WO2015049826A1 - 物体検出装置、物体検出方法および学習装置 - Google Patents
物体検出装置、物体検出方法および学習装置 Download PDFInfo
- Publication number
- WO2015049826A1 WO2015049826A1 PCT/JP2014/004284 JP2014004284W WO2015049826A1 WO 2015049826 A1 WO2015049826 A1 WO 2015049826A1 JP 2014004284 W JP2014004284 W JP 2014004284W WO 2015049826 A1 WO2015049826 A1 WO 2015049826A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- probability distribution
- area
- appearance probability
- region
- appearance
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20072—Graph-based image processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20076—Probabilistic image processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Definitions
- the present invention relates to an object detection apparatus and method for detecting an object from an image.
- object detection is used for a purpose of finding a desired object that is a monitoring purpose from a captured image of a camera, a purpose of focusing on a specific object in order to improve image quality, and the like.
- a technique called a sliding window method is used as a form of object detection.
- the processing concept of the sliding window method is shown in FIG.
- a rectangular area (Window) is set for an image to be detected, and the presence / absence of a detection target object in each rectangular area is changed while changing the position and size of the window. Is evaluated by an evaluation function.
- the position of the detection target cannot be obtained accurately when the detection target object is partially hidden behind another object or when the window size is significantly different from the size of the detection target.
- Non-Patent Document 1 proposes a technique for detecting a person using an ensemble learning method constructed by a set of a plurality of decision trees.
- a decision tree is constructed from a group of images called local parts, in which the local position of an object is reflected, and the part into which the input image is classified is evaluated by a score. Since parts are local areas, there are reports of cases where there is a high possibility that parts can be detected based on areas that are not occluded even if some objects are occluded.
- a person is detected using an average value of scores calculated from a result of recognizing a plurality of parts (FIG. 16).
- Non-Patent Document 2 proposes a constellation model that compares the relationships between parts with constellations.
- the constellation model is a model that expresses a probability distribution indicating the appearance, relative position, rotation angle, size, and other parameters of each part in a two-dimensional image.
- a model composed of an average value and variance of parts positions and sizes is created, and the likelihood of whether or not the combination of all the candidate parts matches the constellation model is calculated, and the likelihood of the background is also calculated. Calculate the degree. In this method, whether or not the object is an object is determined based on whether or not the likelihood ratio between the two is greater than or equal to a threshold value.
- Non-Patent Document 1 an image showing a human body is decomposed into a plurality of parts areas, and it is determined whether or not a human body has been detected from the average value of scores calculated for each part area.
- this method does not consider the positional relationship of the part areas, as shown in FIG. 17, even if the position of the parts area is wrong, it is detected as a human body. Therefore, there is a risk of erroneous detection.
- Non-Patent Document 2 the evaluation is performed based on the positional relationship between the modeled part areas.
- modeling is performed based only on the total value of the appearance probability distributions of the part regions, and thus the region is not adjusted with respect to the region where the appearance probability is zero. For this reason, since an area larger than the original object area may be detected, the position accuracy is lowered.
- the present invention has been made in view of the above problems, and an object thereof is to provide an object detection device, an object detection method, and a learning device that can detect an object region with high accuracy.
- An object detection apparatus includes a parts area designating unit that designates a part area, which is an area in which a target part is captured, from a plurality of images in which the object is captured among parts constituting an object that is a detection target; An appearance probability distribution generating means for generating an appearance probability distribution of the part area and a non-appearance probability distribution based on an appearance frequency associated with a position in the image of the part area, and an appearance probability distribution of the part area , With reference to the non-appearance probability distribution, an object determination unit that determines an area in which the object appears in the input image.
- a part area that is an area in which a target part is captured among parts constituting an object that is a detection target is specified from a plurality of images in which the object is captured, Based on the appearance frequency associated with the position in the image, generate an appearance probability distribution of the part region and a non-appearance probability distribution, and refer to the appearance probability distribution of the part region and the non-appearance probability distribution, A region where the object is shown in the input image is determined.
- the learning apparatus includes a parts region designating unit that designates a part region that is a region in which the target part is captured from a plurality of images in which the object is captured, among the parts that configure the object that is a detection target; Appearance probability distribution generating means for generating an appearance probability distribution of the part region and a non-appearance probability distribution based on the appearance frequency of the part region associated with the position in the image.
- the object is also achieved by a computer program that implements the object detection device, the object detection method, or the learning device having the above-described configurations by a computer, and a computer-readable recording medium that stores the computer program. Achieved.
- an object region can be detected with high accuracy.
- FIG. 1 is a block diagram showing the configuration of the object detection apparatus according to the first embodiment of the present invention.
- FIG. 2 is a flowchart showing the flow of object detection.
- FIG. 3 is a flowchart showing the flow of object detection.
- FIG. 4 is a diagram for explaining the generation of the appearance frequency distribution.
- FIG. 5 is a diagram for explaining the generation of the appearance frequency distribution.
- FIG. 6 is a diagram for explaining generation of a non-appearance frequency distribution.
- FIG. 7 is a diagram for explaining another generation method of the appearance frequency distribution.
- FIG. 8 is a diagram illustrating an example of a tire region detection result.
- FIG. 9 is a diagram for explaining the determination of the object region.
- FIG. 10 is a diagram for explaining the determination of the object region.
