JP2014164461A - Pedestrian detector and pedestrian detection method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately detect a pedestrian from an image picked up by an on-vehicle camera.SOLUTION: A pedestrian detector extracts candidate images for a pedestrian from a picked-up image by retrieving the picked-up image with a human type. The detector detects a travel lane from the picked-up image. The detector detects a lane width at a location of each candidate image in the picked-up image and detects as a pedestrian a piece of picked-up image in which the ratio of the size of a candidate image to the lane width is within a prescribed range. In principle, an image picked up by an on-vehicle camera contains travel lane(s) and the width of each travel lane is almost determined. Therefore, when the width of a travel lane is used as a reference, a pedestrian in a picked-up image can be accurately detected by determining whether a candidate image has an appropriate size as a pedestrian.

Description

  The present invention relates to a technique for detecting a pedestrian from an image taken by an in-vehicle camera.

  In recent vehicles, a pedestrian is detected from an image around a vehicle photographed by an in-vehicle camera, and the result is used for various driving assistances. Moreover, in order to detect the pedestrian reflected in the image, the technique which searches the shape of a pedestrian as follows is used. First, the pedestrian's external shape has several features that most pedestrians have but no objects other than pedestrians. Therefore, such features of the outer shape are stored as features of the outer shape of the pedestrian (that is, a human figure). Then, by setting search windows of various sizes in the image and moving the search window, the pedestrian is extracted by searching for an image (humanoid image) that has the characteristics of the pedestrian's outer shape. To do.

However, in this method, for example, an image in which a road sign appears to be illuminated by illumination from a vehicle headlamp or an image in which illumination is reflected on a window glass may be erroneously detected as an image of a pedestrian.
Therefore, focusing on the fact that pedestrians appear smaller because the distant landscape appears in the upper part of the image, a technology that searches for pedestrians using a smaller search window from the bottom to the top of the image. Has been proposed (Patent Document 1). In this way, it is possible to avoid erroneously detecting a pedestrian in an image that appears smaller in the lower part of the image or an image that appears larger in the upper part of the image.

JP 2008-152672 A

  However, even with the proposed technique, there is a problem that it is difficult to detect a pedestrian in an image with sufficient accuracy. This is because it is true that the pedestrian appears smaller on the upper side of the image, but the size of the pedestrian depends on the position on the image. It will be different for each vehicle due to slight differences. In addition, even when the front side of the vehicle is tilted in the vertical direction or there is a slope in the traveling direction during traveling, the size of the pedestrian that appears varies depending on the position on the image. For this reason, it is necessary to search for pedestrians with a considerable width in the size of the search window. As a result, images that appear extremely small at the bottom of the images (images that are confusing with pedestrians) Only an image that is extremely large in the upper part (an image that is confusing with a pedestrian) can be determined not to be a pedestrian. For these reasons, the proposed technique has a problem that it is difficult to detect a pedestrian in an image with sufficient accuracy.

  The present invention has been made in view of the above-described problems, and an object thereof is to provide a technique capable of accurately detecting a pedestrian in an image.

  In order to solve the above-described problem, the pedestrian detection device and the pedestrian detection method of the present invention extract a candidate image of a pedestrian from a captured image, and the size of the candidate image and the position of the candidate image in the captured image. And remember. Further, the traveling lane is detected from the captured image. Then, the lane width at the position of the candidate image is detected, and a captured image in which the ratio between the size of the candidate image and the lane width is within a predetermined range is detected as a pedestrian.

  Since the vehicle travels on the road, the traveling lane is shown in principle in the captured image. Moreover, the driving lane is continuously captured from near to far from the vehicle, and the lane width is roughly determined. Therefore, if the lane width of the traveling lane is used as a reference, it can be accurately determined whether or not the size of the candidate image is an appropriate size for a pedestrian, so that the pedestrian in the captured image can be accurately determined. It becomes possible to detect.

  In the pedestrian detection apparatus of the present invention described above, the feature amount of the outer shape of the pedestrian is stored, and the candidate image may be extracted by searching for a captured image using the feature amount. . And the representative point corresponding to the position where the pedestrian standing searched using the feature amount is also stored as the feature amount, and the position of the representative point when the candidate image is extracted is stored as the candidate image position. It is good to do.

  In this way, if the candidate image is a pedestrian, the lane width at the position where the pedestrian is standing can be compared with the size of the candidate image, so is the candidate image a pedestrian? It is possible to accurately determine whether or not. As a result, pedestrian detection accuracy can be further improved.

