WO2021227723A1 - 目标检测方法、装置、计算机设备及可读存储介质 - Google Patents

目标检测方法、装置、计算机设备及可读存储介质 Download PDF

Info

Publication number
WO2021227723A1
WO2021227723A1 PCT/CN2021/086077 CN2021086077W WO2021227723A1 WO 2021227723 A1 WO2021227723 A1 WO 2021227723A1 CN 2021086077 W CN2021086077 W CN 2021086077W WO 2021227723 A1 WO2021227723 A1 WO 2021227723A1
Authority
WO
WIPO (PCT)
Prior art keywords
detection point
target
probability value
detection
detected
Prior art date
Application number
PCT/CN2021/086077
Other languages
English (en)
French (fr)
Inventor
杨静林
Original Assignee
京东方科技集团股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 京东方科技集团股份有限公司 filed Critical 京东方科技集团股份有限公司
Priority to US17/765,366 priority Critical patent/US20220351413A1/en
Publication of WO2021227723A1 publication Critical patent/WO2021227723A1/zh

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
    • G06V10/20—Image preprocessing
    • G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/70—Determining position or orientation of objects or cameras
    • G06T7/77—Determining position or orientation of objects or cameras using statistical methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/70—Determining position or orientation of objects or cameras
    • G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/70—Determining position or orientation of objects or cameras
    • G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
    • G06T7/75—Determining position or orientation of objects or cameras using feature-based methods involving models
    • 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
    • G06T2207/10004—Still image; Photographic image
    • 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
    • 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
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection

Definitions

  • the embodiments of the present disclosure relate to a target detection method, a target detection device, a computer device, and a computer-readable storage medium.
  • target detection is a basic and practical research direction.
  • the target objects such as people, animals, plants, vehicles, etc.
  • the identified targets can be marked.
  • a target detection method includes: determining, according to the image to be detected, a plurality of detection points corresponding to a plurality of regions on the image to be detected, and the probability value of the target object in the region corresponding to each detection point; from all the detection points Screen out the first detection point with the largest probability value, and at least one second detection point whose probability value is less than the probability value of the first detection point and greater than or equal to the probability threshold; wherein, the area corresponding to the first detection point There is a target; it is determined whether the first distance between each second detection point and the first detection point is greater than or equal to the distance threshold, and when the first distance is greater than or equal to the distance threshold, the corresponding The probability value of the second detection point is updated to obtain the updated probability value; the updated probability value is compared with the probability threshold to obtain a comparison result, and the second detection point corresponding to the second detection point is determined according to the comparison result. Whether there is a new target in the area.
  • the updating the original probability value of the corresponding second detection point to obtain the updated probability value includes: determining the probability reduction amount of the corresponding second detection point according to each of the first distances , Wherein the probability reduction amount is positively correlated with the first distance; the difference between the original probability value of the second detection point and the probability reduction amount is taken as the updated probability of the second detection point value.
  • the updating the original probability value of the corresponding second detection point to obtain the updated probability value includes: inputting the probability value of the second detection point as a dependent variable into a preset function, and Obtain the first parameter; wherein, the preset function is a monotonically decreasing function, and is located in the first quadrant of the rectangular coordinate system; the first distance from the second detection point to the first detection point and Adding the first parameters to obtain a second parameter; and inputting the second parameter as an independent variable into the preset function to obtain an updated probability value of the second detection point.
  • the preset function is the part of the following Gaussian function in the first quadrant of the rectangular coordinate system:
  • a, b, and c are real constants, and a>0, x is the second parameter, and f(x) is the updated probability value of the second detection point.
  • the preset function is the part of the following linear function in the first quadrant of the rectangular coordinate system:
  • k and b are real constants, and k ⁇ 0, x is the second parameter, and f(x) is the updated probability value of the second detection point.
  • the comparing the updated probability value with the probability threshold value to obtain a comparison result, and determining whether there is a new target in the area corresponding to the second detection point according to the comparison result includes : Determine whether the updated probability value is greater than or equal to the probability threshold; if it is, it is determined that there is a new target in the area corresponding to the second detection point; if not, it is determined that the area corresponding to the second detection point is not There are new targets.
  • the method further includes: when the first distance is less than the distance threshold, detecting that the area corresponding to the second detection point and the area corresponding to the first detection point have the same target.
