JP2011198006A - Object detecting apparatus, object detecting method, and object detecting program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an object detecting method and program for detecting an object with high accuracy within a predetermined time period without depending on the scene of the images input continuously.SOLUTION: The object detecting apparatus includes: a scan window setting section for setting a set-up scan window in each of the images continuously captured; a first object detecting section for executing first object detection on an image in the set-up scan window to output a result of prestage processing; a past processing result storing unit for storing a past processing result detected by a second object detection section; a scan window selecting unit for selecting a selection scan window on the basis of the prestage processing result and the past processing result; and the second object detecting section for detecting an object from the selection scan window.

Description

  The present invention relates to an object detection apparatus, an object detection method, and an object detection program for detecting a target object from continuously photographed images.

  A technique for detecting a specific object from continuously photographed images is called an object detection technique, and examples of target objects include a person, a car, a two-wheeled vehicle, and an animal.

  In a general object detection technique, scanning is performed with a rectangle of a predetermined size (referred to as a scan window), and whether or not a specific object is included in each scan window is determined using an object detector.

  The object detection technique includes a feature amount extractor that extracts a feature amount from each scan window, and an object detector that receives the extracted feature amount as input and determines whether or not a specific object is included in each scan window. Consists of

  For example, in Patent Document 1, a HOG (Hitogram of Oriented Gradient) feature value obtained by histogramating the luminance gradient strength in a local region for each direction is extracted from each scan window as a feature value and used as an input to the object detector. The object detector is constructed by using the boosting method, which is one of the statistical learning methods. By combining many object detectors with low detection accuracy (called weak classifiers) and ensemble these, An object detector (called a strong classifier) with a low error rate is constructed.

  Furthermore, Patent Document 1 uses a cascade type object detector in which multiple strong classifiers having different accuracy are connected to a strong classifier constructed using the boosting method (one stage as a stage). The scan window that has passed through all the stages is determined to contain a specific object. In cascaded object detectors, scan detection windows that are not subject to detection are rejected at the first stage of the cascaded object detector, and object detection processing in the latter half of the stage is omitted, thereby speeding up object detection processing. Realized.

  In Non-Patent Document 1, HOG feature values, color frequency feature values, and covariance matrix feature values are extracted from the scan window, and a 170,820-dimensional feature value vector is generated. From these feature quantities, feature quantities useful for object detection are selected, and a high-speed object detector that uses only a small number of selected feature quantities as input to the object detector is realized. The previous stage. On the other hand, an object detector that realizes high-precision object detection processing by inputting all the extracted feature quantities of 170 and 820 dimensions without selecting feature quantities is set as a subsequent stage of the cascade type object detector. Thus, a cascade type object detector having a two-stage configuration in which the two object detectors are connected is constructed. With this configuration, it is possible to reject scan windows that are not to be detected at the previous stage of the high-speed cascade-type object detector, and to omit the object detection process in the subsequent stage. Has been realized.

FIG. 10 shows a conventional object detection apparatus. In FIG. 10, a scan window is set by the scan window setting unit 1002 for an image photographed by the camera unit 1001, and an object is detected by the first object detection unit 1003. The first object detection unit 1003 is a high-speed object detection unit, and object detection is performed at all positions of an image captured by the camera unit.
The result is stored in the pre-processing result storage unit 1004, the object detection is performed on the result by the second object detection unit 1005, and the result is displayed on the detection result display unit 1006. The second object detection unit 1005 is highly accurate object detection, and is performed based on the result detected by the first object detection unit 1003.

US Patent Application Publication No. 2007/0237387

W. R. Schwartz, A. Kembhavi, D. Harwood and L. S. Davis, `` Human Detection Using Partial Least Squares Analysis '', The 12th IEEE International Conference on Computer Vision, Kyoto, JAPAN, pages 24-31, September 2009.

  However, when the cascade type object detector proposed in the non-patent document 1, the patent document 1, or the conventional object detection device shown in FIG. 10 is used, there are the following problems.

