JP2015056004A - Object detection device, method and program, and feature quantity deriving method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To increase processing speed and improve object detection accuracy.SOLUTION: An object detection device generates an integral image of a predetermined scale, for each of channels including a predetermined number of luminance gradient direction and gradient intensity, on the basis of an input image (31). On the basis of the generated integral image of the predetermined scale, the object detection device calculates a gradient feature quantity for each of the luminance gradient directions and an intensity feature quantity for the gradient intensity, by a rectangular area (51, 61) unit corresponding to a search window (40). The object detection device identifies an object by use of a normalized gradient feature quantity obtained by dividing the gradient feature quantity calculated based on the gradient integral image of the predetermined scale, as an identification feature quantity of the luminance gradient direction, by the intensity feature quantity calculated based on the intensity integral image of the predetermined scale in the same rectangular area.

Description

  The present invention relates to an object detection device, a method and a program, and a feature amount derivation method and a program, and in particular, an object detection device for detecting an object such as a pedestrian from an input image using a luminance gradient, The present invention relates to a method and a program, and a feature quantity derivation method and a program for deriving an identification feature quantity for detecting an object from an input image.

  There are various methods for detecting pedestrians based on image processing technology. Currently, there is a method called “Integral Channel Features” as one of the methods showing the world's highest level of detection accuracy (Non-Patent Document 1). In this method, various features such as the direction of brightness gradient, intensity, and color information are generated as an integrated image from an input image. Then, a tree-structure weak classifier having a rectangular feature calculated using these integral images as a node is configured, and this weak classifier is learned by boosting to construct a strong classifier.

  In the above method, in order to deal with a pedestrian having an arbitrary size, a plurality of integral images of different scales are generated for each channel (feature). As described above, in order to detect an object having an arbitrary size, it is well known that processing on a resized image is required as disclosed in, for example, Japanese Patent Application Laid-Open No. 2008-27275 (Patent Document 1). is there.

  One of the methods for speeding up the calculation of Integral Channel Features is “FPDW (The Fastest Pedestrian Detector in the West)” (Non-patent Document 2). In this method, a multi-scale integral image is generated at a sparser scale interval than that of Integral Channel Features. Then, by multiplying the generated integral image by a fixed constant obtained experimentally, the rectangular feature amount of the integral image of a predetermined scale is approximately calculated (estimated).

JP 2008-27275 A

Piotr Dollar and three others, “Integral Channel Features”, [online], 2009, [searched July 3, 2013], Internet <HYPERLINK "http://www.loni.ucla.edu/~ztu/publication /dollarBMVC09ChnFtrs#0.pdf "http://www.loni.ucla.edu/~ztu/publication/dollarBMVC09ChnFtrs#0.pdf> Piotr Dollar and two others, “FPDW (The Fastest Pedestrian Detector in the West)”, [online], 2010, [searched July 3, 2013], Internet <http://vision.ucsd.edu/sites /default/files/FPDW#0.pdf>

  In Integral Channel Features (hereinafter abbreviated as “ICF”), since it is necessary to generate integrated images of many scales, it is difficult to detect pedestrians in real time. In addition, although the FPDW achieves a certain degree of scale independence, a fixed constant used for estimation of the rectangular feature value is obtained by a prior experiment in order to maintain detection accuracy.

  An object of the present invention is to provide an object detection device, a method and a program, and a feature quantity deriving method and a program capable of increasing the processing speed and improving the detection accuracy.

  An object detection device according to an aspect of the present invention is an object detection device for detecting an object from an input image, and each of a plurality of channels including a predetermined number of luminance gradient directions and gradient intensities. And a rectangular area unit corresponding to the search window based on the gradient integration image and the intensity integral image of the predetermined scale generated by the generation means And calculating means for calculating the gradient feature quantity for each luminance gradient direction and the intensity feature quantity for the gradient intensity, and the object in the rectangular area based on the gradient feature quantity and the intensity feature quantity calculated by the calculation means. Identification means for identifying the presence or absence of an object. The identifying means includes a normalized gradient feature amount obtained by dividing the gradient feature amount calculated based on the gradient integral image of a predetermined scale by the intensity feature amount calculated based on the intensity integral image of the predetermined scale in the same rectangular region. The feature quantity is used to identify the presence or absence of an object.

