DBo is shown in FIG.

Hereinafter, the configuration of the navigation device 100 equipped with the image recognition device 101 according to the present application will be described with reference to FIG.
FIG. 3 is a block diagram showing an outline of the configuration of the navigation device 100 according to the present application.

  As shown in the figure, the navigation device 100 has a current position detection device 5, an information storage device 6, a display input device 7, a remote control input device 8 as peripheral devices for the arithmetic processing device 4 as the device body, A remote control receiver 9 and a VICS receiver 10 are provided.

Here, the current position detection device 5 includes a GPS receiver 5a, an orientation detection sensor 5b, and a distance detection sensor 5c.
The GPS receiver 5a is a device that receives a signal from an artificial satellite and has various information such as a signal transmission time, position information of the GPS receiver 5a, a moving speed of the GPS receiver 5a, and a traveling direction of the GPS receiver 5a. Can be obtained.

The azimuth detection sensor 5b includes a geomagnetic sensor, a gyro sensor, an optical rotation sensor attached to the rotating part of the steering wheel, a rotary resistance volume, an angle sensor attached to the wheel part, etc., and detects the traveling direction of the vehicle. can do.
The distance detection sensor 5c is configured by a combination of a sensor that detects the number of rotations of the wheel, a sensor that detects the acceleration of the vehicle, and a circuit that integrates the detected acceleration twice, and can detect the moving distance of the vehicle. it can.
These devices work together with the travel point determination means 4a provided in the arithmetic processing unit 4 to determine the current position of the own vehicle. This current position is determined, for example, as east longitude and north latitude information of the own vehicle.

  A VICS (Vehicle Information and Communication System: registered trademark) receiver 10 is a device that receives a signal transmitted from a signal transmitter installed at a predetermined point on a road. The apparatus can receive a VICS signal multiplexed in FM broadcast. The arithmetic processing device 4 can acquire various traffic information such as traffic jam information, current position information, and parking lot information from the signal received by the VICS receiver 10.

  The information storage device 6 includes a recording medium capable of storing information, such as a hard disk drive, a DVD drive equipped with a DVD-ROM, a CD drive equipped with a CD-ROM, and a drive means for the recording medium. Is done. Here, the information storage device 6 stores a recognition failure point database DBo, a map database DBm, and the like according to the present application.

The recognition failure point database DBo is a database that forms a key in the present invention, and is a database that is automatically generated in the apparatus 101 by continuing to use the image recognition apparatus 101. The purpose of this database is to identify the recognition failure point O in the present application.
FIG. 2 shows a configuration example of the database DBo. As shown in the figure, this database DBo has “registration point ro”, “position (east longitude / north latitude)” of the registration point ro, “unrecognizable rate” of the registration point ro, registration thereof, in the horizontal direction. Each column includes “determination result of whether or not the point ro is a recognition failure point”, “recognition logic” to be applied at the registered point ro, and “recognition level” to be applied at the registered point ro. It is configured.
The vertical direction indicates that the information is related to each registration point ro. These registered points ro are points on the travel route R, and are registered in descending order of points where image recognition is determined to be impossible even once.

In the “position (east longitude / north latitude)” column, the east longitude and north latitude of the registration point ro are registered.
In the “recognition impossible rate” column, the recognition failure rate of the registration point ro is recorded, and this recognition failure rate means that when the vehicle travels the registration point ro a plurality of times, the image recognition is impossible. Determined number of times] / [number of times of traveling at the registration point (number of times of image recognition)]. This unrecognizable rate is a kind of statistical probability result in the present application.

  In the description in this example, the recognition difficulty of image recognition is determined in two states of recognizable and unrecognizable, and the probability on the unrecognizable side is used, but the probability on the recognizable side is used. Also good. Furthermore, when adopting a configuration in which recognition difficulty of image recognition is obtained with continuous probability values such as the probability of recognition, an average value of probability values indicating the certainty of recognition can be used.

  In the “determination result whether or not it is a recognition failure point” column, a determination is made as to whether or not the registration point ro corresponds to the recognition failure point O in the present application. On the display, the recognition failure point O is indicated by a circle, and the other points are not marked. This recognition failure point O is a point where the unrecognizable rate described above is a predetermined value (for example, 80%) or more. At this recognition failure point O, there are frequent failures to image recognition, which indicates that recognition is difficult when image recognition is performed with normal recognition logic.

