JP2010218569A - Feature information collection device and feature information collecting method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a feature information collection device and feature information collecting method for collecting highly accurate feature information at a low cost. <P>SOLUTION: The feature information collection device includes: a vehicle position information acquiring means 16 for acquiring vehicle position information P representing the present position of a vehicle; an image information acquiring means 12 for acquiring image information G around the vehicle; an image recognizing means 18 for performing image recognition of a recognition object included in the image information G; a recognition result storing means DB 3 for storing recognition information representing a recognition result of the recognition object by the image recognizing means 18 in association with information Aa of a recognition position of the recognition object on the basis of the vehicle position information P to store them; and a learning feature extracting means 31 for extracting a recognition object that can be repeatedly recognized as a learning feature on the basis of a plurality of recognition information items of the same place, stored in the recognition result storing means DB 3 by performing image recognition of image information of the same place a plurality of times, and outputting the recognition object along with the position information Aa of the learning feature. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a feature information collecting apparatus and a feature information collecting method for collecting feature information included in image information around a vehicle acquired by an imaging device or the like mounted on the vehicle by performing image recognition processing.

  There is known a navigation device that corrects the position of the host vehicle by recognizing a feature included in image information around the vehicle acquired by an imaging device or the like mounted on the vehicle (for example, Patent Document 1 shown below) reference.). This device is mounted on a vehicle, detects a crossing from a camera that prompts a landscape along a traveling road ahead of the vehicle, identifies a crossing symbol from a landscape image prompted by the camera, and detects an intersection from the vehicle position (intersection) And a vehicle position correcting means for correcting the vehicle position to a point on the traveling road that is separated from the intersection position obtained from map data by the distance. The intersection symbol is, for example, a traffic light, a pedestrian crossing, a white line for a median strip, or the like. When the position of the vehicle is corrected using the apparatus of Patent Document 1, it is desirable that information such as the type and location of these intersection symbols corresponding to the feature is clear.

  Patent Document 2 shown below shows a technique for constructing a database relating to such a feature. According to this, instead of the conventional method in which the operator manually rewrites data, the feature is photographed while the measurement vehicle is running, and the feature is automatically recognized from the photographed image.

  Further, in Patent Document 3 shown below, there is a navigation device that collects feature information based on an image recognition result of a target feature included in image information around the vehicle acquired by an imaging device or the like mounted on the vehicle. It is disclosed. In Patent Document 3, a traffic sign or a road traffic information display board installed on a road is a target feature. And this apparatus performs image recognition of the target feature included in the image information acquired by the imaging device mounted on the vehicle, and uses the feature information such as the sign information extracted from the recognition result as the position information and the section. Store it in the map database in association with the information. At this time, position information and section information associated with the feature information are determined based on information from a GPS receiver, a gyro, a vehicle speed sensor, and the like that are generally used in navigation devices.

JP-A-9-243389 JP 2000-293670 A JP 2006-038558 A

  However, position information acquired by a GPS receiver, a gyro, a vehicle speed sensor, or the like used in a navigation device generally has an error of several meters to several tens of meters. Therefore, the position information when the feature information is stored in the map database may be information having such an error. In particular, when using feature information for vehicle control, vehicle position correction, etc., it is desired to prepare feature information having highly accurate position information. However, as described in Patent Document 2, on all roads, it is not realistic in terms of time and cost to collect highly accurate feature information using a measurement vehicle or the like. Therefore, highly accurate feature information is preferentially prepared for roads with many users such as main roads. On the other hand, it is preferable for the user that high-precision feature information is maintained on the road where the person frequently passes, and various fine services including navigation can be enjoyed. Therefore, it is desired to maintain high-precision feature information at a low cost according to the use situation of the user regardless of whether it is a main road or not. In addition, it is difficult to collect feature information for such road markings on roads such as mountain roads where there are few regular road markings on the road surface.

  The present invention has been developed in view of the above problems, and collects highly accurate feature information at a low cost according to the use status of the road by the user for roads for which feature information is not maintained. It is an object of the present invention to provide a feature information collecting apparatus and a feature information collecting method.

In order to achieve the above object, the feature configuration of the feature information collecting apparatus according to the present invention is as follows:
Own vehicle position information acquisition means for acquiring own vehicle position information representing the current position of the own vehicle;
Image information acquisition means for acquiring image information around the host vehicle;
Image recognition means for performing image recognition processing of a recognition target which is a feature included in the image information;
When the feature amount of the recognition result of the recognition target by the previous SL image recognition means is equal to or higher than a predetermined threshold value, the recognition target extracting means for extracting the recognition target as the recognition target capable of image recognition,
Wherein the recognition target extracted by the recognition target extraction means, said in that it comprises a recognition target storage means for storing in association with the information of the recognition position of the recognition target based on the vehicle position information.

According to this characteristic structure, without being limited to the specific features, such as road signs and road display, the recognition target extracting means, to extract a recognition target with a characteristic capable of recognizing images. In other words, even without a fixed object such as road markings, it is possible to collect an ambiguous subject, such as "What is in the road surface." Accordingly, recognition objects can be collected even on roads with few road markings such as mountain roads, and the feature information collection device can be used over a wide area . Et al is, since the information of the recognition target when only the vehicle is utilized, the optical system of the imaging apparatus related to the collection and use of the recognition target is the same, an imaging device for imaging the image information of the vehicle It does not have to be precisely calibrated at the time of production. As a result, the recognition target can be collected and used at low cost .

The recognition result of the recognition target is a feature amount obtained by image recognition processing for the recognition target. Thus, feature quantity, for example if it exceeds a predetermined threshold value, the recognition result is that satisfying the predetermined learning condition.

  Then, it is possible to determine whether or not the recognition target (feature) is sufficiently characterized with respect to the recognition result (feature amount) indicated quantitatively.

  In the feature information collecting apparatus according to the present invention, the recognition target includes any one of noise, an edge, and a predetermined color included in the image information, and the feature amount includes the noise amount and the number of extracted edges. And any one of the types of colors and the number of extractions corresponding thereto.

The recognition result of the recognition target is a feature amount obtained by image recognition processing for the recognition target. When there is no feature on the road surface, the image information on the road surface is uniform, but if any feature is present, the uniformity is lost. The factors disturbing the uniformity and noise, if the noise recognition target, some of the noise amount becomes characteristic amount as the recognition result. In addition, stains, dirt, cracks, and the like on the road surface can be extracted as edges by performing known filter processing on the image information. If this edge is a recognition target, the number of extracted edges is a feature amount as a recognition result. In addition, a road sign painted in white, yellow, or orange can be extracted by performing a known window comparator process on the image information. If each predetermined color is a recognition target, the type of color and the number of extractions corresponding thereto are the feature amounts as a recognition result . As this, the recognition target, noise, edges, by a etc. predetermined color can be obtained a quantitative feature value. In particular, the noise, the edge, the recognition target such as a predetermined color, need not have a fixed resistance, such as road signs and road markings, as the recognition target in an ambiguous state, such as "something is on the road" Can be collected. Therefore, even on roads with few road markings, recognition targets can be collected smoothly, and the main feature information collection device can be used over a wide area.

Here, before Symbol recognition target extraction means, with respect to at least one of weather recognition time and the time of recognition, the recognition target capable of recognizing images when predetermined conditions, effective only if the condition is satisfied Extract as a limited recognition target ,
The recognition target storage means includes the limited recognition target extracted by the recognition target extraction means, information on the recognition position of the limited recognition target based on the vehicle position information, the recognition time, and the weather at the time of recognition. It is preferable that the information is stored in association with at least one of the limiting conditions .