  Preferably, for the gradient intensity channel, the generation means further generates a reduced intensity integrated image having a scale different from the predetermined scale based on the input image, and the calculation means calculates the reduced intensity integration for the gradient intensity channel. An intensity feature quantity based on the image is further calculated. The identification means identifies the presence / absence of an object based on intensity feature quantities for a plurality of scales in the same rectangular area when the channel to be identified is gradient intensity.

  Preferably, the identification unit includes a normalization unit that divides the gradient feature amount by the intensity feature amount of a predetermined scale when the channel to be identified corresponds to the luminance gradient direction.

  Preferably, the plurality of channels further include color information, and the generation unit generates a multi-scale color integrated image including a predetermined scale based on the input image for the color information channel, and the calculation unit includes the color information. For the channel, a color feature amount is calculated in units of rectangular areas based on each of the generated multi-scale color integrated images. The identification means identifies the presence / absence of an object based on color feature values for a plurality of scales in the same rectangular area when the channel to be identified is color information.

  Alternatively, the plurality of channels further include color information, and the calculating unit calculates color feature amounts based on the color integration image of a predetermined scale based on the input image based on the input image with respect to the color information channel in units of rectangular areas. It is also preferable that the means identify the presence / absence of an object using a normalized color feature amount obtained by dividing the color feature amount by a calculated value indicating the sum of colors in the target rectangular area.

  The calculated value may be a pixel value in a rectangular area.

  Preferably, a determination unit that determines the position of the object in the input image based on the identification result for each rectangular area by the identification unit, and an output unit that outputs the determination result by the determination unit are further provided.

  An object detection method according to another aspect of the present invention is an object detection method for detecting an object from an input image, wherein each of a plurality of channels including a predetermined number of luminance gradient directions and gradient intensities is input. A step of generating a gradient integrated image and an intensity integrated image of a predetermined scale based on the image, and each luminance in a rectangular area unit corresponding to the search window based on the generated gradient integrated image and the intensity integrated image of the predetermined scale. Calculating a gradient feature amount for the gradient direction and an intensity feature amount for the gradient strength, and identifying the presence or absence of an object in the rectangular region based on the calculated gradient feature amount and the intensity feature amount; Is provided. The identifying step includes a normalized gradient feature amount obtained by dividing a gradient feature amount calculated based on a gradient integral image of a predetermined scale by an intensity feature amount calculated based on an intensity integral image of a predetermined scale in the same rectangular region. The step of identifying the presence / absence of an object using the included feature amount is included.

  An object detection program according to still another aspect of the present invention causes a computer to execute each step included in the object detection method described above.

  A feature value deriving method according to still another aspect of the present invention is a method for deriving an identification feature value for detecting an object from an input image, and includes a plurality of channels indicating predetermined characteristic information based on the input image. The step of generating each integrated image, the step of calculating the rectangular feature amount of each channel for each rectangular region corresponding to the search window based on the generated integrated image, and the calculated rectangular feature amount are the same. And calculating a feature value for identification by dividing by a calculated value indicating the sum of predetermined characteristic information in the rectangular area.

  A feature quantity deriving program according to still another aspect of the present invention causes a computer to execute each step included in the feature quantity deriving method described above.

  According to the present invention, the presence / absence of an object is identified by a normalized gradient feature quantity independent of the scale. Therefore, the processing speed can be increased. Moreover, the detection accuracy of the object can be improved.

It is a block diagram which shows the hardware structural example of the target object detection apparatus which concerns on embodiment of this invention. It is a functional block diagram which shows the function structure of the target object detection apparatus which concerns on embodiment of this invention. It is a conceptual diagram which shows notionally the calculation method of the feature-value for identification of color information and gradient strength. It is a conceptual diagram which shows notionally the calculation method of the feature-value for identification of a brightness | luminance gradient. It is a figure which shows notionally the identification process which concerns on embodiment of this invention. It is a flowchart which shows the target object detection process which concerns on embodiment of this invention. It is a flowchart which shows the identification process in embodiment of this invention. It is a graph which shows the pedestrian detection accuracy at the time of evaluating using an INRIA data set.