  • the determining, according to the image to be detected, a plurality of detection points corresponding to a plurality of regions on the image to be detected, and the probability value of the target object in the region corresponding to each detection point includes:
  • the image to be detected is input into a trained target network model for feature extraction, and corresponding feature information is obtained.
  • the feature information includes a feature map and the probability value of the target object at each detection point on the feature map; the resolution of the feature map It is 1/n times the resolution of the image to be detected, and n>1.
  • the feature information includes the position compensation accuracy of each detection point; the target detection method further includes: determining the target object corresponding to the first detection point according to the first detection point and its position compensation accuracy The center point on the image to be detected; and, according to the second detection point where a new target exists in the corresponding area and its position compensation accuracy, it is determined that the new target corresponding to the second detection point is in the to-be-detected The second center point on the image.
  • the feature information includes the regression size of each target; the target detection method further includes: determining that each detected target is in the image to be detected according to the regression size of each target. The area covered by the above.
  • the target network model includes a down-sampling module, an up-sampling module, and a residual module;
  • the down-sampling module is configured to perform a down-sampling operation on the image to be detected to obtain the feature map;
  • the up-sampling module is configured to perform an up-sampling operation on the feature map;
  • the residual module is configured to extract multiple input image features from the to-be-detected image, so that the up-sampling module can combine the multiple The input image feature performs the upsampling operation on the feature map.
  • the probability threshold is 0.4 to 0.6.
  • a target detection device in another aspect, includes a detection point determination module, a screening module, a judgment update module, and a comparison determination module.
  • the detection point determination module is configured to determine, according to the image to be detected, a plurality of detection points corresponding to a plurality of regions on the image to be detected, and a probability value of the target object in the region corresponding to each detection point.
  • the screening module is configured to screen out the first detection point with the largest probability value from all the detection points, and at least one second detection point whose probability value is less than the probability value of the first detection point and greater than or equal to the probability threshold; wherein , There is a target in the area corresponding to the first detection point.
  • the judgment update module is configured to judge whether the first distance between each second detection point and the first detection point is greater than or equal to a distance threshold, and when the first distance is greater than or equal to the distance threshold, The original probability value of the corresponding second detection point is updated to obtain the updated probability value.
  • the comparison and determination module is configured to compare the updated probability value with the probability threshold to obtain a comparison result, and determine whether there is a new target in the area corresponding to the second detection point according to the comparison result.
  • a computer device in another aspect, includes: a memory; a processor; and, computer program instructions stored in the memory and capable of being run on the processor, and the processor implements any one of the foregoing when the computer program instructions are executed.
  • the processor implements any one of the foregoing when the computer program instructions are executed.
  • a computer-readable storage medium stores computer program instructions, and when the computer program instructions run on a processor, the processor executes the target detection method according to any one of the foregoing embodiments.
  • a computer program product includes computer program instructions, and when the computer program instructions are executed on a computer, the computer program instructions cause the computer to execute the target detection method as described in any of the foregoing embodiments.
  • a computer program is provided.
  • the computer program When the computer program is executed on a computer, the computer program causes the computer to execute the target detection method as described in any of the above embodiments.
  • Fig. 1 is a flowchart of a target detection method according to some embodiments of the present disclosure
  • Figure 2 is a target network model according to some embodiments of the present disclosure
  • Figure 3 is an image to be detected and its corresponding feature map according to some embodiments of the present disclosure
  • Figure 4 is another target network model according to some embodiments of the present disclosure.
  • Fig. 5 is a residual error module in a target network model according to some embodiments of the present disclosure.
  • Fig. 6 is a characteristic diagram according to some embodiments of the present disclosure.
  • Fig. 7 is a flowchart of another target detection method according to some embodiments of the present disclosure.
  • FIG. 8 is a flowchart of still another target detection method according to some embodiments of the present disclosure.
  • FIG. 9 is a flowchart of yet another target detection method according to some embodiments of the present disclosure.
  • FIG. 10 is a flowchart of yet another target detection method according to some embodiments of the present disclosure.
  • Fig. 11 is a block diagram of a target detection device according to some embodiments of the present disclosure.
  • FIG. 12 is a block diagram of another target detection device according to some embodiments of the present disclosure.
  • Fig. 13 is a block diagram of a computer device according to some embodiments of the present disclosure.