Now, in the present application, with respect to image recognition, a point where there are many illegal parkings (recognition failure point O) and a point where there is no illegal parking (a point other than the recognition failure point) will be described. In addition to performing the process accompanied by the deletion, the white line lane mark 3 is extracted using the Hough transform, while a determination with a high recognition level is performed. In addition, the normal determination with a low recognition level is performed to extract the white line lane mark 3 by using the Hough transform without performing the process accompanied by the deletion.

Therefore, in the “recognition logic” and “recognition level” columns, the recognition logic to be applied at the registration point ro and the recognition level employed therein are registered.
That is, regarding the recognition failure point O, extraction / removal processing of the body image candidate is added, and the white line lane mark 3 becomes a broken line, so that the recognition level is set to a second recognition level having a high level.
On the other hand, at points other than the recognition failure point, the body level image candidate deletion process is not performed, and the recognition level is set to the normal first recognition level with a low recognition level.

  The map database DBm is a database that stores map information to be displayed on a display unit 7a described later of the display input device 7. The map information includes a road layer, a background layer having background information such as facilities, a character layer for displaying characters such as a city name, and the like.

  The in-vehicle camera 2 is configured to be able to image a scene in front of the host vehicle, and a captured image from the in-vehicle camera 2 is sent to an image recognition processing unit 4B to be described later and subjected to image recognition.

  The arithmetic processing unit 4 includes, for example, a CPU that controls various arithmetic processes and operation of each unit of the navigation device 100, a RAM that is used as a working memory when the CPU performs arithmetic processing, and a CPU for operating the CPU. It can be configured to include a ROM or the like in which software such as various operation programs and control programs is stored.

  The arithmetic processing unit 4 stores software for achieving a predetermined purpose, and is configured to work with hardware to achieve a certain purpose. That is, the software and hardware constitute a processing unit that realizes a predetermined purpose by working together.

  As shown in FIG. 3, the arithmetic processing unit 4 includes a travel point determination unit 4A, an image recognition processing unit 4B, and a recognition failure point determination unit 4C as its main functional blocks. In addition to these parts 4A, 4B, and 4C, a guide route search unit 4D that executes a search for a guide route, which is the original role of the navigation device, is also provided.

The travel point determination unit 4A is a part that plays a role as the travel point determination means 4a, and includes each device (GPS receiver 5a, direction detection sensor 5b, distance detection sensor) provided in the current position detection device 5 described above. Based on the information from 5c) and the information from the VICS receiver 10, the positions of the devices 100, 101 are determined.
Here, the positions of the devices 100 and 101 are determined as east longitude and north latitude values, and are used to specify the position of the registration point ro described above. In other words, when it is determined that the image recognition is impossible in the image recognition that is performed separately, it is used for specifying the point where the image recognition is performed.

The image recognition processing unit 4B is a part that performs image processing on a captured image from the vehicle-mounted camera 2 that is an imaging unit. This part is basically provided with image recognition means 4b1 for recognizing a recognition target existing on the road R from a captured image taken by the imaging means 2 in accordance with a predetermined recognition logic. In the case of the white line lane mark 3 provided on the travel path R, the white line lane mark 3 information is output on an output image as shown in FIGS.
The captured image i1 described above with reference to FIG. 1 shows an example in which there is an obstacle (a car 11 parked in this example) that is an obstacle to image recognition on the travel path R. When the obstacle 11 is not present, a captured image as shown in FIG. 6 is captured, and after the image recognition of the white line lane mark 3 is executed, an image in which the white line lane mark 3 is emphasized becomes an output image. .
In addition to the output image, the white line lane mark 3 is recognized and identified, so that the distance from the white line lane mark 3 of the vehicle, the direction with respect to the white line lane mark 3 and the like are also found.

  As described above, the processing in the image recognition unit 4b1 is configured such that different processing can be performed on the recognition failure point O and a point other than the recognition failure point. In addition to performing a process involving extraction and deletion of body image candidates, a determination with a high recognition level is performed in order to extract the white line lane mark 3 using the Hough transform. On the other hand, at other points, the normal recognition with a low recognition level is performed to extract the white line lane mark 3 using the Hough transform without performing the process accompanied by the extraction and deletion of the body image candidate. To do.

  The recognition process changing means 4b2 provided in the image recognition processing unit 4B is a means for instructing the change of the recognition process in the image recognition means 4b1, and specifically, the point that is the current image recognition target is the first one. It is determined whether or not the registration point ro is registered in the recognition failure point database DBo described in the above, and if the registration point ro is the recognition failure point O, the recognition failure point O is determined. The image recognition means 4b1 is caused to execute image recognition, and normal image recognition is executed at a point that is not regarded as the recognition failure point O. That is, in the present application, the recognition processing in the image recognition unit 4b1 is different between the recognition failure point O, which is a point where the possibility of occurrence of a failure that makes image recognition difficult, is high and a point other than the recognition failure point. .