  Embodiments of the present invention will be described in detail with reference to the drawings. In the drawings, the same or corresponding parts are denoted by the same reference numerals and description thereof will not be repeated.

  The object detection device according to the present embodiment detects, for example, a pedestrian as an object from the input image. Similar to the ICF described above, the object detection device according to the present embodiment extracts color information, luminance gradient, and luminance gradient intensity as characteristic information effective for detection of a pedestrian from an input image. . The color information is classified into three channels based on, for example, the LUV color system. The luminance gradient is classified into, for example, six channels of 0 °, 30 °, 60 °, 90 °, 120 °, and 180 ° obtained by dividing 180 ° into six. The intensity gradient intensity is only for one channel. The object detection device generates an integrated image for each channel, extracts rectangular features from the integrated image, and calculates a score of the strong classifier constructed based on boosting from the score of the weak classifier.

  Note that the weak classifiers are, for example, soft cascade connected. A method for detecting an object from an image using a soft cascade is described in US Pat. No. 7,634,142, and a method for speeding up an image identification process using a soft cascade is disclosed in US Pat. No. 2009/0018980. A method for constructing a soft cascade is also known.

  Here, in the ICF and the FPDW, a multi-scale integrated image is generated for a detection window of the same size for each channel. On the other hand, in the present embodiment, a scale-independent feature amount is constructed by performing an appropriate correction when calculating a rectangular feature from an integrated image corresponding to a luminance gradient. Hereinafter, the configuration and operation of an object detection apparatus for deriving an identification feature quantity independent of the scale and detecting the object using such an identification feature quantity will be described in detail.

<About configuration>
(Hardware configuration)
FIG. 1 is a block diagram illustrating a hardware configuration example of an object detection device 1 according to an embodiment of the present invention. As shown in FIG. 1, the object detection device 1 can be realized by a general-purpose computer such as a PC (Personal Computer).

  The object detection apparatus 1 includes a CPU (Central Processing Unit) 11 for performing various arithmetic processes, a ROM (Read Only Memory) 12 for storing various data and programs, and a RAM (Random Access Memory) for storing work data and the like. ) 13, a hard disk 14 which is a nonvolatile storage device, an operation unit 15 including a keyboard, a display unit 16 for displaying various information, and data and programs from the recording medium 17a can be read and written. The drive device 17 includes a communication I / F (interface) 18 for network communication, and an image input unit 19 for inputting images (including still images and moving images). The image input unit 19 is realized by a camera that captures an image. The recording medium 17a may be, for example, a CD-ROM (Compact Disc-ROM) or a memory card.

  Note that the object detection device 1 may not include the image input unit 19. In that case, for example, the object detection process may be performed on the image data obtained from the communication I / F 18 or the image data read from the recording medium 17a. In the present embodiment, it is assumed that an image obtained from the outside is also an “input image”.

(Functional configuration)
FIG. 2 is a functional block diagram showing a functional configuration of the object detection device 1 according to the embodiment of the present invention. Referring to FIG. 2, the object detection apparatus 1 includes, as its functional configuration, an image storage unit 110, a conversion unit 120, a generation unit 130, a setting unit 140, a calculation unit 150, an identification unit 160, A determination unit 170 and an output unit 180 are included.

  The image storage unit 110 stores an input image. The input image is, for example, an image input from the image input unit 19. The image storage unit 110 is realized by the RAM 13 or the hard disk 14, for example.

  The conversion unit 120 generates a plurality of converted images having a predetermined scale for each of a plurality of channels from the input image stored in the image storage unit 110. The plurality of channels includes three color information, six luminance gradient directions, and one gradient intensity as described above. Therefore, the conversion unit 120 includes a color information extraction unit 122, a luminance gradient extraction unit 124, and a gradient strength extraction unit 126. The color information extraction unit 122 extracts color information from the input image and converts the input image into an LUV image. The luminance gradient extraction unit 124 extracts a luminance gradient from the input image, and converts the input image into a gradient image in each direction. The gradient strength extraction unit 126 extracts gradient strength from the input image, and converts the input image into a gradient strength image. The gradient strength indicates the total in the luminance gradient direction.

