JP5333862B2 - Vehicle position detection system using landscape image recognition - Google Patents

Description

  The present invention relates to a vehicle position detection system that obtains the current position of a running vehicle using a landscape image taken from the vehicle.

  Conventionally, in the technical field of car navigation, as a method of calculating the current position of a vehicle, a method of using information acquired from a sensor such as a gyroscope or a geomagnetic sensor (autonomous navigation), a method of using a signal from a GPS satellite, Or the method of combining autonomous navigation and GPS is adopted. Furthermore, in order to calculate the current position with high accuracy, the provisional current position is obtained using a signal from a positioning satellite, and the provisional current position is used as a reference using the captured image in front of the vehicle. The coordinate of the feature point of the road marking (automotive coordinate system feature point) in the coordinate system (car coordinate system) is calculated, and the calculated coordinate point of the car coordinate system and the coordinate of the feature point of the stored road sign (world 2. Description of the Related Art Position positioning devices configured to calculate the current position of a vehicle using a coordinate system) are known (for example, see Patent Document 1). In this apparatus, even if positioning based on signals from positioning satellites and signals from various sensors includes an error, it is possible to calculate a current position with high accuracy. However, in the positioning device according to Patent Document 1, the spatial coordinates of the feature points of the road marking on the road are obtained from the stereo image, and obtained from the latitude / longitude of the road marking having the feature point stored in the road marking information database. Since the vehicle position is calculated using the coordinates obtained, it cannot be used in a place where there is no road marking. In addition, since it is necessary to calculate the spatial coordinates of the feature points recognized by the image processing, a high calculation capability is required for the apparatus, resulting in a cost increase.

  Also, in the same car navigation technology field, images with a high similarity are searched from an image database using a photographed image of a building or landscape as a search key, and one or more searched images are displayed together with corresponding point specifying information. There is known a navigation device that presents to a user and sets position information associated with an image selected by the user viewing this display as a destination (see, for example, Patent Document 2). When this navigation device is used, if a captured image taken at a certain point is input and a search is executed, an image that matches the input image and the corresponding point specifying information are hit in the search and presented to the user, The point is set as the destination by the user's selection. This makes it possible to set the destination using the captured image of the point even if the information such as the position of the point to be set as the destination and the name of the facility is not known at all. However, in specifying the position using such a landscape photographed image, the navigation device presents an image that matches the input image as a search key, and the user needs to confirm it. It cannot be applied to vehicle position detection.

JP 2007-108043 A (paragraph number 0009-0013, FIG. 1) Japanese Patent Laying-Open No. 2004-333233 (paragraph numbers 0010-0043, FIG. 1)

  From the above situation, it is desired to realize a vehicle position detection system that can efficiently detect the vehicle position without requiring human judgment while using the landscape image recognition technology.

  The feature configuration of the vehicle position detection system using landscape image recognition according to the present invention includes image feature point data generated by extracting image feature points from a captured image obtained by photographing a landscape from a vehicle, and the image feature points. By inputting the reference data database storing reference data relating the shooting position of the shot image corresponding to the data and the actual shot image of the landscape by the in-vehicle camera, and extracting the image feature points from the actual shot image The captured image processing unit that outputs the generated matching data, the vehicle behavior acquisition unit that acquires the vehicle behavior data indicating the actual vehicle behavior that is the vehicle behavior of the vehicle, and the estimated vehicle position of the vehicle as an extraction condition Preliminary reference data having a shooting position around the estimated vehicle position is extracted from the reference data database, and the preliminary data is extracted based on the actual vehicle behavior. A reference data output unit that narrows down the reference data from the reference data and outputs it as a matching partner candidate for the matching data, a matching execution unit that performs matching between the output reference data and the matching data, and the matching And a vehicle position determination unit that determines the vehicle position based on the photographing position related to the reference data that has been successfully completed.

  Since an error occurs when the position of the vehicle is estimated in the measurement of position coordinates by GPS or dead reckoning, the actual vehicle position (actual vehicle position) is a predetermined radius region (hereinafter referred to as this) based on the estimated vehicle position. Is referred to as an estimation error range). Assuming that the vehicle exists only on the road or an equivalent position, the predetermined radius area can be reduced to a shape along the road, for example. The term “estimated error range” (usually simply abbreviated as “error range”) is used. When this estimated error range is large, it is necessary to extract all the reference data belonging to the range from the reference data database and make it a matching partner. However, in the vehicle position detection system according to the present invention, vehicle behavior data indicating the actual vehicle behavior that is the vehicle behavior of the vehicle in real time is acquired, and even the reference data belonging to the estimated error range matches the actual vehicle behavior. Reference data that is considered not to be, for example, reference data that is not considered to belong to the shooting position where the actual vehicle behavior occurs can be excluded from the matching partner. In this way, efficient landscape image recognition is realized by excluding reference data that is not suitable as a matching partner.

