JP4922642B2 - Vehicle detection device and vehicle detection method - Google Patents

Description

  The present invention relates to a vehicle detection device and a vehicle detection method for detecting a vehicle from an image obtained by photographing with a camera or the like.

  Conventionally, for vehicle candidates obtained by performing background difference, frame difference processing, or light detection processing on images captured by a video camera, vehicle detection processing is performed by combining time-series tracking processing, and the trajectory of individual vehicles It is possible to measure the traffic volume or detect sudden abnormal events such as stoppage.

  As a device for performing such vehicle detection processing, it is a vehicle region detection device that detects a vehicle region from a vehicle candidate region with a high probability that a vehicle in the photographed image exists, and determines the left-right symmetry of the gray value of the image in the region. Symmetry evaluation unit for evaluation, dispersion degree evaluation unit for evaluating the variance of the gray value of the image in the region, gray value of the image in the lower region of the region, and the shade of the image in the region close to the lower portion of the region There is known a vehicle area detection device having a lower gray level average value evaluation unit that evaluates a difference from a value and an area determination unit that determines whether a vehicle candidate area is a vehicle area based on an evaluation result by each evaluation unit Yes.

In the area determination unit of this apparatus, the left / right symmetry evaluated by the left / right symmetry evaluation unit, the variance value evaluated by the variance value evaluation unit, and the lower gray level average value evaluated by the lower gray level average value evaluation unit are evaluated. The vehicle region is determined by comparing the lower grayscale average value with a predetermined threshold (see, for example, Patent Document 1).
Japanese Patent Laid-Open No. 2000-039306

  However, if some disturbance (such as a change in weather) occurs in the location where the vehicle is traveling, traffic flow measurement and sudden event detection may not be detected correctly if the threshold is fixed as described above. There was a situation. In addition, when a predetermined threshold value is used, it is necessary to prepare a plurality of threshold values for each environment (outdoor, tunnel, lighting around, etc.) where the vehicle is traveling when performing vehicle detection processing. There was a situation that it took time to set the threshold.

  The present invention has been made in view of the above-described conventional circumstances, and does not require time and effort to set a threshold for vehicle detection, and can reliably detect a vehicle candidate even if some disturbance occurs. An object of the present invention is to provide a vehicle detection device and a vehicle detection method.

In order to achieve the above object, a vehicle detection apparatus according to the present invention includes an image input unit that captures and inputs sequential images to obtain a current image, and a background based on a plurality of current images that are sequentially input by the image input unit. A background image creation unit that creates an image, a difference image creation unit that creates a difference image based on the current image and the background image, and a vehicle candidate that detects a vehicle candidate region having a characteristic region from the difference image An area detection unit, a vehicle recognition unit that recognizes a characteristic region of the vehicle candidate region as a vehicle, and a vehicle determination unit that determines that the recognized vehicle recognized by the vehicle recognition unit is a vehicle, vehicle determination unit is calculated from the variance of the plurality of luminance signals in a plurality of vehicle candidate regions which are sequentially obtained with a dispersion value and time elements of the luminance signal in the current vehicle candidate region is sequentially updated Comparing the dynamic variance threshold value, the result of the comparison, it has a configuration in which the dispersion value of the luminance signal in the current vehicle candidate region is determined to be the vehicle of the recognition vehicle when the larger variance threshold.

  This configuration eliminates the effects of environmental changes and camera aging, and eliminates false detections and undetections even when disturbances occur, enabling more stable vehicle detection and vehicle determination. it can. In addition, for example, when the vehicle has been stopped for a long time, even if the vehicle has fallen into the background image, it is possible to easily prevent erroneous detection of the vehicle when the vehicle restarts and there is no vehicle in the current image. Can do.

The vehicle detection method of the present invention includes an image input step in which images are sequentially captured and input to form a current image, and a background image is created based on a plurality of current images sequentially input in the image input step; A step of creating a difference image based on the current image and the background image, a vehicle candidate region detection step of detecting a vehicle candidate region having a characteristic region from the difference image, and a feature of the vehicle candidate region A vehicle recognition step in which a certain region is recognized as a vehicle, and a vehicle determination step in which the recognized vehicle recognized in the vehicle recognition step is determined to be a vehicle. calculated from the variance value of a plurality of luminance signals in a plurality of vehicle candidate regions which are sequentially obtained with a dispersion value and time elements of the luminance signal in the region Re successive approximation of the updated are dynamic, distributed threshold is performed, the result of the comparison, said the dispersion value of the current luminance signal in the vehicle candidate region is the recognition vehicle vehicle when the larger variance threshold value The method is determined.