  When a cascade-type object detector is used, the scan window that is not to be detected is rejected at the previous stage of the cascade-type object detector, and the object detection process at the subsequent stage of the cascade-type object detector is omitted. The object detection process can be speeded up. However, depending on the input image, the number of scan windows passing through the previous stage of the cascade type object detector varies. Along with this, the number of scan windows for object detection processing in the subsequent stage also varies, and as a result, the object detection processing time of the entire cascade type object detector also varies depending on the input image.

  For example, if the target object is a person, cascading object detection with low detection accuracy can be achieved by performing object detection in a scene that contains many objects with similar shapes or colors (trees, telephone poles, dolls, etc.). The scan window cannot be rejected only with the front stage of the instrument, and many scan windows are delivered to the subsequent stage with high detection accuracy. In general, the latter stage of the cascade type object detector has the property that the object detection accuracy is high while the object detection speed is slow. Therefore, the cascade type object detection is increased as the number of scan windows to be detected is increased. The increase in the object detection processing time of the entire vessel is also noticeable.

Further, when the object detection technique is incorporated in hardware such as a camera, real-time processing is essential. The object detection process must always end within a certain time so as not to affect other processing processes of the hardware. However, when the object detection processing time varies depending on the input image as in the cascade type object detector, there is a possibility that the object detection processing does not end within a certain time. As a result, if priority is given to real-time processing, depending on the input image, object detection processing may be forcibly interrupted during the scan window scan, resulting in an increase in undetected results and increased object detection accuracy. It becomes difficult to keep.

  The conventional object detection apparatus shown in FIG. 10 will be specifically described with reference to FIG. When a scan window is set by the scan window setting unit 1002 for the image input from the camera unit 1001 and an object is detected by the first object detection unit 1003 which is the previous stage of the cascade type object detector, A scan window group 1101 in which is detected is output. At this time, the number of scan window groups 1101 output by the input image varies. When this number is large, in order to finish the object detection process within a fixed time, the second object detection unit 1005 which is the subsequent stage of the cascade type object detector has limited scan windows in the order of scan window scans. Only the object detection process is performed (1102). In such a case, the object detection process is forcibly interrupted during the scan window scan, and the object detection process for the detection target object (here, a person) such as 1103 is not performed in the second object detection unit 1005. As a result, it becomes difficult to maintain the object detection accuracy.

  The present invention has been made to solve the conventional problems, and can perform high-speed processing, which is an original characteristic of a cascade type object discriminator, while being highly accurate without depending on an input image. An object of the present invention is to provide a cascade-type object detection device, an object detection method, and an object detection program that can finish object detection processing within a predetermined time.

  The object detection apparatus according to the present invention executes scan object setting means for setting a set scan window in each of continuously shot images, and executes first object detection for the image in the set scan window. A first object detection unit that outputs a previous processing result; a past processing result storage unit that holds a past processing result detected by the second object detection unit from past images of the continuously captured images; And a scan window selection unit that selects a selected scan window based on the previous process result and the past process result, and a second object detection unit that detects an object from the selected scan window.

  With this configuration, the scan window transferred to the second object detection means corresponding to the subsequent stage of the cascade type object detector is input by the scan window selection means and the processing result of the first object detection means in the past. Based on the processing result of the second object detecting means for the processed image, a predetermined number is preferentially selected from the scan windows that are likely to include the detection target object. This can be done within the object detection processing time.

  In the object detection device of the present invention, the past processing result held in the past processing result storage unit may be the second for the image of the selected scan window taken before the continuously photographed image. This is a detection result of the object detection means.

  With this configuration, the object detection is performed by reflecting the object detection result from the image processed in the past, so that highly accurate object detection can be performed.

Furthermore, the object detection device of the present invention is such that the pre-processing result is the position and size of the set scan window of the continuously captured images, and the past processing result is captured before the images. It is characterized by the position and size of the selected scan window.

  With this configuration, the detection position of the object can be accurately determined.

  Furthermore, in the object detection apparatus of the present invention, the scan window selection means selects the selected scan window according to the overlapping position of the selected scan window taken before the image and the set scan window of the image. It is characterized by that.

  With this configuration, it is possible to select an effective object detection result from the detection results of the current object using the detection results of the objects detected in the past.

  Furthermore, the object detection apparatus of the present invention has scale image generation means for creating a scale image by enlarging or reducing each continuously captured image to at least two or more images having different sizes. The window setting means sets a setting scan window for the scale image.