  In the pedestrian detection device of the present invention described above, it is determined whether or not the captured image is an image captured at night, and when the captured image is an image captured at night, the lane width and A pedestrian may be detected from the candidate images based on the ratio with the candidate image size.

  In this way, a pedestrian can be detected with high accuracy by using lane width information at night when it is difficult to clearly recognize the outer shape of the object. Further, during daytime when it is easy to clearly recognize the outer shape of the object, it is not necessary to detect the traveling lane, so the processing load of the process for detecting the pedestrian can be reduced.

It is explanatory drawing which shows the vehicle 1 carrying the pedestrian detection apparatus 100 of a present Example. It is a block diagram which shows the internal structure of the pedestrian detection apparatus 100 of a present Example. 4 is a flowchart of a pedestrian detection process in which the pedestrian detection device 100 of the present embodiment detects a pedestrian from a captured image of the vehicle 1. It is explanatory drawing which illustrated the picked-up image image | photographed at night. It is explanatory drawing about the method of extracting a candidate image from a picked-up image using a human figure. It is explanatory drawing which illustrated the traveling lane detected from the picked-up image. It is explanatory drawing about the method of detecting a pedestrian based on the ratio of the lane width in a candidate image position, and the size of the humanoid which extracted the candidate image. It is explanatory drawing about the reason why a pedestrian can be detected accurately even if the vehicle leans back and forth. It is explanatory drawing about the reason which can detect a pedestrian accurately even if a road gradient changes ahead of a vehicle. It is a block diagram which shows the internal structure of the pedestrian detection apparatus 200 of a modification.

Hereinafter, examples will be described in order to clarify the contents of the present invention described above.
A. Apparatus configuration of this embodiment:
FIG. 1 shows a vehicle 1 equipped with a pedestrian detection device 100 of the present embodiment. As shown in the figure, a pedestrian detection device 100 is connected to a vehicle-mounted camera 120 mounted on the vehicle 1, and the pedestrian detection device 100 analyzes a captured image obtained by the vehicle-mounted camera 120 to generate an image. Extract pedestrians inside. Further, when driving at night, the driver or a control device (not shown) turns on the headlamp 10 and the vehicle-mounted camera 120 captures an image illuminated by the headlamp 10.
The pedestrian detection device 100 is a microcomputer configured by connecting a CPU, ROM, RAM, and the like to each other via a bus.

FIG. 2 is a block diagram showing the internal configuration of the pedestrian detection device 100, focusing on the function for detecting a pedestrian in a captured image. As illustrated, the pedestrian detection device 100 can be divided into a human type storage module 102, a candidate image extraction module 104, a travel lane detection module 106, a candidate image storage module 108, and a pedestrian detection module 110. it can.
The “module” is an abstract concept in which the pedestrian detection device 100 is divided for convenience, focusing on functions, and does not necessarily indicate that the pedestrian detection device 100 is physically divided. Absent. Therefore, the module can be realized by a computer program, or can be realized by an electronic circuit or a RAM in which a plurality of IC chips are combined.

The human-type storage module 102 stores feature value data used when detecting a pedestrian from a captured image. That is, the pedestrian in the captured image can be captured in various aspects such as a walking aspect, a running aspect, and a stationary aspect. However, as long as the pedestrian is a pedestrian, its outer shape is an object other than a pedestrian. Have common features that cannot be taken. Or conversely, a feature quantity that a pedestrian cannot take can be assumed. Data obtained by collecting a plurality of such feature quantities is called a “humanoid”. The human type storage module 102 stores such human type data.
Candidate image extraction module 104 searches for a pedestrian in a captured image by searching for an image (an image having a feature amount of a pedestrian) that matches human-type data in the captured image. Are extracted as candidate images (candidate images). Here, the pedestrian image is not the pedestrian image but the pedestrian candidate image because the non-pedestrian object appears to have the pedestrian feature amount, or two objects happen to happen. This is because the pedestrian candidates are determined at this point in consideration of the fact that the pedestrians may appear to have the feature amount.
The travel lane detection module 106 detects a travel lane in the captured image. Since the vehicle 1 travels on the road in principle, the travel lane in which the vehicle 1 travels is reflected in the image taken by the in-vehicle camera 120. Therefore, the travel lane detection module 106 detects the travel lane from the captured image.
For each extracted candidate image, the candidate image storage module 108 stores the size of the candidate image and the position of the candidate image on the captured image.
The pedestrian detection module 110 uses the candidate image size and the position of the candidate image stored in the candidate image storage module 108 and the information on the travel lane detected by the travel lane detection module 106 to generate a candidate image (pedestrian Pedestrians are detected from (candidate images). A method for detecting a pedestrian from candidate images will be described later.