  The generation unit 130 generates an integral image for each channel based on each converted image generated by the conversion unit 120. The integral image is a well-known arrangement in which the result of calculating each pixel value of the converted image is arranged according to the arrangement of the pixels of the converted image so that the feature amount is derived by adding and subtracting the four pixel values of the rectangular area. It is data. The generation unit 130 generates a multi-scale integrated image according to the FPDW described above for the color information and gradient intensity channels. For the luminance gradient channel, only an integral image of a predetermined scale is generated based on the input image. The “predetermined scale” is preferably the same scale as the input image. In the present embodiment, an integral image is generated for the entire screen, but the integral image may be partially generated.

  The setting unit 140 sets a search window for the input image. The search order by the search window is determined in advance.

  The calculation unit 150 calculates the feature amount (rectangular feature amount) of each channel in units of a rectangular area corresponding to the search window set by the setting unit 140. The calculation unit 150 calculates a rectangular feature amount for each channel based on one or more integrated images generated by the generation unit 130. A specific calculation method of the rectangular feature amount by the calculation unit 150 will be described later.

  The identification unit 160 identifies the presence or absence of an object in the search window based on the rectangular feature amount for each channel calculated by the calculation unit 150. In the present embodiment, identification unit 160 includes a normalization unit 162 and a discrimination processing unit 164. The normalization unit 162 normalizes the rectangular feature amount of the brightness gradient calculated by the calculation unit 150, that is, the gradient feature amount. That is, the normalization unit 162 derives a normalized gradient feature value as an identification feature value for the luminance gradient. A specific normalization method by the normalization unit 162 will also be described later.

  The discrimination processing unit 164 performs discrimination processing based on the rectangular feature amount of each channel. The discrimination process is a process performed at each node of the weak classifier, and is performed based on a threshold value set in advance by learning. The discrimination processing unit 164 performs discrimination processing using the normalized gradient feature amount for the luminance gradient channel, and performs discrimination processing based on the rectangular feature amount calculated by the calculation unit 150 for the other channels. Do.

  The determination unit 170 determines the position of the pedestrian in the input image based on the identification result for each search window by the identification unit 160. The determination unit 170 outputs information specifying the position of the pedestrian to the output unit 180.

  The output unit 180 outputs the determination result by the determination unit 170. The output unit 180 is realized by the display unit 16, for example. In that case, the search window determined (identified) that there is a pedestrian is indicated by a rectangular frame on the display unit 16. The output unit 180 may be realized by the communication I / F 18, for example. In this case, the determination result may be transmitted to an external device connected via a network.

  The functional blocks other than the image storage unit 110 and the output unit 180 illustrated in FIG. 2 may be realized by the CPU 11 illustrated in FIG. 1 executing, for example, software stored in the hard disk 14. At least one of them may be realized by hardware.

  Here, in the present embodiment, the feature amount calculation (derivation) method used for the identification processing by the identification unit 160 is different between the case of the luminance gradient and the other case. This will be specifically described with reference to FIG. 3 and FIG.

  FIG. 3 is a conceptual diagram conceptually showing a method for calculating feature information for identifying color information and gradient strength.

  Referring to FIG. 3, an image 31 is an input image. Images 32 to 36,... Conceptually represent reduced images generated at fine scale intervals in accordance with the above-described ICF. In the present embodiment, reduced images 36,... Are generated at sparse scale intervals in accordance with the FPDW described above. That is, in this embodiment, the reduced images 32 to 35 are not generated. The generation unit 130 extracts color information (3 channels) and gradient intensity information (1 channel) from the input image 31 and generates “integrated image 1” for each channel. Similarly for the reduced image 36, an “integrated image 6” is generated for each channel of color information and gradient intensity. In the ICF, all the integral images 1 to 6 corresponding to each scale are generated.