  With this configuration, an object can be detected by setting a scan window for the scale image.

  Furthermore, the object detection apparatus of the present invention is characterized in that the past processing result holding unit holds a result of extending the detection result detected by the second object detection unit to the scale image.

  With this configuration, the object can be detected by reflecting the past object detection result on the scale image.

  The present invention provides an object detection method and an object detection program capable of performing object detection processing with high accuracy and within a predetermined time without depending on scenes of continuously input images. It can be done.

The block diagram of the object detection apparatus in the 1st Embodiment of this invention Schematic diagram of scan window selection processing of the object detection device according to the first embodiment of the present invention Explanatory drawing which shows the scan window overlap condition calculation method of the object detection apparatus in the 1st Embodiment of this invention. Overview of integrated processing in past processing result storage unit of object detection device according to first exemplary embodiment of the present invention Flow chart for explaining the operation of the object detection apparatus according to the first embodiment of the present invention. The block diagram of the object detection apparatus in the 2nd Embodiment of this invention Overview of detection target movement model setting in the past processing result storage unit of the object detection device according to the second exemplary embodiment of the present invention Schematic diagram showing an overview of the processing of the second embodiment of the present invention Flow chart for explaining the operation of the object detection apparatus in the second embodiment of the present invention Block diagram of a conventional object detection device Conventional Object Detection Device Flowchart

Hereinafter, an object detection device according to an embodiment of the present invention will be described with reference to the drawings.
An object detection apparatus according to a first embodiment of the present invention is shown in FIG.

  In FIG. 1, an object detection apparatus according to an embodiment of the present invention includes a scan window setting unit 102 that sets a set scan window in each of images continuously captured from a camera unit 101, and a set scan. A first object detection unit 103 that performs first object detection on a window; a pre-stage processing result storage unit 104 that stores a detection result of the first object detection unit 103; and A scan window selection unit 105 that selects a selected scan window from the setting scan window based on the result, a second object detection unit 106 that performs second object detection on the image of the selected scan window, and a second The past processing result storage for storing the result of the object detection unit 106 It has a section 108, a detection result display unit 107 for displaying the results of the second object detection unit 106.

  The scan window setting unit 102 sets, for example, a plurality of scan windows whose positions are changed for each pixel in the vertical and horizontal directions in each of the continuously captured images.

  The first object detection unit 103 that executes the first object detection extracts, for example, HOG feature amounts described in Patent Document 1 for each of the plurality of scan windows set by the scan window setting unit 102. Then, using the feature amount as an input, a score representing the likelihood of detection is output.

  The first object detection unit 103 extracts feature amounts from detection target samples and non-detection target samples (referred to as learning samples) in advance, and based on the distribution of learning samples in the feature amount space, SVM (Support By using a typical statistical learning method such as Vector Machine) or boosting method, the boundary line between the detection target sample and the non-detection target sample in the feature amount space may be learned and constructed.

  Next, the first object detection unit 103 is a space between the feature amount extracted from each scan window and the boundary line on the feature amount space that is pre-constructed to determine the likelihood of detection of the object in each scan window. Output as a score based on the relative positional relationship.

  The first object detection unit 103 corresponds to the preceding stage in the conventional cascade type object detector. Then, the second object detection unit 106 that performs second object detection corresponding to the subsequent stage is constructed so that high-speed processing is possible.

  Specifically, the first object detection unit 103 may be constructed using the following method. Based on the distribution in the feature value space extracted from the learning sample, for example, select the feature value useful for object detection using VIP (Variable Importance on Projection) described in Non-Patent Document 1 or the boosting method. Then, based on each feature quantity selected by the boosting method, as in the method of constructing a high-speed object detector by using SVM that receives only the selected feature quantity, For example, a high-speed object detector can be constructed by connecting multiple stages of strong classifiers with different accuracy on a cascade.

  Further, the first object detection unit 103 may be constructed using another typical object detection technique other than the construction method using the SVM or the boosting method mentioned above.

  Since the object detection process in the first object detection unit 103 is always performed for all of the plurality of scan windows created by the scan window setting unit 102, the object detection processing time in the first object detection unit 103 is the scan time. This is determined by the total number of scan windows created by the window setting unit 102 and the object detection processing time per scan window.