  In this embodiment, the human type storage module 102 corresponds to “feature amount storage means” in the present invention, and the candidate image extraction module 104 corresponds to “candidate image extraction means”. Further, the travel lane detection module 106 corresponds to “travel lane detection means” in the present invention, the candidate image storage module 108 corresponds to “candidate image storage means”, and the pedestrian detection module 110 corresponds to “pedestrian detection means”. Correspond.

B. Pedestrian detection process:
FIG. 3 shows a flowchart of processing (pedestrian detection processing) in which the pedestrian detection device 100 detects a pedestrian from a captured image. When the pedestrian detection process is started, first, a captured image of the vehicle 1 is captured using the in-vehicle camera 120 (S100).
FIG. 4 illustrates a captured image obtained by capturing the front of the vehicle 1 at night. As shown in the figure, a pedestrian walking on the sidewalk, a plurality of streetlights that illuminate the surroundings, a reflector that reflects the light of the headlamp 10 of the vehicle 1 and a road sign are reflected in the photographed image. .

Next, candidate images are extracted from the photographed captured images using a humanoid (S102 in FIG. 3). Here, as described above, the human figure is a collection of a plurality of data indicating the feature amount of the outer shape of the pedestrian, and does not indicate a human shape. However, since it is difficult to generate an image with the abstract concept of “outline shape feature quantity”, in the following, for convenience, a human figure is displayed using a figure representing a human shape. FIG. 5A illustrates a human figure displayed for convenience using a human shape. When such a humanoid is used, for example, even when a pedestrian's limbs are hidden behind a building or a pedestrian is facing sideways and the limb cannot be confirmed, it can be extracted as a candidate image.
Of course, when a pedestrian is detected using a human figure, an object other than a human is extracted as a candidate for a pedestrian if the shape is close to that of a human. However, in the pedestrian detection apparatus 100 of the present embodiment, only the pedestrian can be detected from the candidate images by detecting the lane width of the traveling lane. For this reason, actually, an object other than a human is not erroneously detected as a pedestrian. This point will be described in detail later.

  In addition, since a pedestrian in the image appears larger as it approaches the vehicle 1 and becomes smaller as it moves away from the vehicle 1, various types of pedestrians are detected when detecting a pedestrian in the image using a human figure. A pedestrian is detected using a search window having a size (eight in the example shown in FIG. 5A, P1 to P8). Correspondingly, in FIG. 5A, a human type is set for each search window size (eight h1 to h8). However, since the human type is abstract data as described above, it is possible to apply one human type to search windows having different sizes. In addition, “representative points” indicating positions where pedestrians are standing are set in the humanoid of the present embodiment.

FIG. 5B conceptually shows how candidate images are extracted by searching for a captured image using the above-described humanoid. When extracting candidate images, the above-mentioned search window is overlaid on the photographed image, and whether the photographed image in the search window matches the human shape (whether it has a feature amount of the pedestrian's external shape) Or not). As a result, when the captured image in the search window matches the human figure, the coordinates of the representative point of the human figure at the position of the search window are detected as the position of the candidate image.
In S102 of FIG. 3, candidate images are extracted from the captured image by moving within the captured image for a plurality of search windows having different sizes.

When the candidate images are searched for the captured image of FIG. 4 in this way, five candidate images a to e are extracted as shown in FIG. In the illustrated example, images other than humans are extracted as candidate images c to e.
In addition, the result of extracting the candidate images in this manner is the size of the search window used for extracting the candidate images (eight types of h1 to h8 in this embodiment), and the representative when the candidate images are extracted. The coordinates of the points are stored in the candidate image storage module 108. FIG. 5C shows how candidate images are stored in this way.
Note that the size of the search window used for extracting the candidate image corresponds to the “size of candidate image” in the present invention. Further, the coordinates of the representative point when the candidate image is extracted correspond to the “position of the candidate image” in the present invention.

  However, the extracted candidate images may include images of objects other than humans (candidate images c to e in FIG. 5B). This is because, in order to avoid the detection omission of the pedestrian, a human shape is used that is extracted if it is as close as possible to the outer shape of the human. However, such candidate images extracted from non-human objects can be shaken out by the following method.

