JP6375911B2 - Curve mirror detector - Google Patents

Description

  The present invention relates to an apparatus for detecting a curve mirror from an image acquired by an in-vehicle camera or the like.

  Conventionally, a technique for detecting a curve mirror by pattern matching from a photographed image of a vehicle-mounted camera has been proposed (see Patent Document 1).

JP 2009-211162 A

However, when the background of the curve mirror in the captured image is complicated, there is a problem that erroneous detection due to pattern matching is likely to occur.
An object of the present invention is to provide a technique for detecting a curved mirror with high accuracy.

The curve mirror detection device according to the first aspect of the present invention includes a curve mirror detection means (21) and a likelihood output means (21).
The curve mirror detection means detects one or more curve mirror candidates, which are areas having the characteristics of a curve mirror, from the picked-up image picked up by the image pickup means (11) for picking up the periphery of the vehicle.

  The likelihood output means includes a saturation parameter that is a parameter based on the saturation of each pixel that constitutes the curve mirror candidate in the captured image, a brightness parameter that is a parameter based on the brightness of each pixel that constitutes the curve mirror candidate, and Two or more selected from the group consisting of shape parameters, which are parameters based on the continuity of the edge points corresponding to the curve mirror candidates detected by the edge detection means (21) for detecting the edge points on the captured image Based on the feature amount, likelihood information relating to the likelihood of the curve mirror candidate as a curve mirror is output.

  In the case of the curve mirror detection device configured as described above, the curve mirror is specified with high accuracy based on the likelihood as the curve mirror output by the likelihood output means from the curve mirror candidates detected from the captured image. be able to.

Moreover, the curve mirror detection apparatus of the 2nd aspect of this invention is provided with a curve mirror detection means (21) and likelihood output means (21) similarly to the 1st aspect mentioned above.
The curve mirror detection means detects one or more curve mirror candidates, which are areas having the characteristics of a curve mirror, from the picked-up image picked up by the image pickup means (11) for picking up the periphery of the vehicle.

  The likelihood output means is a likelihood of the curve mirror candidate as a curve mirror based on a ratio of pixels in which at least one of saturation and brightness belongs to a predetermined range among the pixels constituting the curve mirror candidate. Output likelihood information about degrees.

If it is the curve mirror detection apparatus comprised in this way, a curve mirror can be specified with high precision similarly to the 1st aspect of this invention mentioned above.
In addition, the code | symbol in the parenthesis described in this column and a claim shows the correspondence with the specific means as described in embodiment mentioned later as one aspect, Comprising: The technical scope of this invention is shown. It is not limited.

It is a block diagram which shows a vehicle-mounted system. It is a flowchart of a curve mirror mirror image output process. It is a schematic diagram explaining the captured image image | photographed with the vehicle-mounted camera. It is a schematic diagram explaining the restriction image which carried out the hue filter of the captured image. FIG. 6A is a diagram in which edge detection is performed, FIG. 5A is a diagram when edge detection is performed on a captured image, and FIG. 5B is a diagram when edge detection is performed on a restricted image. It is a schematic diagram which shows a curve mirror, Comprising: (A) is a figure when the ratio of the road and sky which occupy for a mirror image is high, (B) is a figure when the ratio of the building which occupies for a mirror image is high. It is a figure explaining the curve mirror arrange | positioned in the condition where it is difficult to detect an edge.

Embodiments to which the present invention is applied will be described below with reference to the drawings.
[1. Embodiment]
[1-1. Constitution]
The in-vehicle system 1 according to the present embodiment is a system that is used by being mounted on a vehicle such as an automobile, and includes an in-vehicle camera 11, a display 13, and a curve mirror detection device 15, as shown in FIG. .