Each scan window score output from the first object detection unit 103 is stored together with the position information of each scan window in the pre-processing result storage unit 104 as shown in Table 1.

  Similar to the first object detection unit 103, the second object detection unit 106 that executes the second object detection extracts feature amounts from a predetermined number of scan windows selected by the scan window selection unit 105. An object detection process is performed in which a feature amount is input and a score representing the likelihood of detection is output.

  The second object detection unit 106 corresponds to the subsequent stage, as in the cascade type object detectors as in Patent Document 1 and Non-Patent Document 1, and more than the first object detection unit 103 corresponding to the previous stage. Build so that high-precision processing is possible.

  Specifically, as described in, for example, Patent Document 1 and Non-Patent Document 1, first object detection is achieved by using only a part of feature values selected by using the boosting method or VIP as input. Using the technique for increasing the object detection accuracy by using the extracted feature quantities as input to an object detector such as a linear SVM for the unit 103 and increasing the number of feature quantities used for object detection processing. Build it.

  Further, the second object detection unit 106 may be constructed using another typical object detection technique in addition to a method for increasing the object detection accuracy by increasing the number of feature amounts.

  The second object detection unit 106 outputs the object detection processing result together with the score for each scan window, and the position information of the scan window to the past processing result storage unit 108 and the detection result display unit 107.

The scan window selection unit 105 performs second object detection for a predetermined number of scan windows in the previously captured image for each scan window score (Table 1) held by the pre-processing result storage unit 104. Weighting is performed using the detection result (past processing result storage unit 108) of the unit 106, a predetermined number of scan windows are selected from the top of the newly calculated score, and the position is sent to the second object detection unit 106 Output information.

  At this time, the past processing result storage unit 108 stores the detection results of the second object detection unit 106 for a predetermined number of scan windows selected by the scan window selection unit 105 from previously captured images. Yes. This result is obtained by comparing the score for each scan window input from the second object detection unit 106 with a threshold value set by the user, so that the scan window (noted as OBJECT) includes the detection target object in the scan window. And a scan window (denoted as NON_OBJECT) in which the detection target object is not included in the scan window. The past processing result storage unit 108 holds internal information as shown in Table 2.

  For example, as shown in FIG. 2, for each scan window score held by the pre-stage processing result storage unit 104, a part of the scan window exceeding the threshold set by the user is stored in the past processing result storage unit 108. Assuming that each scan window is 202, as a result of the scan window selection process, a predetermined number of scan windows are selected as 205, and transferred to the second object detection unit 106.

  Specifically, as shown in FIG. 2, for example, when there is a result 204 of the previous process result storage unit 104 and 203 held by the past process result storage unit 108, the score of 204 depends on the overlap between 203 and 204. Weighting is performed and a new score is calculated. In this way, when the same weighting calculation is repeatedly executed for all scan windows in 201, and 204 is positioned within a predetermined number from the top in the newly calculated score, 204 is 205 Are selected and passed to the second object detection unit 106.

As a weighting method for the scan window of the pre-processing result storage unit in the scan window selection unit 105, first, as shown in 301 and 302 of FIG. 3, the degree of overlap of the scan windows is calculated. For example, the following IOU (Intersection Over Union) may be used for the degree of overlap of the scan windows.
If the overlap degree is IOU_SCORE and area (A) is the area of A,

IOU_SCORE = area (301∩302) / area (301∪302)

The degree of overlap may be defined as follows.

Further, the scan window 201 of the pre-processing result storage unit 104 calculates a predetermined number of scan windows 202 and IOUs in a previously captured image, and a combination of scan windows that minimizes the distance between the centers of gravity of the scan windows. IOU is calculated for each other.

  Next, the weighting method for the scan window of the pre-processing result storage unit 104 in the scan window selection unit 105 is calculated as follows according to the degree of overlap of the scan windows, as indicated by 301 and 302 in FIG. do it.