  • first and second are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Thus, the features defined with “first” and “second” may explicitly or implicitly include one or more of these features. In the description of the embodiments of the present disclosure, unless otherwise specified, “plurality” means two or more.
  • the expressions “coupled” and “connected” and their extensions may be used.
  • the term “connected” may be used when describing some embodiments to indicate that two or more components are in direct physical or electrical contact with each other.
  • the term “coupled” may be used when describing some embodiments to indicate that two or more components have direct physical or electrical contact.
  • the term “coupled” or “communicatively coupled” may also mean that two or more components are not in direct contact with each other, but still cooperate or interact with each other.
  • the embodiments disclosed herein are not necessarily limited to the content of this document.
  • the exemplary embodiments are described herein with reference to cross-sectional views and/or plan views as idealized exemplary drawings.
  • the thickness of layers and regions are exaggerated for clarity. Therefore, variations in the shape with respect to the drawings due to, for example, manufacturing technology and/or tolerances can be envisaged. Therefore, the exemplary embodiments should not be construed as being limited to the shape of the area shown herein, but include shape deviations due to, for example, manufacturing.
  • an etched area shown as a rectangle will generally have curved features. Therefore, the areas shown in the drawings are schematic in nature, and their shapes are not intended to show the actual shape of the area of the device, and are not intended to limit the scope of the exemplary embodiments.
  • Image target detection is one of the most common and practical research directions in the field of computer vision.
  • convolutional neural networks have shown great advantages in the field of image processing, especially in the detection and recognition of targets.
  • the algorithms in the related art all need to set the anchor frame in the algorithm in advance. This method requires a priori knowledge of the target to be inspected, increases the hyperparameters that need to be set in the algorithm, and generates a large number of invalid anchor frames to be inspected, which wastes computing resources.
  • the method in the related technology is to directly remove potential target points near the target with a larger probability value, which makes the detection ability poor in scenes containing dense targets.
  • the target detection method includes steps 101 to 105.
  • Step 101 Determine, according to the image to be detected, multiple detection points corresponding to multiple regions on the image to be detected, and the probability value of the target object in the region corresponding to each detection point.
  • the image to be detected can be input into a trained target network model for feature extraction to obtain corresponding feature information.
  • the feature information includes the target feature map and the probability value of the target object at each detection point on the target feature map.
  • the resolution of the target feature map is 1/n times the resolution of the image to be detected, and n>1.
  • the target network model is an hourglass network model.
  • the target network model includes a down-sampling module 11 and an up-sampling module 12, where a down-sampling module 11 and an up-sampling module 12 form an hourglass structure 1.
  • the down-sampling module 11 may be configured to perform a down-sampling operation on the image A to be detected to obtain a feature map.
  • the down-sampling operation may extract data from the image A to be detected at uniform intervals, thereby reducing the size of the image A to be detected. For example, if the down-sampling operation of 2 times is performed, the size of the image A to be detected can be reduced from 4 ⁇ 4 to 2 ⁇ 2 (that is, the resolution of the feature map obtained at this time is 1/2 of the image A to be detected. Times).
  • the size of the image A to be detected can be reduced from 8 ⁇ 8 to 2 ⁇ 2 (that is, the resolution of the feature map obtained at this time is 1/4 times that of the image A to be detected).
  • the downsampling operation can map all pixels in the 2 ⁇ 2 area 01 in the image A to be detected into a pixel 02 in the feature map A', for example, The average value of all pixels in the area is taken as the pixel value of the one pixel in the down-sampled image.
  • each pixel 02 corresponds to a detection point.
  • the up-sampling module 12 may be configured to perform an up-sampling operation on the obtained feature map.
  • the upsampling operation may be to interpolate the feature map at even intervals. For example, if a 2 times upsampling operation is performed, the size of the image feature can be increased from 2 ⁇ 2 to 4 ⁇ 4 (that is, the feature map can be better restored to the image to be detected A at this time). If a 4-fold upsampling operation is performed, the size of the feature map can be increased from 2 ⁇ 2 to 8 ⁇ 8 (that is, the feature map can be better restored to the image to be detected A at this time).
  • the upsampling operation may include performing interpolation operations on the image, for example, neighbor interpolation (such as bilinear interpolation, bicubic interpolation, spline interpolation, etc.), edge-based interpolation, and/or region-based interpolation.
  • neighbor interpolation such as bilinear interpolation, bicubic interpolation, spline interpolation, etc.