The in-vehicle system 1 is a system that detects a curve mirror installed on a road and notifies a driver of a mirror image reflected on the mirror surface of the curve mirror.
The in-vehicle camera 11 is an image pickup device that picks up an image of the periphery of the vehicle, in particular, the front side. For example, a known CCD image sensor or CMOS image sensor can be used. The in-vehicle camera 11 images the front of the vehicle at a predetermined time interval (1/15 s as an example), and outputs the captured image to the curve mirror detection device 15. The in-vehicle camera 11 is an example of an imaging unit in the present invention.

  The display 13 is a display means such as a liquid crystal display for displaying an image, and displays an image according to a signal input from the curve mirror detection device 15. In the present embodiment, the display 13 enlarges and displays a mirror image of the curve mirror captured by the in-vehicle camera 11.

The curve mirror detection device 15 detects the curve mirror from the captured image captured by the in-vehicle camera 11, and controls the display 13 to display the image.
The curve mirror detection device 15 includes a CPU 21, a ROM 22 that stores programs executed by the CPU 21, a RAM 23 that is used as a work area when the CPU 21 executes programs, a flash memory that can electrically rewrite data, an EEPROM, and the like The non-volatile memory 24 is configured as a computer system, and executes predetermined processing by executing a program.

The CPU 21 is an example of a curve mirror detecting unit, an edge detecting unit, a likelihood output unit, a limiting unit, and a specifying unit in the present invention.
[1-2. Processing by Curve Mirror Detection Device 15]
The curve mirror mirror image output process executed by the CPU 21 of the curve mirror detection device 15 will be described based on the flowchart shown in FIG. This process is started when an ignition switch of the vehicle is turned on as an example.

  In S <b> 1, the CPU 21 acquires a captured image captured by the in-vehicle camera 11 from the in-vehicle camera 11. FIG. 3 is an example of a captured image. The captured image 31 is an image obtained by photographing the front of the vehicle when the vehicle enters a straight road composed of a straight road and a road intersecting with the straight road from the intersecting road side. In the captured image 31, a road 33, a telegraph pole 35, a plant 37, a curve mirror 39 arranged adjacent to the telegraph pole 35, and the like are shown.

  In S2, the CPU 21 filters the HSV color space of the captured image 31 acquired in S1 with a hue to generate a restricted image. The limited image is an image in which the hue of the captured image 31 is limited to a value within a preset range. Here, processing is performed so that the orange color that is the color of the outer edge and the back surface of the curved mirror remains.

  When expressed in the HSV color space of a cone, the hue space can be expressed as 0 to 360 °. Therefore, the captured image 31 is limited to display only red space of 0 to 60 ° and 300 to 360 °. This parameter is an example and can be set as appropriate.

  FIG. 4 schematically shows a limited image 41 obtained by performing hue filter processing on the captured image 31. The curve mirror 43 and the colors close thereto are shown relatively clearly, but the color region far from the color of the curve mirror 43 is removed by a filter. Thereby, unnecessary backgrounds other than the curve mirror can be removed, and the accuracy of the Hough (HOUGH) conversion in S3 can be improved.

  In S3, the CPU 21 detects a curve mirror candidate that is a region having the characteristics of a curve mirror from the restricted image 41 that has been subjected to the hue filter process in S2. Here, the CPU 21 performs edge detection on the restricted image 41, performs Hough transform on the detected edge (edge point), and extracts a circle or an ellipse that may correspond to the outer edge of the curve mirror.

  FIG. 5A shows the detection result when the edge is detected without executing the filtering process of S2, and FIG. 5B shows the edge detection result when the filtering process is executed. Compared with the case where the filter process is not executed, the edge detection result when the filter process is executed has fewer edges that are unrelated to the curve mirror. Therefore, curve mirror candidates can be extracted with high accuracy. Note that the lower part of each image is not subjected to edge detection because there is a low possibility that a curved mirror exists in the captured image due to the shooting area of the in-vehicle camera 11.