First, a travel lane in which the vehicle 1 is traveling is detected from the captured image (S104 in FIG. 3). Various well-known methods can be used for detecting the traveling lane. For example, a traveling lane can be detected by detecting two parallel white lines in a captured image using edge detection. Of course, a white broken line may be detected instead of the white line. In addition, even when a part of the white line is hidden by a vehicle that has stopped on the shoulder, it is possible to estimate the position of the traveling lane in consideration that the lane width changes only continuously.
FIG. 6 illustrates a state where a traveling lane is detected by detecting two white lines (L and R) in the captured image of FIG.

As described above, candidate images in the captured image are extracted (S102 in FIG. 3), and when a traveling lane is detected (S104), one candidate image is selected from the extracted candidate images (S106). Subsequently, the lane width at the position of the representative point of the selected candidate image is detected (S108). Here, the “lane width at the position of the representative point” is the lane width when the traveling lane is viewed in the horizontal direction from the position of the representative point. The lane width can be obtained by scanning the captured image in the horizontal direction from the position of the representative point and detecting the distance between the intersection points obtained for the white line L on the left side and the white line R on the right side of the traveling lane. .
Referring to the example shown in FIG. 7, for the candidate image a in the figure, the captured image is scanned in the horizontal direction from the position of the representative point, and each of the white line L on the left side and the white line R on the right side of the driving lane. The lane width wa is detected from the interval between the intersections. Further, the lane width wc is detected for the candidate image displayed as “c” in the drawing, and the lane width we is detected for the candidate image displayed as “e” in the drawing.
Further, as described above with reference to FIG. 5A, the representative point of the candidate image is set at a position where a pedestrian detected by a human figure stands. Therefore, the process of detecting the lane width at the position of the representative point (S108 in FIG. 3) assumes that the candidate image is a pedestrian and detects the lane width at the position where the pedestrian stands. It corresponds to.

Subsequently, it is determined whether the ratio between the size of the candidate image and the lane width at the representative point is within a predetermined range (S110 in FIG. 3). This is due to the following reason. First, the height of a human being is within a predetermined range (typically 1 to 2 m), and the lane width is not significantly different from the predetermined length (around 2.5 m). Therefore, if the candidate image is a pedestrian, the ratio between the size of the candidate image (the size of the search window when the candidate image is extracted) and the lane width at the representative point of the candidate image is within a predetermined range. There is a certain trap. Conversely, if this ratio is not within the predetermined range, it may be considered that the candidate image is not a pedestrian.
Since the vehicle 1 travels on the road in principle, the traveling lane can be detected from the captured image captured by the in-vehicle camera 120 in most cases. In addition, since the traveling lane continuously exists from near to far from the vehicle 1, the lane width at the position of the candidate image can be detected regardless of the position where the candidate image is detected in the captured image. Therefore, if it is determined whether the candidate image is a pedestrian based on the ratio with the lane width using the lane width of the traveling lane as a reference, the pedestrian can be detected with high accuracy.

In the pedestrian detection process of the present embodiment, for the reasons described above, it is determined whether the ratio between the size of the candidate image and the lane width is within a predetermined range (S110). S110: yes), the candidate image is detected as a pedestrian (S112).
On the other hand, when the ratio between the size of the candidate image and the lane width is not within the predetermined range (S110: no), the candidate image is not detected as a pedestrian. In the example shown in FIG. 7, the candidate image a is detected as a pedestrian, but the candidate images c and e are not detected as pedestrians.

  As described above, when a series of processing (S108 to S112 in FIG. 3) for one candidate image is completed, whether or not the above-described processing has been performed for all candidate images extracted from the captured image. Is determined (S114). As a result, if an unprocessed candidate image remains (S114: no), the process returns to S106 to select one new candidate image, and then the above-described series of processing (S108 to S112) is repeated. As a result, when processing is completed for all candidate images (S114: yes), the pedestrian detection processing shown in FIG. 3 is terminated.

  As described above, the pedestrian detection apparatus 100 according to the present embodiment can shake off an image of an object other than a human being from candidate images by using the lane width. For this reason, at the stage of extracting as a candidate image, an image that looks like a human being can be extracted as much as possible, so that detection errors of pedestrians can be reduced. Nevertheless, since the candidate image in which an object other than a human is captured can be shaken off, the detection accuracy of a pedestrian can be greatly improved.