For example, the shadow of an object appears in the daytime or on fine weather. Since the shadow has a strong contrast with the road surface or the like, it is easy to recognize if the edge or specific color is a recognition target. However, since the appearance of shadows is affected by time and weather, the recognition result is not stable. According to this configuration, since the limited recognition target is effective only when a predetermined condition such as the recognition time or the weather at the time of recognition is satisfied, it is possible to expand the targets that can be used as suitable recognition targets .

The feature configuration of the feature information collecting method according to the present invention is as follows.
Own vehicle position information acquisition step for acquiring own vehicle position information indicating the current position of the own vehicle;
An image information acquisition step of acquiring image information around the host vehicle;
An image recognition step for performing an image recognition process of a recognition target that is a feature included in the image information;
When the feature amount of the recognition result of the recognition target by the previous SL image recognition process is equal to or greater than a predetermined threshold value, the recognition target extraction step of extracting the recognition target as the recognition target capable of image recognition,
A recognition target storage step of storing the recognition target extracted in the recognition target extraction step in a recognition target storage unit in association with information on the recognition position of the recognition target based on the vehicle position information ;
It is in the point provided with.

According to this characteristic structure, without being limited to the specific features, such as road signs and road displayed in the recognition target extraction step, the recognition object capable of recognizing images is issued extracted. That is, an ambiguous object such as “something is on the road surface” can be collected as a recognition object even if it is not a fixed object such as a road marking. Therefore, recognition objects can be collected even on roads with few road markings such as mountain roads, and the feature information collection method can be used over a wide area.

The block diagram which shows typically the structure of the navigation apparatus which concerns on embodiment of this invention. The figure which shows the example of a structure of the map information memorize | stored in the map database The figure which shows the example of the feature information of the road marking memorize | stored in the feature database The figure which shows an example of the arrangement configuration of the imaging device to the own vehicle Explanatory drawing for explaining the outline of the feature information collection process when the vehicle travels on a road where the feature information is not maintained FIG. 5B is an enlarged view of the main part of the learning value stored in the learning database shown in FIG. Explanatory drawing which shows the example of the calculation method of the image processing with respect to recognition object, and the information of a position Explanatory drawing which shows the example of the image processing with respect to recognition object FIG. 8 is an enlarged view of the main part of the learning value stored in the feature learning database shown in FIG. The flowchart which shows an example of the procedure of the feature information collection method by the navigation apparatus which concerns on this embodiment. The figure which shows the example in which the one part structure of the feature information collection apparatus which concerns on embodiment of this invention was installed in the server apparatus.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a schematic configuration of a navigation device 1 according to the present embodiment. The navigation device 1 includes a feature information collection device 2, an image recognition device 3, and a vehicle position recognition device 4 as an embodiment of the present invention. The navigation device 1 performs an image recognition process of a recognition target included in the image information G, and uses a recognition target that can be repeatedly recognized as a learning feature based on recognition information that represents a recognition result of the recognition target. Extract. Then, by reflecting the feature information F of the learned feature in the feature database DB2, the correction of the vehicle position information P using the feature information F is performed even on the road where the feature information F is not maintained. Make it possible to do. The navigation device 1 performs navigation processing such as route guidance based on the corrected vehicle position information P.

  Each functional unit of the navigation device 1 shown in FIG. 1, specifically, an image information acquisition unit 12, a vehicle position information acquisition unit 16, a feature information acquisition unit 17, an image recognition unit 18, a vehicle position correction unit 19, The navigation calculation unit 20, the learned feature extraction unit 31, the estimated position determination unit 34, the feature information management unit 35, and the road attribute information acquisition unit 41 are input data using a calculation processing device such as a CPU as a core member. A functional unit for performing various processes is implemented by hardware or software (program) or both. Each of these functional units is configured to exchange information with each other. Further, each of the databases DB1 to DB3 of the navigation device 1 is a recording medium that can store information, such as a hard disk drive, a DVD drive equipped with a DVD-ROM, a CD drive equipped with a CD-ROM, and its drive. And a device having a means as a hardware configuration. Hereinafter, the configuration of each part of the navigation device 1 according to the present embodiment will be described in detail.

1. Map database The map database DB1 is a database in which map information M divided into predetermined sections is stored. FIG. 2 is an explanatory diagram showing an example of the configuration of the map information M stored in the map database DB1. As shown in this figure, the map information M has road information Ra representing a road network by connection relations between a large number of nodes n corresponding to intersections and links k corresponding to roads connecting the intersections. . Each node n has information on the position (coordinates) on the map expressed by latitude and longitude. Each link k is connected via a node n. Each link k has information such as a road type, a region type, a link length, a road width, and a shape interpolation point for expressing a link shape as attribute information. Here, the road type information is information on the road type when the road is divided into a plurality of types, such as an automobile-only road, a city road, a narrow street, a mountain road, and the like. In addition, the region type information is divided into a plurality of regions where roads corresponding to the link k are provided, such as regions such as the Kanto region and the Kansai region, and administrative divisions such as prefectures and municipalities. Information on the type of the area. The attribute information of these links k corresponds to the road attribute information Rb (see FIG. 1). And in this embodiment, this map database DB1 is corresponded to the map information storage means in this invention. In FIG. 2, only the road information Ra for one section is shown, and the road information Ra for other sections is omitted.

2. Feature Database The feature database DB2 is a database in which information on various features provided on or around the road, that is, feature information F is stored. As shown in FIG. 1, in this embodiment, two types of information, initial feature information Fa and learning feature information Fb, are stored in the feature database DB2. Here, the initial feature information Fa is feature information F about a plurality of features that are prepared and stored in advance in the feature database DB2. Such initial feature information Fa is maintained only for some areas and roads such as the periphery of major cities and main roads among all areas where map information M including road information Ra is maintained. . On the other hand, the learned feature information Fb is feature information F stored in the feature database DB2 as a result of learning using the image recognition result of the target feature by the image recognition unit 18, as will be described later. In the following description, when “feature information F” is simply referred to, the initial feature information Fa and the learned feature information Fb are collectively referred to. In the present embodiment, this feature database DB2 corresponds to the feature information storage means in the present invention.

  The features whose feature information F (particularly the initial feature information Fa) is stored in the feature database DB2 include road markings (paint markings) provided on the road surface. FIG. 3 is an explanatory diagram showing an example of the feature information F of the road marking stored in the feature database DB2. Examples of such features related to road markings include pedestrian crossings, stop lines, speed markings, and lane lines that divide lanes along the road (including various lane markings such as solid lines, broken lines, and double lines). Included are travel direction indications (including arrow indications such as straight arrows, right-turn arrows, etc.) that specify the direction of travel of each lane. Note that the features in which the feature information F is stored can include various features such as traffic lights, signs, overpasses, and tunnels in addition to such road markings.