- FIG. 10 is a diagram for explaining the determination of the object region.
- FIG. 11 is a block diagram showing a configuration of an object detection device according to the second exemplary embodiment of the present invention.
- FIG. 12 is a diagram for explaining determination of an object area in the second embodiment.
- FIG. 13 is a block diagram which shows the structure of the object detection apparatus concerning the 3rd Embodiment of this invention.
- FIG. 14 is a diagram illustrating a hardware configuration for realizing the object detection device according to each embodiment of the present invention.
- FIG. 15 is a diagram for explaining the processing concept of the sliding window method.
- FIG. 16 is a diagram for explaining a technique for detecting a person using an ensemble learning method constructed by a set of a plurality of decision trees.
- FIG. 17 is a diagram for explaining a technique for detecting a person using an ensemble learning method constructed by a set of a plurality of decision trees.
- FIG. 1 is a block diagram showing a configuration of an object detection apparatus according to an embodiment of the present invention.
- the object detection apparatus 100 includes an image supply unit 110, a detection unit 120, a learning unit 130, and a position detection result output unit 140.
- the detection unit 120 includes a part area detection unit 121 and an object determination unit 122.
- the learning unit 130 includes a part region designation unit 131, an appearance probability distribution generation unit 132, and an appearance probability distribution storage unit 133.
- the position detection result output unit 140 outputs the detection result obtained by the detection unit 120.
- the object detection device 100 may be any device that processes an image, and includes, for example, a general computer device, an imaging device (a mobile phone, a smartphone, a digital still camera), and the like.
- the image supply unit 110 acquires an image captured by an arbitrary imaging unit. Further, the image supply unit 110 supplies the image to the part region designation unit 131 at the time of learning, and supplies the image to the part region detection unit 121 at the time of detection.
- the image supply unit 110 only needs to supply an image to be processed.
- the image supply unit 110 reads an image from a storage device (for example, a USB (Universal Serial Bus) memory) that is detachable from the object detection device 100 and detects the image. 120 or the learning unit 130 may be supplied.
- a storage device for example, a USB (Universal Serial Bus) memory
- the image supplied by the image supply unit 110 is assumed to be a still image. However, the image supply unit 110 may cut out each frame constituting the moving image and sequentially supply it to the detection unit 120 or the learning unit 130. Good.
- the learning unit 130 includes a part region specifying unit 131 that specifies a part region, an appearance probability distribution generating unit 132 that generates an appearance probability distribution of the part region, and an appearance probability distribution storage unit 133.
- the learning unit 130 may be detachable from the object detection device 100.
- the part area designating unit 131 designates the part area of the object shown in the image, and passes the coordinate value indicating the part area to the appearance probability distribution generating unit 132.
- the part region designation method may be designated manually or by a known part region detection technique described later.
- the parts area is an area in which an image of an object shows a part constituting the object.
- the appearance probability distribution generation unit 132 generates an appearance frequency histogram by counting the number of times for the positions of a plurality of part regions input from the part region specifying unit 131.
- the appearance probability distribution generation unit 132 generates an appearance probability distribution and a non-appearance probability distribution for each part region constituting the object based on the generated appearance frequency histogram.
- the appearance probability distribution generation unit 132 converts, for example, the position and size of the tire region in the image into the position and size on the xy plane (shown in FIG. 4).
- the appearance probability distribution generation unit 132 can obtain the appearance frequency of the tire region at the position in the z-axis direction by performing this conversion on a large number of images.
- the appearance probability distribution generation unit 132 calculates the appearance frequency associated with the position of the tire region in the image.
- the appearance probability distribution generation unit 132 calculates the total appearance frequency of the tire areas converted into the xy plane, and divides the appearance frequency of the tire areas at each position by the total to obtain the appearance probability of the tire area. You may ask for it.
- the appearance probability distribution generation unit 132 may obtain a local maximum value of the appearance frequency, and may fit a spread from the position of the local maximum value to a certain appearance frequency value to a Gaussian distribution as a distribution variance.
- the tire non-appearance probability distribution is equal to the appearance probability distribution generated when the appearance frequency of parts other than the tire is calculated.
- the appearance probability distribution storage unit 133 stores the appearance probability distribution and the non-appearance probability distribution generated by the appearance probability distribution generation unit 132 from a large number of images in advance.
- the appearance probability distribution storage unit 133 may hold the distribution value in a file format or a database table or the like. Further, the appearance probability distribution storage unit 133 may store appearance probability distributions related to the positions of the parts regions of a plurality of types of objects. For example, the appearance probability distribution storage unit 133 may store an appearance probability distribution related to the position of each part region for each type of large vehicle, medium vehicle, small vehicle, scooter, off-road motorcycle, and naked motorcycle.
- the parts area means an area where functional unit parts such as tires and headlights are copied.
- the part area detection unit 121 divides the image supplied from the image supply unit 110 into partial areas, and expresses the partial areas as evaluation values (scores) indicating the likelihood of the part areas.
- the part area detection unit 121 generates a score map in which a score is stored for each partial area in the image.
- the object determination unit 122 determines whether or not an object to be detected is captured in a certain area of the image supplied from the image supply unit 110.