As a result of the Hough transform, if there are a plurality of circles or ellipses that satisfy the condition (the number of votes is a predetermined number or more), a plurality of curve mirror candidates are extracted.
In S4, the CPU 21 calculates a score for each of the curve mirror candidates detected in S3. The score is obtained by converting each feature amount calculated for the three features of saturation, shape, and lightness into a value indicating the likelihood as a curve mirror. The purpose of calculating the score and the specific method of calculating the score will be described below.

<Saturation score>
The mirror image reflected on the mirror surface of the curved mirror as viewed from the driver includes an object indicated by high saturation and an object indicated by low saturation. For example, buildings, color pavements, white lines, vehicles, and the like correspond to the objects shown with high saturation, and sky, roads, and the like correspond to the objects shown with low saturation, for example.

  Here, the area of the high-saturation part (the number of pixels whose saturation is greater than or equal to a predetermined threshold) and the area of the low-saturation part (the number of pixels whose saturation is less than the predetermined threshold) in the mirror image The ratio is obtained as a feature value, and a score corresponding to the feature value is output. The ratio here is an example of a saturation parameter which is a parameter based on the saturation of each pixel constituting the curve mirror candidate in the present invention.

The score is output as a higher value as the feature amount is closer to a preset matching value. The method for setting the adaptive value will be described below.
When the distribution of the above ratios in a plurality of actual curve mirrors was taken, it was found that there was a bias depending on the region, time, and the like. For example, in a certain area, as shown in FIG. 6A, the road image 61 and the sky 63 occupy a large proportion of the mirror image 51a of the curve mirror 51, and there are relatively many curve mirrors with a high proportion of low saturation. .

  On the other hand, in the area, as shown in FIG. 6B, a curved mirror having a high proportion of the high-saturation portion is relatively relative to the road 61 and the sky 63 in the mirror image 51a. Very few. In addition, when the road 61 is hidden by the vehicle 67, the ratio of a portion with higher saturation increases. However, if the road is not a road where traffic congestion frequently occurs, the possibility that the vehicle 67 is included in the mirror image is not high.

In consideration of such a situation, a range having a high probability that can be taken as the above-described ratio is set as a suitable value.
A specific process for calculating the saturation score will be described.

  First, the captured image 31 is filtered in the HSV color space. Here, unlike S2, pixels with low saturation are removed by filtering to generate an image with pixels with high saturation. Specifically, for each pixel of the captured image, a threshold value is provided to binarize the saturation, and a binarized image in which the gradation value is set to 0 or 255 is generated.

  Next, in the generated image, the number of pixels where the binary value of one curve mirror candidate extracted in S3 is 255 is divided by the number of pixels of the entire curve mirror candidate. The value becomes the feature amount of the curve mirror candidate.

  The score corresponding to the feature amount is determined in advance by a conversion table (not shown), and the score is uniquely determined when the feature amount is determined. In the present embodiment, when the number of pixels whose binary value is 255 is 10 to 20% of the whole, that is, when the feature amount is 0.1 to 0.2, the score becomes the highest, and the feature amount deviates from that value. The lower the score. This score is calculated for each curve mirror candidate.

<Shape score>
When a captured image includes a curved mirror, a large number of edges are likely to exist on the circumference of a circle or ellipse of a curved mirror candidate detected by the Hough transform. Therefore, the degree of edge continuity on the circumference of the curve mirror candidate extracted in S3 is obtained as a feature amount, and a higher score is output as the continuity is higher.

A specific process for calculating the shape score will be described.
First, the captured image 31 is filtered in the HSV color space. Here, pixels with low saturation are removed by filtering to generate an image with pixels with high saturation. Specifically, for each pixel of the captured image, a threshold value is provided to binarize the saturation, and a binarized image in which the gradation value is set to 0 or 255 is generated.

  Next, an edge is detected from the image. Then, the point where the edge is detected by searching the circumference of the curve mirror candidate extracted in S3 is set as a point of interest, and the surrounding 5 × 5 pixels are searched for each point of interest, and the existing edge is searched. Count.