  As one preferred embodiment of the present invention, the reference data output unit obtains a shooting position range where the actual vehicle behavior occurs based on map data, and uses the shooting position range from the preliminary reference data. Some are configured to narrow down the reference data. This configuration is advantageous because the shooting position range where the actual vehicle behavior occurs is obtained based on the map data, and dedicated data for defining the relationship between the shooting position range and the vehicle behavior is unnecessary.

  As a specific vehicle behavior handled in the present invention, it is advantageous that the vehicle behavior can be specified according to the road condition or the road condition. Therefore, in one preferred embodiment of the present invention, the vehicle behavior is a vehicle behavior resulting from a specific road condition on which the vehicle travels. Examples of such vehicle behavior include curve traveling, uphill downhill traveling, deceleration traveling, and acceleration traveling. In curve driving, a predetermined steering angle can be detected, and in uphill driving and downhill driving, an acceleration increase / decrease or a vehicle body inclination angle by a level meter can be detected. Further, in deceleration traveling and acceleration traveling, a predetermined acceleration can be detected by an acceleration sensor, and it is more preferable to stop the vehicle. This vehicle stop can be easily detected by a distance sensor or the like. Furthermore, the vehicle stop has a relationship with an intersection, a railroad crossing, and a pedestrian crossing, and can be suitably used.

It is a schematic diagram explaining the basic concept of the own vehicle position detection technique in the own vehicle position detection system using landscape image recognition according to the present invention. It is a functional block diagram which shows the main functions of the image processing system which produces | generates the reference data used for the own vehicle position detection system by this invention. It is a schematic diagram which shows typically the adjustment of the weighting coefficient using an adjustment coefficient. It is a functional block diagram which shows the function of the car navigation system which employ | adopted the own vehicle position detection system by this invention. It is a schematic diagram for demonstrating typically the process which narrows down the reference data group which belongs to an estimation error range based on a vehicle behavior comparison for a curved road as an example. It is a schematic diagram for demonstrating typically the process which narrows down the reference data group which belongs to an estimation error range based on a vehicle behavior comparison for an intersection. It is a schematic diagram for demonstrating typically the process which narrows down the reference data group which belongs to an estimation error range based on a vehicle behavior comparison for an uphill road. It is a functional block diagram which shows the function of a scenery matching part.

  Hereinafter, the present invention will be described in detail with reference to the drawings. FIG. 1 shows an image recognition of a landscape image from an in-vehicle camera arranged so as to photograph the front in the vehicle traveling direction through a matching process using reference data created by the image processing system according to the present invention. 3 schematically shows the basic concept of position positioning technology for determining the position where the landscape image is taken, that is, the position of the vehicle.

  First, the construction procedure of the reference data database 92 (hereinafter simply referred to as “reference data DB”) will be described. As shown in FIG. 1, a captured image obtained by capturing a landscape from a vehicle during traveling, photographing attribute information including a photographing position and a photographing direction at the time of photographing, and vehicle behavior at the time of photographing (for traveling on a curved road) A predetermined steering angle or the like is input (# 01). A feature point detection process for detecting an image feature point, for example, an edge detection process is performed on the input photographed image (# 02). Here, an edge point corresponding to one pixel or a plurality of pixels constitutes a line segment such as an outline is called an edge, and an intersection point where a plurality of edges intersect is called a corner. However, an example of the image feature point is the edge and the corner. An edge including a corner is extracted as an image feature point from the edge detection image obtained by the edge detection process (# 03).

  In consideration of being used for landscape image recognition, not all extracted image feature points have the same importance. For example, the importance of each image feature point may differ depending on the coordinate position in the corresponding captured image. Therefore, while applying rules such as reducing the importance of image feature points that are not suitable for landscape image recognition and increasing the importance of image feature points that are important for landscape image recognition, the importance of each image feature point It is preferable to determine the degree (# 04). When the importance of each image feature point is determined, a weighting coefficient matrix that defines the assignment of the weighting coefficient to each image feature point according to the importance is generated (# 05).

  Subsequently, using the weighting coefficient matrix, the image feature points are organized to generate image feature point data for each captured image (# 07). In this image feature point data generation process, image feature points having a weighting factor equal to or lower than a predetermined threshold level are discarded, or image feature points having a weighting factor equal to or higher than a predetermined threshold level and surrounding image feature points Sorting process to discard other than is performed. Since the image feature point data generated here is used as a pattern when pattern matching is adopted for the purpose of landscape image recognition, it is necessary to provide only image feature points effective in pattern matching of landscape images. Important for speed and accuracy. The generated image feature point data can be used as a search key by relating the shooting position, shooting direction, and vehicle behavior of the shot image corresponding to the image feature point data to the vehicle behavior. Data for the database is obtained (# 08). That is, the image feature point data is stored in the reference data used for landscape image recognition, that is, the reference data DB 92 to be extracted as a matching partner (# 09).