- the object determination unit 122 calculates an evaluation value from the part region score calculated by the part region detection unit 121 and the appearance probability distribution of the part region stored in the appearance probability distribution storage unit 133. Note that the object determination unit 122 may integrate evaluation values of a plurality of parts regions and determine whether the object is an object based on a threshold value.
- FIG. 2 and 3 are flowcharts showing the flow of object detection by the object detection apparatus 100.
- FIG. 2 and 3 are flowcharts showing the flow of object detection by the object detection apparatus 100.
- the image supply unit 110 supplies a plurality of images (object images) in which an object to be detected is captured in advance to the part region designation unit 131 in the learning unit 130.
- the part area designation unit 131 designates the position and size of the part area from the input image (S1). Based on the position and size of the part area, the appearance probability distribution generation unit 132 calculates how many times the part area appears in the image and generates a distribution (S2).
- the position of the part area may be an absolute position or a relative position.
- the part area designating unit 131 generates an appearance probability distribution and a non-appearance probability distribution from the appearance frequency distribution generated for each part area, and stores them in the appearance probability distribution storage unit 133 (S3).
- the image supply unit 110 supplies the detection target image to the part region detection unit 121 in the detection unit 120.
- the part area detection unit 121 detects a part area for each partial area obtained by dividing the image, and stores a score indicating the part area uniqueness together with the position of the part area (S4).
- the object determination unit 122 determines whether the object is an object from the appearance probability distribution and the non-appearance probability distribution stored in the appearance probability distribution storage unit 133 and the score indicating the part area likelihood calculated by the part area detection unit 121. Do. For example, the object determination unit 122 sets the appearance probability distribution as a positive value and the non-appearance probability distribution as a negative value, and calculates a score for each partial region of the part region and the appearance probability distribution or the non-appearance probability distribution. Integration may be performed. Then, the object determination unit 122 obtains an evaluation value by taking the sum of the integrated values within a certain area, and determines that an object has been detected when the evaluation value is equal to or greater than a threshold value (S5).
- S5 a threshold value
- the object detection apparatus 100 will be described using a specific embodiment.
- the following description will be divided into a probability distribution generation phase (operation of the learning unit 130) for generating an appearance probability distribution in advance and a detection phase (operation of the detection unit 120) for detecting an object position from an input image.
- the part area designating unit 131 determines an area such as a headlight or a tire, which is a part area of a functional unit, from an object shown in the image, for example, a car or a motorcycle.
- the user may designate a region such as a headlight or a tire by operating the parts region designation unit 131 in advance.
- the part area designating unit 131 may detect an area such as a headlight or a tire using a known detection technique to be described later, and specify the area using the detection result.
- the designated part area is input to the appearance probability distribution generation unit 132.
- the appearance probability distribution generation unit 132 generates, for each designated part area, a histogram indicating the appearance frequency from the position and size of the part area. For example, as shown in FIG. 4, the appearance probability distribution generation unit 132 determines which position in the image the tire region has in relation to the position and size of the tire region that appears in the image 1 and the position and size of the tire region that appears in the image 2. How many times it appears in Based on the calculated result, the appearance probability distribution generation unit 132 can generate the appearance frequency distributions P1 and P2 related to the tire region in the XY plane.
- the appearance probability distribution generation unit 132 generates the tire appearance frequency probability distribution by normalizing the appearance frequency of the tire at a certain position with the sum of the appearance frequencies of the tires. In this manner, the appearance probability distribution generation unit 132 similarly generates an appearance probability distribution for a part region such as a car headlight or a front grill from a plurality of image groups.
- FIG. 5 shows an example of an appearance probability distribution of a car headlight and a tire.
- distributions P3 and P4 indicate the appearance probability distribution of the vehicle headlight
- distributions P5 and P6 indicate the appearance probability distribution of the vehicle tire.
- the appearance probability distribution generation unit 132 generates a non-appearance probability.
- the non-appearance probability indicates the probability that the part region of interest is unlikely to appear at that position in the image. For example, since a tire has a low probability of appearing in the headlight region, the appearance probability of the headlight is defined as a tire non-appearance probability.
- the non-appearance probability of a tire is defined as a distribution obtained by multiplying the appearance probability of a headlight by “ ⁇ 1”.
- FIG. 6 is a diagram illustrating an example of a distribution of tire appearance probability and non-appearance probability. In FIG.
- distributions P ⁇ b> 7 and P ⁇ b> 8 indicate tire appearance probabilities
- distributions P ⁇ b> 9 and P ⁇ b> 10 indicate tire non-appearance probabilities. Therefore, when the tire appears at the headlight position, the appearance probability has a negative value as a penalty.
- a value obtained by converting the appearance probability below the threshold to a negative value may be defined as the non-appearance probability. That is, as shown in FIG. 7, the appearance probability below the threshold value may be converted into a negative value, and the converted value may be set as the non-appearance probability.
- the generated appearance probability distribution may be stored in the appearance probability distribution storage unit 133 as a file or a database.
- detection is performed using the appearance probability generated in the probability distribution generation phase.
