Feature point processing apparatus

Abstract

To provide a technique capable of excluding a feature point with low accuracy.SOLUTION: The feature point processing device 100 includes a resolution hierarchical image generation unit 120 that uses a processing target image obtained by the camera 20 mounted on the moving body 10 to generate a plurality of resolution hierarchical images having sizes obtained by reducing the processing target image at a plurality of reduction ratios, a feature point detection unit 130 that executes a feature point detection process of detecting a plurality of feature points from the plurality of resolution hierarchical images, and a feature point exclusion processing unit 140 that selects and excludes some of the plurality of feature points as exclusion feature points. The feature point exclusion processing unit selects, from among the plurality of feature points, a feature point whose pixel accuracy, which is accuracy per pixel determined according to a reduction ratio of each hierarchical image, exceeds an allowable value, as an exclusion feature point.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a feature point processing device that performs feature point exclusion processing. [Background technology]

[0002] Patent Document 1 discloses a technology for detecting multiple feature points from an image and estimating the position of a moving object using the multiple feature points. Position estimation technology requires scale invariance, which is a property that allows accurate detection of the features of an object even if the scale of the object changes. To ensure scale invariance, the feature point detection process uses a technology that creates multiple hierarchical images in a pyramid hierarchy and detects feature points from each of the multiple hierarchical images. Note that hierarchical images are also called "pyramid images" or "reduced images."

[0003] Detecting feature points using multiple hierarchical images enables robust feature point detection and tracking against changes in the scale of the camera and the environment. For example, when the camera moves closer to or farther away from the object, pyramid image processing makes it possible to detect and track feature points at an appropriate resolution. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2020-67439 Summary of the Invention [Problem to be solved by the invention]

[0005] Among feature points detected from an image, those located far away tend to have a large error per pixel. This error value increases in small-sized layered images with a large reduction ratio among multiple layered images, resulting in a large error in the position of feature points located far away. As a result, the reliability of the position and orientation of feature points in 3D coordinates decreases, deteriorating the accuracy of feature point tracking between image frames and the accuracy of estimating the position of moving objects. Therefore, a technology that can exclude feature points with low accuracy is desired. [Means for solving the problem]

[0006] According to one embodiment of the present disclosure, there is provided a feature point processing device (100). This feature point processing device includes: a hierarchical image generation unit (120) that uses a processing target image acquired by a camera (20) mounted on a moving object (10) to generate a plurality of hierarchical images each having a size obtained by reducing the processing target image at a plurality of reduction ratios; a feature point detection unit (130) that executes a feature point detection process to detect a plurality of feature points from the plurality of hierarchical images; and a feature point exclusion processing unit (140) that selects and excludes some of the plurality of feature points as exclusion feature points. The feature point exclusion processing unit selects, from the plurality of feature points, feature points whose pixel precision, which is the precision per pixel determined according to the reduction ratio of each hierarchical image, exceeds a tolerance value as the exclusion feature points.

[0007] According to this feature point processing device, it is possible to exclude feature points with low accuracy and use only feature points with high accuracy. [Brief explanation of the drawings]