  This configuration eliminates the effects of environmental changes and camera aging, and eliminates false detections and undetections even when disturbances occur, enabling more stable vehicle detection and vehicle determination. it can. In addition, for example, when the vehicle has been stopped for a long time, even if the vehicle has fallen into the background image, it is possible to easily prevent erroneous detection of the vehicle when the vehicle restarts and there is no vehicle in the current image. Can do.

  The present invention can provide a vehicle detection device and a vehicle detection method capable of reliably detecting a vehicle candidate even if some disturbance occurs without taking time and effort to set a threshold for vehicle detection. .

  Hereinafter, a vehicle detection device and a vehicle detection method according to embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 is a block diagram of a vehicle detection device 100 according to an embodiment of the present invention. The vehicle detection device 100 includes a current image input unit 1, a background image generation unit 2, a difference image generation unit 3, a binary image generation unit 4, a circumscribed rectangle detection unit 5, a vehicle recognition unit 6, a vehicle tracking unit 7, a vehicle A determination unit 8 and an image storage unit 10;

  Here, the current image input means 1 is an example of an image input unit. The background image creation unit 2 is an example of the background image creation unit 2. Moreover, the difference image creation means 3 is an example of a difference image creation unit. The circumscribed rectangle detecting means 5 is an example of a vehicle candidate area detecting unit. The vehicle recognition means 6 is an example of a vehicle recognition unit. The vehicle determination means 8 is an example of a vehicle determination unit.

  The current image input means 1 sequentially inputs the current image with a video camera or the like. The background image creation means 2 creates a background image by averaging or accumulation based on the current image stored by the current image storage means 11 and the original background image stored by the background image storage means 11. The difference image creation means 3 takes the difference between the current image stored by the current image storage means 11 and the background image stored by the background image storage means 12 and creates a difference image using background difference processing, frame difference processing, or the like. To do. The binarized image creating unit 4 creates a binarized image of the difference image stored by the difference image storing unit 13.

  The circumscribed rectangle detecting unit 5 detects only a characteristic part from the binarized image stored by the binarized image storing unit 14 as a circumscribed rectangle of the vehicle candidate. Normally, when the time zone in which the vehicle is traveling is daytime, a difference image is created by frame difference or background difference method as a normal difference process, and a vehicle candidate rectangle is extracted based on this difference image. When the time zone during which the vehicle is traveling is at night or the like, a method of extracting the vehicle candidate rectangle by the vehicle light may be used. In this method (also referred to as light detection processing), a vehicle candidate is detected by detecting a light pair that is lit in the vehicle.

  The vehicle recognition means 6 recognizes the image in the circumscribed rectangle as a vehicle by processing the image data included in the circumscribed rectangle area detected by the circumscribed rectangle detecting means 5 using a technique such as pattern recognition. .

  The vehicle determination means 8 calculates a variance value in the circumscribed rectangle recognized by the vehicle recognition means 6 of the current image stored by the current image storage means 1 and compares the calculated variance value with a threshold value to determine the vehicle. Or in the rectangle in the background image stored by the background image storage means 12 located at the same coordinates as the circumscribed rectangle recognized as the vehicle by the vehicle recognition means 6 in the current image stored by the current image storage means 1 The correlation coefficient is calculated, and the vehicle is determined based on the result. Detailed description regarding the vehicle determination means 8 will be described later. The vehicle tracking unit 7 performs a tracking process of the vehicle determined by the vehicle determination unit 8.

  The image storage means 10 includes a current image storage means 11 for storing the current image input by the current image input means 1, a background image storage means 12 for storing the background image created by the background image creation means 2, and a difference image. A difference image storage unit 13 for storing the difference image created by the creation unit 3 and a binarized image storage unit 14 for storing the binarized image created by the binarized image creation unit 4 are provided.