  The feature information F includes feature position information of each feature and feature attribute information associated therewith as its contents. Here, the feature position information includes information on the position (coordinates) of the representative point of each feature associated with the link k or the node n constituting the road information Ra on the map, and the orientation of each feature. Yes. The representative point of the initial feature information Fa is set, for example, near the center in the length direction and width direction of each feature. The representative points of the learned feature information Fb will be described later. The feature attribute information of the initial feature information Fa includes feature type information, feature form information such as the shape, size, and color of the feature. For example, the feature types of the initial feature information Fa are basically the same forms such as “pedestrian crossing”, “stop line”, “speed indication (30 km / hour)”, and the like. This is information indicating the type of the object. The feature attribute information of the learned feature information Fb will be described later. In the present embodiment, the feature information F is related information that represents a relationship with other nearby features, or distance information between features that represents the distance between the features. have. Here, the association information is information for allowing other features existing ahead to be predicted by recognizing an image of one feature while the host vehicle 50 is traveling along the road. . Further, the distance information between features is information for accurately predicting the distance from the own vehicle 50 to such a feature existing ahead.

3. Learning Database The learning database DB3 is a database that stores the recognized position information Aa derived by the learned feature extracting unit 31 in a state where the recognition target corresponding to each recognized position information Aa can be identified. Here, in order to be able to identify the recognition target corresponding to each recognition position information Aa, each recognition position information Aa and the corresponding feature attribute information Ab of the recognition target are stored in association with each other. Specific contents of the recognized position information Aa and the feature attribute information Ab stored in the learning database DB3 will be described in detail later. In the present embodiment, this learning database DB3 corresponds to the recognition result storage means in the present invention.

4). Image Information Acquisition Unit The image information acquisition unit 12 functions as an image information acquisition unit that acquires image information G around the host vehicle captured by the imaging device 11. Here, the imaging device 11 is an in-vehicle camera or the like provided with an imaging element, and is provided at a position where at least a road surface of the road around the host vehicle 50 can be imaged. As such an imaging device 11, for example, it is preferable to use a back camera that images the road surface behind the host vehicle 50 as shown in FIG. The image information acquisition unit 12 captures imaging information captured by the imaging device 11 at predetermined time intervals via a frame memory (not shown). The time interval for capturing the image information G at this time can be set to, for example, about 10 to 50 ms. Thereby, the image information acquisition unit 12 can continuously acquire the image information G of a plurality of frames captured by the imaging device 11. The acquired image information G is output to the image recognition unit 18.

5). Own vehicle position information acquisition unit The own vehicle position information acquisition unit 16 functions as own vehicle position information acquisition means for acquiring own vehicle position information P indicating the current position of the own vehicle 50. Here, the vehicle position information acquisition unit 16 is connected to the GPS receiver 13, the direction sensor 14, and the distance sensor 15. Here, the GPS receiver 13 is a device that receives GPS signals from GPS (Global Positioning System) satellites. This GPS signal is normally received every second and is output to the vehicle position information acquisition unit 16. The own vehicle position information acquisition unit 16 analyzes signals from GPS satellites received by the GPS receiver 13 and acquires information such as the current position (latitude and longitude), traveling direction, moving speed, and time of the own vehicle 50. can do. The direction sensor 14 is a sensor that detects the traveling direction of the host vehicle 50 or a change in the traveling direction. The azimuth sensor 14 includes, for example, a gyro sensor, a geomagnetic sensor, an optical rotation sensor attached to the rotating part of the handle, a rotary resistance volume, an angle sensor attached to the wheel part, and the like. Then, the direction sensor 14 outputs the detection result to the vehicle position information acquisition unit 16. The distance sensor 15 is a sensor that detects the vehicle speed and the moving distance of the host vehicle 50. The distance sensor 15 is, for example, a vehicle speed pulse sensor that outputs a pulse signal each time a drive shaft or wheel of the vehicle rotates by a certain amount, a yaw / G sensor that detects acceleration of the host vehicle 50, and integration of the detected acceleration. Configured by a circuit or the like. Then, the distance sensor 15 outputs information on the vehicle speed and movement distance as the detection result to the vehicle position information acquisition unit 16.

  And the own vehicle position information acquisition part 16 performs the calculation which pinpoints an own vehicle position by a well-known method based on the output from these GPS receiver 13, the direction sensor 14, and the distance sensor 15. FIG. In addition, the vehicle position information acquisition unit 16 acquires road information Ra around the vehicle position extracted from the map database DB1, and indicates the vehicle position in the road information Ra by performing known map matching based on the road information Ra. Also make corrections to fit on the road. In this way, the host vehicle position information acquisition unit 16 acquires host vehicle position information P including information on the current position of the host vehicle 50 represented by latitude and longitude and information on the traveling direction of the host vehicle 50.

6). Feature information acquisition unit The feature information acquisition unit 17 functioning as a feature information acquisition unit, based on the vehicle position information P acquired by the vehicle position information acquisition unit 16 and the like, from the feature database DB2 to the host vehicle 50. The feature information F of features existing around is extracted and acquired. The feature information F includes both the initial feature information Fa and the learned feature information Fb. In the present embodiment, as an example, the feature information acquisition unit 17 is based on the vehicle position information P, and the vehicle 50 is traveling from the current position of the vehicle 50 indicated by the vehicle position information P. The feature information F of the target feature existing before the end of the link k representing the road is extracted from the feature database DB2. The target feature is a feature that is a target of image correction processing for position correction (details will be described later) by the image recognition unit 5. Therefore, the feature information F acquired by the feature information acquisition unit 17 is output to the image recognition unit 18 and the vehicle position information correction unit 19. In this embodiment, the road surface collected as a learning feature in addition to various road markings provided on the road surface such as pedestrian crossings, stop lines, speed markings, etc., which are prepared in advance as initial features. Stains and stains, cracks, joints of pavements, manhole covers, etc. are also subject features. Also, various types of road markings that are not registered as initial features due to low usage frequency are also target features as learning features. Further, what appears when a predetermined condition such as a specific time or weather such as a shadow is satisfied becomes a target feature as a limited learning feature effective only when the condition is satisfied.

  Note that the feature information acquisition unit 17 refers to the feature database DB2 based on the own vehicle position information P, and determines whether or not the target feature stored as the feature information F exists around the own vehicle 50. It also functions as presence determination means for determining. When it is determined that the target feature does not exist around the host vehicle 50, an information collection image recognition process is performed as described later. That is, when it is determined that the target feature does not exist in the vicinity of the host vehicle 50, it is determined that the host vehicle 50 is not in a high-accuracy host vehicle position recognition state to be described later, and the learned feature information Fb is collected. The information collection image recognition process is performed. Note that the feature information F when functioning as the presence determination unit may include the learned feature information Fb or may be limited to the initial feature information Fa. In this embodiment, both are included.

7). Image Recognition Unit In this embodiment, the image recognition unit 18 learns position correction image recognition processing for use in correcting the vehicle position information P and the image recognition result of the feature and reflects it in the feature database DB2. Two types of image recognition processing are performed, namely, information collection image recognition processing for generating information. That is, the image recognition unit 18 performs the position correction image recognition processing based on the feature information F of the target feature existing around the host vehicle 50 extracted from the feature database DB2 based on the host vehicle position information P. It functions as image recognition means for performing image recognition processing of the target feature indicated by the feature information F included in the image information G. The image recognition unit 18 functions as an image recognition unit that performs image recognition processing of a recognition target included in the image information G acquired by the image information acquisition unit 12.

  In the present embodiment, the image recognition unit 18 is determined by the feature information acquisition unit 17 that also functions as presence determination means that the target feature stored as the feature information F exists around the host vehicle 50. In this case, a position correction image recognition process is performed. On the other hand, when the feature information acquisition unit 17 determines that the target feature stored as the feature information F does not exist around the host vehicle 50, the information collection image recognition process is performed. When the vehicle 50 is traveling on a road on which the feature information F has already been prepared, the correction of the vehicle position information P is performed based on the image recognition result by the position correction image recognition process and the feature information F. Processing is performed. On the other hand, when the host vehicle 50 is traveling on a road where the feature information F is not maintained, an information collection image recognition process is performed, and a recognition target included in the image information G is extracted as a learning feature. A material collection process is performed.