- a known detection technique may be used for the detection method. For example, the part area detection unit 121 calculates a feature amount in advance by a known feature extraction method using a luminance gradient such as SIFT (Scale-Invariant FeatureformTransform) or HOG (Histograms of Oriented Gradients) for each part region. Then, the part region detection unit 121 calculates the same feature amount for the detection target region, and determines which part region the calculated feature amount is close to the feature amount between the feature amounts. You may obtain
- the part area detection unit 121 determines the part area having the closest feature quantity to the previously calculated feature quantity as the detected part area, and expresses the distance representing the proximity as a score. For example, the part region detection unit 121 may set the score to 100 because the distance between the feature amounts is 0 when the feature amounts are the same. Further, the part region detection unit 121 may determine a region where the distance between the feature amounts is smaller than a predetermined value as the detected part region.
- FIG. 8 is a diagram illustrating an example of a result of detecting a tire.
- the partial area 2000 which is an area where the input image 1000 is divided, three areas where tires are detected are illustrated as shaded areas.
- the scores of the three areas are 40, 80, and 50, respectively.
- the score is held in a score map.
- the area indicating the score 40 is erroneous detection.
- the object determination unit 122 determines whether a certain area is an object area (an area in which an object is shown) based on a product sum value of the detected probability for each part area and the appearance probability distribution of the part area.
- the product-sum value is an evaluation value.
- a rectangular area 3000 shown in FIG. 9 is set as the fixed area.
- the appearance probability of the area of score 40 is “ ⁇ 0.3” obtained by multiplying “0.3” by a negative value
- the appearance probability of the area of score 80 is “0.8”
- the appearance probability of the area of score 50 is set.
- the probability is assumed to be “0.5”.
- the rectangular area 3500 in FIG. 10 having the highest product-sum value is determined as the detected object area.
- the certain area may be determined by generating one or a plurality of patterns of a rectangular area including the detected part area, for example. Further, the area size of the fixed area may be determined based on the size of the part area.
- the object determination unit 122 can also determine a single object region by integrating detection results for each part region.
- the object detection apparatus 100 can give a penalty to a falsely detected part region by generating an object appearance probability distribution composed of a part region appearance probability distribution and a non-appearance probability distribution. Therefore, the object region can be determined with high accuracy.
- the object detection apparatus 100 according to the present embodiment can obtain an effect that the object region can be detected with high accuracy by using the object appearance probability distribution in consideration of the non-appearance probability distribution.
- the size of the determination region (that is, the region for determining whether or not the object is reflected) is changed based on the increase rate of the objectness.
- the area of the determination region is increased or decreased, and the region is changed in a direction in which the probability that the object is determined to be captured (object probability) increases. Then, when the object probability is equal to or greater than a set threshold value in a region where the object probability is a maximum value, it may be determined that an object is reflected in the region.
- the object detection apparatus 101 according to the second embodiment of the present invention is characterized by having an object position search unit as compared with the configuration of the first embodiment. The difference between the object detection apparatus 101 according to the present embodiment and the first embodiment will be described below.
- FIG. 11 is a block diagram showing a configuration of the object detection apparatus 101 according to the second embodiment.
- the object detection apparatus 101 according to the second embodiment includes an object position search unit 123 in the detection unit 120 in addition to the configuration illustrated in FIG.
- the object position searching unit 123 searches for an area where the product sum value described in the first embodiment is highest while changing the area where the object area is detected.
- the object detection apparatus 101 has a product-sum value when the right side of the determination area is expanded or narrowed by one divided area, or a case where the upper side of the determination area is expanded by one divided area, The product-sum value is calculated in the case of narrowing, and a region where the product-sum value is maximum is searched.
- the tire area that should originally be included in the object area may protrude from the rectangular area.
- FIG. 12 is a diagram illustrating an example of a result of determining an object region by the object detection device 101 according to the second embodiment.
- the object detection apparatus 101 can prevent an area exceeding the tire area from being included in the object area, as indicated by a rectangular area 3600 in FIG.
- the object detection apparatus 101 calculates, for example, a product-sum value when the area is enlarged or reduced in a certain direction, and increases or decreases the product-sum value compared to the product-sum value of the previous area.
- the area may be enlarged / reduced in the direction of the image.
- the object position search unit 123 searches for a region with the highest product sum value while changing the region in which the object region is detected. The effect that the region can be detected is obtained.
- the object detection apparatus 102 includes a part region designation unit 103, an appearance probability distribution generation unit 104, and an object determination unit 105.
- the part area designating unit 103 designates a part area, which is an area in which the target part is captured, from a plurality of images in which the object is captured, among the parts constituting the object that is the detection target.
- the appearance probability distribution generation unit 104 generates an appearance probability distribution and a non-appearance probability distribution of the part area based on the appearance frequency associated with the position of the part area in the image.
- the object determination hand unit 105 refers to the appearance probability distribution of the part area and the non-appearance probability distribution to determine an area in which an object appears in the input image.
- the third embodiment it is determined whether or not the input image is an area in which an object is captured with reference to the appearance probability distribution and the non-appearance probability distribution of the part area. Therefore, the effect that the object region can be detected with high accuracy is obtained.
- each part of the object detection apparatus shown in FIGS. 1 and 11 is realized by the hardware resources illustrated in FIG. That is, the configuration shown in FIG. 14 includes a CPU (Central Processing Unit) 20, a RAM (Random Access Memory) 21, a ROM (Read Only Memory) 22, an external connection interface 23, and a storage device 24.
- the CPU 20 controls the overall operation of the object detection device by reading various software programs (computer programs) stored in the ROM 22 or the storage device 24 into the RAM 21 and executing them. That is, in each of the embodiments described above, the CPU 20 executes a software program that implements each function (each unit) included in the object detection device while referring to the ROM 22 or the storage device 24 as appropriate.