  Then, the total number of edges of all the attention points in the curve mirror candidate is divided by the size of the circumference. The value is a feature amount indicating the degree of edge continuity, and is an example of a shape parameter in the present invention.

Similarly to the saturation score, the shape score is determined in advance in a conversion table (not shown) corresponding to the feature amount. The shape score increases as the feature amount increases.
<Lightness score>
The mirror image reflected on the mirror surface of the curved mirror as viewed from the driver includes an object having a high brightness and an object having a low brightness. For example, a pedestrian or a vehicle corresponds to the object indicated with high brightness, and the sky or road corresponds to the object indicated with low brightness.

  Here, the ratio of the area of the high brightness portion (number of pixels whose brightness is equal to or higher than a predetermined threshold) to the area of the low brightness portion (number of pixels whose brightness is lower than the predetermined threshold) in the mirror image is a feature amount. And outputs a score corresponding to the feature amount. The ratio here is an example of a brightness parameter that is a parameter based on the brightness of each pixel constituting the curve mirror candidate in the present invention.

  The score is output as a higher value as the feature amount is closer to a preset matching value. Similar to the saturation, in an actual curved mirror, the distribution of the ratio of the area of the high brightness portion and the area of the low brightness portion in the mirror image is biased depending on the region, time, and the like. In consideration of such a situation, a range having a high probability that can be taken as the above-described ratio is set as a suitable value.

A specific process for calculating the brightness score will be described.
Specifically, first, the captured image 31 is filtered in the HSV color space. Here, a pixel with high saturation and a pixel with low lightness are removed by filtering, and an image with pixels with low saturation and high lightness is generated. Specifically, for each pixel of the captured image, a threshold is provided for each of saturation and lightness, and the gradation value of a pixel whose saturation is equal to or less than the threshold and whose brightness is equal to or greater than the threshold is 255. Generates a binarized image in which the gradation value of a pixel whose brightness exceeds the threshold or whose brightness is less than the threshold is 0.

  Next, in the generated image, the number of pixels where the binary value of one curve mirror candidate extracted in S3 is 255 is divided by the number of pixels of the entire curve mirror candidate. The value becomes the feature amount of the curve mirror candidate.

  Similarly to the saturation score, the brightness score is determined by a conversion table (not shown) corresponding to the feature amount. In the present embodiment, when the number of pixels where the binary value is 255 is 60% of the total, that is, when the feature amount is 0.6, the score is highest, and the score is lower as the feature amount is farther from the value. . This score is calculated for each curve mirror candidate.

  In S5, CPU21 calculates the likelihood information (likelihood parameter L) regarding the likelihood as a curve mirror of a curve mirror candidate from the three scores calculated in S4. Then, by performing threshold processing on the likelihood parameter L, the curve mirror candidates are limited to those having a high likelihood. In the present embodiment, the following equation is used to calculate the likelihood parameter L.

L = (saturation score) × α ₁ + (shape score) × α ₂ + (lightness score) × α ₃
α ₁ + α ₂ + α ₃ = 1.0 (α _{1 to} α ₃ are arbitrary)
In this way, the likelihood parameter L is calculated for each curve mirror candidate, and a curve mirror candidate whose likelihood parameter L is equal to or greater than a predetermined threshold is extracted.

In S6, the CPU 21 specifies a curve mirror by clustering from the curve mirror candidates extracted in S5. In the present embodiment, clustering is executed in the following steps.
(I) One of the curve mirror candidates extracted in S5 is selected. This is a reference candidate.
(Ii) Extract other curve mirror candidates that intersect the reference candidate on the restricted image 41.
(Iii) A curve mirror candidate having the largest likelihood parameter L is selected from the reference candidates and the curve mirror candidates extracted in (ii) above.
(Iv) When a reference candidate is selected in (iii) above, the reference candidate is specified as a curve mirror. On the other hand, when a curve mirror candidate that is not a reference candidate is selected in (iii), (ii) is executed again using the selected curve mirror candidate as a reference candidate.