Next, the vehicle determination means 8 will be described in detail.
First, the condition that the vehicle exists in the vehicle candidate rectangle detected by the circumscribed rectangle detecting means 5 is (1) the correlation coefficient between the current image and the background image in the vehicle candidate rectangle is low, or (2) the vehicle Either the variance value of the current image in the candidate rectangle is large. When obtaining the variance value from the current image in the vehicle candidate rectangle, for example, the variance value is calculated from the lower portion of the vehicle candidate rectangle in the differential image of the current image in FIG. The reason for this calculation is that the vehicle exists in the vehicle candidate rectangle and there are many edges such as license plates, vehicle shadows and lights, and the variance value becomes a large value.

  This correlation coefficient is a coefficient normalized by 0 to 1, and can be used quantitatively as it is. The correlation coefficient between the image I and the image T, whose image size is M × N (M and N are natural numbers), which is a correlation coefficient calculation target, is expressed by the following equation.

  Although the correlation coefficient takes a value between −1 and 1, in this embodiment, if the correlation coefficient value between the current image and the background image is less than 0.7 in the vehicle candidate rectangle, the correlation is not so high. It is determined that a vehicle exists in the current image.

  On the other hand, the variance value is affected by the environment or aging. As a result, if the threshold value for comparison with the variance value is a predetermined value (fixed value), for example, the camera or the like is blurred due to secular change, that is, the contrast tends to decrease. The variance value becomes small, and even if a vehicle candidate exists, it may not be detected as a vehicle. Therefore, if this threshold value can be dynamically calculated from a traveling vehicle instead of a fixed value, it is possible to cope with environmental changes, secular changes, and the like.

  As described above, the vehicle determination means 8 calculates a variance value based on the current image of the circumscribed rectangle recognized as a vehicle, and compares the calculated variance value with a threshold value to determine the vehicle. The threshold value is dynamically changed by tracking the time series of the vehicle. Specifically, in a tunnel as shown in FIG. 3A, for example, 80 meters of one vehicle is tracked, and a vehicle candidate rectangle is extracted about 20 times during that time to recognize the vehicle. By repeating this work for 100 vehicles at the same place, about 2000 vehicle candidates can be acquired. A variance value is calculated based on the 2000 vehicle candidates.

  In the present embodiment, differential images are used for the current image and the background image in order to calculate a threshold value and a correlation coefficient for comparison with the variance value. FIG. 3 is a diagram showing vehicle candidate rectangles extracted by the vehicle detection device 100 according to the embodiment of the present invention. FIG. 3A shows a differential image (edge image) of the current image, and a thick frame shows a vehicle candidate rectangle. FIG. 3B shows a differential image (edge image) of the background image, and the thick frame is a rectangle at the same position as the vehicle candidate rectangle of FIG.

  FIG. 4 is a diagram showing vehicle candidate rectangles extracted by the vehicle detection device 100 according to the embodiment of the present invention (when the vehicle falls in the background image). FIG. 4A shows a differential image (edge image) of a background image, and a thick frame shows a vehicle candidate rectangle. FIG. 4B shows a differential image (edge image) of the current image, and the thick frame is a rectangle at the same position as the vehicle candidate rectangle of FIG.

  When creating these differential images, for example, a Sobel filter used for edge detection is used. By using the differential image in this way, the edge is enhanced and the robustness is improved. Although the more stable vehicle determination process can be performed using the differential image, the vehicle determination process may be performed without using the differential image.

Next, a vehicle detection method in this embodiment will be described.
FIG. 2 is an operation flowchart when the vehicle detection apparatus 100 according to the present embodiment of the present invention detects the vehicle. In the present embodiment, background detection, frame difference processing, and the like are normally used for vehicle detection processing. However, when it is difficult to detect a vehicle by these processing methods at night or the like, light detection processing (vehicle light Is also useful for detecting vehicle candidates when lit.

  Further, the threshold value dynamically calculated in the present embodiment is obtained in each of the normal difference process and the light detection process. This is because the variance value in the vehicle candidate rectangle is greatly different between the case where the vehicle is lit and the case where the vehicle is not lit even when the same aging has occurred.

  First, the current image input unit 1 sequentially inputs the current image, and the background image generation unit 2 generates a background image from the current image input in time series (step S1). Specifically, for example, a product-sum operation is performed on the current image for 30 seconds as an initial background. Then, the difference image creating means 3 creates a difference image between the current image and the background image (step S2). Subsequently, the binarized image creating means 4 creates a binarized image from the difference image (step S3). Then, the circumscribed rectangle detecting means 5 detects a portion recognized as a feature amount in the binarized image as a circumscribed rectangle (step S4). Although it is detected as a rectangle here, it may be detected by other than a rectangle. And the vehicle recognition means 6 recognizes the detected image in the circumscribed rectangle as a vehicle. (Step S5).

  If the light is not lit (step S6), the vehicle determination means 8 sequentially adds the differential value Sd of the differential image of the current image in the lower area of the circumscribed rectangle recognized as a vehicle, and squares the variance value Sd. The values are added sequentially (step S7). Then, when the number of added variance values reaches a predetermined sample number N (step S8), the vehicle determination means 8 calculates an average value Sd_m of the variance values Sd and a variance value Sd_m2 of the variance values Sd, and this average A new threshold value Sd_th is calculated by dividing the value Sd_m by the weighting factor K to the variance value Sd_m2 (step S9). Then, the vehicle determination unit 8 determines whether the variance value Sd is equal to or greater than the threshold value Sd_th, or whether the correlation coefficient between the current image and the background image in the rectangle is less than a predetermined value (for example, less than 0.7). (Step S10). If the variance value is greater than or equal to the threshold value or the correlation coefficient is less than the predetermined value, the vehicle determination unit 8 determines that the image in the circumscribed rectangle is a vehicle (step S11).

  If the light is on (step S6), similarly, the vehicle determination means 8 sequentially adds the variance value Sd of the differential image of the current image in the lower area of the circumscribed rectangle recognized as the vehicle, and the variance value. The square value of Sd is sequentially added (step S13). Then, when the number of added variance values reaches a predetermined sample number N (step S14), the vehicle determination means 8 calculates an average value Sd_m_l of the variance value Sd and a variance value Sd_m_l2 of the variance value Sd, and this average A new threshold value Sd_th_l obtained by dividing the variance value Sd_m_l2 by the weighting coefficient K from the value Sd_m_l is calculated (step S15). Then, the vehicle determination unit 8 determines whether the variance value Sd is equal to or greater than the threshold value Sd_th_l or whether the correlation coefficient between the current image and the background image in the rectangle is less than a predetermined value (for example, less than 0.7). (Step S16). If the variance value is greater than or equal to the threshold value or the correlation coefficient is less than the predetermined value, the vehicle determination unit 8 determines that the image in the circumscribed rectangle is a vehicle (step S11).

  In this way, the vehicle determination means 8 determines the vehicle by comparing the variance value of the circumscribed rectangular current image recognized as a vehicle with the threshold value. When the determination of the vehicle is completed, the vehicle tracking means 7 performs a tracking process for the vehicle (step S12). As a result, the trajectory of the individual vehicle can be accurately determined, and sudden abnormal events such as measurement of traffic volume or stoppage of the place where the video camera is installed can be detected.

  In the present embodiment, the vehicle detection device 100 includes vehicle determination means 8 that determines a vehicle by comparing a variance value of a circumscribed rectangular current image recognized as a vehicle with a threshold value. By dynamically changing the threshold to be compared by tracking the time series of the vehicle, it is not affected by environmental changes and camera aging, and it eliminates false detections and non-detections even when disturbances occur. And more stable vehicle detection can be performed.

  Further, since the circumscribed rectangle of the vehicle candidate is detected from the difference image between the differential images of the current image and the background image, the robustness can be improved. Furthermore, since the vehicle is determined based on at least one of the correlation function between the current image and the background image in the circumscribed rectangle and the variance value of the current image, a more stable vehicle can be determined. Simple and stable determination of the vehicle by detecting the circumscribed rectangle of the vehicle candidate by performing the light detection process from the differential images of the image and the background image when the normal difference process or the normal difference process is difficult to detect. It can be carried out.

  Further, even when the vehicle has fallen into the background image as shown in FIG. 4A, if the vehicle no longer exists in the current image as shown in FIG. Since the value is smaller than the threshold value, the vehicle candidate rectangle is not determined as a vehicle, and erroneous detection can be prevented.