  Next, a procedure for determining the position of the vehicle (vehicle position) during actual vehicle travel will be described using the reference data DB constructed by the above-described procedure. As shown in FIG. 1, an estimated own vehicle position estimated using a GPS measurement unit, an actual captured image captured by an in-vehicle camera with the estimated own vehicle position as a capturing position, its capturing orientation, and at the time of capturing Vehicle behavior data that is actual vehicle behavior is input (# 11). First, matching data, which is image feature point data, is generated from the input photographed image through the processing procedures of steps # 02 to # 07 (# 12).

  Further, an estimated error range (hereinafter simply referred to as an error range) defined by the error of the estimated own vehicle position that is also the input photographing position is calculated (# 13a). Further, the input vehicle behavior data is evaluated, and vehicle behavior information (hereinafter simply referred to as actual vehicle behavior) that can be used as a search condition in the reference data DB 92 is generated (# 13b). The error range and the actual vehicle behavior are used to extract and output a reference data group appropriate as a matching partner of the matching data from the reference data DB 92 (# 13c) (# 13d). The output reference data is set for matching processing that is landscape image recognition (# 14).

  Reference data is set as a pattern one by one from the extracted matching candidate reference data string, and pattern matching processing with the currently generated matching data is executed as landscape image recognition (# 15). If the matching is successful, the shooting position associated with the reference data that is the target is read (# 16), and this shooting position is determined as the official vehicle position that replaces the estimated vehicle position (#). 17).

Next, an example of an image processing system that creates reference data used in the above-described vehicle position detection system from a photographed image obtained by photographing a landscape from a vehicle will be described. FIG. 2 is a functional block diagram schematically showing the functions of such an image processing system.
This image processing system includes a data input unit 51, a feature point extraction unit 52, a feature point importance degree determination unit 53, a weighting unit 55, an adjustment coefficient setting unit 54, an image feature point data generation unit 56, and a reference data database generation unit 57. , Etc., and each of these functions can be created by hardware, software, or a combination thereof.

  The data input unit 51 includes a photographic image obtained by photographing a landscape by a camera mounted on a vehicle traveling for reference data creation, photographing attribute information including a photographing position and a photographing direction at the time of photographing, and a photographing situation. Information is input as shooting-related information. Furthermore, vehicle behavior data indicating the vehicle behavior at the time of shooting is also input. The vehicle behavior data includes, for example, a steering angle (steering angle), a vehicle front-rear direction inclination (that is, a road gradient) calculated based on a signal from an acceleration sensor or a signal from a level. In the form in which the image processing system is mounted on the traveling vehicle, the captured image, the photographing attribute information, and the photographing situation information are input to the input unit 51 in real time. In such a configuration, the photographed image, photographing attribute information, and photographing state information are temporarily recorded on a recording medium, and these data inputs are performed in a batch process. Since the method for generating the photographed image and the photographing attribute information is well known, the description thereof is omitted.

  The shooting situation information is information indicating the possibility that a specific subject is included in the shot image, and the contents included in the shooting situation information of this embodiment include travel lane data, moving object data, and area attribute data. It is. The traveling lane data is data indicating the traveling lane region of the own vehicle and the region outside the road in the photographed image obtained from the recognition result of the white line, the guide rail, and the safety zone obtained through image processing on the photographed image. The moving object data is data indicating an existing area in a captured image of a moving object present around the vehicle that is recognized by an in-vehicle sensor that detects an obstacle such as a radar. Area attribute data is data indicating area attributes such as mountain area, suburban area, urban area, high-rise city area, etc. is there.

  The feature point extraction unit 52 extracts edge points as image feature points from the captured image using an appropriate operator. The feature point importance determination unit 53 determines the importance of the image feature points extracted by the feature point extraction unit 52 based on the contents of each data included in the shooting situation information. For example, when the content of the travel lane data is used, higher importance is assigned to image feature points belonging to a region further away from the travel lane closer to the shoulder in the captured image. When moving object data is used, a low importance is assigned to image feature points belonging to a region where a moving object exists in a captured image. Further, when the contents of the area attribute data are used, the importance assigning rule corresponding to the position in the captured image is changed according to the area attribute. For example, in a captured image of a mountain area, since there is a high possibility that the upper part of the photographing center optical axis is empty and the left and right are forests, high importance is set for the central region around the photographing central optical axis. In the photographed image of the suburban area, there is little traffic of cars, and structures such as houses spread around, so a high importance is set for the area below the photographing center optical axis. In the captured image of the urban area, since there is a lot of traffic, a high importance is set for the area above the optical axis of the image capturing center. In the photographed image of the high-rise city area, since there are many elevated roads and viaducts, high importance is set for the region above the photographing center optical axis.