  In the image recognition processing for position correction, the image recognition unit 18 performs image processing based on the feature information F of the target feature existing around the host vehicle 50 acquired from the feature database DB2 based on the host vehicle position information P. Image recognition processing of the target feature indicated by the feature information F included in the information G is performed. At this time, the image recognition unit 18 sets a predetermined recognition area for requesting recognition of the target feature indicated by the feature information F, and performs image recognition processing of the target feature with respect to the image information G in the recognition area. I do. This recognition area is set as a position range on which the target feature is estimated to exist on the link k indicated by the road information Ra. Here, the imaging area of the actual road included in each image information G is the vehicle position and imaging area calculated in advance based on the mounting position, mounting angle, field angle, and the like of the imaging device 11 to the host vehicle 50. Can be obtained based on the vehicle position information P. Therefore, the image recognition unit 18 extracts the image information G corresponding to the recognition area set for each target feature based on the information of the imaging area of each image information G obtained in this way, and performs image recognition processing. Do. When the target feature is an initial feature, the image recognition unit 18 performs known binarization processing, edge detection processing, and the like on the image information G, and the target feature included in the image information G The contour information of the object (road marking) is extracted. Then, by performing pattern matching between the extracted contour information and the form of the target feature, the target feature included in the image information G is recognized. If the target feature is a learned feature, the target feature is recognized by performing a process similar to the image recognition process for the recognition target at the time of learning as described later. Then, the image recognition result of the target feature by the position correction image recognition process is used for correcting the vehicle position information P by the vehicle position information correction unit 19.

  In the information collection image recognition process, the image recognition unit 18 performs an image recognition process of a recognition target included in the image information G. This image recognition process does not use the feature information F stored in the feature database DB2, and thus performs the image recognition process of the recognition target without setting the target feature and the recognition area as described above. Here, the recognition target is a feature included in the image information, and includes any one or more of noise, an edge, a predetermined color, and a predetermined shape. The recognition result of these recognition targets is a feature amount obtained by a predetermined image recognition process for the recognition target. For example, in the state where there is no feature on the road surface, the image information on the road surface is uniform, but if any feature exists, the uniformity is lost. If the factor that destroys the uniformity is noise, and this noise is a recognition target, the amount of noise becomes a feature amount as a recognition result. Since such noise becomes a high frequency component in the spatial frequency of the image information G, the amount of noise can be extracted by a filtering process using a high-pass filter or the like.

  Features that can be learned on the road surface include road markings on the road surface, stains on the road surface, dirt on the road surface, cracks on the road surface, pavement joints, manholes There are shadows on the road surface. The boundary between the road surface stains, dirt, cracks, and the like and the road surface can be extracted as an edge component by subjecting the image information G to a known Gaussian filter process or the like. If this edge component is a recognition target, the number of extracted edge components is a feature amount as a recognition result. Further, the color components of road markings painted in white, yellow, and orange can be extracted by performing known window comparator processing on the image information G. If each color is a recognition target, the type of color and the number of extractions corresponding to it become the feature value as a recognition result. Alternatively, a specific shape, for example, a specific shape such as a triangle, a circle, a quadrangle, or a number, may be used as a recognition target, and pattern matching processing may be performed, and the degree of adaptation may be used as a feature amount. The recognition result (feature amount) of the recognition target by the information collection image recognition process is stored in the learning database DB3 in association with the information on the recognition position.

8). Self-vehicle position information correction unit The self-vehicle position information correction unit 19 is a result of the image recognition processing of the target feature by the image recognition unit 18 and the position of the target feature included in the feature information F about the target feature. It functions as own vehicle position information correcting means for correcting own vehicle position information P based on the information. In the present embodiment, the vehicle position information correction unit 19 includes the image recognition result of the target feature by the image recognition process for position correction based on the feature information F in the image recognition unit 18 and the feature information F included in the feature information F. The own vehicle position information P is corrected along the traveling direction of the own vehicle 50 using the position information of the target feature. Specifically, the own vehicle position information correction unit 19 first, based on the result of the position correction image recognition processing by the image recognition unit 18 and the mounting position, mounting angle, angle of view, and the like of the imaging device 11, The positional relationship between the host vehicle 50 and the target feature when the image information G including the image of the target feature is acquired is calculated. Next, the own vehicle position information correction unit 19 calculates the positional relationship between the own vehicle 50 and the target feature, and the position information of the target feature included in the feature information F about the target feature. Based on the above, the position information of the subject vehicle 50 in the traveling direction of the subject vehicle 50 (feature information F) is calculated and acquired with high accuracy. Then, the own vehicle position information correction unit 19 includes the own vehicle position information P acquired by the own vehicle position information acquisition unit 16 based on the highly accurate position information of the own vehicle 50 acquired in this way. Information on the current position in the traveling direction of the vehicle 50 is corrected. As a result, the host vehicle position information acquisition unit 16 acquires the corrected host vehicle position information P with high accuracy, and the host vehicle 50 is in a highly accurate host vehicle position recognition state.

9. Navigation Calculation Unit The navigation calculation unit 20 operates in accordance with the application program 23 in order to execute navigation functions such as display of the vehicle position, route search from the departure place to the destination, route guidance to the destination, and destination search. Arithmetic processing means. For example, the navigation calculation unit 20 acquires map information M around the host vehicle 50 from the map database DB1 based on the host vehicle position information P, displays a map image on the display input device 21, and displays the map image. On top of this, based on the vehicle position information P, a process of displaying the vehicle position mark in an overlapping manner is performed. Further, the navigation calculation unit 20 searches for a route from a predetermined departure point to a destination based on the map information M stored in the map database DB1. Further, the navigation calculation unit 20 uses one or both of the display input device 21 and the voice output device 22 based on the searched route from the departure point to the destination and the own vehicle position information P, and the driver. Provides route guidance for. In the present embodiment, the navigation calculation unit 20 is connected to the display input device 21 and the audio output device 22. The display input device 21 is a combination of a display device such as a liquid crystal display device and an input device such as a touch panel. The audio output device includes a speaker and the like. In the present embodiment, the navigation calculation unit 20, the display input device 21, and the audio output device 22 function as the guidance information output means 24 in the present invention.

10. Learning Feature Extraction Unit The learning feature extraction unit 31 repeatedly performs image recognition based on a plurality of pieces of recognition information regarding the same location stored in the learning database DB3 by image recognition of the image information G at the same location multiple times. It can function as learning feature extraction means for extracting a recognition target that can be learned as a learning feature and outputting it together with position information of the learning feature. Hereinafter, the detail of the process which the learning feature extraction part 31 performs is demonstrated using FIGS. Here, the learning features are road markings provided on the road surface as described above, stains on the road surface, dirt on the road surface, cracks on the road surface, pavement joints, manhole covers, etc. And shadows on the road surface. In FIG. 5, “stain” is set as the features f1 to f3 as an example of such a learning feature. The recognition information is information representing the recognition result of the recognition target by the image recognition unit 18 and is a so-called feature amount as described above. The recognition target is noise, edge, predetermined color, predetermined shape, or the like included in the image information G, and is a feature on the image information from which a quantitative recognition result (feature amount) can be obtained by image processing.