  A search window 40 having the same position and size is set for each of the images 31 and 36. It is assumed that the rectangular areas 51 and 52 of the integrated images 1 and 6 correspond to the search window 40, respectively. In this case, the calculation unit 150 calculates the first intensity feature amount from the rectangular region 51 of the gradient intensity integral image 1 (intensity integral image having the same scale as the input image), for example. That is, the total (integrated value) of the rectangular areas 51 is calculated as the first intensity feature amount. Similarly, the second intensity feature quantity is calculated from the rectangular area 52 of the gradient intensity integral image 6 (intensity integral image having a different scale from the input image). That is, the total (integrated value) of the rectangular areas 52 is calculated as the second intensity feature amount. The calculation unit 150 estimates a multiplier from these two intensity feature values, and processes the “integrated image 3” from the integrated image 1. The integral image 3 is an integral image having a scale between the scale of the integral image 1 and the scale of the integral image 6 (hereinafter referred to as “intermediate scale”). That is, by multiplying the first intensity feature quantity by an experimentally obtained fixed multiplier, the intensity feature quantity (third intensity feature quantity) of the integral image 3 that is not actually generated is approximately approximated. calculate. Similarly, color feature values for a plurality of scales are calculated for each channel of color information. In the color information and gradient strength channels, the rectangular feature values for a plurality of scales calculated as described above are the identification feature values. Thus, by estimating the rectangular feature amount of the integral image 3 that is not actually generated, the number of generated integral images can be reduced as compared with the ICF.

  FIG. 4 is a conceptual diagram conceptually showing a method for calculating a luminance gradient identifying feature value.

  Referring to FIG. 4, an image 31 is an input image. For the luminance gradient, none of the reduced images 32-36,... Is generated. As for the luminance gradient, an integral image corresponding to the angle is generated from only the input image 31. It is assumed that the rectangular area 61 of the gradient integral image having the same scale as the input image corresponds to the same search window 40 as in FIG. In this case, the calculation unit 150 calculates the total (integrated value) of the rectangular regions 61 of the gradient integral image as the gradient feature amount. In the present embodiment, this gradient rectangular feature amount is the sum (integrated value) of rectangular regions 51 having the same position and size calculated from the intensity integrated image (corresponding to the integrated image 1 in FIG. 3) of the same scale as the input image. ). As a result, a normalized gradient feature value is obtained as a feature value for identifying the luminance gradient.

  The intensity feature amount of the rectangular area 51 indicates the sum of the luminance gradients of the search window 40. Therefore, the normalized gradient feature quantity in a certain gradient direction indicates the component ratio in the gradient direction in the rectangular area 61. The composition ratio does not depend on the size of the rectangular area, that is, the search window. Therefore, for a luminance gradient channel, it is possible to obtain an identification feature quantity independent of scale without generating a plurality of scaled reduced images and calculating a plurality of rectangular feature quantities.

  FIG. 5 is a diagram conceptually showing the identification processing according to the embodiment of the present invention.

  Referring to FIG. 5, in the weak classifiers 70A, 70B,..., 70N having a tree structure, the nodes of the luminance gradient channels are indicated by hatching. For example, in the tree-structure weak classifier 70N, among the nodes N1 to N7, N3 and N7 are channels of luminance gradient, and the other nodes N1, N2, and N4 to N6 are channels of any one of color information and gradient intensity. It is.

<About operation>
FIG. 6 is a flowchart showing the object detection processing according to the embodiment of the present invention. The object detection process shown in FIG. 6 is realized by the CPU 11 reading and executing a program stored in advance in the hard disk 14, for example.

  Referring to FIGS. 2 and 6, first, each extraction unit 122, 124, 126 included in conversion unit 120 obtains color information, luminance gradient, and gradient intensity from the input image stored in image storage unit 110. Extract (step S2). Thereby, an LUV image (3 channels), a luminance gradient image in 6 directions (6 channels), and an image of gradient intensity (1 channel) are generated from the input image. Next, the production | generation part 130 produces | generates the integral image of the same scale as an input image for every channel (step S4). In addition, for the color and gradient intensity channels, the generation unit 130 further generates a plurality of reduced integrated images, for example, at a scale interval according to the FPDW described above (step S6). The generated integral image is temporarily stored in the RAM 13, for example.