  The weighting unit 55 assigns a weighting factor to the image feature point according to the importance determined by the feature point importance determining unit 53. High importance is set for image feature points that are considered important for accurate image recognition (pattern matching), so a large weighting factor is assigned to image feature points with high importance. Considering that image feature points with low weighting factors are not likely to be used in actual image recognition or deleted from reference data, they can be used to determine the selection of image feature points Calculated.

  The adjustment coefficient setting unit 54 calculates an adjustment coefficient for changing the weighting coefficient assigned by the weighting unit 55 from the viewpoint of the distribution state in the corresponding captured image region. In other words, the importance determined based on the shooting situation information for the image feature points extracted by the feature point extraction unit 52 includes a certain degree of error, and image feature points having a certain degree of importance are also randomly generated. It is also possible that For this reason, when there is an uneven distribution of image feature points, in other words, an uneven distribution of the weighting coefficient assigned by the weighting unit 55, the adjustment coefficient setting unit 54 is used for the purpose of easing the uneven distribution. When the distribution of image feature points obtained by calculation processing indicates the uneven distribution of image feature points, the adjustment coefficient is set so that the weighting factor of image feature points belonging to the region where the density of image feature points is small is increased. The adjustment coefficient is set so that the weighting coefficient of the image feature point belonging to the region where the density of the image feature point is large is small.

  The image feature point data generation unit 56 organizes the image feature points based on the weighting coefficient assigned by the weighting unit 55 and the adjustment coefficient assigned in some cases, and generates image feature point data for each captured image. To do. At this time, the image feature points can be narrowed down so that the matching process is efficiently performed by deleting the image feature points having a weighting coefficient equal to or less than a predetermined threshold value. In addition, this weighting factor is attached to the image feature point data so that it can be related to each image feature point in the reference data as it is, and for weighting similarity calculation at the time of matching processing using the reference data with the weighting factor. It may be used for.

  Here, a process of spreading image feature points as widely as possible over the entire photographed image area using the above-described adjustment coefficient will be described with reference to a schematic explanatory diagram shown in FIG. A feature point image (FIG. 3B) is generated by extracting image feature points from the photographed image (FIG. 3A), and importance is given to each image feature point of the feature point image. 3C, the importance corresponding to each image feature point is shown in the form of the importance layer corresponding to the feature point image so that the degree to which the degree is given can be schematically understood. 3D, a weighting factor is assigned to each image feature point, and in FIG.3 (d), the weighting factor is in the form of a feature point image in which image feature points are drawn such that the larger the weighting factor, the larger the point. Here, when image feature points are arranged such that image feature points assigned weighting factors equal to or less than a predetermined threshold are removed, for example, FIG. Only image feature points that are large in 3 (d) are selected. Then, the image feature points located in the lower region of the feature point image are excluded, and the distribution of the remaining image feature points is greatly uneven. The adjustment coefficient is set so as to increase the weighting coefficient of the image feature points in the region where the density of the image feature points to be selected is reduced as a result. As can be understood, in FIG. 3 (e), the adjustment coefficient group is shown in the form of an adjustment coefficient layer arranged in a matrix (here, in units of a plurality of pixel areas) so as to correspond to the feature point image. The image feature point data generation unit 56 organizes each image feature point using the weighting coefficient finally set based on such weighting coefficient and adjustment coefficient, and is shown in FIG. Image feature point data It is generated for each shadow image.

  The reference data database unit 57 associates the image feature point data generated by the image feature point data generation unit 56 with the shooting attribute information of the shot image corresponding to the image feature point data, and recognizes the landscape image. Create a database as reference data to be used. At this time, in order to assist landscape image recognition, information indicating the vehicle behavior at the time of shooting, which is extracted from the vehicle behavior data as the recorded vehicle behavior, is also related to the reference data. The reference data stored in the database in this way is stored in the reference data DB 92.

  In the above description, the degree of importance is determined for each image feature point, and as a result, a weighting factor is set for each image feature point. However, these processes can be performed in units of groups. is there. In that case, for example, the captured image area is divided into a plurality of image sections, and the feature point importance determination unit 53 groups image feature points belonging to the same image section as an image feature point group and handles them in a unified manner. The image feature points included in the image feature point group may be given the same importance, and the weighting unit 55 may similarly set the weight coefficient in units of image feature point groups. In addition, the image sections handled here may be handled in units of one pixel constituting the captured image, but may be handled in units of a plurality of pixels. Therefore, in the present invention, the image section is composed of one or a plurality of pixels.

  Next, an in-vehicle car that corrects the vehicle position by landscape image recognition (image feature point pattern matching) using the basic concept of the vehicle position detection system described above and the reference data DB 92 created by the image processing system described above. A navigation system will be described. FIG. 4 is a functional block showing such a car navigation system incorporated in an in-vehicle LAN. This car navigation system includes an input operation module 21, a navigation control module 3, a vehicle position detection module 4, a photographing situation information generation unit 7, a road map DB 91 (hereinafter simply referred to as “car map navigation road map data”). A database 9 having a vehicle map and a vehicle behavior detection module 24. The vehicle behavior detection module 24 receives sensor signals and control signals indicating various behaviors of the vehicle such as vehicle braking, acceleration, and steering through the in-vehicle LAN, and generates and outputs vehicle behavior data including various vehicle behaviors. It has a function.