[0008] [Figure 1] 1 is a block diagram showing the configuration of a mobile object position estimation device according to a first embodiment. [Figure 2] FIG. 10 is an explanatory diagram showing an example of a plurality of hierarchical images. [Figure 3] FIG. 10 is an explanatory diagram showing an example of a plurality of feature points. [Figure 4] FIG. 2 is an explanatory diagram showing the horizontal resolution and vertical resolution of a hierarchical image. [Figure 5] FIG. 3 is an explanatory diagram showing the number of pixels and pixel precision of each layer image in the first embodiment. [Figure 6] FIG. 4 is an explanatory diagram showing layer images to be excluded in the first embodiment. [Figure 7] 5 is a flowchart showing the procedure of a position estimation process in the first embodiment. [Figure 8] FIG. 10 is a block diagram showing the configuration of a mobile object position estimation device according to a second embodiment. [Figure 9] FIG. 10 is an explanatory diagram showing an example of a method for estimating distances to feature points by road surface estimation. [Figure 10] FIG. 10 is an explanatory diagram showing the relationship between the distance of a feature point and pixel accuracy in the second embodiment. [Figure 11] FIG. 10 is an explanatory diagram showing feature points to be excluded in the second embodiment. [Figure 12] 10 is a flowchart showing the procedure of a position estimation process in the second embodiment. [Figure 13] FIG. 10 is a block diagram showing the configuration of a mobile object position estimation device according to a third embodiment. [Figure 14] FIG. 11 is an explanatory diagram showing deviation tolerances in the third embodiment. [Figure 15] FIG. 11 is an explanatory diagram showing feature points to be excluded in the third embodiment. [Figure 16] 10 is a flowchart showing the procedure of a position estimation process in the third embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0009] A. First embodiment: 1, a position estimation device 30 of the first embodiment executes a process of estimating the self-position of a moving object 10 using an image captured by a camera 20 mounted on the moving object 10. The position estimation device 30 includes a feature point processing device 100 that extracts a plurality of feature points using a frame image IM0 captured by the camera 20, a self-position estimation unit 200 that estimates the self-position of the moving object 10 using the plurality of feature points, and a running state detection unit 300 that detects the running state of the moving object 10.

[0010] A monocular camera may be used as the camera 20, or a stereo camera or an RGBD camera including an RGB camera and a depth camera may be used. In this embodiment, a monocular camera for capturing color images is used.

[0011] The position estimation device 30 can be realized as an ECU (Electronic Control Unit) having a processor and memory. The functions of each part of the position estimation device 30 can be realized by the processor executing a computer program stored in the memory. Alternatively, some or all of the parts may be realized by hardware circuits.

[0012] The feature point processing device 100 includes a grayscale processing unit 110 , a hierarchical image generating unit 120 , a feature point detecting unit 130 , a feature point exclusion processing unit 140 , a tolerance setting unit 150 , and a pixel precision calculating unit 160 .

[0013] The grayscale processing unit 110 performs grayscale processing on the frame image IM0, which is a full-color image captured by the camera 20. The grayscale processing is a process for converting a full-color image into, for example, an 8-bit grayscale image (multi-tone gray image). In this embodiment, the grayscale image is also referred to as the "processing target image IM1." The processing target image IM1 is an image used in the feature point detection process.

[0014] When a RAW image is supplied from the camera 20 to the feature point processing device 100, a full-color frame image IM0 may be obtained by performing preprocessing such as demosaicing, noise removal, and distortion correction on the RAW image. Demosaicing is a process that interpolates the pixel values ​​of R, G, and B pixels arranged in a checkerboard pattern in the RAW image to generate a full-color image in which each pixel has an RGB pixel value.

[0015] The hierarchical image generator 120 uses the processing target image IM1 to generate a plurality of hierarchical images PMj each having a size obtained by reducing the processing target image IM1 at a plurality of reduction ratios.

[0016] As shown in Figure 2, the image to be processed IM1 is used as the lowest layer image PM0 as is. When j is an integer equal to or greater than 1, the jth layer image PMj is an image obtained by reducing the j-1th layer image PMj-1 by a scale factor S. Therefore, the jth layer image PMj is obtained by reducing the image to be processed IM1 by a reduction ratio of 1 / S. j The scale factor S is set to a value greater than 1. In the example of FIG. 2, a fixed scale factor S is used to generate multiple hierarchical images PM0 to PM7 in a pyramid hierarchical structure, but it is not necessary to use a fixed scale factor S; it is sufficient to generate multiple hierarchical images PMj each having a size obtained by reducing the processing target image IM1 at multiple reduction rates. Including the lowest hierarchical image PM0, the variable j that distinguishes the hierarchical images PMj is an integer greater than or equal to 0.

[0017] The feature point detection unit 130 executes a feature point detection process to detect a plurality of feature points from a plurality of hierarchical images PMj.