  • edge-based interpolation such as edge-based interpolation, and/or region-based interpolation.
  • the target network model further includes a residual module 13.
  • Fig. 4 shows another structural form of the target network model
  • Fig. 5 shows a residual module in Fig. 4.
  • the residual module 13 can extract multiple input image features from the image to be detected, and the up-sampling module 12 can combine the multiple input image features to perform an up-sampling operation on the feature map, so that the feature map can be better Restore to the image A to be detected.
  • the size shown above may not be its actual size, but only used to represent the proportional relationship between the size of the image A to be detected and the feature map.
  • the size of the feature map obtained after the 2x downsampling operation can be 512 ⁇ 512 (that is, the resolution of the feature map obtained at this time is the image to be detected)
  • the size of the feature map obtained after performing the 4-fold downsampling operation can be 256 ⁇ 256 (that is, the resolution of the feature map obtained at this time is 1/4 times of the image A to be detected). That is, the resolution of the feature map obtained by the down-sampling module 11 may be 1/n times the resolution of the image A to be detected, and n>1.
  • the target network model can process the obtained one or more feature maps, and then realize the output of the above-mentioned feature information (the feature information includes the target feature map and the presence of the target at each detection point on the target feature map. Probability value).
  • each detection point on the target feature map can correspond to a region on the image to be detected, for example, a block of 2 ⁇ 2 Area, or, a 4 ⁇ 4 area.
  • the target network model can output the probability value P (i, j) of the c-th target in the detection point (i.e., the area corresponding to the detection point on the image to be detected) at the position (i, j) on the target feature map. .
  • Step 102 Screen out the first detection point with the largest probability value from all the detection points, and at least one second detection point whose probability value is less than the probability value of the first detection point and greater than or equal to the probability threshold; where the first detection There is a target in the area corresponding to the point.
  • Figure 6 shows a target feature map.
  • the aforementioned feature information includes width and height information of the target feature map, and the number of target categories in the target feature map can be displayed.
  • the probability value of the existence of the target object (ie polar bear) at the first detection point T1 is 1; the probability value of the existence of the target object (ie polar bear) at the second detection point T2 is greater than 0 and less than 1. That is, by inputting the image to be detected into the trained target network model, it can be directly detected that the target object exists in the area corresponding to the first detection point T1 (for example, there is one area corresponding to the two first detection points T1 in FIG. Polar bear). Moreover, in some examples, the above-mentioned feature information also includes the position compensation accuracy of each detection point.
  • the center point of the target corresponding to the first detection point T1 on the image A to be detected can be determined according to the first detection point T1 and its position compensation accuracy.
  • the above-mentioned feature information further includes the regression size of each target object.
  • the width of the target object on the target feature map can be obtained from the image features around the first detection point T1 (for example, the probability value of the surrounding detection points) (for example, the width of the target object corresponding to the first detection point T1 on the left is W ) And height (for example, the height of the target object corresponding to the first detection point T1 on the left is H), and then according to the regression size, the area covered by the target object on the image to be detected can be determined, so as to accurately detect The location of the target.
  • the target detection method of this embodiment is used to detect the target in the image without setting an anchor frame in advance. Therefore, there is no need to set a priori knowledge of the target to be detected, no hyperparameters set in the algorithm, and a large number of invalid anchor frames to be checked are not generated, which saves computing resources.
  • the second detection point T2 when detecting overlapping targets (for example, detecting two polar bears corresponding to the first detection point T1 and the second detection point T2 on the left in FIG. 5), the second detection point T2 The distance to the first detection point T1 on the left is relatively close, and the probability value of the second detection point T2 is less than the probability value of the first detection point T1 on the left, resulting in the target object corresponding to the second detection point T2 It cannot be detected, that is, in the related art, for the first detection point T1 and the second detection point T2 on the left side, only one polar bear can be detected, which makes it difficult to effectively detect dense targets.
  • Step 103 Determine whether the first distance between each second detection point and the first detection point is greater than or equal to a distance threshold, and when the first distance is greater than or equal to the distance threshold, step 104 is executed.
  • Step 104 Update the original probability value of the corresponding second detection point to obtain the updated probability value.
  • Step 105 Compare the updated probability value with the probability threshold value to obtain a comparison result, and determine whether there is a new target in the area corresponding to the second detection point according to the comparison result.