  The navigation control module 3 includes a route setting unit 31, a route search unit 32, and a route guide unit 33. The route setting unit 31 sets, for example, a departure point such as the vehicle position, an input destination, a passing point, and traveling conditions (whether or not an expressway is used). The route search unit 32 is a processing unit that performs arithmetic processing for searching for a guide route from the departure point to the destination based on the conditions set by the route setting unit 31. The route guidance unit 33 provides appropriate route guidance to the driver by guidance display on the display screen of the monitor 12 or voice guidance by the speaker 13 according to the route from the departure place to the destination searched by the route search unit 32. It is a process part which performs the arithmetic processing for performing.

  The own vehicle position detection module 4 corrects the estimated own vehicle position obtained by the conventional position calculation by GPS and the position calculation by dead reckoning with the own vehicle position determined by landscape image recognition using the estimated own vehicle position. It has the function to do. The own vehicle position detection module 4 includes a GPS processing unit 41, dead reckoning processing unit 42, own vehicle position coordinate calculation unit 43, map matching unit 44, own vehicle position determination unit 45, captured image processing unit 5, and landscape matching unit 6. I have. A GPS measurement unit 15 that receives GPS signals from GPS satellites is connected to the GPS processing unit 41. The GPS processing unit 41 analyzes the signal from the GPS satellite received by the GPS measurement unit 15, calculates the current position (latitude and longitude) of the vehicle, and sends it to the own vehicle position coordinate calculation unit 43 as GPS position coordinate data. The dead reckoning processing unit 42 is connected to the distance sensor 16 and the azimuth sensor 17. The distance sensor 16 is a sensor that detects a vehicle speed or a moving distance of the vehicle. For example, a vehicle speed pulse sensor that outputs a pulse signal every time a drive shaft, a wheel, or the like of the vehicle rotates by a certain amount detects the acceleration of the host vehicle C. And a circuit that integrates the detected acceleration. The distance sensor 16 outputs information on the vehicle speed and moving distance as the detection result to the dead reckoning processing unit 42. The direction sensor 17 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, and the direction of the detection result Information is output to dead reckoning processing unit 42. The dead reckoning processing unit 42 calculates dead reckoning position coordinates based on the moving distance information and the direction information sent every moment, and sends the dead reckoning position coordinate data to the own vehicle position coordinate calculating unit 43 as dead reckoning position coordinate data. The own vehicle position coordinate calculation unit 43 performs an operation for specifying the position of the vehicle from the GPS position coordinate data and the dead reckoning position coordinate data by a known method. The calculated vehicle position information is information including measurement errors and the like, and in some cases, the vehicle position information is off the road. Therefore, the map matching unit 44 determines the vehicle position on the road indicated on the road map. Correction is performed. The vehicle position coordinates are sent to the vehicle position determination unit 45 as the estimated vehicle position.

  The photographed image processing unit 5 substantially includes most of the functional units that constitute the image processing system shown in FIG. The captured image processing unit 5 includes a data input unit 51, a feature point extraction unit 52, a feature point importance degree determination unit 53 and a weighting unit 55, an adjustment coefficient setting unit 54, and an image feature point data generation unit 56. When the front landscape photographed image from the vehicle photographed by the above is input to the data input unit 51, the image feature point data is output from the image feature point data generation unit 56 through the procedure described above. Note that the shooting situation information used in the feature point importance determination unit 53 is generated by the shooting situation information generation unit 7 mounted on the vehicle and sent to the shot image processing unit 5. The shooting state information generation unit 7 is connected to the in-vehicle camera 14 to generate the travel lane data, and receives the same shot image sent to the shot image processing unit 5. Traveling lane data is created by image processing the received captured image using a known algorithm. The photographing situation information generation unit 7 is connected to an obstacle detection sensor group 18 in order to create the moving object data. Moving object data is created based on sensor information from the sensor group 18. Further, the shooting situation information generation unit 7 is connected to the vehicle position determination unit 45 and the database 9 in order to create the above. The database 9 is searched using the vehicle position coordinates and the search key from the vehicle position determination unit 45 to acquire area attributes (mountainous area, city area, etc.) of the current traveling location, and the area attribute data is based on the area attributes. Created.