[0018] As shown in Fig. 3, a plurality of feature points CP are detected from each hierarchical image PMj. In this embodiment, FAST (Features from Accelerated Segment Test) is used as the feature point detection algorithm. However, other feature point detection algorithms such as Harris corner detection, Shi-Tomashi corner detection, GFTT, SIFT, AKAZE, and ORB may also be used.

[0019] The feature point exclusion processing unit 140 selects and excludes some of the feature points CP as exclusion feature points from among the plurality of feature points CP. The tolerance setting unit 150 sets a tolerance for determining the exclusion feature points. The tolerance may be a fixed value, or may be changed depending on the running state of the moving object 10 detected by the running state detection unit 300. The pixel accuracy calculation unit 160 executes processing to calculate pixel accuracy, which is the accuracy per pixel. As will be described later, the pixel accuracy is determined depending on the reduction ratio of each hierarchical image PMj. The processing details of the feature point exclusion processing unit 140, tolerance setting unit 150, and pixel accuracy calculation unit 160 will be described later.

[0020] 4, in the first embodiment, for each layer image PMj, a horizontal resolution α is calculated according to the horizontal angle of view of the camera 20, and a vertical resolution β is calculated according to the vertical angle of view. These resolutions α and β are angular resolutions per pixel, and are used as pixel accuracy in the first embodiment.

[0021] As shown in FIG. 5, the number of horizontal pixels and the number of vertical pixels of each layer image PMj are determined according to the respective reduction ratios. The pixel precision calculation unit 160 calculates the horizontal resolution α and the vertical resolution β according to the number of horizontal pixels and the number of vertical pixels of each layer image PMj. The horizontal resolution α is the value obtained by dividing the horizontal angle of view by the number of horizontal pixels. The vertical resolution β is the value obtained by dividing the vertical angle of view by the number of vertical pixels. In this disclosure, "horizontal" and "vertical" refer to the lateral and vertical directions, respectively, on the image plane of the camera 20.

[0022] As shown in FIG. 6 , the horizontal resolution α and vertical resolution β, which are used to measure pixel accuracy, are determined for each layer image PMj. That is, all feature points detected from the same layer image PMj have the same horizontal resolution α and vertical resolution β. In the example of FIG. 6 , the accuracy tolerance Pt1 for determining feature points to be excluded is set to 0.35 degrees. In the example of FIG. 6 , the vertical resolution β of layer image PM7 is greater than the accuracy tolerance Pt1. Therefore, the feature point exclusion processor 140 selects and excludes all feature points detected from layer image PM7 as feature points to be excluded. As described above, in the first embodiment, pixel accuracy is the angular resolution per pixel determined for each layer image PMj. If at least one of the horizontal resolution α and vertical resolution β of a layer image PMj exceeds the accuracy tolerance, all feature points detected from that layer image PMj are selected as feature points to be excluded. In the example of FIG. 6 , the accuracy tolerance Pt1 for the horizontal resolution α and the vertical resolution β is set to the same value, but they may also be set to different values.

[0023] The accuracy tolerance Pt1 may be a fixed value, or may be changed depending on the traveling state of the moving body 10 detected by the traveling state detection unit 300. The traveling state detection unit 300 detects the traveling speed of the moving body 10, the position of the moving body 10 on the map, the traffic congestion state of the traveling road of the moving body 10, etc. as the traveling state of the moving body 10.

[0024] <Setting the accuracy tolerance Pt1 according to the running speed> The accuracy tolerance Pt1 may be set to a lower value as the traveling speed of the moving body 10 increases. In this way, it is possible to perform more accurate position estimation when the moving body 10 is traveling at high speed. The traveling speed of the moving body 10 can be obtained from the output of the vehicle speed sensor of the moving body 10.

[0025] <Setting the accuracy tolerance Pt1 according to the location or region of the mobile object> The accuracy tolerance value Pt1 may be changed depending on the position of the mobile object 10 on the map and the area in which the mobile object 10 is traveling. For example, when the mobile object 10 is traveling on an ordinary road, it may be set to a lower value than when the mobile object 10 is traveling on an expressway. In this way, highly accurate position estimation can be performed even on ordinary roads where the detection accuracy of feature points is low. Furthermore, for roads with low traffic volume, such as gravel roads, the accuracy tolerance value Pt1 may be set to a higher value than for other roads. The position and area of ​​the mobile object 10 on the map can be detected using, for example, a GNSS (Global Navigation Satellite System) receiver or a navigation device.