  Next, when the setting unit 140 sets a search window, the calculation unit 150 extracts a rectangular feature amount in the search window (step S8). For each channel of the luminance gradient, the calculation unit 150 calculates a gradient feature amount from the one integrated image generated in step S4. For each color information channel, color feature quantities for a plurality of scales are calculated from the integrated image generated in step S4 and at least one reduced integrated image generated in step S6. Further, from the calculated color feature values, the color feature values of the intermediate scale are approximately calculated (estimated). In the gradient intensity channel, as in the case of color information, intensity feature quantities for a plurality of scales are calculated and estimated.

  When the rectangular feature amount is extracted, identification processing by the identification unit 160 is executed (step S10). The identification process will be described with reference to a subroutine in FIG. The identification process here corresponds to the process performed by each of the weak classifiers 70A to 70N in FIG.

  FIG. 7 is a flowchart showing identification processing in the embodiment of the present invention.

  With reference to FIG. 7, the identifying unit 160 determines whether or not the type of the determination target channel is a luminance gradient (step S <b> 102). If the determination target channel is a luminance gradient, the process proceeds to step S104. Otherwise, the process proceeds to step S108.

  In step S104, the normalization unit 162 divides the gradient feature amount calculated from the gradient integral image having the same scale as that of the input image in step S8 by the intensity feature amount corresponding to the same scale of the same rectangular region, and normalizes. Get features. The discrimination processing unit 164 executes discrimination processing based on the normalized feature amount (step S106). The discrimination process is a process performed in the nodes N1 to N7 shown in FIG.

  In step 108, the discrimination processing unit 164 executes discrimination processing using the feature amounts of each of the multiple scale integrated images based on the FPDW method.

  The above processing is repeated until the discrimination processing for all layers by the discrimination processing unit 164 is completed (NO in step S110). When the discrimination process for all layers is completed (YES in step S110), identification unit 160 identifies whether or not there is a pedestrian in the search window (rectangular region) (step S112). If all nodes (weak classifiers) have not been identified (NO in step S114), the process returns to step S102 and the above process is repeated. When all nodes have been identified (YES in step S114), a score is calculated using a predetermined calculation formula (step S116). Once the score is calculated, processing returns to the main routine. The nodes representing the weak classifiers are soft-cascade connected as described above, and the weak classification processing shown in steps S102 to S112 is performed according to the connection order.

  Referring to FIG. 3 again, when the identification process for the set search window is completed, it is determined whether or not the entire area is covered (step S12). That is, it is determined whether or not the identification processing for all search windows has been completed. If not yet (NO in step S12), the search window is moved by setting unit 140 (step S14), and the above-described feature amount extraction processing (step S8) and identification processing (step S10) are repeated.

  When the identification process for all the search windows has been completed (YES in step S12), determination unit 170 determines the presence or absence of a pedestrian in the input image based on the result of the identification process (step S16). If it is determined that there is a pedestrian, the position (the position of the search window) is specified and given to the output unit 180. Thereby, the determination result by the determination part 170 is output by the output part 180 (step S18).

  This completes the object detection process.

  As described above, in the present embodiment, the gradient feature amount is corrected (normalized) by a numerical value that is physically related instead of a simple fixed constant. Therefore, according to the target object detection apparatus 1 according to the present embodiment, it is possible to cope with any input without conducting a prior experiment. In addition, since only one integral image needs to be generated for the luminance gradient, it is possible to realize a speed increase equivalent to or higher than that of the FPDW. The divisor used for normalization of the gradient feature value is a feature value of one channel for pedestrian detection. Therefore, it is possible to reduce the load of normalization processing.

  FIG. 8 is a graph showing the pedestrian detection accuracy when an evaluation is performed using the INRIA data set. FIG. 8 shows a graph in which the horizontal axis represents the false detection rate (False Positive Per Image) and the vertical axis represents the detection error rate. In the graph, a graph line L1 indicates a false detection rate when pedestrian detection is performed according to the present embodiment. A graph line L2 indicates a false detection rate of the FPDW method. A graph line L3 indicates the false detection rate of the ICF method. The graph lines L2 and L3 are quoted from the 2012 INRIA dataset.