In subsequent S <b> 7, the CPU 21 extracts a mirror image of the curve mirror identified by the clustering in S <b> 6 from the captured image 31.
In S <b> 8, the CPU 21 displays a mirror image extracted from the captured image 31 on the display 13. Although the display method is not particularly limited, as described in FIG. 7 of Japanese Patent Application Laid-Open No. 2009-212624, a mirror image may be displayed together with the travel guide screen, or only a mirror image may be displayed. After S8, the process returns to S1.

[1-3. effect]
[1A] In the curve mirror detection device 15 of the present embodiment, the CPU 21 selects one curve mirror candidate that is a region having the characteristics of a curve mirror from the captured image 31 captured by the in-vehicle camera 11 that captures the periphery of the vehicle. This is detected.

  Then, the likelihood parameter L is output based on the above-described three feature amounts of the saturation parameter, the brightness parameter, and the shape parameter, and the curve mirror candidate having a large likelihood parameter L is specified as a curve mirror.

  In the captured image 31, the area where the curve mirror is actually projected is a curve mirror candidate detected by the Hough transform, and the brightness parameter and the saturation parameter match when compared with the area where the curve mirror is not actually projected. Probability that is close to the value and the degree of continuity of the edge points is high.

  Therefore, the curve mirror detection device 15 can appropriately output the likelihood of the curve mirror candidate as the curve mirror by referring to the above-described feature amount of each of the curve mirror candidates. Whether it is a mirror or not can be determined with high accuracy.

  [1B] In the curve mirror detection device 15 of the present embodiment, the CPU 21 calculates a score indicating likelihood from each of the three feature quantities, and calculates a likelihood parameter L based on these scores. Specifically, the saturation score and the lightness score are output such that the closer to the matching value, the higher the likelihood, and the shape score is output so that the likelihood increases as the value increases.

  Therefore, even if the score of any one of the feature quantities is low, the likelihood as a curve mirror becomes high if the scores of other feature quantities are high. Therefore, it is possible to detect a curve mirror that is resistant to noise.

  As an example of a low saturation score, the case shown in FIG. 6B can be considered. Further, as an example where the score of lightness is low, as shown in FIG. 7 (A), a large vehicle 69 may be included in the mirror image so as to cover the road 61.

  Further, examples where the edge continuity is low include a case where the mirror has no edge, a case where the edge of the mirror is not photographed, or a case where the edge is peeled off. Specifically, as shown in FIG. 7B, the edge of the curve mirror 51 is small, and the plant 71 reflected in the mirror image 51a and the plant 73 positioned behind the curve mirror 51 overlap. Can be considered. In such a case, the edge detection accuracy at the left end 51b of the mirror image 51a is lowered, and the edge continuity is lowered.

  In this way, depending on the arrangement status of the curve mirror, one score may be lowered, but a high likelihood parameter can be obtained depending on the result of the other scores, so the curve mirror can be determined with high accuracy. Can do.

  [1C] In the present embodiment, the captured image 31 is shown in the HSV color space including hue, saturation, and brightness. The CPU 21 generates a limited image 41 that is an image in which the hue of the captured image 31 is limited to a value within a preset range so that the detection accuracy of the curve mirror is high, and uses the limited image 41. Detect curve mirror candidates.

  The curve mirror detection device 15 configured in this way is difficult to detect an edge in a region showing a color different from that of the curve mirror in the captured image 31, and therefore edge detection for detecting a curve mirror candidate (S3 in FIG. 2). ) Can be improved. As a result, the detection accuracy of the curve mirror can be increased.

  In the present embodiment, the configuration in which the hue filter is performed so that the orange color is emphasized on the assumption that the edge of the curve mirror is orange is illustrated. However, if the accuracy of edge detection for detecting a curve mirror candidate is increased, not only the hue but also the saturation and the lightness may be limited, or two or more of them may be combined and limited.