  • the distance threshold can be set according to application scenarios, that is, different application scenarios (for example, an image to be detected corresponding to an intersection, and an image to be detected corresponding to a school gate), the corresponding distance threshold is also different. This embodiment does not limit this.
  • the original probability value of the second detection point T2 is updated to obtain the updated probability value, and the updated probability value is compared with the probability threshold to obtain the comparison result, which can be based on The comparison result determines whether there is a new target in the area corresponding to the second detection point. Therefore, compared with related technologies, this embodiment can detect dense targets more effectively.
  • the probability threshold is 0.4 to 0.6.
  • the probability threshold may be 0.5.
  • the above step 105 includes step 1051 to step 1053.
  • Step 1051 determine whether the updated probability value is greater than or equal to the probability threshold; if yes, go to step 1052, determine that there is a new target in the area corresponding to the second detection point; if not, go to step 1053, determine the second detection point There is no new target in the corresponding area.
  • the aforementioned step 104 includes step 104A and step 104B.
  • step 104A the probability reduction amount of the corresponding second detection point is determined according to each of the first distances, where the probability reduction amount is positively correlated with the first distance. Among them, each first distance has its corresponding probability reduction.
  • Step 104B Use the difference between the original probability value of the second detection point and the probability reduction amount as the updated probability value of the second detection point.
  • the probability value after the update of the second detection point T2 can be made smaller than the probability value before the update of the second detection point T2. This will help reduce the amount of subsequent data processing and improve computing efficiency.
  • the second detection point with a larger probability value before the update can be screened out, so that the adjacent target can be more accurately achieved. Detection.
  • the updated probability value is compared with The greater the reduction of the probability value before the update, the greater the amount of subsequent data processing can be effectively reduced, the calculation efficiency can be improved, and the detection and recognition of adjacent targets can be more accurately achieved.
  • step 104 includes step 1041 to step 1043.
  • step 1041 input the original probability value of the second detection point as the dependent variable into the preset function to obtain the first parameter; wherein the preset function is a monotonically decreasing function, and is located in the first parameter of the rectangular coordinate system.
  • Step 1042 Add the first distance from the second detection point to the first detection point and the first parameter to obtain the second parameter;
  • Step 1043 Input the second parameter as an independent variable into the preset function to obtain the updated probability value of the second detection point.
  • the second detection point with the farther distance from the first detection point that is, the greater the first distance
  • the second detection point with the farther distance from the first detection point that is, the greater the first distance
  • this can effectively reduce the amount of subsequent data processing, improve computing efficiency, and achieve more accurate adjacent goals Detection and identification.
  • the preset function is the part of the following Gaussian function located in the first quadrant of the rectangular coordinate system:
  • a, b, and c are real constants, and a>0, x is the above-mentioned second parameter, and f(x) is the updated probability value of the second detection point.
  • the part of the Gaussian function in the first quadrant of the rectangular coordinate system is a monotonically decreasing function, and according to the Gaussian function, the larger the second parameter, the smaller the updated probability value obtained, that is, For a second detection point with the same probability value before the update but a different distance from the first detection point, the greater the distance value from the second detection point to the first detection point, the greater the value of the second detection point after the update.
  • the probability value is lower than the probability value before the update, so the detection and recognition of adjacent targets can be achieved more accurately.
  • the preset function is the part of the following linear function in the first quadrant of the rectangular coordinate system:
  • k and b are real constants, and k ⁇ 0, x is the above-mentioned second parameter, and f(x) is the updated probability value of the second detection point.
  • the part of the linear function in the first quadrant of the rectangular coordinate system is a monotonically decreasing function, and according to the linear function, the larger the second parameter, the smaller the updated probability value obtained. That is, the greater the distance value from the second detection point to the first detection point, the greater the reduction in the obtained probability value after the update compared with the probability value before the update, so that the detection and recognition of adjacent targets can be more accurately achieved.
  • the target detection method further includes: if it is determined that the distance between the second detection point and the first detection point is less than the distance threshold, step 106 is executed.
  • Step 106 Determine that the same target object exists in the area corresponding to the second detection point and the area corresponding to the first detection point. For example, referring to FIG. 6, if the second detection point T2 falls within the rectangular area covering the polar bear on the left side corresponding to the first detection point, it can be considered that the two correspond to the same target.
  • some embodiments of the present disclosure can divide the function modules of the target detection device according to the above method examples.