As shown in FIG. 8, when the accuracy of pedestrian detection is evaluated using the false detection rate as an evaluation index, the pedestrian detection method according to the present embodiment exceeds the conventional method at any false detection rate. Achieves the highest detection accuracy in the world. For example, in the case of a detection threshold value of “10 ⁻¹ ”, that is, a detection threshold value that generates one false detection when a full screen search is performed on 10 input images, a detection accuracy of approximately 85% is achieved. ing. Therefore, according to the object detection device of the present embodiment, high detection accuracy that can withstand practical use for in-vehicle applications that is strongly demanded from the market is realized. In particular, in recent years, there has been a trend toward mounting a pedestrian detection system in many automobiles, and the economic effect is high. Moreover, it can be easily applied not only to in-vehicle use but also to monitoring use and human machine interface, and it is expected that the possibility of being adopted in such a market is high.

  In this embodiment, in the identification process of the identification unit 160, the normalization process is performed when the channel to be identified is a luminance gradient. However, the normalization process is performed in advance before the process proceeds to the identification process. May be. That is, the normalization calculation process performed in step S104 in FIG. 4 may be performed between step S8 (rectangular feature amount calculation) and step S10 (identification process) in FIG.

  In this embodiment, the scale-free identification feature amount is derived only for the luminance gradient channel. Similarly, for the rectangular feature amount of the color information, the proportion of each color in the rectangular region is calculated. However, this value may be used as a scale-free identification feature quantity. In this case, a normalized color feature amount obtained by dividing the color feature amount by the sum of the colors in the same rectangular region, for example, the pixel value in the rectangular region can be used as the distinguishing feature amount.

  Or color information is auxiliary information for pedestrian detection, and it is known that only 2-3% contributes to pedestrian detection. Therefore, color information may not be included in the pedestrian detection channel. In that case, further speeding up of the pedestrian detection process can be expected.

  In the present embodiment, with respect to the luminance gradient channel, one integrated image having the same scale as that of the input image is generated for one input image. . In this case, the gradient feature amount calculated from the gradient integral image may be normalized by the intensity feature amount calculated from the intensity integral image having the same scale as the gradient integral image (that is, a predetermined scale).

  In addition, when there is characteristic information effective for pedestrian detection in addition to the color, the luminance gradient, and the gradient intensity, the characteristic information may be corrected based on the same concept as the luminance gradient. Specifically, the rectangular feature value calculated from the integral image of a certain channel may be corrected to a normalized feature value indicating the proportion of the channel in the rectangular region. That is, the identification feature amount may be derived by dividing the rectangular feature amount calculated by the calculation unit by the calculated value indicating the sum of predetermined characteristic information in the same rectangular area. In that sense, according to the present embodiment, it is also possible to provide a method for deriving an identification feature amount for detecting a pedestrian.

  In the present embodiment, the object is a pedestrian, but other objects such as a vehicle may be used as long as the detection device uses a luminance gradient.

  Note that the object detection method and the identification feature amount derivation method executed by the object detection device of the present embodiment can also be provided as a program. Such a program can be provided by being recorded on an optical medium such as a CD-ROM (Compact Disc-ROM) or a computer-readable non-transitory recording medium such as a memory card. A program can also be provided by downloading via a network.

  The program according to the present invention is a program module that is provided as a part of a computer operating system (OS) and calls necessary modules in a predetermined arrangement at a predetermined timing to execute processing. Also good. In that case, the program itself does not include the module, and the process is executed in cooperation with the OS. A program that does not include such a module can also be included in the program according to the present invention.

  The program according to the present invention may be provided by being incorporated in a part of another program. Even in this case, the program itself does not include the module included in the other program, and the process is executed in cooperation with the other program. Such a program incorporated in another program can also be included in the program according to the present invention.

  The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

  DESCRIPTION OF SYMBOLS 1 Object detection apparatus, 11 CPU, 14 Hard disk, 15 Operation part, 16 Display part, 17 Drive apparatus, 17a Recording medium, 18 Communication I / F, 19 Image input part, 110 Image storage part, 120 Conversion part, 122 colors Information extraction unit, 124 luminance gradient extraction unit, 126 gradient intensity extraction unit, 130 generation unit, 140 setting unit, 150 calculation unit, 160 identification unit, 162 normalization unit, 164 discrimination processing unit, 170 determination unit, 180 output unit.