  According to the vehicle detection apparatus 100 in the embodiment of the present invention as described above, based on the current image input unit 1 that sequentially captures images and inputs them as the current image, and a plurality of current images that are sequentially input by the current image input unit 1. A background image creating means 2 for creating a background image, a difference image creating means 3 for creating a difference image based on the current image and the background image, and a circumscribing for detecting a vehicle candidate area having a characteristic area from the difference image It has a rectangular detection means 5, vehicle recognition means 6 for recognizing a characteristic area of the vehicle candidate area as a vehicle, and vehicle confirmation means 8 for confirming that the recognized vehicle recognized by the vehicle recognition means 6 is a vehicle. The vehicle determination means 8 calculates a current variance value indicating the variance value in the vehicle candidate area, calculates a variance threshold based on the variance values in the vehicle candidate area of a plurality of current images, The vehicle is detected by a configuration in which the vehicle is determined by the first vehicle determination that determines that the recognized vehicle is a vehicle when the current dispersion value is larger than the dispersion threshold as a result of the comparison. Setting a threshold value for performing the operation does not take time, and the vehicle candidate can be reliably detected even if some disturbance occurs.

  In addition, the threshold value of the dispersion value used in normal difference processing (processing that does not use light such as frame difference method and background difference method) and light detection processing (processing to the vehicle recognized by the light) is dynamically changed. By changing it, stable vehicle detection becomes possible regardless of whether the light is on or off.

  In addition, it is possible to detect a vehicle more stably by using a dynamic threshold value and / or a correlation function of a variance value in a vehicle candidate rectangle in the current image and the background image to determine whether to recognize the vehicle. .

  INDUSTRIAL APPLICABILITY The present invention is useful as a vehicle detection apparatus that can detect a vehicle candidate reliably even if some disturbance occurs without taking time and effort to set a threshold for vehicle detection.

The block diagram of the vehicle detection apparatus in embodiment of this invention Operational flow diagram at the time of vehicle detection of the vehicle detection device in the embodiment of the present invention The figure which shows the vehicle candidate rectangle which the vehicle detection apparatus in embodiment of this invention extracts. The figure which shows the vehicle candidate rectangle which the vehicle detection apparatus in embodiment of this invention extracts (when a vehicle falls in a background image)

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Vehicle detection apparatus 1 Current image input means 2 Background image creation means 3 Difference image creation means 4 Binary image creation means 5 circumscribed rectangle detection means 6 Vehicle recognition means 7 Vehicle tracking means 8 Vehicle determination means 10 Image storage means 11 Current image Storage means 12 Background image storage means 13 Difference image storage means 14 Binary image storage means

Claims (2)

An image input unit that captures and inputs sequential images to form a current image;
A background image creation unit that creates a background image based on a plurality of current images sequentially input by the image input unit;
A difference image creation unit for creating a difference image based on the current image and the background image;
A vehicle candidate area detection unit for detecting a vehicle candidate area having a characteristic area from the difference image;
A vehicle recognition unit for recognizing a characteristic area of the vehicle candidate area as a vehicle;
A vehicle determining unit that determines that the recognized vehicle recognized by the vehicle recognition unit is a vehicle,
The vehicle determination unit is
Comparing the variance value of the luminance signal in the current vehicle candidate region with the dynamic variance threshold value that is calculated and sequentially updated from the variance values of the multiple luminance signals in the plurality of candidate vehicle regions that are sequentially acquired with time elements And the vehicle detection apparatus which determines the said recognition vehicle to be a vehicle when the dispersion value of the luminance signal in the said current vehicle candidate area is larger than the dispersion threshold as a result of the comparison. An image input step in which sequential images are captured and input to make the current image;
A step of creating a background image based on a plurality of current images sequentially input in the image input step;
A difference image is created based on the current image and the background image;
A vehicle candidate area detection step in which a vehicle candidate area having a characteristic area is detected from the difference image;
A vehicle recognition step in which a characteristic region of the vehicle candidate region is recognized as a vehicle;
A vehicle determining step in which the recognized vehicle recognized in the vehicle recognition step is determined to be a vehicle,
The vehicle determination step includes
Comparison between the variance value of the luminance signal in the current vehicle candidate region and a dynamic variance threshold value that is calculated and sequentially updated from the variance values of the plurality of luminance signals in a plurality of candidate vehicle regions that are sequentially acquired with time elements The vehicle detection method in which the recognized vehicle is determined to be a vehicle when the variance value of the luminance signal in the current vehicle candidate region is larger than the variance threshold as a result of the comparison.

Images