  The scenery matching unit 6 is based on an extraction algorithm that will be described in detail later from the reference data DB 92 based on the estimated vehicle position sent from the vehicle position determination unit 45 and the actual vehicle behavior sent from the vehicle behavior detection module. A reference data string consisting of a predetermined number of reference data is extracted. Further, reference data is sequentially set as a pattern from the extracted reference data string, and pattern matching processing is performed on the image feature point data sent from the captured image processing unit 5. If the pattern matching is successful, the shooting position associated with the reference data that is the matching pattern is read out. This photographing position is transferred to the own vehicle position determination unit 45 as the own vehicle position. The own vehicle position determination unit 45 corrects the own vehicle position by replacing the transferred own vehicle position with the estimated own vehicle position.

  This car navigation system also includes an input device 11 such as a touch panel 11 or a switch as a peripheral device, and an operation input evaluation unit 21a that changes an operation input through the input device 11 into an appropriate operation signal and transfers the operation signal to the inside. The input operation module 21 has a display module 22 for displaying image information necessary for car navigation on the monitor 12, and a sound generation module 23 for flowing sound information necessary for car navigation from the speaker 13 and the buzzer.

Next, calculation of an error range centered on the estimated own vehicle position performed by the landscape matching unit 6 and a measurement data group included in the error range are shown as actual vehicle behavior sent from the vehicle behavior detection module 24. The process of narrowing down based on the vehicle behavior data will be described with reference to the schematic diagrams of FIGS.
In FIG. 5, the estimated own vehicle position is set as a reference point: Pn, and the shooting positions of the reference data located on the front side in the vehicle traveling direction from there are Pn + 1, Pn + 2,. (K is a positive integer), and from there the shooting positions of the reference data located on the rear side in the vehicle traveling direction are Pn-1, Pn-2,... Pn + k (k is negative) Integer). Reference point: The error range calculated based on the vehicle position measurement error around Pn is indicated by a dotted line along the road. In this example, the photographing positions: Pn + 1, Pn + 2, Pn-1, and Pn-2 belong to the error range together with the reference point: Pn. The shooting position means the shooting position of the reference data stored in the reference data DB 92. Further, the steering angle is schematically shown as the vehicle behavior associated with the reference data of the shooting position. The reference point: Pn is 0 °, the shooting positions: Pn + 1, Pn-1 are 0 °, and shooting is performed. Position: Pn + 2 is 15 ° and Pn-2 is −15 °. From this, if the accuracy of the reference point: Pn, which is the estimated vehicle position, is quite high, the matching partner has a steering angle that is substantially the same as the steering angle that is the vehicle behavior at this reference point: Pn. It becomes possible to narrow down to data only. Therefore, in the example of FIG. 5, the photographing positions: Pn + 2 and Pn-2 can be excluded from matching partners. That is, of the reference data included in the error range, the vehicle behavior at the reference point: Pn, here, the reference data outside the predetermined allowable range with the steering angle can be excluded. By narrowing down such matching partners, matching processing efficiency is improved.

  FIG. 6 exemplifies intersections, and the way of signing, the way of expressing the error range, vehicle behavior, etc. are the same. It can be considered that the steering angle of the vehicle entering the intersection is also maintained within a predetermined range when entering from a predetermined direction and going out in a predetermined direction. In this way, it is possible to narrow down the reference data belonging to the error range, as in the description using FIG. In this example, since the steering angle as the vehicle behavior at the reference point: Pn is 30 °, the reference data having substantially the same steering angle (here, indicated by the shooting positions: Pn + 1 and Pn + 2) Only) is left as a matching partner, and others (here, the shooting position: indicated by Pn + k) are excluded. In particular, at a large intersection, a lot of reference data exists depending on a combination of an approaching direction and an exiting direction. Therefore, such narrowing of the reference data is advantageous.

  Further, FIG. 7 illustrates an uphill road and is a schematic view from the side so that the road gradient is easy to understand. However, the way of signing, the way of expressing the error range, vehicle behavior, etc. are the same. . As can be easily understood from FIG. 7, if there is a slope with a large slope in front, the landscape photographed image at that point changes greatly. In consideration of this phenomenon, reference data of a shooting position where the gradient of the traveling road as the vehicle behavior is clearly different from the gradient at the reference point: Pn is excluded from the matching partner. The gradient of the traveling road can be detected by evaluating a signal from the G sensor, using a signal from the vehicle body tilt sensor, or the like. In FIG. 7, shooting positions: Pn + 1, Pn-1, and Pn-2 that have substantially the same vehicle behavior corresponding to the gradient of the traveling road at the reference point: Pn are left, and the shooting position: Pn + 2. Is excluded. Thus, even on a road with a slope, the reference data belonging to the error range can be narrowed down from the viewpoint of the similarity of the vehicle behavior by the same algorithm as described above.

  As described above, the landscape matching unit 6 uses the estimated vehicle position of the vehicle as an extraction condition, and the reference data that uses the reference data DB 92 as a shooting position around the estimated vehicle position, specifically within the error range of the estimated vehicle position. Are extracted as preliminary reference data. Then, the landscape matching unit 6 narrows down the reference data from the preliminary reference data based on actual vehicle behavior such as a course change by steering or a gradient running, and outputs the reference data as matching partner candidates for the matching data. Note that such processing of the landscape matching unit 6 is performed by a reference data output unit 63 described later.