When the score of each scan window of the pre-stage processing result storage unit 104 is SCORE, the score of the weighted scan window is FINAL_SCORE, the score of the IOU is IOU_SCORE, and the score held by the past processing result storage unit 108 is PRE_SCORE, FINAL_SCORE is

FINAL_SCORE = a * SCORE + b * IOU_SCORE * PRE_SCORE

The scan window selection unit 105 selects a predetermined number of scan windows from the top based on the value of FINAL_SCORE, and outputs the position information to the second object detection unit 106.
Note that a and b are arbitrary parameters set by the user.

  As illustrated in FIG. 4, the past processing result storage unit 108 may integrate at least two or more scan windows within a distance equal to or less than a predetermined threshold into one rectangle in order to reduce the amount of stored information. For the integration of such rectangles, for example, a rectangle whose vertical and horizontal lengths are the average values of the vertical and horizontal lengths of the two rectangles at the midpoint between the centroids of the two rectangles is the result of the integration process. And may be stored as new position information. Two or more scan windows to be integrated are executed between OBJECTs or NON_OBJECTs. Further, as the integrated scan window score, an average value of the scan window scores used for the integration may be adopted.

When the scan window selection unit 105 calculates IOU_SCORE, OBJECT and NON_OBJECT information is used.
Specifically, the expression in paragraph (0052) is

IOU_SCORE_OBJECT: IOU between the scan window of the pre-processing result storage unit 104 and the scan scan window in which the object was included in the scan scan window of the past process result storage unit 108
IOU_SCORE_NON_OBJECT: IOU between the scan window of the previous process result storage unit 104 and the scan scan window of the past process result storage unit 108 that does not include an object

When

IOU_SCORE = c * IOU_SCORE_OBJECT + d * IOU_SCORE_NON_OBJECT

You may expand like this.
Here, c and d are arbitrary parameters set by the user.

As described above, in the scan window selection unit 105, scan windows having a high probability that the detection target object exists are always preferentially selected based on the newly calculated score from the top of the score. Therefore, the second object detection unit 106 can always finish the object detection process within a predetermined time without depending on the input image, and depends on the scan window scanning order as in the past. As a result, no detection failure occurs.

The object detection apparatus configured as described above will be further described with reference to the flowchart of FIG.
In step S501, the scan window setting unit 102 performs scan window setting processing. A plurality of scan windows are set in each of the images photographed continuously.
In S502, the first object detection unit 103 performs a first object detection process. For example, the HOG feature amount described in Patent Document 1 is extracted from the plurality of scan windows set in S501, and the feature amount is used as an input to output a score representing the likelihood of detection.

In S503, the scan window selection 105 performs a scan window selection process. A predetermined number of scan windows are selected based on the score output in S502 and the information held in the past process result storage process in S505.
In S504, the second object detection unit 106 performs a second object detection process. For the predetermined number of scan windows selected in S503, feature amounts are extracted in the same manner as in S502, and the feature amount is input and a score representing the likelihood of detection is output. It is determined whether or not a detection target is included in the scan window based on a threshold for a score predetermined by the user.
In step S505, the past processing result storage unit 108 performs past processing result storage processing. The second object detection processing result in S504 is held together with the position and score information for each scan window.

  According to the object detection apparatus of the first embodiment of the present invention as described above, the second object detection processing result for the scan window photographed in the past and the first object for the scan window photographed at the current time. Based on the detection processing result, a predetermined number of scan windows to be processed with high accuracy are provided. By providing a scan window selection unit, it is highly accurate and within a predetermined time without depending on the scenes of images input continuously. The object detection process can be performed.

Next, an object detection apparatus according to a second embodiment of the present invention is shown in FIG.
In FIG. 6, the object detection apparatus according to the embodiment of the present invention includes a scale image generation unit 601 that generates a pyramid image from continuously captured images, and a setting scan window in each hierarchical image of the pyramid image. Scan window setting unit 102 to be set, first object detection unit 103 that performs first object detection on the set scan window, and pre-stage processing for storing detection results of the first object detection unit 103 A result storage unit 104, a scan window selection unit 105 that selects a selected scan window from the set scan window based on the result of the first object detection, and a second object detection for the image of the selected scan window. A second object detection unit 106 to execute, With the past processing result storage unit 108 for storing the results of the object detection unit 106, a detection result display unit 107 for displaying the results of the second object detector.