  The learning feature extraction unit 31 derives the recognition position of the recognition target as information representing the position of the recognition target on the road with reference to the vehicle position information P. That is, the learning feature extraction unit 31 includes an image including a recognition target (feature) based on the image recognition result of the recognition target by the image recognition unit 18 and the mounting position, mounting angle, and angle of view of the imaging device 11. The positional relationship between the host vehicle 50 and the recognition target when the information G is acquired is calculated. Next, the learning feature extraction unit 31 uses the own vehicle position information P based on the calculation result of the positional relationship between the own vehicle 50 and the recognition target and the own vehicle position information P when the image information G is acquired. The position of the recognition target is calculated with reference to. In this example, the learned feature extraction unit 31 obtains this position as a position in a direction along the link k representing the road on which the host vehicle 50 is traveling. Then, the learning feature extraction unit 31 derives the position calculated in this way as the recognition position of the recognition target. Since this recognition position is derived on the basis of the own vehicle position information P when the image information G including the recognition target is acquired, the recognition position becomes position information including an error included in the own vehicle position information P. Therefore, the position information is determined based on the recognition result (recognition information) regarding the same location stored in the learning database DB3 by recognizing the image information G of the same location multiple times.

  FIG. 5 is an example of the case where the host vehicle 50 travels on a road (link k) where the feature information F is not maintained, and outline of the collection processing of the feature information F using the result of the information collection image recognition processing. It is explanatory drawing for demonstrating. As shown in FIG. 5A, on the road on which the host vehicle 50 travels, for example, features f1 to f3 that are stains, dirt, cracks, and the like exist. The learning feature extraction unit 31 uses the recognition position information Aa for the recognition target representing the features of the features f1 to f3 on the image information G, and a predetermined position range (described later) to which the recognition position of the recognition target by the image recognition unit 18 belongs. Is generated as a learning value. And as shown in FIG.5 (b), whenever it recognizes the said recognition target, a learning value is added for every said position range, and it memorize | stores in learning database DB3. The learning feature extraction unit 31 derives an estimated position pg (details will be described later) of the learning feature based on the addition result of the learning value. This estimated position pg becomes position information included in the learned feature information Fa.

  As described above, the learned feature extraction unit 31 derives the recognition position of the recognition target as information representing the position of the recognition target on the road with reference to the own vehicle position information P. The learning feature extraction unit 31 adds a learning value to be described later for each predetermined position range to which the recognition position belongs, and stores it in the learning database DB3. Here, the learning value is a value given when the recognition result (feature amount) of the recognition target satisfies a predetermined learning condition. For example, when the recognition target (feature) is “a color different from a predetermined road surface color due to a stain” and the feature amount is the number of color extractions as a result of color extraction, the number of extractions is equal to or greater than a predetermined threshold value. This is a predetermined learning condition. That is, the condition for determining whether or not the recognition target (feature) is sufficiently characterized with respect to the recognition result (feature value) indicated quantitatively is the learning condition.

  In this example, the predetermined position range is a range that is divided and set at a certain distance in the direction along the link k representing the road, for example, every 0.5 [m] in the direction along the link k. The range is divided. The learning value is a value added to the position range to which the recognition position of the feature to be recognized belongs in the learning database DB3 every time image recognition is successful once. For example, every time image recognition is successful, one point is given. That is, in this example, the recognition position information Aa is configured by information indicating a position range including the recognition position of the feature to be recognized and information of the learning value (for example, “1”). As described above, the recognition position is derived on the basis of the own vehicle position information P when the image information G including the recognition target is acquired, and therefore includes an error included in the own vehicle position information P. However, in this example, a minute error at a position that satisfies the learning condition is absorbed within a predetermined position range.

  Here, FIG. 6 is an enlarged view of a part related to the feature f1 to be recognized of the learning value stored in the learning database DB3 shown in FIG. 5B. For example, in the example of FIG. 5A, the recognition position of the recognition target (feature of the feature f1) derived by the learning feature extraction unit 31 is the position range indicated as “a4” in FIG. In this case, as indicated by the broken line in FIG. 6, 1 is added to the learning value of the position range a4. And if the same recognition object is image-recognized several times by the own vehicle 50 passing the same road several times, as shown in FIG.5 (b) and FIG. The learning values as the plurality of recognition position information Aa generated each time is recognized are accumulated and accumulated for each position range representing the recognition position of the recognition target. As will be described later, when the learning value of the position range a4 is equal to or greater than a predetermined learning threshold T1, the feature information generating unit 36 uses the position representing the position range a4 as a representative point to learn about the recognition target. Object information Fb is generated and stored in the feature database DB2. By performing statistical processing in this way, large errors in positions that satisfy the learning conditions are also suppressed. In the example of FIG. 5, learning feature information Fb1 and Fb3 corresponding to the feature f1 and the feature f3 are stored in the feature database DB2. Note that the learning feature information Fb is not generated for the feature f2 shown in FIG. 5 because the learning value of the image recognition result for the feature of the feature f2 to be recognized does not reach the predetermined learning threshold T1. It is not stored in the feature database DB2.

  As described above, according to the present embodiment, an ambiguous object such as “something is on the road surface” can be used as a learning feature even if it is not a fixed object such as a road marking. Therefore, learning features can be collected at a low cost. Of course, clearer image recognition may be performed, and shapes such as stains, dirt, and cracks may be identified from the results of color extraction and edge extraction, and learning features with reduced ambiguity may be collected. . FIG. 7 shows an example of further recognizing the shape from the results of color extraction and edge extraction, FIG. 7A shows an example of recognizing a road surface stain and dirt, and FIG. 7B shows an example of recognizing a road surface crack. Is shown. As shown in FIG. 7A, when the edge e1 has a closed loop, the attribute of the feature is estimated as a stain or a dirt, and a recognition position corresponding to the attribute is calculated. For example, a rectangular area parallel to the direction along the link k and circumscribing the feature can be set, and the center of gravity position x1 can be obtained based on the vertices y1 to y4, which can be used as the recognition position. As shown in FIG. 7B, when the edge e2 does not have a closed loop, or has a closed loop and the width is narrow, it is estimated that the edge e2 is cracked, and the recognition position corresponding to the attribute is calculated. For example, two end points y5 and y6 of the edge e2 can be obtained, and a midpoint x2 of a straight line connecting y5 and y6 can be set as a recognition position.

  As described above, the features present on the road surface as learning features include road markings provided on the road surface, and the type thereof may be the same as the initial feature. Naturally, various types of road markings that are not registered as initial feature information due to reasons such as low usage frequency, a high degree of freedom in shape, and a large area with respect to the recognition target area are also included. FIG. 8 is an explanatory diagram showing an example in which the characteristics of the entry prohibition area (so-called zebra zone) are recognized. Here, it is assumed that the feature to be recognized is “white area in the predetermined region”, and the feature amount as the recognition result is “ratio of white area in the predetermined region”. The image recognition unit 18 performs, for example, window comparator processing so that the white paint portion W and the road surface can be identified in a predetermined small region r. The learning feature extraction unit 31 sets the learning value to “1” when the paint area W is included in the small area r by a predetermined ratio or more. As described above, the learning value is added to the position range to which the recognition position of the recognition target belongs in the learning database DB3 every time image recognition is successful. The recognition position can be, for example, the center of a circular small region r, and the learning value is added to the position range to which the center of the small region r belongs. FIG. 9 is a diagram showing an example of integration of learning values for each predetermined position range with respect to the zebra zone as shown in FIG. Since the zebra zone has a large area, as shown in FIG. 9, learning values are accumulated in a plurality of small regions r and exceed the learning threshold value T1 with the same score in a plurality of position ranges. Become. In this case, for example, the intermediate position range a7 may be used as position information (estimated position pg) of the learned feature.

  Note that features with features to be recognized are not limited to stains, dirt, cracks, etc. that are regularly present on the road surface, but also include features that appear when certain conditions such as specific times and weather conditions are met. be able to. For example, this corresponds to a shadow. In this case, the recognition information is stored in the learning database DB3 in association with at least one of the recognition time and the weather at the time of recognition. And the learning feature extraction part 31 is the case where the recognition target (shadow feature) capable of recognizing the image repeatedly under a predetermined condition with respect to at least one of the recognition time and the weather at the time of the recognition satisfies the condition. It is extracted as a limited learning feature that is effective only for. Specifically, a limited learning feature can be obtained by including this limitation condition in the feature attribute information Fb described later. A feature such as a shadow has a strong feature as a recognition target due to a strong contrast with a road surface or the like, but an unstable factor is that its appearance is affected by time and weather. However, it can be used as a suitable learning feature by using limited learning features that take into account predetermined conditions such as the time at which such features appear and the weather.

  By the way, the learning feature extraction unit 31 represents the generated recognition position information Aa as various attributes of the recognition target so that the recognition target that has succeeded in the image recognition can be identified with respect to other recognition targets. It is stored in association with the feature attribute information Ab. Here, the recognition target attribute included in the feature attribute information Ab may be anything that can identify the one recognition target from the other recognition target. Therefore, for example, the feature attribute information Ab includes the type of feature as a recognition target, that is, the classification of color, edge, noise, etc., and the shape, size, and recognition target when a shape recognition is performed. 1 or two or more pieces of information selected from the link ID of the link k to be provided. In addition, the condition attribute information Ab may include conditions for limiting limited learning features such as time and weather. Information constituting the feature attribute information Ab for such a feature includes the image recognition result of the recognition target by the image recognition unit 18, the own vehicle position information P at the time of obtaining the image information G related to the image recognition processing, and the like. Is generated based on

11. Estimated position determination unit The estimated position determination unit 34 stores a plurality of pieces of recognition position information Aa for the same location stored in the learning database DB3 by recognizing the image multiple times as image information at the same location corresponds to the recognition target. Based on this, it functions as an estimated position determination means for determining the estimated position pg (see FIG. 5) of the recognition target. Here, the estimated position determination unit 34 determines the representative value of the distribution as the estimated position pg of the recognition target based on the distribution of the plurality of recognized position information Aa regarding the same place. In the present embodiment, the mode value is used as the representative value of the distribution. That is, the estimated position determination unit 34 uses the position representing the position range where the learning value as the recognition position information Aa for each recognition target is initially equal to or greater than the predetermined learning threshold T1 as a representative point. The target estimated position pg is determined. That is, the estimated position determination unit 34 extracts the recognition target in the position range where the learning value is equal to or greater than the predetermined learning threshold T1 as the learning feature, and sets the position representing the position range as the estimated position pg. As an example, a determination method for determining the estimated position pg of the recognition target (feature of the feature f1) in the example of FIG. 5 will be described. As shown in FIG. 6, the learning value as the recognition position information Aa for the feature f1 is initially equal to or greater than the learning threshold value T1 in the position range a4. Therefore, the estimated position determination unit 34 determines a position representing the position range a4, for example, the center position pg4 of the position range a4, as the estimated position pg related to the feature f1.

12 Feature information management unit The feature information management unit 35 manages the feature information F stored in the feature database DB2 based on the learning result of the target feature stored in the learning database DB3. Functions as a means. In the present embodiment, the feature information management unit 35 includes a feature information generation unit 36 and a feature distance determination unit 37. Hereinafter, it demonstrates individually.

12-1. Feature Information Generation Unit The feature information generation unit 36 functions as a feature information generation unit that generates learning feature information Fb based on the learning result of the feature stored in the learning database DB3. In other words, the feature information generation unit 36 includes the position information indicating the estimated position pg of each recognition target determined by the estimated position determination unit 34 and the feature attribute based on the image recognition result by the image recognition unit 18 for the recognition target. Learning feature information Fb associated with the information is generated. Here, the feature attribute information constituting the learned feature information Fb is generated using the content of the feature attribute information Ab stored in the learning database DB3 in association with the recognition position information Aa for the recognition target. . Thereby, the learning feature information Fb is generated as information including the position information and the feature attribute information associated therewith like the initial feature information Fa. The learned feature information Fb generated by the feature information generating unit 36 is stored in the feature database DB2. In the present embodiment, learning feature information Fb5 including attributes of features such as types such as stains, dirt, cracks, road markings (paint), and limited learning feature limitation conditions is generated as feature information generation. Generated by the unit 36 and stored in the feature database DB2.

12-2. Feature distance determination unit The feature distance determination unit 37 determines a feature distance D representing the distance between the two learned features based on a plurality of recognition position information Aa for the two learned features. It functions as a distance determination means between features. In the present embodiment, the feature distance determination unit 37 uses the information on the estimated position pg for the two learned features determined by the estimated position determination unit 34 and uses the estimated position pg of one of the learned features. The distance D between the two learning features is obtained by calculating the distance between the estimated position pg of the other learning features. In the example shown in FIGS. 5B and 5C, the feature distance determination unit 37 calculates a distance between two learning features adjacent to each other. That is, in this example, the feature distance determination unit 37 determines and determines the feature distance D between the feature f1 and the feature f3. Information on the feature distance D determined by the feature distance determination unit 37 is used as information constituting the learned feature information Fb generated by the feature information generation unit 36. That is, in the present embodiment, the learning feature information Fb for each learning feature includes association information indicating the relationship between the learning feature and another learning feature that is in proximity, and the other learning feature. Is generated as having feature distance information representing the feature distance D between the two. These pieces of information also have the same contents as the association information and the inter-feature distance information included in the initial feature information Fa.

13. Operation Process of Navigation Device Next, a feature information collecting method executed in the navigation device 1 according to the present embodiment will be described. FIG. 10 is a flowchart showing an example of the procedure of the feature information collecting method by the navigation device 1 according to the present embodiment. Here, the case where the learning method as described above is used based on FIGS. 5A and 5B will be described as an example.

  As shown in FIG. 10, in the navigation device 1, first, the vehicle position information acquisition unit 16 acquires the vehicle position information P (own vehicle position information acquisition step: # 1). Next, the image information acquisition unit 12 acquires image information G around the host vehicle 50 captured by the imaging device 11 (image information acquisition step: # 2). Thereafter, the feature information acquisition unit 17 performs a period from the current position (vehicle position) of the host vehicle 50 indicated in the host vehicle position information P to the end of the link k representing the road on which the host vehicle 50 is traveling. The feature information F of the existing target feature is acquired from the feature database DB2 (feature information acquisition step: # 3), and the presence / absence of the feature information F is determined (presence determination step: # 4). The feature information F may be either the initial feature information Fa or the learned feature information Fb.

  If it is determined that there is the feature information F, the host vehicle 50 is traveling on the road where the feature database DB2 is maintained. Therefore, the image recognition unit 18 performs the above-described position correction image recognition process (image recognition process (position correction image recognition process): # 5). That is, the image recognition unit 18 performs the image recognition process by setting a recognition area for each target feature indicated by the acquired feature information F. When the target feature is image-recognized in the recognition area, the target feature is automatically recognized based on the image recognition result of the target feature and the position information of the target feature included in the feature information F about the target feature. The vehicle position information correction unit 19 corrects the vehicle position information P (own vehicle position information correction step (not shown)). Thereby, the own vehicle 50 will be in a highly accurate own vehicle position recognition state.

  On the other hand, if it is determined that there is no feature information F (# 4), the navigation device 1 executes a feature information collection image recognition process (# 6). First, the image recognition unit 18 performs image processing of a recognition target included in the image information G. When the recognition target is image-recognized, it is determined that there is a feature that is a candidate for the learning feature, and when the recognition target is not image-recognized, the process is terminated (# 7). As described above, steps # 6 and # 7 correspond to the image recognition process (image recognition process for collecting feature information) of the present invention.

  When the recognition target is image-recognized (# 7), the learning feature extraction unit 31 derives the recognition position of the feature based on the own vehicle position information P acquired in step # 1. And the positional information and learning value of the said recognition object are memorize | stored in learning database DB3 (recognition result memory | storage process: # 8). That is, the recognition position information Aa representing the recognized recognition position of the feature is generated as a learning value for the predetermined position range to which the recognition position belongs, and as shown in FIG. 5B and FIG. It is added every time and stored in the learning database DB3.

  Next, it is determined whether or not the learning value as the recognition position information Aa for the target feature stored in the learning database DB3 is equal to or greater than a predetermined learning threshold T1 (# 9). If the learning value is less than the predetermined learning threshold value T1, the process ends. On the other hand, when the learning value as the recognition position information Aa for the target feature stored in the learning database DB3 is equal to or greater than a predetermined learning threshold T1 (# 9), the estimated position determination unit 34 The estimated position pg of the feature is determined. After that, the feature information generation unit 36 generates learning feature information Fb in which the estimated position pg for the recognition target is associated with the feature attribute information based on the image recognition result (# 10). The generated learned feature information Fb is stored in the feature database DB2 (# 11). As described above, steps # 9 to # 11 correspond to the learned feature extraction step of the present invention.

14 Other Embodiments (1) In the above embodiment, as shown in FIG. 5B, the position information of the learned features is estimated using a method of adding learning values for a predetermined position range. However, without being limited to this method, the learning feature extraction unit 31 may estimate the position information of the learning feature using the following method. The learning database DB3 stores the quantitative recognition result (feature value) obtained by the image recognition processing performed on the recognition target as a raw value as shown in FIG. 5C. That is, the learning database DB3 continuously stores the raw value of the feature value as the recognition result of the recognition target in association with the position information. Then, the learning feature extraction unit 31 extracts the recognition target at the position where the peak exists as a learning feature based on the distribution of the recognition result of the stored recognition target, and obtains information on the position where the peak exists. The position information (estimated position pg) of the learned feature is assumed.

  In FIG. 5, stains as learning features are illustrated as the features f1 to f3, and the features on the image to be recognized are, for example, a color different from a predetermined road surface color or a boundary with a road surface color. is there. When window comparator processing is performed on the road surface color as an extraction target as the image recognition processing, a color different from the road surface color can be extracted. FIG. 5C shows the relationship with the recognition position using the extracted number of colors different from the road surface extracted as described above as the feature amount (recognition result). Since the features f1 and f3 are large stains, the feature amount is large, and the feature f2 is a small stain, so the feature amount is small. For example, the learning feature extraction unit 31 has a feature amount equal to or greater than a predetermined feature amount threshold value T2 in the distribution of feature amounts, and if the peak exists, (Stains) are extracted as learning features. Further, information on the position where the peak exists is assumed to be position information (estimated position pg) of the learned feature. As described above, the recognition position of the recognition result is derived on the basis of the own vehicle position information P at the time of acquisition of the image information G including the recognition target. Therefore, the position information including the error included in the own vehicle position information P. It is. In the learning database DB3, the raw values of the feature values of each time when traveling a plurality of times are accumulated and stored. Therefore, the position where the peak of the feature value exists absorbs the error, and the learning feature is accurately obtained. Position information can be estimated.

  This method can also be used to determine whether or not to learn. For example, when water droplets are attached to the lens of the imaging device 11 during stormy weather, the image information G is disturbed. As a result, the distribution of recognition results is poorly changed, and the recognition results at this time may be excluded from the storage target for learning.

(2) As a feature having a feature as a recognition target, a road sign having a simple shape can be used although the frequency of use is low and it is not registered as initial feature information. Such a road marking shape can be set as a predetermined shape, and the predetermined shape can be a recognition target. For example, a triangular mark indicating “there is a priority road ahead”, and individual characters of an instruction display drawn on the road surface correspond to this road marking.

(3) The feature as the recognition target may be a road marking based on the result of performing extraction processing of a predetermined color with the feature as the recognition target being the “ratio of white on the road surface” May be determined. If there is a high possibility of being a road marking, a shape recognition algorithm for a road marking with a high appearance frequency (such as a pedestrian crossing and a stop line) is executed to collect learning features with higher accuracy. May be.

(4) In the above embodiment, the case where the recognition position of the recognition target represented by the recognition position information Aa is derived as information representing the position of the feature on the road with the vehicle position information P as a reference will be described as an example. did. However, the recognized position information Aa according to the present invention is not limited to such position information. Therefore, for example, the recognition position represented by the recognition position information Aa may be the position of the own vehicle 50 indicated by the own vehicle position information P when the image information G including the recognition target is acquired. One of the forms. In this case, when the learned feature information Fb is generated by the feature information generation unit 36, the image information G in the image information G with respect to the position of the host vehicle 50 is based on the mounting position, mounting angle, angle of view, and the like of the imaging device 11. It is preferable that the position of the feature to be recognized on the road is calculated and the position of the feature having the feature as the recognition target on the road is the position information included in the learned feature information Fb.

(5) In the above-described embodiment, an example has been described in which one learning value is added each time image recognition is successful. However, the embodiment of the present invention is not limited to this, and different points may be added according to the recognition result. For example, as described above with reference to FIG. 8, when “white area in a predetermined area” is a feature to be recognized and the feature amount as a recognition result is “ratio of white area in the predetermined area” Depending on the “ratio”, different points may be added. For example, the score may be varied according to the recognition result, such as 1 point for 50% or more and 3 points for 70% or more.

(6) In the above embodiment, the case where all the configurations of the feature information collection device 2 according to the present invention are mounted on the host vehicle 50 has been described as an example. However, the embodiment of the present invention is not limited to this. That is, a part of the configuration of the feature information collection device 2 including the learning database DB3 as the recognition result storage means of the present invention communicates with a plurality of vehicles 50 via a wireless communication line or the like as shown in FIG. One of the preferred embodiments of the present invention is a configuration in which the server devices 60 are connected to each other. If the feature information collection device 2 has such a configuration, the learning results of features by the plurality of vehicles 50 can be accumulated in the learning database DB3 installed in the server device 60. Therefore, it is possible to generate the feature information F having more accurate position information by using more recognized position information Aa. Note that the configuration of the feature information collection device 2 installed in the server device 60 is not limited to the learning database DB3, and needs to be installed in the host vehicle 50, such as the imaging device 11 or the host vehicle position information acquisition unit 16. All configurations other than the configuration can be installed in the server device 60. Similarly, it is also one preferred embodiment of the present invention that part of the configuration of the image recognition device 3, the vehicle position recognition device 4, and the navigation device 1 is installed in the server device 60.

(7) In each of the above embodiments, an example in which the image recognition device 3 and the vehicle position recognition device 4 are used for the navigation device 1 has been described. However, the application range of the present invention is not limited to this, and can naturally be used for other purposes such as a vehicle travel control device.

  The present invention can be applied to a feature information collection device that collects information on features included in image information around a vehicle acquired by an imaging device or the like mounted on the vehicle by performing image recognition processing. The device can be used in a vehicle position recognition device, a navigation device, a vehicle control system, and the like.

1: Navigation device 2: Feature information collection device 3: Image recognition device 4: Own vehicle position recognition device 12: Image information acquisition unit (image information acquisition means)
16: Own vehicle position information acquisition unit (own vehicle position information acquisition means)
17: Feature information acquisition unit (feature information acquisition means)
18: Image recognition unit (image recognition means)
19: Own vehicle position information correction unit (own vehicle position information correction means)
23: Application program 24: Guidance information output means 31: Learning feature extraction unit (learning feature extraction means)
34: Estimated position determination unit (estimated position determination means)
35: Feature information management department (feature information management means)
36: Feature information generation unit (feature information generation means)
37: Distance determination unit between features (distance determination unit between features)
39: Presence determining unit (presence determining means)
50: Own vehicle Aa: Recognition position information (position information of learned features)
Ab: feature attribute information DB1: map database (road information storage means)
DB2: Feature database (feature information storage means)
DB3: Learning database (recognition result storage means)
D: feature distance F: feature information G: image information P: own vehicle position information T1: learning thresholds a1 to a12: position range

Claims (15)

Own vehicle position information acquisition means for acquiring own vehicle position information representing the current position of the own vehicle;
Image information acquisition means for acquiring image information around the host vehicle;
Image recognition means for performing image recognition processing of a recognition target included in the image information;
Recognition result storage means for storing recognition information representing the recognition result of the recognition target by the image recognition means in association with the recognition position information of the recognition target based on the vehicle position information;
Based on a plurality of pieces of recognition information relating to the same place, which are stored in the recognition result storage means by recognizing the image information of the same place a plurality of times, a recognition target capable of repeatedly recognizing an image is learned. Learning feature extraction means for extracting as and output together with position information of the learning feature;
A feature information collecting apparatus comprising: The recognition result storage means continuously stores the recognition result of the recognition target in association with the information of the recognition position;
The learning feature extracting unit extracts a recognition target at a position where a peak exists as the learning feature based on a recognition result distribution of the recognition target stored in the recognition result storage unit, and the peak exists. The feature information collecting apparatus according to claim 1, wherein the position information is the position information of the learned feature. The recognition result storage means obtains a learning value given when a recognition result of a recognition target satisfies a predetermined learning condition, and a position of a point that satisfies the learning condition every time the recognition result of the recognition target satisfies the learning condition. Is added and stored for each predetermined position range to which
The learning feature extraction unit extracts a recognition target of a position range where the learning value is equal to or greater than a predetermined learning threshold as the learning feature, and acquires information on a position representing the position range as the learning place. The feature information collecting apparatus according to claim 1, wherein the feature information collecting device is position information of an object. The recognition information is stored in association with information of at least one of recognition time and weather at the time of recognition,
The learning feature extraction means is a limited learning that is effective only when a recognition target capable of repeatedly recognizing an image under a predetermined condition with respect to at least one of recognition time and weather at the time of recognition is satisfied only when the condition is satisfied. The feature information collecting apparatus according to any one of claims 1 to 3, wherein the feature information collecting device is extracted as a feature.   5. The inter-feature distance determination means for determining a distance between features that represents a distance between the two learned features based on position information of the two learned features. 5. The feature information collection device described.   The feature information collection device according to any one of claims 1 to 5, wherein the recognition result storage unit is communicably connected to a plurality of vehicles and stores the recognition information by the plurality of vehicles.   The learning feature includes at least one of a road marking provided on a road surface, a stain on the road surface, dirt on the road surface, cracks on the road surface, and shadows on the road surface. The feature information collection device according to any one of claims 1 to 6.   The feature information collection device according to claim 1, wherein the recognition target includes at least one of noise, an edge, a predetermined color, and a predetermined shape included in image information. Feature information storage means for storing, as feature information of a target feature, feature information that associates attribute information related to the recognition information about the learned feature and position information of the learned feature;
Or
Attribute information related to the recognition information about the learned feature and feature information that associates the position information of the learned feature, and feature information including position information and attribute information prepared in advance for a plurality of features The feature information collection device according to any one of claims 1 to 8, further comprising feature information storage means for storing both as feature information of the target feature. Further, presence determination means for determining whether or not the target feature stored as the feature information exists in the vicinity of the own vehicle by referring to the feature information storage means based on the own vehicle position information. ,
The feature information collection device according to claim 9, wherein the recognition information is stored in the recognition result storage unit when the presence determination unit determines that the target feature does not exist around the host vehicle. . A feature information collecting device according to claim 9 or 10, and own vehicle position information correcting means for correcting the own vehicle position information based on an image recognition result of the target feature,
The image recognizing means includes the feature information included in the image information based on the feature information of the target feature existing around the subject vehicle extracted from the feature information storage means based on the own vehicle position information. Perform image recognition processing of the target feature indicated by the object information,
The own vehicle position information correction unit corrects the own vehicle position information based on the image recognition result of the target feature by the image recognition unit and the position information included in the feature information of the target feature. The vehicle position recognition device.   12. The vehicle position recognition device according to claim 11, map information storage means storing map information, an application program operating with reference to the map information, and operating according to the application program to output guidance information A navigation information output means. Own vehicle position information acquisition step for acquiring own vehicle position information indicating the current position of the own vehicle;
An image information acquisition step of acquiring image information around the host vehicle;
An image recognition process for performing image recognition processing of a recognition target included in the image information;
A recognition result storage step of storing recognition information representing a recognition result of a recognition target in the image recognition step in a recognition result storage unit in association with information of the recognition position of the recognition target based on the vehicle position information;
Based on a plurality of pieces of recognition information relating to the same place, which are stored in the recognition result storage means by recognizing the image information of the same place a plurality of times, a recognition target capable of repeatedly recognizing an image is learned. And a learning feature extraction step of outputting with the position information of the learning feature,
Feature information collection method comprising: In the recognition result storage step, the recognition result of the recognition target is continuously stored in the recognition result storage unit in association with the information of the recognition position,
In the learning feature extraction step, the recognition target at the position where the peak exists is extracted as the learning feature based on the distribution of the recognition result of the recognition target stored in the recognition result storage unit, and the peak exists. The feature information collecting method according to claim 13, wherein the position information is the position information of the learned feature. In the recognition result storage step, the learning value given by the recognition result of the recognition target satisfying the predetermined learning condition is obtained, and the position of the point that satisfies the learning condition every time the recognition result of the recognition target satisfies the learning condition. Is added to each predetermined position range to which the
In the learning feature extraction step, a recognition target in a position range where the learning value is equal to or greater than a predetermined learning threshold is extracted as the learning feature, and position information representing the position range is extracted from the learning place. The feature information collecting method according to claim 13, wherein the feature information is the position information of an object.

Images