FIG. 8 includes a landscape matching unit 6 that performs landscape image recognition between reference data that is narrowed down and output using the above-described error range and vehicle behavior and matching data generated from a captured image. A functional block diagram for explaining each functional unit is shown. The landscape matching unit 6 includes a vehicle behavior evaluation unit 61, an error range calculation unit 62, a reference data output unit 63, a matching execution unit 64, and a matching photographing position extraction unit 65. The vehicle behavior evaluation unit 61 evaluates the vehicle behavior data sent from the vehicle behavior detection module 24, and values indicating actual vehicle behavior such as a steering angle, a vehicle inclination angle, and an acceleration value are associated with the reference data. The vehicle behavior evaluation information recorded in a format comparable to the recorded vehicle behavior is output. The vehicle behavior evaluation unit 61 may be configured to function as a vehicle behavior acquisition unit that receives the vehicle behavior data sent from the vehicle behavior detection module 24 and transfers it to other functional units.
The error range calculation unit 62 calculates the error range in consideration of the error of the estimated vehicle position with the estimated vehicle position as a reference. The reference data output unit 63 is based on the actual vehicle behavior recorded in the vehicle behavior evaluation information sent from the vehicle behavior evaluation unit 61 and the error range sent from the error range calculation unit 62. A search condition for extracting an appropriate matching partner candidate for the matching data generated by 5 is created. Further, the reference data output unit 63 transfers the reference data extracted from the reference data DB 92 according to this search condition to the matching execution unit 64. Note that the reference data output unit 63 is configured to extract reference data from the reference data database as the matching partner candidate using an AND condition that the photographing position belongs to the error range and the vehicle behavior has a predetermined allowable range. Alternatively, reference data is extracted from the reference data database using the imaging position belonging to the error range as a search condition, and only the reference data having a vehicle behavior within a predetermined allowable range among the extracted reference data. May be output as the matching partner candidate. The matching execution unit 64 performs pattern matching on the matching data while sequentially setting the extracted reference data string as a pattern. When the matching is successful, the matching photographing position extraction unit 65 reads out the photographing position (own vehicle position) related to the reference image as a successful matching partner, and uses this as a high-accuracy own vehicle position. Send to decision unit 45.

  In addition, the process of narrowing down to only those similar to the actual vehicle behavior at the estimated vehicle position from the reference data belonging to the error range is required to have high reliability of the estimated vehicle position. For this reason, when an attribute value such as reliability is attached to the estimated own vehicle position, it is preferable to determine whether or not to narrow down the reference data based on the vehicle behavior. When there is no such attribute value, when it is determined that the road traveling to the estimated vehicle position is prone to slip error or cornering error, or other factors are likely to cause distance error However, the reference data based on vehicle behavior may not be narrowed down.

  In the embodiment described above, the reference data stored in the reference data DB 92 relates the image feature point data, the shooting position of the shot image corresponding to the image feature point data, and the recorded vehicle behavior indicating the vehicle behavior at the time of shooting. However, in the present invention, the reference data need only relate at least the image feature point data and the shooting position thereof. From the map data read from the road map DB 91, it is possible to know the shooting position where a specific vehicle behavior occurs. For example, braking behavior occurs before a sharp curve, and steering behavior occurs when entering a sharp curve. In this way, the reference data can be narrowed down using the possibility of vehicle behavior that can be read from the map data. In this case, the reference data output unit 63 obtains a shooting position range where the actual vehicle behavior occurs based on the map data, and uses the shooting position range as a preliminary reference data with the error range of the estimated vehicle position as the shooting position. The reference data is narrowed down from the list. A feature of the present invention is that the actual vehicle behavior is detected and the reference data to be matched is narrowed down by the actual vehicle behavior.