[0026] <Setting the accuracy tolerance Pt1 according to traffic congestion conditions> The accuracy tolerance value Pt1 may be changed depending on the congestion state of the road on which the mobile object 10 is traveling. For example, when congestion occurs on the road, the accuracy tolerance value Pt1 may be set to a higher value than when there is no congestion. The congestion state of the road can be detected, for example, using a wireless communication device that performs wireless communication with an Intelligent Transport System.

[0027] The position estimation process shown in FIG. 7 is performed periodically after the moving body 10 starts up. In step S11, the grayscale processing unit 110 performs grayscale processing on the frame image IM0 to create a processing target image IM1. In step S12, the hierarchical image generation unit 120 generates multiple hierarchical images PMj each having a size obtained by reducing the processing target image IM1 at multiple reduction ratios. In step S13, the feature point detection unit 130 detects feature points from each hierarchical image PMj. In step S14, pixel accuracy, which is the accuracy per pixel, is calculated according to the reduction ratio of each hierarchical image PMj. As described above, in the first embodiment, pixel accuracy is the angular resolution obtained by dividing the angle of view of the camera 20 by the number of pixels in the hierarchical image PMj.

[0028] In step S15, the tolerance setting unit 150 sets the accuracy tolerance Pt1. As described above, the accuracy tolerance Pt1 may be a fixed value, or may be changed depending on the traveling state of the moving object 10 detected by the traveling state detection unit 300. If the preset accuracy tolerance Pt1 is used as is, step S15 is omitted. In step S16, the feature point exclusion processing unit 140 selects and excludes feature points that exceed the accuracy tolerance as excluded feature points. As described in FIG. 6, in the first embodiment, all feature points detected from the hierarchical image PM7 whose pixel accuracy exceeds the accuracy tolerance Pt1 are selected as excluded feature points.

[0029] In step S17, the self-location estimation unit 200 executes a process of estimating the self-location of the moving body 10 using the multiple feature points obtained after the feature elimination process. This self-location estimation process is executed using, for example, a SLAM (Simultaneous Localization and Mapping) algorithm. SLAM is an algorithm that simultaneously estimates the self-location of where the moving body is located on a map and creates an environmental map of the surrounding environment. However, the feature points may also be used for processes other than self-location estimation. For example, the feature points may be used for detecting lanes or other moving bodies or objects.

[0030] According to the first embodiment described above, feature points whose pixel accuracy, which is the accuracy per pixel determined according to the reduction ratio of each layer image PMj, exceeds an allowable value are selected and excluded as excluded feature points, so that feature points with low accuracy can be excluded and only feature points with high accuracy can be used. Also, feature points that exceed the allowable value can be excluded from tracking targets across multiple frame images. Furthermore, because feature points with low accuracy are not handed over to the processing of the self-location estimation unit 200, it is possible to simultaneously achieve scale invariance, faster processing time, and improved self-location estimation accuracy.

[0031] B. Second embodiment: The position estimation device 30 of the second embodiment shown in FIG. 8 is obtained by adding a distance estimation unit 170 to the feature point processing device 100 of the position estimation device 30 of the first embodiment shown in FIG. 1, and other configurations are the same as those of the first embodiment.

[0032] The distance estimation unit 170 estimates the distance from the camera 20 to the three-dimensional position of each feature point CP according to the position of each feature point CP in each hierarchical image PMj. The "three-dimensional position of the feature point CP" means the position of the point in the real world corresponding to the feature point CP detected from the hierarchical image PMj. As a method for estimating the distance, for example, one or more of the following methods can be used.

[0033] <Method M1 based on road surface estimation> A method for estimating the distance from the camera 20 to the three-dimensional position of the feature point CP by assuming that the feature point CP is on the road surface.

[0034] <Method M2 based on monocular depth estimation> A method for estimating the depth of each part in an image from a single image captured by a monocular camera mounted on the moving body 10. In this method M2, the distance from the camera 20 to the three-dimensional position of the feature point CP can be estimated using the position of the feature point CP in each hierarchical image PMj and the depth at that position.

[0035] <Method M3 using a Depth camera> A method for estimating the depth using a Depth camera mounted on the moving body 10. This method M3 can be used when using an RGBD camera as the camera 20. <0000-167> <Method M4 using a stereo camera> A method for estimating the distance of the feature point CP by triangulation using a stereo camera mounted on the moving body 10. This method M3 can be used when using a stereo camera as the camera 20.

[0037] <Method M5 using a distance measuring device> This method estimates the distance to the feature point CP using the measurement results of a distance measuring device mounted on the moving object 10. The distance measuring device can be a millimeter-wave radar or LiDAR (Light Detection and Ranging). This method M3 can estimate the distance from the camera 20 to the three-dimensional position of the feature point CP using the position of the feature point CP in each layer image PMj and the distance measurement value at the point on the screen of the distance measuring device that corresponds to that position.

[0038] 9, in method M1 based on road surface estimation, it is estimated that feature point CP is located on road surface SF of the travel path of the mobile object. In this case, the distance L from camera 20 to feature point CP can be calculated using the depression angle δ determined from the position of feature point CP on the image plane IP of camera 20 and the height Hc of camera 20 from the road surface SF, as follows: L=Hc / tanδ(q1) Since the image plane IP of the camera 20 corresponds to the image plane of each layer image PMj, the depression angle δ is determined from the position of the feature point CP in each layer image PMj. As the road surface estimation method M1, methods other than the method shown in Fig. 9 can also be used. For example, a method of estimating the distance L using triangulation using the focal length of the camera 20 may be used.

[0039] The above-described distance estimation methods M1 to M5 are all similar in that they estimate the distance from the camera 20 to the three-dimensional position of each feature point CP in accordance with the position of each feature point CP in each layer image PMj. Note that distance estimation methods other than the above-described methods may also be used.

[0040] As shown in FIG. 10, in the second embodiment, the pixel accuracy calculation unit 160 calculates the pixel accuracy for each layer image PMj according to the distance L from the camera 20 to the three-dimensional position of the feature point CP. The pixel accuracy in the second embodiment is the amount of positional deviation per pixel determined according to the reduction ratio of each layer image PMj and the distance L of each feature point CP. The unit of the positional deviation [m] is a length in real three-dimensional space. The horizontal deviation amount Xpix and vertical deviation amount Ypix as pixel accuracy are calculated using the following equations. Xpix=L*tanα(q2) Ypix=L*tanβ(q3) Here, α is the angular resolution per pixel in the horizontal direction, and β is the angular resolution per pixel in the vertical direction.

[0041] FIG. 10 illustrates the values ​​of the horizontal shift amount Xpix and the vertical shift amount Ypix according to the distance L for the two hierarchical images PM0 and PM7.

[0042] As shown in FIG. 11, the horizontal and vertical misalignment amounts Xpix and Ypix, which are expressed as pixel precision, are determined for each layer image PMj according to the distance L to the feature point CP. In the example of FIG. 11, the precision tolerance Pt2 for determining feature points to be excluded is set to 0.3 m. For layer image PM7, the horizontal misalignment amount Xpix exceeds the precision tolerance Pt2 when the distance L to the feature point CP is 52 m or greater. The vertical misalignment amount Ypix also exceeds the precision tolerance Pt2 when the distance L to the feature point CP is 42 m or greater. Therefore, for layer image PM7, feature points CP whose distance L is 42 m or greater are selected as feature points to be excluded. Similarly, for the other layer images PMj, feature points CP whose at least one of the horizontal misalignment amount Xpix and the vertical misalignment amount Ypix exceeds the precision tolerance Pt2 are selected as feature points to be excluded.

[0043] Like the accuracy tolerance Pt1 in the first embodiment, the accuracy tolerance Pt2 may be a fixed value, or may be changed depending on the traveling state of the moving object 10 detected by the traveling state detection unit 300. The accuracy tolerance Pt2 for the horizontal deviation amount Xpix and the vertical deviation amount Ypix may be set to different values.

[0044] The procedure of the position estimation process shown in Fig. 12 is obtained by replacing step S14 in Fig. 7 with step S20, with the other steps being the same as those in the first embodiment. In step S20, the pixel accuracy calculation unit 160 calculates pixel accuracy in accordance with the reduction ratio of each hierarchical image PMj and the distance L of each feature point CP. The other steps are almost the same as those in the first embodiment, and therefore will not be described here.

[0045] The second embodiment also provides the same effects as the first embodiment. Furthermore, in the second embodiment, it is possible to exclude feature points with low pixel accuracy according to the distance L from the camera 20 to the three-dimensional position of the feature point CP.

[0046] C. Third embodiment: The position estimation device 30 of the third embodiment shown in Fig. 13 is configured by adding a travel path detection unit 180 to the feature point processing device 100 of the position estimation device 30 of the second embodiment shown in Fig. 8, and the other configurations are the same as those of the second embodiment. The travel path detection unit 180 detects the width of the travel path of the moving object 10 using the frame image IM0.

[0047] 14, both sides of the roadway on which the moving object 10 is traveling are defined by white lines WL1 and WL2, and the distance between the white lines WL1 and WL2 is detected as the roadway width Wr. The roadway width Wr may be detected using the shoulders of the roadway instead of the white lines. The white lines WL1 and WL2 and the roadway shoulders can be detected, for example, by detecting characteristic points on the road surface from the frame image IM0 and estimating a curve that extends approximately parallel to the traveling direction of the moving object 10 from among the curves connecting those characteristic points.

[0048] The tolerance for pixel accuracy in the third embodiment is a deviation tolerance η that is dynamically determined according to the difference between the road width Wr and the width Wv of the moving body 10. This deviation tolerance η is calculated, for example, by the following equation. η=(Wr-Wv) / (2*Rs) (q4) Here, Rs is a safety factor, and is preferably set to a value greater than 1.

[0049] As can be seen from the above formula (q4), it is preferable that the deviation tolerance η is proportional to (Wr-Wv) / 2, which is half the value of the difference between the travel path width Wr and the width Wv of the moving body 10. It is more preferable that the deviation tolerance η is set to a value smaller than (Wr-Wv) / 2.

[0050] As shown in FIG. 15 , in the third embodiment, the deviation tolerance η for determining feature points to be excluded is set for the horizontal deviation amount Xpix. On the other hand, the vertical deviation amount Ypix is ​​not used for selecting feature points to be excluded. For layer image PM7, when the distance L to feature point CP is 85 m or more, the horizontal deviation amount Xpix exceeds the deviation tolerance η. Therefore, for layer image PM7, feature points CP whose distance L is 85 m or more are selected as feature points to be excluded. Similarly, for other layer images PMj, feature points CP whose horizontal deviation amount Xpix exceeds the deviation tolerance η are selected as feature points to be excluded.

[0051] The process of dynamically determining the deviation tolerance η for determining excluded feature points is preferably performed when the distribution of pixel values ​​in the processing target image IM1 is flat and the number of detected feature points CP is equal to or less than the number threshold. Generally, the fewer the number of feature points CP, the lower the accuracy of estimating the self-position of the moving body 10 tends to be. Therefore, when the number of detected feature points CP is equal to or less than the number threshold, it is preferable to dynamically determine the deviation tolerance η according to the difference between the road width Wr and the width Wv of the moving body 10, so that more feature points can be used in the self-position estimation process. However, the deviation tolerance η may also be dynamically determined regardless of the number of feature points CP.

[0052] Whether or not to use the deviation tolerance value η may be determined in advance depending on, for example, the location or region of the mobile object 10 on the map. That is, for each region through which the mobile object 10 may travel, it may be investigated in advance whether or not the number of detected characteristic points CP is small, and the region in which the deviation tolerance value is to be used may be determined in accordance with the investigation results.

[0053] The procedure of the position estimation process shown in FIG. 16 is the same as that of the second embodiment, except that steps S31 to S33 are added between steps S12 and S13 in FIG. 12 and step S15 is replaced with step S34.

[0054] In step S31, the traveling condition detection unit 300 determines whether the deviation tolerance use condition is satisfied. For example, if the position of the moving object 10 on the map is in a predetermined deviation tolerance use area, it is determined that the deviation tolerance use condition is satisfied. If the deviation tolerance use condition is not satisfied, steps S32 and S33 are skipped and step S13 is executed. On the other hand, if the deviation tolerance use condition is satisfied, the process proceeds to step S32, where the traveling path detection unit 180 detects the traveling path width Wr of the moving object 10. In step S33, the feature point detection unit 130 relaxes the detection threshold for the feature point detection process compared to when the deviation tolerance use condition is not satisfied. The relaxed, lower detection threshold is used in the feature point detection process in the next step S13. Therefore, when the deviation tolerance use condition is satisfied, more feature points can be detected than when the deviation tolerance use condition is not satisfied. Note that step S33 may be omitted.

[0055] If the deviation tolerance use conditions are met, then in step S34, the tolerance setting unit 150 sets the deviation tolerance η according to the difference between the road width Wr and the moving object width Wv, and in step S16, the feature point exclusion processing unit 140 executes the feature point exclusion process using the deviation tolerance η. On the other hand, if the deviation tolerance use conditions are not met, then the same accuracy tolerance Pt2 as in the second embodiment is used, and the feature point exclusion process described in Fig. 11 is executed.

[0056] The third embodiment also achieves the same effects as the first and second embodiments. In addition, in the third embodiment, the deviation tolerance η is dynamically determined according to the difference between the road width Wr and the moving body width Wv, so that a large number of feature points can be used according to the traveling state of the moving body 10.

[0057] The present disclosure is not limited to the above-described embodiments and their modifications, and can be implemented in various forms without departing from the spirit and scope of the present disclosure. [Explanation of symbols]

[0058] 10...mobile body, 20...camera, 30...position estimation device, 100...feature point processing device, 110...grayscale processing unit, 120...hierarchical image generation unit, 130...feature point detection unit, 140...feature point exclusion processing unit, 150...tolerance value setting unit, 160...pixel accuracy calculation unit, 170...distance estimation unit, 180...roadway detection unit, 200...self-position estimation unit, 300...driving state detection unit

Claims

1. A feature point processing device (100), a hierarchical image generating unit (120) that uses a processing target image acquired by a camera mounted on a moving body to generate a plurality of hierarchical images each having a size obtained by reducing the processing target image at a plurality of reduction ratios; a feature point detection unit (130) that executes a feature point detection process to detect a plurality of feature points from the plurality of hierarchical images; a feature point exclusion processing unit (140) that selects and excludes some feature points from the plurality of feature points as exclusion feature points; Equipped with the feature point exclusion processing unit selects, from among the plurality of feature points, feature points whose pixel accuracy, which is accuracy per pixel determined according to the reduction ratio of each hierarchical image, exceeds an allowable value as the excluded feature points.

2. The feature point processing device according to claim 1 , the pixel accuracy is an angular resolution per pixel determined for each of the plurality of layer images, The feature point exclusion processing unit selects, as the excluded feature points, all feature points detected from the hierarchical images whose pixel precision exceeds the allowable value.

3. The feature point processing device according to claim 1 , further comprising: a distance estimation unit (170) for estimating a distance from the camera to a three-dimensional position of each feature point according to the position of each feature point in each layer image; The pixel precision is a positional deviation amount per pixel that is determined according to the reduction ratio of each layer image and the distance of each feature point.

4. The feature point processing device according to claim 3, further comprising: a travel path detection unit (180) that detects the width of the travel path of the moving object from the processing target image; The feature point processing device, wherein the tolerance is a deviation tolerance that is dynamically determined depending on the difference between the width of the road and the width of the moving object.

Images