  In addition to the two means 4b1 and 4b2, the image recognition processing unit 4B is provided with a recognition difficulty determination means 4b3 for determining the difficulty of image recognition by the image recognition means 4b1 at a point on the road R.

  Here, the recognition difficulty determination means 4b3 is a means for determining whether or not recognition is possible. When the recognition target is the white line lane mark 3 described so far, edge point extraction is performed on the captured image. Processing, pairing extraction processing, the degree of the voting value on the voting processing when the Hough transformation processing is performed, further, the degree of convergence as a straight line of the edge point after the recognition processing, or the enhancement of the edge point, Whether or not recognition is possible is determined by comparing a threshold set in advance for each image recognition. For example, when the captured image is the one shown in FIG. 1, the vote value after the Hough conversion of the imaging line corresponding to the white line lane mark 3 after the Hough conversion is low, and is high in the case of the one shown in FIG. By setting a threshold value in the middle, it is possible to determine whether image recognition is possible when a normal image recognition method is employed.

  In the image recognition apparatus 101 according to the present application, the use of the apparatus is started on the assumption that all points are points other than the recognition failure point, and points where image recognition has been difficult are found sequentially. In this case, when the point is set as the registered point ro and, further, when the image recognition at the registered point is repeatedly performed, the recognition failure rate at the registered point ro is maintained to be higher than a certain value, The registered point is set as a recognition failure point O.

  Here, with respect to the recognition failure point O, an image recognition method having sufficiently high recognition ability is adopted in terms of processing logic, so that the recognition difficulty determination unit 4b3 determines that recognition is possible. . That is, the determination in the recognition difficulty determination unit 4b3 in this example is a determination of whether or not recognition is possible according to normal processing logic.

The travel point information determined by the travel point determination unit 4a and the recognition difficulty determination information are sent to the recognition failure point determination unit 4C and subjected to processing by the recognition failure point determination unit 4c. That is, in this means 4c, the unrecognizable rate is obtained for the same point (registered point ro). Then, when the unrecognizable rate is maintained in a state higher than a predetermined threshold, the registration point ro is set as a recognition failure point O. At a later date, when the vehicle travels the same point ro and the recognition failure rate at the point ro decreases, the registration with the recognition failure point O is deleted for the registration point ro.
As described above, for each registered point ro registered in the recognition failure point database DBo, it is determined whether the point is a recognition failure point O in a state that is statistically consistent with the current state.
In the guide route search unit 4D, a guide route to the destination is searched according to a conventionally known method. The information related to the white line lane mark from the image recognition apparatus 101 described so far is used, for example, for specifying the current driving lane position of the own vehicle. An appropriate guide route is searched from the lane position.

Hereinafter, based on FIG. 4, the processing flow in the image recognition apparatus 101 according to the present application will be described.
This processing flow repeats the image recognition and the processing flow in the case where the recognition failure point O and other points are subject to image recognition (the processing flow unit F1 on the front stage side), and the recognition failure at a specific point occurs. In this case, the process is established from a flow part (a process flow unit F2 on the rear stage side) that executes a process required for the point (a process for determining whether or not the point corresponds to the recognition failure point O).

  When executing this processing flow, first, it is determined whether or not the current travel point corresponds to the recognition failure point O by the recognition process changing means 4b2 (# 41), and image recognition is performed based on the determination result. Recognition logic is changed. If it corresponds to the recognition failure point O (# 41: yes), as described above, for example, as shown in FIG. 5 (b), the edge extraction (# 51), the vehicle body is optimized to optimize the recognition logic. Image candidate extraction / removal (# 51-2), pairing processing (# 52), Hough transform (# 53), and edge extraction near the lane mark (# 54) are performed. That is, the vehicle body image candidate extraction / removal process is performed, and in the Hough conversion process, a process of the second recognition level having a high recognition level is executed (# 42, # 43).

On the other hand, when it corresponds to a point other than the recognition failure point (# 41: no), processing based on normal recognition logic is executed (# 44). That is, as explained so far, for example,
As shown in FIG. 5A, edge extraction (# 51), pairing processing (# 52), Hough transform (# 53), and edge extraction (# 54) near the lane mark are performed. That is, the first recognition level process with a low recognition level is executed in the Hough conversion process without the extraction / removal process of the body image candidate (# 44).

  In this way, after the image recognition process is completed at the current point, the recognition difficulty determination unit 4b3 determines whether image recognition is possible at the local point. As described above, when the feature quantity to be recognized is lower than the predetermined threshold (for example, when the number of votes in the Hough transform is lower than the predetermined vote threshold), image recognition is impossible. (# 45).

In this determination, if image recognition is possible (# 45: yes), the subsequent processing flow unit F2 is omitted and a new recognition process is executed if necessary.
On the other hand, in this determination, when the image recognition is impossible (# 45: no), the process moves to the processing flow unit F2 on the rear stage side to determine whether or not the recognition failure point O is referred to in the present application.

  First, registration processing for a point that is currently the target of processing is executed (# 46). This registration is executed only when it is determined in the image recognition apparatus 101 that the image recognition is impossible for the first time, and if it is registered in advance in the recognition failure point database DBo as the registration point ro, it is duplicated. It will never be registered.

Next, the recognition failure point determination unit 4c calculates a recognition failure rate (an example of statistical probability) at the point (# 47), and determines whether the value exceeds a predetermined threshold (#). 48). If the recognition failure rate is lower than the threshold (# 48: no), it is determined that there is no need to register as a recognition failure point O, and the process returns to new image recognition.
On the other hand, when the unrecognized rate is maintained higher than the threshold (# 48: yes), the point is registered as a recognition failure point O (# 49).

  The processing flow in the image recognition apparatus 101 ends the process in a state where an end command is separately input and is recognized by the end determination (# 410).

  By sequentially repeating this process as the traveling body 1 travels, for example, points where illegal parking frequently occurs and obstacle situations that cause obstacles in image recognition always occur are registered as recognition obstacle points O. The traveling points in the expressway are assumed to be points other than the recognition failure point, which is sufficient by normal processing.

  The above is the processing flow executed by the image recognition apparatus 101 of the present application when the travel path R includes the recognition failure point O. The recognition processing at the recognition failure point O in FIG. 4 and the recognition processing at points other than the recognition failure point are shown in FIG. Will be described with reference to the drawings. In the description, for the sake of easy understanding, an example in which there is no illegal parking shown in FIG. 6 and there are low obstacles in image recognition and an example in which there are many obstacles shown in FIG. 1 will be described.

Image recognition when there are few obstacles in image recognition In this case, a captured image captured by the in-vehicle camera 2 is as shown in FIG.
Edge point extraction processing is executed for this captured image. The image subjected to this processing is the image shown in FIG. 7A (# 51). Then, pairing processing is performed as shown in FIG. 7B to remove isolated points (# 52), and then Hough transform is performed on the image (# 53). In the θ-ρ plane obtained after the Hough transform, the number of votes of coordinates corresponding to the four boundary lines of the white line lane mark 3 on the captured image is high. That is, the process of the first recognition level having a low recognition level is sufficient. Therefore, returning to the imaging side again, as shown in FIGS. 7C and 7D, only pixel points corresponding to the edges of the four white line lane marks 3 can be left (# 54). In the figure, edge lines extracted by the Hough transform are shown as L1 and L2, their up edges are shown as L1u and L2u, and their down edges are shown as L1d and L2d.
In the case of this example, the edge e of the white line lane mark 3 is imaged as a continuous edge line on the entire captured image, so the number of votes in the Hough transform is high and the discrete state on the image is also low. As a result, the recognition difficulty referred to in the present application is low and recognition is easy.

Image recognition when there are many obstacles in image recognition In this case, a captured image captured by the vehicle-mounted camera 2 is as shown in FIG. Edge point extraction processing is executed for this captured image (# 51). The image subjected to this processing is the image shown in FIG. In the case of the present application, advanced image recognition is adopted in advance in order to advance image recognition in a state where this point is determined to be a recognition failure point O.

That is, in the image shown in FIG. 8A, an image part i11 that is highly likely to be an image of the parked car 11 is searched. In this example, a set of edge points that form a certain closed region is extracted, and an image portion that can be regarded as an image of the car 11 is extracted from the shape of the closed region, the color in the region, or the shade information. Further, these image parts are deleted from the captured image (# 51-2). Thereafter, a pairing process is performed to remove independent points (# 52). The image obtained in this way is the image shown in FIG. As can be seen from this figure, the edge e related to the white line lane mark 3 appears on the image as a broken broken line.
The Hough transform is also applied to this image (# 53). In the converted θ-ρ plane, the number of votes of coordinates corresponding to the four boundary lines of the white line lane mark 3 on the captured image is lower than that in the case where there is no obstacle described above. Therefore, recognition at the second recognition level having a higher recognition level than that described above is executed. A straight line candidate is extracted using a lower threshold as the number of votes after the Hough transform. Thereafter, returning to the imaging side again, as shown in FIGS. 8C and 8D, only the pixel points corresponding to the edge 3 of the four white line lane marks can be left (# 54).