  In the following description, with respect to the scan window output from each block, the output of the scan window setting unit 102 is set as the set scan window, the output of the first object detection unit 103 is set as the pre-processing result scan window, and the scan window. The output of the selection unit 105 is referred to as a selection scan window, the output of the second object detection unit 106 is referred to as a post-processing result scan window, and the output from the past processing result storage unit 108 is referred to as an extended scan window.

  The scale image generation unit 601 generates a pyramid image while sequentially resizing successively photographed images at a predetermined scale.

  The scan window setting unit 102 sets a plurality of scan windows and outputs a set scan window for each layer image of the pyramid image generated by the scale image generation unit 601.

  As shown in FIG. 7, the past processing result storage unit 108 includes the second object detection unit 106 for a predetermined number of selected scan windows selected by the scan window selection unit 102 from the previously captured pyramid images. Based on the movement model of the detection target object constructed in advance, the post-processing result scan window 706 that is the detection result is the position of the scan window so as to straddle the pyramid image hierarchy with respect to the pyramid image photographed at the current time. The information is extended and held as an extended scan window.

  The movement model of the detection target object in the past processing result storage unit 108 can be set by the user according to the movement speed of the detection target in the image scene to be shot. Specifically, the movement model is, for example, a point in which the scan window frame 705 is extended to which layer of the pyramid image when the detection result of the second object detection unit 106 is the layer 701 of the pyramid image (see FIG. In the example of FIG. 7, the scan window frame is expanded to the upper and lower layers such as 702 and 703, and how much the scan window frame is expanded vertically and horizontally on a certain layer of the pyramid image (in FIG. In the example, the scan window of 704 is expanded to 705).

When the scan window frame 705 is expanded to the hierarchy of 702 and 703 in the movement model setting by the user, the scan window frame in 702 and 703 is changed in the scale (vertical and horizontal) of the scan window frame 705 and the scale change of the pyramid image. Is considered.
Specifically, the vertical and horizontal lengths of the scan window frame 705 are HEIGHT_705, WIDTH_705, and the vertical and horizontal lengths of the scan window frame 702 are HEIGHT_702 and WIDTH 702, respectively. Is scale,

HEIGHT_702 = scale * HEIGHT_705
WIDTH_702 = scale * WIDTH_702

May be calculated.

  The setting of the movement model of the detection target object in the past processing result storage unit 108 is, for example, the movement of the detection target at a short distance according to the scan window position on the pyramid image that is the detection result of the second object detection unit 106 means. The range may be set adaptively such that the range is large and the moving range of the detection target at a long distance is small.

  In the scan window selection unit 105, the score of each previous process result scan window in the pyramid image held by the previous process result storage unit 104 is the past process result storage unit 108 on the hierarchy of the pyramid image in which the scan window exists. The weighting calculation is executed based on the information held by the.

FIG. 8 is an explanatory diagram for explaining an object detection procedure according to the second embodiment. In FIG. 8, reference numerals 801, 803, 805, and 807 are obtained by arranging the pyramid images shown in FIG. Reference numeral 801 denotes a state in which detection by the second object detection unit 106 is performed on a past pyramid image, and reference numeral 802 denotes a post-processing result scan window that is a detection result.
Next, the post-processing result scan window is expanded with respect to the result of the second object detection means, and the expanded scan window 804 is stored in the past processing result storage unit 108.
In this state, when the first object detection is performed on the current pyramid image by the first object detection unit 103, a pre-processing result scan window 806 is obtained.

Then, the scan window selection unit 105 obtains a selected scan window 808 from the preceding process result scan window 806 and the extended scan window 804.
As to how to select the selected scan window 808 at this time, as in the first embodiment, the IOU_SCORE can be calculated for 804 and 806 as the overlapping degree of the windows, and the FINAL_SCORE can be calculated based on this.

The second object detection unit 106 performs object detection processing on the selected scan window 808.
The operation of the object detection apparatus configured as described above will be further described with reference to the flowchart of FIG.

In step S901, the scale image generation processing unit 601 performs scale image generation processing. A pyramid image is generated while sequentially resizing successively photographed images at a predetermined scale.
In step S902, the scan window setting unit 102 performs scan window setting processing. A plurality of setting scan windows are set in each hierarchical image of the pyramid image generated in S9801.
In step S903, the first object detection unit 103 performs a first object detection process. For the plurality of setting scan windows set in S902, for example, HOG feature amounts described in Patent Document 1 are extracted, the feature amounts are input, a score representing the likelihood of detection is output, and the object is What is determined to exist is output as a pre-processing result scan window.

In S904, the scan window selection 105 performs a scan window selection process. A predetermined number of selected scan windows are selected based on the scored preceding process result scan window output in S903 and the extended scan window information held by the past process result storage process in S906.
In step S905, the second object detection unit 106 performs a second object detection process. For the predetermined number of selected scan windows selected in S904, feature amounts are extracted in the same manner as in S9803, and the feature amount is input, and a score representing the likelihood of detection is output. It is determined whether or not a detection target is included in the scan window based on a threshold for a score predetermined by the user. As a result, the presence / absence of an object is output to the detection result display unit 107. The result is output to the past processing result storage unit 108 as a subsequent processing result scan window.
In step S906, the past processing result storage unit 108 performs past processing result storage processing. The post-processing result scan window obtained as a result of the second object detection process in S905 is expanded to an extended scan window and held together with position and score information.

As described above, according to the object detection device of the second exemplary embodiment of the present invention, by providing the scale image generation unit 601, the past processing result storage unit 108 considers the movement range of the detection target at the current time. Scan window selection can be performed, and only the scan window that is more likely to have a detection target object is processed by the second object detection unit 106.

  As described above, the object detection apparatus according to the present invention is capable of performing object detection processing with high accuracy and within a predetermined time without depending on a scene of continuously input images. The present invention has an effect that a method and an object detection program can be provided, and the present invention provides an object detection apparatus, an object detection method, and an object detection method for detecting a target object from continuously captured images. This is useful as an object detection program.

DESCRIPTION OF SYMBOLS 101 Camera part 102 Scan window setting part 103 1st object detection part 104 Previous stage process result storage part 105 Scan window selection part 106 Second object detection part 107 Detection result display part 108 Past process result storage part

Claims (8)

A scan window setting means for setting a setting scan window in each of the continuously shot images;
First object detection means for executing first object detection on an image in the setting scan window and outputting a pre-processing result;
A past processing result storage unit for storing a past processing result detected by a second object detection unit from past images of the continuously shot images;
A scan window selection unit for selecting a selected scan window based on the previous process result and the past process result;
An object detection apparatus comprising: the second object detection unit that detects an object from the selected scan window.   The past processing result held in the past processing result storage unit is a detection result of the second object detection unit with respect to the image of the selected scan window taken before the images taken continuously. An object detection device characterized by. The pre-processing result is the position and size of the set scan window of the continuously captured images.
The object detection apparatus according to claim 2, wherein the past processing result is a position and a size of a selected scan window taken before the image.   The said scan window selection means selects a selection scan window according to the overlapping degree of the position of the selection scan window image | photographed before the said image, and the setting scan window of this image, It is characterized by the above-mentioned. Object detection device. Scale image generating means for creating a scale image by enlarging or reducing each of the continuously captured images into at least two or more images of different sizes;
2. The object detection apparatus according to claim 1, wherein the scan window setting means sets a set scan window for the scale image.   The object detection apparatus according to claim 1, wherein the past processing result holding unit holds a result of extending the detection result detected by the second object detection unit to the scale image. A scan window setting step for setting a setting scan window in each of the continuously shot images;
A first object detection step of executing a first object detection on an image in the setting scan window and outputting a pre-processing result;
A past processing result storage step for storing a past processing result detected by a second object detection step from past images of the continuously photographed images;
A scan window selection step of selecting a selected scan window based on the previous process result and the past process result;
An object detection apparatus comprising: a second object detection step of detecting an object from the selected scan window. On the computer,
A scan window setting step for setting a setting scan window in each of the continuously shot images;
A first object detection step of executing a first object detection on an image in the setting scan window and outputting a pre-processing result;
A past processing result storing step for holding a past processing result detected by the second object detecting means;
A scan window selection step of selecting a selected scan window based on the previous process result and the past process result;
An object program for executing a second object detection step of detecting an object from the selected scan window.

Images