Claims (11)

An object detection device for detecting an object from an input image,
For each of a plurality of channels including a predetermined number of luminance gradient directions and gradient intensities, generation means for generating a gradient integral image and an intensity integral image of a predetermined scale based on the input image;
On the basis of the gradient integral image and intensity integral image of the predetermined scale generated by the generation means, the gradient feature amount for each of the luminance gradient directions and the gradient intensity for each rectangular area corresponding to the search window. A calculation means for calculating an intensity feature amount;
Identification means for identifying the presence or absence of the object in the rectangular region based on the gradient feature value and the intensity feature value calculated by the calculation means,
The identification means is a normal value obtained by dividing the gradient feature amount calculated based on the gradient integral image of the predetermined scale by the intensity feature amount calculated based on the intensity integral image of the predetermined scale in the same rectangular region. An object detection apparatus that identifies the presence or absence of the object using a feature amount including a grading gradient feature amount. The generation means further generates a reduced intensity integrated image having a scale different from the predetermined scale based on the input image for the gradient intensity channel,
The calculation means further calculates an intensity feature amount based on the reduced intensity integrated image for the gradient intensity channel,
The said identification means identifies the presence or absence of the said object based on the said intensity | strength feature-value for several scales in the same said rectangular area, when the channel of discrimination | determination object is the said gradient intensity | strength. Object detection device.   3. The object detection according to claim 2, wherein the identification unit includes a normalization unit that divides the gradient feature amount by an intensity feature amount of the predetermined scale when a channel to be identified corresponds to the luminance gradient direction. apparatus. The plurality of channels further include color information;
The generation unit generates a multi-scale color integration image including the predetermined scale based on the input image for the color information channel;
For the color information channel, the calculation means calculates a color feature amount in units of the rectangular area based on each of the generated multi-scale color integrated images.
The said identification means identifies the presence or absence of the said object based on the said color feature-value for several scales in the same said rectangular area, when the channel of discrimination | determination object is the said color information. The object detection apparatus as described. The plurality of channels further include color information;
The calculation means calculates a color feature amount based on the color integration image of the predetermined scale based on the input image in the rectangular area unit for the color information channel;
The said identification means identifies the presence or absence of the said object using the normalized color feature-value which divided the said color feature-value with the calculated value which shows the sum total of the color in the said rectangular area | region of object. 2. The object detection apparatus described in 1.   The object detection apparatus according to claim 5, wherein the calculated value is a pixel value in the rectangular area. Determination means for determining a position of the object in the input image based on an identification result for each rectangular area by the identification means;
The object detection apparatus according to claim 1, further comprising an output unit that outputs a determination result by the determination unit. An object detection method for detecting an object from an input image,
Generating a gradient integrated image and an intensity integrated image of a predetermined scale based on the input image for each of a plurality of channels including a predetermined number of luminance gradient directions and gradient intensities;
Based on the generated gradient integral image and intensity integral image of the predetermined scale, a gradient feature amount for each of the luminance gradient directions, and an intensity feature amount for the gradient strength in units of rectangular regions corresponding to the search window, Calculating steps,
Identifying the presence or absence of the object in the rectangular region based on the calculated gradient feature amount and the strength feature amount,
In the identifying step, the gradient feature amount calculated based on the gradient integral image of the predetermined scale is divided by the intensity feature amount calculated based on the intensity integral image of the predetermined scale in the same rectangular region. An object detection method comprising the step of identifying the presence or absence of the object using a feature quantity including a normalized gradient feature quantity.   An object detection program for causing a computer to execute each step included in the object detection method according to claim 8. A method for deriving a feature for identification for detecting an object from an input image,
Generating an integral image of each of a plurality of channels indicating predetermined characteristic information based on the input image;
Calculating a rectangular feature amount of each channel for each rectangular region corresponding to the search window based on the generated integral image;
A feature amount deriving method comprising: calculating the identification feature amount by dividing the calculated rectangular feature amount by an operation value indicating a sum of the predetermined characteristic information in the same rectangular region.   A feature quantity deriving program for causing a computer to execute each step included in the feature quantity deriving method according to claim 10.

Images