- the hardware configuration of the object detection device and each functional block thereof is not limited to the above-described configuration.
- the present invention described by taking each embodiment as an example is achieved by supplying a computer program capable of realizing the functions described above to the object detection apparatus, and then executing the computer program by the CPU 20. Is done.
- the supplied computer program may be stored in a computer-readable storage device such as a readable / writable memory (temporary storage medium) or a hard disk device.
- a computer-readable storage device such as a readable / writable memory (temporary storage medium) or a hard disk device.
- the present invention can be understood as being configured by a code representing the computer program or a storage medium storing the computer program.
- a part or all of the above-described embodiment can be described as in the following supplementary notes, but is not limited thereto.
- Appendix 1 A part region designating unit that designates a part region that is a region in which the target part is captured from a plurality of images in which the object is captured, among the parts that configure the object that is the detection target;
- An appearance probability distribution generating means for generating an appearance probability distribution of the part area and a non-appearance probability distribution based on an appearance frequency associated with the position of the part area in the image;
- An object detection apparatus comprising: an object determination unit that determines an area in which an object is captured in an input image with reference to the appearance probability distribution and the non-appearance probability distribution of the part area.
- Appendix 2 The object detection apparatus according to appendix 1, wherein the appearance probability distribution generation unit creates the appearance probability distribution by creating a histogram of appearance frequencies associated with positions of the part regions in the image.
- Appendix 4 The object detection device according to any one of appendix 1 to appendix 3, wherein the appearance probability distribution generation unit calculates an appearance frequency of an area in which a part other than the target part is captured as a non-appearance frequency of the part area.
- the appearance probability distribution generation means calculates, as the non-appearance probability, a value obtained by adding a negative value to a value equal to or less than a certain value of the appearance probability associated with the position in the image of the part region. 4.
- the object detection device according to any one of 3.
- a feature amount composed of a luminance change of a pixel is extracted with respect to a partial region that is a region into which the input image is divided, and a score value indicating the likelihood of the part region regarding the partial region is obtained based on the extracted feature amount. It further comprises a part area detection means for calculating, The object determination means is based on a result of product-sum operation between the score value calculated by the part area detection means and the appearance probability distribution of the part area generated by the appearance probability distribution generation means.
- the object detection device according to any one of appendix 1 to appendix 5, which determines an area in which an object is captured in an image.
- Appendix 7 The object detection device according to appendix 6, wherein the appearance probability distribution includes a distribution having a negative value indicating a position and a size at which the part region is unlikely to appear in the image.
- appendix 8 The object detection device according to appendix 6 or appendix 7, further comprising object position search means for changing a size of a determination area, which is an area for determining an area in which the object appears in the input image, based on an increase rate of the object likeness.
- the object position search means increases or decreases the area of the determination region, changes the region in a direction in which the probability that the object is determined to be captured increases,
- the object detection unit according to any one of appendix 6 to appendix 8, wherein the object determination unit determines that an object appears in the region when the probability is equal to or greater than a threshold value in the region where the probability is a maximum value. apparatus.
- a part region designating unit that designates a part region that is a region in which the target part is captured from a plurality of images in which the object is captured, among the parts that configure the object that is the detection target;
- An appearance probability distribution generating means for generating an appearance probability distribution of the part area and a non-appearance probability distribution based on an appearance frequency associated with the position of the part area in the image;
- a learning device for generating an appearance probability distribution of the part area and a non-appearance probability distribution based on an appearance frequency associated with the position of the part area in the image.
- Appendix 11 Among the parts that make up the object that is the detection target, specify the part area that is the area where the target part is shown from a plurality of images where the object is shown, Based on the appearance frequency of the part area associated with the position in the image, an appearance probability distribution of the part area and a non-appearance probability distribution are generated, An object detection method for determining an area where the object appears in an input image with reference to an appearance probability distribution and a non-appearance probability distribution of the part area.
- Appendix 12 A process of designating a part area, which is an area where the target part is captured, from a plurality of images where the object is captured, among the parts constituting the object that is the detection target; A process of generating an appearance probability distribution of the part region and a non-appearance probability distribution based on the appearance frequency of the part region associated with the position in the image.
- An object detection program for causing a computer to execute a process of determining an area in which an object appears in an input image with reference to the appearance probability distribution and the non-appearance probability distribution of the part area.
- the present invention can be applied to, for example, a computer device, a mobile phone, a smartphone, a digital still camera, and the like having a function of processing an image.
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
Description
第1の実施形態
図1は、本発明の実施の形態にかかる物体検出装置の構成を示すブロック図である。物体検出装置100は、画像供給部110、検出部120、学習部130、及び位置検出結果出力部140を有する。
第2の実施の形態
次に、上述した第1の実施形態を基礎とする第2の実施形態について説明する。以下の説明では、第1の実施形態と同様の構成については同じ参照番号を付与することにより、重複する説明は省略する。
次に、上記実施形態を包含する第3の実施形態について説明する。図13は、本発明の第3の実施の形態にかかる物体検出装置102は、パーツ領域指定部103、出現確率分布生成部104および物体判定部105を備える。
(付記1)
検出対象である物体を構成するパーツのうち、対象パーツが写る領域であるパーツ領域を、前記物体が写る複数の画像から指定するパーツ領域指定手段と、
前記パーツ領域の、前記画像における位置に関連付けられる出現頻度に基づいて、前記パーツ領域の出現確率分布と、非出現確率分布とを生成する出現確率分布生成手段と、
前記パーツ領域の出現確率分布と、非出現確率分布とを参照して、入力画像において前記物体が写る領域を判定する物体判定手段と
を備える物体検出装置。
前記出現確率分布生成手段は、前記パーツ領域の、前記画像における位置に関連付けられる出現頻度についてのヒストグラムを作成することにより前記出現確率分布を作成する
付記1に記載の物体検出装置。
前記出現確率分布生成手段は、前記パーツ領域の、前記物体が写る画像における出現位置と大きさを表す正の値を持つ分布を前記出現確率分布として生成する
付記1または付記2記載の物体検出装置。
前記出現確率分布生成手段は、前記対象パーツ以外のパーツが写る領域の出現頻度を、前記パーツ領域の非出現頻度として算出する
付記1ないし付記3のいずれか1に記載の物体検出装置。
前記出現確率分布生成手段は、前記パーツ領域の、前記画像における位置に関連付けられる出現確率の、一定値以下の値に負の値を積算した値を、前記非出現確率として算出する
付記1ないし付記3のいずれか1に記載の物体検出装置。
前記入力画像が分割された領域である部分領域に対して画素の輝度変化から構成される特徴量を抽出し、抽出した特徴量に基づいて、前記部分領域に関する前記パーツ領域らしさを示すスコア値を算出するパーツ領域検出手段をさらに備え、
前記物体判定手段は、前記パーツ領域検出手段により算出されたスコア値と、前記出現確率分布生成手段が生成した、前記パーツ領域の前記出現確率分布との積和演算の結果に基づいて、前記入力画像における物体が写る領域を判定する
付記1から付記5のいずれか1に記載の物体検出装置。
前記出現確率分布は、前記パーツ領域が前記画像における出現しにくい位置と大きさを表す負の値を持つ分布を含む
付記6に記載の物体検出装置。
物体らしさの増加率に基づいて、前記入力画像における物体が写る領域を判定する領域である判定領域の大きさを変化させる物***置探索手段を更に備える
付記6又は付記7に記載の物体検出装置。
前記物***置探索手段は、前記判定領域の面積を増減させ、前記物体が写ると判定される確率が大きくなる方向に領域を変化させ、
前記物体判定手段は、前記確率が極大値となる領域において、前記確率が、しきい値以上の場合、前記領域に物体が写ると判定する
付記6から付記8のいずれか1に記載の物体検出装置。
検出対象である物体を構成するパーツのうち、対象パーツが写る領域であるパーツ領域を、前記物体が写る複数の画像から指定するパーツ領域指定手段と、
前記パーツ領域の、前記画像における位置に関連付けられる出現頻度に基づいて、前記パーツ領域の出現確率分布と、非出現確率分布とを生成する出現確率分布生成手段と、
を備えた学習装置。
検出対象である物体を構成するパーツのうち、対象パーツが写る領域であるパーツ領域を、前記物体が写る複数の画像から指定し、
前記パーツ領域の、前記画像における位置に関連付けられる出現頻度に基づいて、前記パーツ領域の出現確率分布と、非出現確率分布とを生成し、
前記パーツ領域の出現確率分布と、非出現確率分布とを参照して、入力画像において前記物体が写る領域を判定する
物体検出方法。
検出対象である物体を構成するパーツのうち、対象パーツが写る領域であるパーツ領域を、前記物体が写る複数の画像から指定する処理と、
前記パーツ領域の、前記画像における位置に関連付けられる出現頻度に基づいて、前記パーツ領域の出現確率分布と、非出現確率分布とを生成する処理と、
前記パーツ領域の出現確率分布と、非出現確率分布とを参照して、入力画像において前記物体が写る領域を判定する処理と
をコンピュータに実行させる物体検出プログラム。
110 画像入力部
120 検出部
121 パーツ領域検出部
122 物体判定部
130 学習部
131 パーツ領域指定部
132 出現確率分布生成部
133 出現確率分布記憶部
140 位置検出結果出力部
Claims (10)
- 検出対象である物体を構成するパーツのうち、対象パーツが写る領域であるパーツ領域を、前記物体が写る複数の画像から指定するパーツ領域指定手段と、
前記パーツ領域の、前記画像における位置に関連付けられる出現頻度に基づいて、前記パーツ領域の出現確率分布と、非出現確率分布とを生成する出現確率分布生成手段と、
前記パーツ領域の出現確率分布と、非出現確率分布とを参照して、入力画像において前記物体が写る領域を判定する物体判定手段と
を備える物体検出装置。 - 前記出現確率分布生成手段は、前記パーツ領域の、前記画像における位置に関連付けられる出現頻度についてのヒストグラムを作成することにより前記出現確率分布を作成する
請求項1に記載の物体検出装置。 - 前記出現確率分布生成手段は、前記パーツ領域の、前記物体が写る画像における出現位置と大きさを表す正の値を持つ分布を前記出現確率分布として生成する
請求項1または請求項2記載の物体検出装置。 - 前記出現確率分布生成手段は、前記対象パーツ以外のパーツが写る領域の出現頻度を、前記パーツ領域の非出現頻度として算出する
請求項1ないし請求項3のいずれか1項に記載の物体検出装置。 - 前記出現確率分布生成手段は、前記パーツ領域の、前記画像における位置に関連付けられる出現確率の、一定値以下の値に負の値を積算した値を、前記非出現確率として算出する
請求項1ないし請求項3のいずれか1項に記載の物体検出装置。 - 前記入力画像が分割された領域である部分領域に対して画素の輝度変化から構成される特徴量を抽出し、抽出した特徴量に基づいて、前記部分領域に関する前記パーツ領域らしさを示すスコア値を算出するパーツ領域検出手段をさらに備え、
前記物体判定手段は、前記パーツ領域検出手段により算出されたスコア値と、前記出現確率分布生成手段が生成した、前記パーツ領域の前記出現確率分布との積和演算の結果に基づいて、前記入力画像における物体が写る領域を判定する
請求項1から請求項5のいずれか1項に記載の物体検出装置。 - 前記出現確率分布は、前記パーツ領域が前記画像における出現しにくい位置と大きさを表す負の値を持つ分布を含む
請求項6に記載の物体検出装置。 - 物体らしさの増加率に基づいて、前記入力画像における物体が写る領域を判定する領域である判定領域の大きさを変化させる物***置探索手段を更に備える