  The captured image 31 is not necessarily processed as an HSV color space, and any one or more elements are appropriately limited in a color space having at least one of hue, saturation, and lightness. The same effect can be obtained.

If the captured image does not have an HSV color space, it may be converted into an HSV color space.
Further, it may be configured not to perform filter processing for edge detection. In this case, there is a possibility that the number of curve mirror candidates that are detected in a region where there is no curve mirror in the captured image may increase, but the possibility that a region where the curve mirror exists is not detected as a curve mirror candidate is reduced. . In other words, the number of false detections of curve mirror candidates increases, while the undetected rate of curve mirrors can be reduced.

  [1D] In the curve mirror detection device 15 of the present embodiment, the CPU 21 detects an edge of the captured image 31 and performs a Hough transform on the detected edge to thereby obtain a curve mirror candidate that is a region having the characteristics of the curve mirror. Detect one or more. Therefore, it is possible to detect the curve mirror candidate with high accuracy by the Hough transform.

  The specific method for detecting the curve mirror candidate from the captured image 31 is not limited to the Hough transform, and various known methods can be employed. For example, it is conceivable to detect a curve mirror candidate by pattern matching using contour data of a curve mirror.

  [1E] In the curve mirror detection device 15 according to the present embodiment, the CPU 21 is based on the ratio of pixels in which at least one of saturation and brightness belongs to a predetermined range among the pixels constituting the curve mirror candidate. The likelihood information (likelihood parameter L) regarding the likelihood of the curve mirror candidate as the curve mirror is output.

  The curve mirror detection device 15 configured as described above can appropriately output the likelihood of the curve mirror candidate as a curve mirror using the saturation and brightness parameters, and can detect the curve mirror with high accuracy.

[2. Other Embodiments]
As mentioned above, although embodiment of this invention was described, this invention can take a various form, without being limited to the said embodiment.

  [2A] In the above-described embodiment, the configuration in which the likelihood parameter L is calculated using the three feature amounts is illustrated, but the configuration is not necessarily limited to the configuration in which the three are calculated. For example, even with a configuration using two of these feature quantities, highly accurate curve mirror detection can be realized. Furthermore, the likelihood parameter L may be calculated in consideration of feature quantities and scores other than the above three feature quantities.

  [2B] In the above embodiment, the likelihood information regarding the likelihood is calculated by calculating one likelihood parameter L for each curve mirror candidate. However, the likelihood information is not limited to this. . For example, it is conceivable that the likelihood of each of the three feature amounts in the curve mirror candidate is determined, and a curve mirror candidate having a large number of feature amounts determined to have a high likelihood is specified as a curve mirror.

  [2C] The adaptive value used when calculating the score in S4 of FIG. 2, the conversion table used when calculating the score from the feature amount, and various threshold values are switched according to the time and the region where the vehicle is located. May be.

  [2D] In the above embodiment, a ratio between a pixel whose saturation is greater than or equal to a predetermined threshold and a pixel that is less than the threshold is used as the saturation parameter, and a pixel whose brightness is greater than or equal to the predetermined threshold is used as the brightness parameter. Although the configuration using the ratio with the pixel that is less than the threshold is illustrated, the configuration may use a value other than these as a feature amount as long as it is a parameter based on saturation or brightness.

  In other words, it is a parameter that represents the saturation and lightness state of the pixels in the curve mirror candidate, and compares the parameter with a previously acquired matching value (a value that is likely to be taken by an actual curve mirror). Any information can be used as long as the information indicates that the curve mirror candidate has a higher likelihood as a curve mirror as the matching value is closer.

For example, it is conceivable to use the average value of the saturation and lightness of all the pixels of one curve mirror candidate, or to use the ratio of pixels whose saturation and lightness fall within a predetermined range as the feature amount.
Similarly, the calculation method of the shape parameter is not particularly limited as long as it is a parameter indicating edge continuity. For example, parameters such as distance between edges and density can be used. The conforming value of the shape parameter can be set to the maximum value that the shape parameter can take.