  Moreover, since the pedestrian detection apparatus 100 of a present Example detects a pedestrian based on the ratio of the magnitude | size of a candidate image and a lane width, even if the vehicle 1 inclines in the front-back direction, it will not receive this influence. A pedestrian can be detected with high accuracy. Here, “the inclination of the vehicle 1 in the front-rear direction” means that the front side of the vehicle 1 is inclined downward or upward as in the case where a heavy load is placed on either the front side or the rear side of the vehicle 1, for example. Say.

  As illustrated in FIG. 8, when the vehicle 1 tilts in the front-rear direction, the shooting range of the in-vehicle camera 120 moves greatly. A rectangle indicated by a thick solid line in the drawing represents a shooting range when the front side of the vehicle 1 is tilted downward, and a rectangle indicated by a thick broken line in the drawing is a shooting range when the vehicle 1 is not tilted. Represents. As described above, when the front side of the vehicle 1 is tilted downward, the photographing range is shifted downward, and as a result, the position where the candidate image appears in the photographed image moves upward.

  Here, since a distant landscape appears in the photographed image as it goes upward, a pedestrian in the photographed image usually appears as small as it goes upward in the image. However, when the position where the candidate image appears moves upward because the front side of the vehicle 1 is tilted downward, the size of the candidate image does not change. Accordingly, in the captured image when the front side of the vehicle 1 is tilted downward, it appears that a pedestrian larger than the actual size appears in the upper part of the image. For this reason, for example, if the candidate image a in FIG. 8 is determined only from the position in the captured image, it may appear too large for a human being and may be determined not to be a pedestrian. There is.

  The same thing can happen when the front side of the vehicle 1 is tilted upward. That is, when the front side of the vehicle 1 is tilted upward, the position where the candidate image on the photographed image is captured at the same size moves downward. Since the photographed image shows the scenery closer to the bottom of the image, pedestrians should be larger in the original image, but the candidate image on the photographed image moves in the downward direction only with its size unchanged. So it looks smaller than the actual size. For this reason, it may be judged that it is not a pedestrian because it is too small for a human being.

  On the other hand, as described above, the pedestrian detection apparatus 100 according to the present embodiment extracts a pedestrian by comparing the size of the candidate image with the lane width. In this way, even if the front side of the vehicle 1 is tilted, the ratio between the size of the candidate image and the lane width does not change, so that it is possible to accurately detect a pedestrian without being affected by the tilt of the vehicle 1. .

  The case where the front side of the vehicle 1 is tilted has been described above. However, a similar phenomenon (that is, a phenomenon that looks larger or smaller than the actual size) can also occur when the road gradient changes in front of the vehicle 1.

FIG. 9 illustrates a captured image when the road gradient changes in front of the vehicle 1. In the illustrated example, the vehicle 1 is traveling on a horizontal portion, and an uphill (a hatched range in the figure) is present ahead. In addition, there is a horizontal road behind it.
In such a case, the position of the pedestrian standing on the other side of the uphill moves upward as compared with the case where there is no uphill. However, since the pedestrian does not appear to be small in association therewith, if it is determined based only on the position on the photographed image, it appears that a pedestrian larger than the actual size is reflected.

Contrary to the case described above, when the vehicle 1 is on the downhill in front of the vehicle 1, the pedestrian standing on the other side of the downhill is on the captured image as compared with the case where there is no downhill. The position where the image appears is moved downward. However, since the pedestrian does not appear to be large, the pedestrian appears smaller than the actual size.
However, even in such a case, the ratio between the size of the candidate image and the lane width does not change. For this reason, in the pedestrian detection apparatus 100 of a present Example, it becomes possible to extract a pedestrian with sufficient precision.

C. Modified example:
In the above-described embodiment, the description has been given on the assumption that the pedestrian detection process of FIG. 3 is always applied to the captured image obtained by the in-vehicle camera 120. However, the above-described pedestrian detection process is particularly effective for an image taken with the headlamp 10 turned on, particularly at night. This is because it is difficult to understand the outer shape of objects including pedestrians at night, so from the viewpoint of preventing pedestrian detection omission, the criteria for determining whether or not they are similar to the pedestrian's outer shape may be slightly relaxed. desirable. On the other hand, at night, light from the headlamp 10 is reflected, or street lights are shined, and it is easy for objects other than pedestrians to be seen by pedestrians. In other words, since it is easy to erroneously detect an object other than a pedestrian as a pedestrian at night, the detection accuracy of the pedestrian can be greatly improved by applying the above-described pedestrian detection process.

  From this, it is possible to take an image with a clear outer shape of an object including a pedestrian as in the daytime, and further, an object other than the pedestrian is affected by the headlamp 10 or the like. Under the condition that there is a low possibility that the pedestrian will appear, a pedestrian may be detected using a general method that does not use lane width information.

  FIG. 10 illustrates a pedestrian detection device 200 of such a modification. In addition to the pedestrian detection device 100 described above with reference to FIG. 2, the pedestrian detection device 200 of the illustrated modification is provided with a night detection module 204 that detects whether or not the vehicle is in a night driving state. In the example shown in FIG. 10, it is determined whether or not the vehicle is in a night driving state based on whether or not the headlamp 10 is lit. You may judge by other methods, such as detecting.

Further, the human-type storage module 202 of the modification stores a human type for daytime and a human type for nighttime. Then, it receives information from the night detection module 204 as to whether or not the vehicle is in the night driving state, and if it is in the night driving state, outputs a human type for night time to the candidate image extraction module 104 to Otherwise, the humanoid for daytime is output to the candidate image extraction module 104.
Data similar to that illustrated in FIG. 5A (a collection of feature amounts) can be used as a humanoid for nighttime use. Further, stricter data can be used as the human type for daytime than the human type for nighttime.

  Information detected by the night detection module 204 (information on whether or not the vehicle is in the night driving state) is also supplied to the traveling lane detection module 106 and the pedestrian detection module 110. In the daytime running state, the running lane detection module 106 stops detecting the running lane. As a result, in the daytime running state, information on the travel lane cannot be obtained from the travel lane detection module 106, so the pedestrian detection module 110 detects a pedestrian without using the information on the travel lane.

In the pedestrian detection device 200 according to the above modification, the pedestrian can be accurately detected by using the lane width information under conditions where it is difficult to clearly recognize the outer shape of the object, such as during night driving. Can be detected. Also, under conditions where it is easy to clearly recognize the outer shape of an object, such as during daytime driving, it is not necessary to detect the driving lane, so the processing burden of processing to detect pedestrians can be reduced. It becomes.
The nighttime detection module 204 of the modification corresponds to “determination means” in the present invention.

DESCRIPTION OF SYMBOLS 1 ... Vehicle, 100 ... Pedestrian detection apparatus, 102 ... Human-type memory module,
104 ... Candidate image extraction module, 106 ... Driving lane detection module,
108 ... Candidate image storage module 110 ... Pedestrian detection module,
120 ... In-vehicle camera, 200 ... Pedestrian detection device,
202 ... Human type memory module, 204 ... Night running detection module.

Claims (4)

A pedestrian detection device that detects a pedestrian from a photographed image taken by an in-vehicle camera,
Candidate image extraction means for extracting a candidate image of the pedestrian by searching for an image in which the pedestrian is captured from the captured image;
When the candidate image is extracted, candidate image storage means for storing the size of the candidate image and the position of the candidate image in the captured image;
Traveling lane detection means for detecting a traveling lane from the captured image;
The candidate image in which the ratio between the lane width of the traveling lane when viewing the traveling lane in the horizontal direction from the position of the candidate image and the size of the candidate image is within a predetermined range is detected as the pedestrian. A pedestrian detection device comprising: pedestrian detection means. The pedestrian detection device according to claim 1,
A feature amount storage means for storing the feature amount of the outer shape of the pedestrian;
The candidate image extracting means extracts the candidate image by searching the captured image using the feature amount,
The feature amount storage means stores a representative point corresponding to a position where the pedestrian is represented by the feature amount,
The candidate image storage means is means for storing, when the candidate image is extracted, a position of the representative point when the candidate image is extracted as a position of the candidate image. apparatus. The pedestrian detection device according to claim 1 or 2,
A determination means for determining whether or not the captured image is an image captured at night;
The pedestrian detection device is used when the photographed image is an image photographed at night. A pedestrian detection method for detecting a pedestrian from a photographed image taken by an in-vehicle camera,
Searching for an image of the pedestrian in the captured image, and extracting candidate images of the pedestrian;
When the candidate image is extracted, storing the size of the candidate image and the position of the candidate image in the captured image;
Detecting a driving lane from the captured image;
The candidate image in which the ratio between the lane width of the traveling lane when viewing the traveling lane in the horizontal direction from the position of the candidate image and the size of the candidate image is within a predetermined range is detected as the pedestrian. A pedestrian detection method comprising the steps of:

Images