請求項6又は請求項7に記載の物体検出装置。 - 検出対象である物体を構成するパーツのうち、対象パーツが写る領域であるパーツ領域を、前記物体が写る複数の画像から指定するパーツ領域指定手段と、
前記パーツ領域の、前記画像における位置に関連付けられる出現頻度に基づいて、前記パーツ領域の出現確率分布と、非出現確率分布とを生成する出現確率分布生成手段と、
を備えた学習装置。 - 検出対象である物体を構成するパーツのうち、対象パーツが写る領域であるパーツ領域を、前記物体が写る複数の画像から指定し、
前記パーツ領域の、前記画像における位置に関連付けられる出現頻度に基づいて、前記パーツ領域の出現確率分布と、非出現確率分布とを生成し、
前記パーツ領域の出現確率分布と、非出現確率分布とを参照して、入力画像において前記物体が写る領域を判定する
物体検出方法。
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2015540368A JP6524910B2 (ja) | 2013-10-01 | 2014-08-21 | 物体検出装置、物体検出方法および学習装置 |
US15/025,337 US10026002B2 (en) | 2013-10-01 | 2014-08-21 | Object detection apparatus, method for detecting object, and learning apparatus |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2013206046 | 2013-10-01 | ||
JP2013-206046 | 2013-10-01 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2015049826A1 true WO2015049826A1 (ja) | 2015-04-09 |
Family
ID=52778433
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2014/004284 WO2015049826A1 (ja) | 2013-10-01 | 2014-08-21 | 物体検出装置、物体検出方法および学習装置 |
Country Status (3)
Country | Link |
---|---|
US (1) | US10026002B2 (ja) |
JP (1) | JP6524910B2 (ja) |
WO (1) | WO2015049826A1 (ja) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR101916832B1 (ko) | 2017-07-11 | 2018-11-08 | 강경수 | 어노테이션 바운딩 박스를 이용한 오브젝트 감지 방법 및 장치 |
JP2019197359A (ja) * | 2018-05-09 | 2019-11-14 | 国際航業株式会社 | 一般物体認識システム |
Families Citing this family (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9727800B2 (en) * | 2015-09-25 | 2017-08-08 | Qualcomm Incorporated | Optimized object detection |
CN109658455B (zh) * | 2017-10-11 | 2023-04-18 | 阿里巴巴集团控股有限公司 | 图像处理方法和处理设备 |
JP2020141288A (ja) * | 2019-02-28 | 2020-09-03 | キヤノン株式会社 | 情報処理装置、情報処理方法、及びプログラム |
DE102019108142A1 (de) * | 2019-03-29 | 2020-10-01 | Bayerische Motoren Werke Aktiengesellschaft | Auswählen einer Handlungsoption für ein automatisiertes Kraftfahrzeug |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2004054442A (ja) * | 2002-07-17 | 2004-02-19 | Glory Ltd | 顔検出装置、顔検出方法および顔検出プログラム |
JP2007034723A (ja) * | 2005-07-27 | 2007-02-08 | Glory Ltd | 顔画像検出装置、顔画像検出方法および顔画像検出プログラム |
WO2013008302A1 (ja) * | 2011-07-11 | 2013-01-17 | トヨタ自動車株式会社 | 赤目判定装置 |
Family Cites Families (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB2409030A (en) * | 2003-12-11 | 2005-06-15 | Sony Uk Ltd | Face detection |
US8131011B2 (en) * | 2006-09-25 | 2012-03-06 | University Of Southern California | Human detection and tracking system |
JP5096211B2 (ja) * | 2008-03-31 | 2012-12-12 | 富士フイルム株式会社 | 確率分布構築方法、確率分布構築装置、および確率分布構築プログラム、並びに被写体検出方法、被写体検出装置、および被写体検出プログラム |
US8913783B2 (en) * | 2009-10-29 | 2014-12-16 | Sri International | 3-D model based method for detecting and classifying vehicles in aerial imagery |
WO2011055772A1 (ja) * | 2009-11-05 | 2011-05-12 | 日本電気株式会社 | 画像目標識別装置、画像目標識別方法、画像目標識別プログラム |
WO2011061905A1 (ja) * | 2009-11-20 | 2011-05-26 | 日本電気株式会社 | 物体領域抽出装置、物体領域抽出方法、及びコンピュータ可読媒体 |
JP5615088B2 (ja) * | 2010-08-18 | 2014-10-29 | キヤノン株式会社 | 画像処理装置及びその方法、プログラム、並びに撮像装置 |
JP5675229B2 (ja) * | 2010-09-02 | 2015-02-25 | キヤノン株式会社 | 画像処理装置及び画像処理方法 |
WO2012046392A1 (ja) * | 2010-10-08 | 2012-04-12 | パナソニック株式会社 | 姿勢推定装置及び姿勢推定方法 |
WO2012077287A1 (ja) * | 2010-12-09 | 2012-06-14 | パナソニック株式会社 | 姿勢状態推定装置および姿勢状態推定方法 |