  [2E] In the above-described embodiment, the configuration in which the curve mirror is specified from the curve mirror candidate by the threshold processing of the likelihood parameter L and the clustering (S5 and S6 in FIG. 2) is illustrated, but the curve mirror is specified by other methods. May be.

  [2F] The functions of one component in the above embodiment may be distributed as a plurality of components, or the functions of a plurality of components may be integrated into one component. In addition, at least a part of the configuration of the above embodiment may be replaced with a known configuration having a similar function. Moreover, you may abbreviate | omit a part of structure of the said embodiment. In addition, at least a part of the configuration of the above embodiment may be added to or replaced with the configuration of the other embodiment. In addition, all the aspects included in the technical idea specified only by the wording described in the claim are embodiment of this invention.

  [2G] In addition to the curve mirror detection device 15 described above, there are various systems such as a system including the curve mirror detection device 15 as a constituent element, a program for causing a computer to function as the curve mirror detection device 15, and a medium on which the program is recorded. The present invention can also be realized in the form of.

DESCRIPTION OF SYMBOLS 11 ... Car-mounted camera, 15 ... Curve mirror detection apparatus, 21 ... CPU, 31 ... Captured image, 41 ... Restriction image

Claims (7)

A curve mirror detection means (21) for detecting one or more curve mirror candidates, which are regions having characteristics of a curve mirror, from a captured image captured by an imaging means (11) for capturing the periphery of the vehicle;
In the captured image, a saturation parameter that is a parameter based on the saturation of each pixel constituting the curve mirror candidate, a brightness parameter that is a parameter based on the brightness of each pixel constituting the curve mirror candidate, and the captured image On the basis of two or more feature quantities selected from the group consisting of shape parameters, which are parameters based on the continuity of the edge points corresponding to the curve mirror candidates detected by the edge detection means for detecting the edge points. And (21) a likelihood output means for outputting likelihood information relating to the likelihood of the curve mirror candidate as a curve mirror. (15) A curve mirror detection device (15). The saturation parameter is a ratio of pixels whose saturation is greater than or equal to a predetermined threshold and pixels less than the threshold,
The curve mirror detection device according to claim 1, wherein the brightness parameter is a ratio of a pixel whose brightness is equal to or higher than a predetermined threshold and a pixel whose brightness is lower than the threshold. The likelihood output means indicates that the likelihood of the curve mirror candidate as a curve mirror is higher as the feature amount is closer to a suitable value acquired in advance as a highly probable value that the curve mirror can take. Information is output. The curve mirror detection device according to claim 1 or 2 characterized by things. A curve mirror detection means (21) for detecting one or more curve mirror candidates, which are regions having characteristics of a curve mirror, from a captured image captured by an imaging means (11) for capturing the periphery of the vehicle;
A likelihood output that outputs likelihood information related to the likelihood of the curve mirror candidate as a curve mirror based on the ratio of pixels that have at least a saturation in a predetermined range among the pixels constituting the curve mirror candidate Means and (21) The curve mirror detection apparatus (15) characterized by the above-mentioned. The captured image is shown in a color space including at least one of hue, saturation, and brightness,
And a limiting unit (21) that generates a limited image that is an image in which at least one of hue, saturation, and brightness of the captured image is limited to a value within a preset range. And
The curve mirror detection unit according to any one of claims 1 to 4, wherein the curve mirror detection unit detects the curve mirror candidate using the limited image generated by the limitation unit. apparatus. The curve mirror detection unit detects an edge point of the captured image, and detects the curve mirror candidate by performing a Hough transform on the detected edge point. The curve mirror detection device according to item 1. A specifying unit (21) for specifying a curve mirror displayed in the captured image based on the curve mirror candidate detected by the curve mirror detection unit and the likelihood information output by the likelihood output unit; The curve mirror detection device according to any one of claims 1 to 6, further comprising: