JP2004180200A - Vehicle recognition apparatus and recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To achieve stable vehicle recognition by adopting a robust voting system in vehicle recognition. <P>SOLUTION: When the apparatus acquires a frame image by a CCD camera (S1), a CPU normalizes the height and the width size of the acquired frame image into the same sizes as those of the frame image of a prototype vehicle and differentiates the density of all pixels of the normalized frame image of a front vehicle (S2). Acquired differentiated values of density are binarized by various kinds of previously set thresholds, to form a plurality of contour line images made up of edge patterns (S3), and dimensional compression is applied to each contour line image, to derive a plurality of features (S4). Each feature is voted on a feature plane stored in a memory (S5), peaks to be appeared are searched (S6), and appeared peaks are updated as the feature of the prototype vehicle (S7). <P>COPYRIGHT: (C)2004,JPO

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a vehicle recognizing device and a recognizing method for recognizing a vehicle ahead of a host vehicle by processing an image.
[0002]
[Prior art]
In recent years, vehicles equipped with a more advanced and more comfortable driving support system, such as a follow-up driving function and a steering assist function on a highway, have been proposed.As one of the methods, a teacher image of the prototype vehicle taken from the rear end face is taken. There is a method of preparing and comparing a captured image of a vehicle in front with a monocular camera with a teacher image to determine whether the captured image is of a vehicle.
[0003]
At this time, the contour image is extracted by processing the captured image of the preceding vehicle. At this time, the threshold value changes dynamically depending on the brightness of the captured scene. There has been a problem that the outline of the vehicle and the outline of the prototype vehicle are so different that the vehicle in front cannot be recognized.
[0004]
In addition, in the case of vehicle recognition, the feature amount of the prototype vehicle is compared with the feature amount of the preceding vehicle.However, when the feature amount of the prototype vehicle is obtained, the lighting conditions may vary depending on, for example, weather, season, time zone, and the like. Therefore, it was practically impossible to prepare a large number of teacher images for each vehicle type. In addition, when the number of teacher images is increased, the matching process becomes very complicated, and it takes a long time to recognize whether or not the vehicle is a vehicle.
[0005]
Further, when temporally tracking the recognized preceding vehicle, an algorithm for predicting the next position of the preceding vehicle is applied using a Kalman filter. There is also a problem that the output of the vehicle is not stable, and vehicle tracking often fails.
[0006]
By the way, when a road white line is detected by performing image processing, it has been proposed to employ a voting method based on Hough transform (for example, see Patent Document 1).
[0007]
[Patent Document 1]
JP-A-11-85999 (paragraphs [0008] to [0010])
[0008]
[Problems to be solved by the invention]
Since many experiments have confirmed that this voting method is robust against noise and robust, in a method of processing a captured image with a single-lens camera and performing vehicle recognition, such a voting method is used from the viewpoint of robustness. However, it is not yet adopted for vehicle recognition processing an image captured by a monocular camera.
[0009]
Therefore, an object of the present invention is to adopt a robust voting method for vehicle recognition and to enable stable vehicle recognition.
[0010]
[Means for Solving the Problems]
In order to achieve the above object, a vehicle recognition device according to the present invention, in a vehicle recognition device that processes an image to recognize a vehicle ahead of a host vehicle, captures a frame image of a vehicle ahead of the host vehicle by an imaging unit. Then, the processing unit binarizes the density differential values of all the pixels obtained by differentiating the densities of all the pixels of the frame image by the image capturing unit of the predetermined prototype vehicle to obtain a contour image formed of an edge pattern. Forming and dimensionally compressing the contour image of the formed prototype vehicle to obtain an eigenvector, create a feature plane or a feature space defined by a feature amount that is the eigenvector, and generate the frame image of the preceding vehicle by the imaging unit. Are binarized at various threshold values to form a plurality of contour images, and each contour image is dimensionally compressed to obtain a plurality of features. Derives the amount, by voting the each feature quantity in the feature plane or the feature space is characterized by updating the feature value of the prototype vehicle (claim 1).
[0011]
According to such a configuration, the contour line image of the prototype vehicle is obtained by binarizing the feature plane or the feature space defined by the feature amount which is an eigenvector obtained by dimensionally compressing the contour image with various threshold values. When voting for a plurality of feature values derived by dimensionally compressing the contour image of the preceding vehicle, the contour image obtained by slightly changing the threshold value for extracting the same vehicle or contour line in successive frame images is characterized by You will be voted very close in the plane or feature space.
[0012]
Therefore, a peak occurs at the optimum position, and this peak is individually set as a feature value of the new prototype vehicle for the preceding vehicle being captured, thereby performing stable vehicle recognition independent of the lighting conditions of the imaging scene. And reliable vehicle tracking can be performed.
[0013]
Further, a vehicle recognition device according to the present invention is a vehicle recognition device that processes an image to recognize a vehicle ahead of the host vehicle, an imaging unit that captures an image of the front of the host vehicle to obtain a frame image of the vehicle ahead, and the imaging unit. A contour image forming unit for binarizing a density differential value of all pixels obtained by differentiating the densities of all the pixels of the frame image to form a contour image composed of an edge pattern, and a predetermined prototype vehicle. A creating unit for dimensionally compressing the contour image with respect to the frame image to obtain an eigenvector and creating a feature plane or a feature space defined by a feature amount that is the eigenvector; and A normalizing unit for normalizing to the same size as the frame image of the vehicle, and the frame of the preceding vehicle normalized by the normalizing unit. A feature amount deriving unit that derives a plurality of feature amounts by dimensionally compressing each of the plurality of contour line images formed by binarizing the image with various thresholds by the contour image forming unit; A voting unit for voting each of the feature amounts by the feature amount deriving unit and updating the feature amount of the prototype vehicle in the feature plane or the feature space by the creating unit; 2).
[0014]
According to such a configuration, the contour image of the prototype vehicle is formed by the contour image forming unit, and the feature plane defined by the feature amount that is an eigenvector obtained by dimensionally compressing the contour image by the creating unit. Alternatively, a feature space is created, the vertical and horizontal sizes of the frame image of the vehicle in front are normalized by the normalization unit to the same size as the frame image of the prototype vehicle, and the feature amount deriving unit outlines the normalized frame image of the vehicle in front. The plurality of contour images formed by the line image forming unit are dimensionally compressed to derive a plurality of feature values, and the voting unit votes each of the derived feature values to a feature plane or feature space to provide a feature of the prototype vehicle. The amount is updated.
[0015]
For this reason, the same vehicle in consecutive frame images or a contour image with a slightly changed threshold value for contour extraction is voted very close in the feature plane or feature space, and becomes the optimal position. A peak is generated, and by setting this peak individually as a feature value of the new prototype vehicle for the vehicle in front of the vehicle being captured, stable vehicle recognition independent of the lighting conditions of the imaging scene can be performed. Vehicle tracking can be performed.
[0016]
Further, the vehicle recognition device according to the present invention is characterized in that the voting unit performs conical weighted voting (claim 3). According to such a configuration, by adopting the cone-type weighted voting, the feature amount of the prototype vehicle can be accurately updated by voting of the feature amount derived by the feature amount deriving unit.
[0017]
Also, the vehicle recognition device according to the present invention is characterized in that the voting unit updates the feature amount of the prototype vehicle by the FIFO method each time the frame image of the preceding vehicle is captured by the imaging unit. (Claim 4).
[0018]
According to such a configuration, it is possible to flexibly cope with a change in brightness of each of the imaging frames of the target vehicle in front and a change in a frame image on a curved road, and an improvement in the capture rate of the vehicle in front. Can be planned.
[0019]
Further, the vehicle recognition method according to the present invention includes an image capturing step of capturing an image ahead of the own vehicle to obtain a frame image of the vehicle ahead, and all pixels obtained by differentiating density of all pixels of the frame image in the image capturing step. And a contour image forming step of forming a contour image composed of an edge pattern by binarizing the density differential value of the contour image, and obtaining the contour image with respect to the frame image of a predetermined prototype vehicle in the imaging step. Creating a feature plane or feature space defined by a feature amount that is the eigenvector obtained by compressing the dimension and estimating the eigenvector, and normalizing the vertical and horizontal sizes of the frame image of the preceding vehicle to the same size as the frame image of the prototype vehicle. Normalizing step, and the frame image of the preceding vehicle normalized by the normalizing step. A feature value deriving step of dimensionally compressing a plurality of the contour images formed by binarization with various threshold values in the contour image forming step to derive a plurality of feature quantities; A voting step of voting each of the feature quantities in the feature quantity deriving step to update the feature quantity of the prototype vehicle in a plane or a feature space (claim 5).
[0020]
According to such a configuration, the contour image of the prototype vehicle is formed by the contour image forming step, and the contour plane image is dimensionally compressed by the creating step. Alternatively, a feature space is created, the vertical and horizontal sizes of the frame image of the preceding vehicle are normalized to the same size as the frame image of the prototype vehicle by the normalization process, and the contour of the normalized frame image of the preceding vehicle is extracted by the feature derivation process. The plurality of contour images formed in the line image forming process are dimensionally compressed to derive a plurality of feature values, and the voting process is performed so that each of the derived feature values is voted on a feature plane or feature space, and the feature of the prototype vehicle is obtained. The amount is updated.
[0021]
For this reason, the same vehicle in consecutive frame images or a contour image with a slightly changed threshold value for contour extraction is voted very close in the feature plane or feature space, and becomes the optimal position. A peak is generated, and by setting this peak individually as a feature value of the new prototype vehicle for the vehicle in front of the vehicle being captured, stable vehicle recognition independent of the lighting conditions of the imaging scene can be performed. Vehicle tracking can be performed.
[0022]
Further, the vehicle recognition method according to the present invention is characterized in that, in the voting step, a cone-type weighted voting is performed (claim 6). According to such a configuration, by adopting the cone-type weighted voting, the feature amount of the prototype vehicle can be accurately updated by voting of the feature amount derived by the feature amount deriving unit.
[0023]
Further, the vehicle recognition method according to the present invention is characterized in that, in the voting step, the updating of the feature amount of the prototype vehicle is performed by a FIFO method each time the frame image of the preceding vehicle is captured in the imaging step. (Claim 7).
[0024]
According to such a configuration, it is possible to flexibly cope with a change in brightness of each of the imaging frames of the target vehicle in front and a change in a frame image on a curved road, and an improvement in the capture rate of the vehicle in front. Can be planned.
[0025]
BEST MODE FOR CARRYING OUT THE INVENTION
An embodiment of the present invention will be described with reference to FIGS. 1 is a block diagram, FIG. 2 is an operation explanatory diagram, and FIG. 3 is a flowchart for operation explanation.
[0026]
As shown in FIG. 1, a monocular CCD camera 1 as an imaging unit is mounted on a vehicle, and an image of the front of the vehicle is taken by the CCD camera 1 to obtain a frame image of the vehicle in front. It is stored in a memory 3 such as a VRAM. Then, the vehicle in front of the own vehicle is recognized by the CPU 2 as a processing unit as follows. Note that imaging by the CCD camera 1 corresponds to an imaging step in the present invention.
[0027]
That is, the CPU 2 differentiates the densities of all the pixels of the frame image captured by the CCD camera 1 and binarizes the obtained density differential values of all the pixels to form a contour image composed of an edge pattern. . At this time, a plurality of contour images are formed by binarizing the frame image of the vehicle in front with various threshold values. Such contour image forming processing by the CPU corresponds to a contour image forming unit and a contour image forming step in the present invention.
[0028]
Further, the CPU 2 similarly forms a contour image for a frame image obtained by imaging the predetermined prototype vehicle with the CCD camera 1, and compresses the formed contour image two-dimensionally to obtain an eigenvector. And a feature plane defined by the eigenvector is created and stored in a predetermined area of the memory 3. Such a process of creating a feature plane by the CPU 2 corresponds to a creating unit and a creating step in the present invention.
[0029]
Further, the CPU 2 normalizes the vertical and horizontal sizes of the frame image of the front vehicle captured by the CCD camera 1 to the same size as the frame image of the prototype vehicle, and performs the above-described processing on the normalized frame image of the front vehicle. In this manner, a plurality of contour images are formed by binarizing with various threshold values, and these are dimensionally compressed to derive a plurality of feature amounts. Such normalization processing by the CPU 2 corresponds to a normalization unit and a normalization step in the present invention, and the processing of deriving a feature quantity by the CPU 2 corresponds to a feature quantity derivation unit and a feature quantity derivation step in the present invention. It is desirable that the various threshold values in this case are set within a predetermined range.
[0030]
Then, the CPU 2 votes the plurality of derived feature amounts for the feature plane, and updates the feature amounts of the prototype vehicle. At this time, cone-type weighted voting is adopted, and the feature amount of the prototype vehicle is updated by the FIFO (First In First Out) method every time the CCD camera 1 captures a frame image of the preceding vehicle. The voting / updating process by the CPU 2 corresponds to a voting unit and a voting process in the present invention.
[0031]
Further, when the CPU 2 forms the contour image, the motion of the preceding vehicle in each frame is predicted by the Kalman filter, and the position of the preceding vehicle in the next frame is predicted, and then the contour image is formed. By doing so, the recognition rate of the preceding vehicle in the frame is improved, so that the position input to the Kalman filter is stabilized, and the accuracy of tracking and running can be improved.
[0032]
For example, as shown in FIG. 2, a feature plane defined by feature quantities e1 and e2, which are eigenvectors obtained based on a contour image of a prototype vehicle, has the same time t, t + 1, and t + 2 in each frame at successive times. When a plurality of feature values derived by changing the threshold value are voted from the image of the preceding vehicle, each feature value is voted at a very close position on the feature plane, and as a result, the feature A peak appears at the optimum position on the plane. White frames in each of the frame images t, t + 1, and t + 2 in FIG. 2 indicate regions where various threshold values are set.
[0033]
When the feature amount (e1_o, e2_o) at the optimum position is obtained by voting in this manner, the preceding vehicle currently being captured is used as the feature amount of the prototype vehicle of the feature amount (e1_o, e2_o) in the next frame and thereafter. It will be used.
[0034]
Next, a series of operations will be described with reference to the flowchart of FIG. Now, as shown in FIG. 3, when the front of the own vehicle is imaged by the CCD camera 1 and a frame image is obtained (S1), the CPU 2 sets the vertical and horizontal sizes of the obtained frame image by the frame image of the prototype vehicle. The density of all pixels is differentiated with respect to the normalized frame image of the preceding vehicle (S2), and the obtained density differential values of all pixels are set to various predetermined threshold values. Are binarized to form a plurality of contour images composed of edge patterns (S3).
[0035]
At this time, a contour image is formed by the CPU 2 on a frame image obtained for a predetermined prototype vehicle, the contour image is dimensionally compressed to obtain an eigenvector, and a feature plane defined by the eigenvector is determined. It is created and already stored in a predetermined area of the memory 3.
[0036]
Further, each contour line image is dimensionally compressed by the CPU 2 to derive a plurality of feature amounts (S4), and these feature amounts are voted on a feature plane (see FIG. 2) stored in the memory 3 (S5). ), A search for an appearing peak is performed (S6), and the appearing peak is updated as a feature amount of the prototype vehicle (S7), and is used in the next frame and thereafter, and thereafter, the process returns to step S1.
[0037]
As described above, the CPU 2 forms the contour image of the prototype vehicle, derives a feature plane defined by the eigenvectors obtained by dimensionally compressing the contour image, and determines the length and width of the frame image of the vehicle ahead by the prototype. The feature image is normalized by the same size as the frame image of the vehicle, and the contour image of the normalized frame image of the preceding vehicle is dimensionally compressed to derive a feature amount. Are updated.
[0038]
Therefore, according to the above-described embodiment, a contour image obtained by slightly changing the threshold value for extracting the same vehicle or contour in successive frame images is voted very close in the feature plane. A peak is generated at the optimum position, and this peak is individually set as a feature amount of the new prototype vehicle for the preceding vehicle being captured, thereby performing stable vehicle recognition independent of the lighting conditions of the imaging scene. And reliable vehicle tracking can be performed.
[0039]
At this time, when deriving the feature amount, since a plurality of feature amounts are derived from the contour image binarized by various threshold values set within a certain range, the influence of the change in brightness is almost eliminated. And the feature amount can be derived.
[0040]
Further, since a prototype is dynamically set for each target vehicle by voting from a continuous frame as shown in FIG. 2, even a vehicle having a feature amount different from a prototype initially prepared can be easily recognized. can do.
[0041]
In addition, since the update of the feature amount of the prototype vehicle is performed by the FIFO method each time the CCD camera 1 captures a frame image of the front vehicle, a change in brightness of each of the target front vehicle imaging frames, It is possible to flexibly cope with a change in a frame image on a curved road, and it is possible to improve an acquisition rate of a vehicle ahead.
[0042]
Furthermore, since the movement of the preceding vehicle is predicted to some extent by the Kalman filter, the capturing rate of the preceding vehicle can be improved, and for example, the performance of tracking and running based on recognition of the preceding vehicle can be improved.
[0043]
The present invention is not limited to the above-described embodiment, and various changes other than those described above can be made without departing from the gist of the present invention.
[0044]
For example, in the above-described embodiment, a case has been described where a feature plane is defined by a feature amount obtained by two-dimensionally compressing a contour image of a prototype vehicle, but the feature space is three-dimensionally or more dimensionally compressed. May be defined, and in this case, the same effect as in the above-described embodiment can be obtained.
[0045]
Further, in the above-described embodiment, a case has been described in which motion prediction is performed using a Kalman filter. However, such motion prediction does not necessarily have to be performed.
[0046]
Further, it goes without saying that the imaging unit to be mounted on the own vehicle is not limited to the above-described CCD camera 1.
[0047]
【The invention's effect】
As described above, according to the first, second, and fifth aspects of the present invention, a feature plane or a feature space defined by a feature amount, which is an eigenvector obtained by dimensionally compressing a contour image of a prototype vehicle, By voting a plurality of feature values derived by dimensionally compressing a contour image of a preceding vehicle obtained by binarizing with various threshold values, the same vehicle in a continuous frame image or the same vehicle at the time of contour extraction The contour image with the threshold slightly changed is voted very close in the feature plane or the feature space, and a peak occurs at the optimum position.
[0048]
Therefore, a peak occurs at the optimum position, and this peak is individually set as a feature value of the new prototype vehicle for the preceding vehicle being captured, thereby performing stable vehicle recognition independent of the lighting conditions of the imaging scene. This enables reliable vehicle tracking, which is very effective as an aid for safe driving.
[0049]
According to the third and sixth aspects of the present invention, by adopting the cone-type weighted voting, the feature amount of the prototype vehicle is accurately updated by voting of the feature amount derived by the feature amount deriving unit. Becomes possible.
[0050]
Further, according to the inventions described in claims 4 and 7, it is possible to flexibly cope with a change in brightness of each of the imaging frames of the target front vehicle and a change in a frame image on a curved road. It is possible to improve the vehicle capture rate.
[Brief description of the drawings]
FIG. 1 is a block diagram of one embodiment of the present invention.
FIG. 2 is an operation explanatory diagram of one embodiment of the present invention.
FIG. 3 is a flowchart for explaining the operation of the embodiment of the present invention;
[Explanation of symbols]
1 CCD camera (imaging unit)
2 CPU (processing unit, contour image forming unit, creating unit, normalizing unit, feature amount deriving unit, voting unit)

Claims (7)

In a vehicle recognition device that processes an image to recognize a vehicle ahead of the own vehicle,
The imaging unit captures a frame image of a vehicle ahead of the own vehicle,
By the processing unit,
Binarizing the density differential value of all pixels obtained by differentiating the density of all pixels of the frame image by the imaging unit of the predetermined prototype vehicle to form a contour image composed of an edge pattern,
The contour image of the formed prototype vehicle is dimensionally compressed to obtain an eigenvector, and a feature plane or feature space defined by a feature amount that is the eigenvector is created.
A plurality of contour images are formed by binarizing the frame image of the vehicle in front by the imaging unit with various thresholds, and a plurality of feature amounts are derived by dimensionally compressing each of the contour images.
A vehicle recognizing device, wherein the feature amount of the prototype vehicle is updated by voting each of the feature amounts to the feature plane or feature space. In a vehicle recognition device that processes an image to recognize a vehicle ahead of the own vehicle,
An imaging unit that captures an image of the front of the vehicle and obtains a frame image of the vehicle in front;
A contour image forming unit that binarizes a density differential value of all pixels obtained by differentiating the densities of all pixels of the frame image by the imaging unit to form a contour image composed of an edge pattern;
A creation unit that creates a feature plane or feature space defined by a feature amount that is an eigenvector obtained by dimensionally compressing the contour image with respect to the frame image of a predetermined prototype vehicle,
A normalization unit that normalizes the vertical and horizontal sizes of the frame image of the preceding vehicle to the same size as the frame image of the prototype vehicle,
Each of the plurality of contour images formed by binarizing the frame image of the preceding vehicle normalized by the normalization unit with various threshold values by the contour image forming unit is dimensionally compressed. A feature value deriving unit that derives a plurality of feature values;
A voting section for voting each of the feature quantities by the feature quantity deriving section and updating the feature quantity of the prototype vehicle in the feature plane or feature space by the creating section; apparatus. 3. The vehicle recognition device according to claim 2, wherein the voting unit performs a cone-type weighted voting. The vehicle recognition according to claim 2, wherein the voting unit updates the feature amount of the prototype vehicle by a FIFO method each time the imaging unit captures the frame image of the preceding vehicle. apparatus. In a vehicle recognition method for processing an image to recognize a vehicle ahead of the own vehicle,
An imaging step of imaging the front of the own vehicle to obtain a frame image of the front vehicle,
A contour image forming step of binarizing a density differential value for all pixels obtained by differentiating the densities of all pixels of the frame image in the image capturing step to form a contour image composed of an edge pattern;
A generating step of obtaining the contour image for the frame image of a predetermined prototype vehicle, compressing the contour image to obtain an eigenvector, obtaining a eigenvector, and creating a feature plane or a feature space defined by a feature amount being the eigenvector in the imaging step; When,
A normalization step of normalizing the vertical and horizontal sizes of the frame image of the preceding vehicle to the same size as the frame image of the prototype vehicle,
A plurality of contour images formed by binarizing the frame images of the preceding vehicle normalized by the normalization step with various thresholds in the contour image forming step are formed. A feature value deriving step of deriving feature values of
A voting step of voting each of the feature quantities in the feature deriving step in the feature plane or feature space in the creating step to update the feature quantity of the prototype vehicle. Method. The vehicle recognition method according to claim 5, wherein in the voting step, a cone-type weighted voting is performed. 7. The vehicle recognition according to claim 5, wherein, in the voting step, updating of the feature amount of the prototype vehicle is performed by a FIFO method each time the frame image of the preceding vehicle is captured in the imaging step. Method.

Images