WO2012077286A1 (ja) * | 2010-12-09 | 2012-06-14 | パナソニック株式会社 | 物体検出装置および物体検出方法 |
US9232128B2 (en) * | 2011-03-28 | 2016-01-05 | Nec Corporation | Image capture position and image capture direction estimation device, image capture device, image capture position and image capture direction estimation method and program |
-
2014
- 2014-08-21 JP JP2015540368A patent/JP6524910B2/ja active Active
- 2014-08-21 WO PCT/JP2014/004284 patent/WO2015049826A1/ja active Application Filing
- 2014-08-21 US US15/025,337 patent/US10026002B2/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2004054442A (ja) * | 2002-07-17 | 2004-02-19 | Glory Ltd | 顔検出装置、顔検出方法および顔検出プログラム |
JP2007034723A (ja) * | 2005-07-27 | 2007-02-08 | Glory Ltd | 顔画像検出装置、顔画像検出方法および顔画像検出プログラム |
WO2013008302A1 (ja) * | 2011-07-11 | 2013-01-17 | トヨタ自動車株式会社 | 赤目判定装置 |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR101916832B1 (ko) | 2017-07-11 | 2018-11-08 | 강경수 | 어노테이션 바운딩 박스를 이용한 오브젝트 감지 방법 및 장치 |
JP2019197359A (ja) * | 2018-05-09 | 2019-11-14 | 国際航業株式会社 | 一般物体認識システム |
JP7180827B2 (ja) | 2018-05-09 | 2022-11-30 | 国際航業株式会社 | 一般物体認識システム |
Also Published As
Publication number | Publication date |
---|---|
JPWO2015049826A1 (ja) | 2017-03-09 |
JP6524910B2 (ja) | 2019-06-05 |
US10026002B2 (en) | 2018-07-17 |
US20160232418A1 (en) | 2016-08-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2015049826A1 (ja) | 物体検出装置、物体検出方法および学習装置 | |
JP6188400B2 (ja) | 画像処理装置、プログラム及び画像処理方法 | |
US9619733B2 (en) | Method for generating a hierarchical structured pattern based descriptor and method and device for recognizing object using the same | |
US20170132493A1 (en) | Image processing apparatus, image processing method, and storage medium | |
KR20170041168A (ko) | 차선 데이터의 처리 방법, 장치, 저장매체 및 기기 | |
US9875253B2 (en) | Color sketch image searching | |
US8929664B2 (en) | Shape detection using chain code states | |
US9256800B2 (en) | Target line detection device and method | |
CN107423306B (zh) | 一种图像检索方法及装置 | |
JP2010205067A (ja) | 領域抽出装置、領域抽出方法及び領域抽出プログラム | |
US9189841B2 (en) | Method of checking the appearance of the surface of a tyre | |
CN106778777B (zh) | 一种车辆匹配方法及*** | |
Zhao et al. | A saliency detection based method for 3d surface simplification | |
CN104700620A (zh) | 一种基于交通卡口的***辆识别方法和装置 | |
JP2017102622A (ja) | 画像処理装置、画像処理方法及びプログラム | |
CN117495891B (zh) | 点云边缘检测方法、装置和电子设备 | |
CN109671098A (zh) | 可适用多重追踪的目标追踪方法与*** | |
JP2020017136A (ja) | 物体検出認識装置、方法、及びプログラム | |
JP6077785B2 (ja) | 対象物検出装置及びプログラム | |
KR101733288B1 (ko) | 방향정보를 이용한 객체 검출기 생성 방법, 이를 이용한 객체 검출 장치 및 방법 | |
JP2006260311A (ja) | マッチング方法およびマッチング装置ならびにプログラム | |
KR101961854B1 (ko) | 3d 객체 모델 기법을 이용한 객체 인식장치 | |
CN103679170B (zh) | 一种基于局部特征的显著区域检测方法 | |
CN115984178A (zh) | 伪造图像检测方法、电子设备和计算机可读存储介质 | |
Bielicki et al. | A method of 3D object recognition and localization in a cloud of points |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 14851346 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2015540368 Country of ref document: JP Kind code of ref document: A |
|
WWE | Wipo information: entry into national phase |
Ref document number: 15025337 Country of ref document: US |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 14851346 Country of ref document: EP Kind code of ref document: A1 |