[Another embodiment]
(1) In the above embodiment, the change process of the recognition process is performed between the two forms of the normal process and the advanced recognition process. It is good also as a change in three or more polymorphisms.
(2) In the above embodiment, in the recognition process at the recognition failure point, the vehicle image candidate extraction / removal and the recognition process based on the second recognition level in the Hough transform are performed. As processing, only one of them may be performed, and as image recognition processing capable of exhibiting high recognition ability, processing such as template matching and combination hub conversion may be employed.
(3) In the determination of the recognition difficulty, a determination method such as a histogram distribution or a frequency determination may be adopted in addition to the determination methods described so far.
(4) Regarding the above image recognition device, an example in which the device is used for a navigation device has been shown, but it can also be used for a millimeter wave radar or the like.
(5) In addition to the white line lane mark, the recognition target can be a pedestrian crossing, a temporary stop line, or the like.

  Image recognition that enables image processing to be executed with a reasonable processing load when there are points where image processing problems occur relatively frequently on the road. Could get the device.

2 In-vehicle camera 3 White line lane mark 4 Arithmetic processing unit 4A Traveling point determination unit 4a Traveling point determination unit 4B Image recognition processing unit 4b1 Image recognition unit 4b2 Recognition processing change unit 4b3 Recognition difficulty determination unit 4C Recognition failure point determination unit 4c Recognition failure Point determination means 4D Guide route search unit 5 Current position detection device 6 Information storage device 7 Display input device 8 Remote control input device 10 VICS receiver DBo Failure situation database DBm Map database N Point other than recognition failure point O Recognition failure point R Runway ro Registration point

Claims (11)

It is configured so that it can be equipped with a traveling body traveling on the traveling road,
An image recognition apparatus comprising image recognition means for recognizing a recognition target existing on the travel path from a captured image captured by an imaging means according to a predetermined recognition logic,
An image recognition apparatus comprising: a recognition process changing unit that makes a recognition process in the image recognition unit different between a recognition failure point that is a point at which a failure occurrence that makes image recognition difficult is highly likely and a point other than the recognition failure point . Recognizing difficulty determining means for determining difficulty of image recognition by the image recognizing means at a point on the traveling road,
A recognition failure point determination unit that determines whether or not a point on the travel route corresponds to the recognition failure point based on a statistical result of recognition difficulty determination by the recognition difficulty determination unit at a point on the travel route. An image recognition apparatus according to claim 1. Whether or not the recognition difficulty determination means can recognize an image;
The image recognition apparatus according to claim 2, wherein the statistical result of the recognition difficulty determination is a statistical probability result related to possible or impossible in the determination of availability. In order to make the recognition process in the image recognition means different between the recognition failure point and a point other than the recognition failure point, the same recognition logic is adopted, and the recognition level is at least the first recognition level, the first recognition level, The image recognition apparatus according to any one of claims 1 to 3, wherein the recognition process is made different between a second recognition level having a recognition ability higher than the recognition level. The image recognition apparatus according to claim 4, wherein recognition processing at the second recognition level is executed at the recognition failure point, and recognition processing at the first recognition level is executed at a point other than the recognition failure point. The image recognition apparatus according to claim 1, wherein the recognition target existing on the travel path is a lane mark on the travel path. The image recognition apparatus according to claim 1, wherein the generated obstacle is an obstacle temporarily positioned on the travel path. 8. The apparatus according to claim 1, further comprising a recognition failure point determination unit that determines the recognition failure point on a map based on travel point information from a travel point determination unit that determines a travel point of the traveling body. The image recognition apparatus according to item. A navigation device that includes the image recognition device according to claim 8 and that operates based on recognition information of the recognition target existing on the travel path obtained from the image recognition device. For the traveling body traveling on the traveling road,
An image recognition method for recognizing a recognition target existing on the travel path from a captured image captured by an imaging unit according to a predetermined recognition logic,
Image recognition is performed by different recognition processing in the image recognition means at a recognition failure point that is a highly likely failure occurrence point that makes the image recognition difficult, a recognition failure point, and a point other than the recognition failure point. The image recognition method to perform. The image recognition method according to claim 10, wherein the recognition failure point and a point other than the recognition failure point are identified from a result of running on the road in the past and executing the image recognition.

Images