In the above-described embodiment, the image feature point is an edge point obtained by edge detection processing, particularly a line segment edge constituting one line segment or an intersection point where such a line segment intersects, preferably approximately orthogonally. A corner edge is treated as an effective image feature point. Further, the matching execution unit 64 employs a general pattern matching algorithm. However, in the matching (matching) of corner points (intersection edge points as intersections of two linear components) having high importance in a landscape image. A weighted pattern matching algorithm that gives a higher matching evaluation than other edge points may be employed. At that time, it is preferable to construct the reference data DB 92 in a form in which corner attribute values indicating the coordinate values of the corner points are attached to the reference data. However, each time the reference data is extracted, the corner points may be detected. Good. The matching data may be given an identification code (label) to a corner point in the process of being generated from the actual captured image.
In the above-described embodiment, the image feature point is an edge point obtained by edge detection processing, particularly a line segment edge constituting one line segment or an intersection point where such a line segment intersects, preferably approximately orthogonally. A corner edge is treated as an effective image feature point. However, the present invention is not limited to such edge points as image feature points. For example, typical edge points that form geometric shapes such as circles and quadrilaterals (three points on the circumference if circles) or the center of gravity of a geometric shape and the point as the center of gravity are also effective depending on the landscape. It is used because it becomes a typical image feature point. It is also preferable to adopt edge strength as a factor for calculating importance. For example, if a line segment is a strong edge, the start point and end point of the line segment are treated as highly important image feature points. be able to. In addition, a specific point in a characteristic geometric shape, for example, an end point of a symmetrical object can be handled as a highly important image feature point.
Furthermore, in addition to the edge points obtained by the edge detection process, it is also possible to regard a captured image as a change in hue or saturation and to employ a point with a large change as an image feature point. Similarly, an end point of an object having a high color temperature can be handled as a highly important image feature point based on color information.
That is, the image feature points handled in the present invention are all used if they are effective for similarity determination between the reference data and the image feature amount data generated from the actual captured image, for example, pattern matching. It becomes.

In the above-described embodiment, the reference data stored in the reference data DB 92 is associated with the shooting position and the shooting direction (camera optical axis direction). The date and time and the weather at the time of shooting may also be related.
The photographing position may be two-dimensional data such as latitude / longitude data at least, but may be three-dimensional data in addition to height data.
Further, it is not essential to relate the shooting direction to the reference data. For example, when creating reference data and when recognizing a landscape image using this reference data, the shooting direction is not required if it is guaranteed that the road is shot with substantially the same shooting direction. It becomes.
Conversely, in the case where reference data in which the shooting direction is appropriately shifted from the reference data in one basic shooting direction can be prepared, based on the traveling direction of the vehicle calculated from information such as a direction sensor, It is also possible to set only the reference data suitable for the traveling direction as the object of landscape image recognition.
The on-vehicle camera handled in the present invention is most suitable for photographing a landscape in front of the vehicle traveling direction. However, the camera may be a camera that takes a landscape oblique to the front, or may be a camera that captures a landscape behind. However, if you use a rear camera that captures the landscape behind, the moving direction of the feature points that move as the vehicle travels is the opposite of that of the front camera, so if the feature points are concentrated in the vicinity, The direction in which the points are offset is not the rear but the front. When the feature points are concentrated in the center, the direction in which the reference point is offset is set to the rear rather than the front. That is, since it can be used with either a front camera or a rear camera, a captured image handled in the present invention is not limited to a photograph of a front landscape in the vehicle traveling direction.

  The functional units shown in the functional block diagrams used in the description of the above-described embodiments are for easy understanding, and the present invention is not limited to the divisions shown here. It is possible to freely combine function parts or further divide one function part

  The image processing system of the present invention can be applied not only to car navigation but also to a technical field that measures the current position and orientation by landscape image recognition.

3: navigation control module 4: own vehicle position detection module 41: GPS processing unit 42: dead reckoning processing unit 43: own vehicle position coordinate calculation unit 44: map matching unit 45: own vehicle position determination unit 5: captured image processing unit 51 : Data input unit 52: feature point extraction unit 53: feature point importance determination unit 54: weighting unit 55: adjustment coefficient setting unit 56: image feature point data generation unit 57: reference data database creation unit 6: landscape matching unit 61: Vehicle behavior evaluation unit 62: error range correction unit 63: reference data output unit 64: matching execution unit 65: matching shooting position extraction unit 14: camera 91: road map DB (road map database)
92: Reference data DB (reference data database)

Claims (3)

Stores image feature point data generated by extracting image feature points from a photographed image obtained by photographing a landscape from a vehicle, and reference data relating the photographing positions of photographed images corresponding to the image feature point data. A reference data database;
A captured image processing unit that outputs a matching image generated by inputting an actual captured image of a landscape by an in-vehicle camera and extracting image feature points from the actual captured image;
A vehicle behavior acquisition unit that acquires vehicle behavior data indicating actual vehicle behavior that is the vehicle behavior of the vehicle;
Using the estimated vehicle position of the vehicle as an extraction condition, preliminary reference data in which the periphery of the estimated vehicle position is a shooting position is extracted from the reference data database, and is referenced from the preliminary reference data based on the actual vehicle behavior A reference data output unit that narrows down the data and outputs it as a matching partner candidate for the matching data;
A matching execution unit that performs matching between the output reference data and the matching data;
A vehicle position detection system using landscape image recognition, comprising: a vehicle position determination unit that determines a vehicle position based on the photographing position related to the reference data that has been successfully matched.   The reference data output unit obtains a shooting position range in which the actual vehicle behavior occurs based on map data, and narrows down the reference data from the preliminary reference data using the shooting position range. Vehicle position detection system.   The host vehicle position detection system according to claim 1, wherein the vehicle behavior is a vehicle behavior caused by a specific road condition on which the vehicle travels.

Images