  • each function module can be divided corresponding to each function, or two or more functions can be integrated into one process.
  • Module The above-mentioned integrated modules can be implemented either in the form of hardware or in the form of software functional components. It should be noted that the division of modules in some embodiments of the present disclosure is illustrative, and is only a logical function division, and there may be other division methods in actual implementation.
  • FIG. 11 shows a possible schematic diagram of the target detection device involved in the foregoing embodiment.
  • the target detection device 200 includes: a detection point The determination module 21, the screening module 22, the judgment update module 23, and the comparison and determination module 24.
  • the detection point determination module 21 is configured to determine, according to the image to be detected, multiple detection points corresponding to multiple areas on the image to be detected, and the probability value of the target object in the area corresponding to each detection point. That is, the detection point determination module 21 can support the target display device 200 to execute the above step 101.
  • the screening module 22 is configured to screen out the first detection point with the largest probability value from all the detection points, and at least one second detection point whose probability value is less than the probability value of the first detection point and greater than or equal to the probability threshold; where , There is a target in the area corresponding to the first detection point. That is, the screening module 22 can support the target display device 200 to execute the above step 102.
  • the judgment update module 23 is configured to judge whether the distance between each second detection point and the first detection point is greater than or equal to the distance threshold, and if so, update the probability value of the corresponding second detection point to obtain the updated The probability value. That is, the judgment update module 23 can support the target display device 200 to execute the above step 103 and step 104.
  • the comparison and determination module 24 is configured to compare the updated probability value with the probability threshold to obtain a comparison result, and determine whether there is a new target in the area corresponding to the second detection point according to the comparison result. That is, the comparison and determination module 24 can support the target display device 200 to execute the above step 105.
  • the target detection device provided by some embodiments of the present disclosure is configured to execute the above-mentioned target detection method, and therefore can achieve the same effect as the above-mentioned target detection method.
  • FIG. 12 shows another possible schematic diagram of the composition of the target detection device involved in the foregoing embodiment.
  • the target detection device 200 includes: a processing component 31, a communication component 32 and a storage component 33.
  • the processing component 31 is configured to control and manage the actions of the target detection device.
  • the processing component 31 is configured to support the target detection device to perform step 101 to step 105 in FIG. 1 and/or is configured as the technology described herein.
  • Other processes are configured to support the target detection device and other network entities.
  • the storage component 33 is configured to store program codes and data of the target detection device.
  • the processing component 31 is a processor. It can implement or execute various exemplary logical blocks, components and circuits described in conjunction with the present disclosure.
  • the processor may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a digital signal processor (Digital Signal Processor, DSP) and a microprocessor, and so on.
  • the communication component 32 may be a communication interface.
  • the storage part 33 may be a memory.
  • some embodiments of the present disclosure also provide a display device, which includes the target detection device described in any of the above-mentioned embodiments.
  • the display device may be any product or component with display function, such as AR helmet, AR glasses, mobile phone, tablet computer, TV, monitor, notebook computer, digital photo frame, navigator, etc.
  • the display device provided by some embodiments of the present disclosure can execute the foregoing target detection method through the target detection device, and therefore can achieve the same effect as the foregoing target detection method.
  • the embodiment of the present disclosure also provides a computer device.
  • the computer device 300 includes: a memory 42; a processor 41; and, computer program instructions 43 stored in the memory 42 and running on the processor 41 When the processor 41 executes the computer program instructions 43, one or more steps in the foregoing target detection method are implemented.
  • Some embodiments of the present disclosure provide a computer-readable storage medium (for example, a non-transitory computer-readable storage medium), the computer-readable storage medium stores computer program instructions, and when the computer program instructions run on a processor , Causing the processor to execute one or more steps in the target detection method described in any one of the foregoing embodiments.
  • a computer-readable storage medium for example, a non-transitory computer-readable storage medium
  • the foregoing computer-readable storage medium may include, but is not limited to: magnetic storage devices (for example, hard disks, floppy disks, or magnetic tapes, etc.), optical disks (for example, CD (Compact Disk), DVD (Digital Versatile Disk, Digital universal disk), etc.), smart cards and flash memory devices (for example, EPROM (Erasable Programmable Read-Only Memory), cards, sticks or key drives, etc.).
  • Various computer-readable storage media described in this disclosure may represent one or more devices and/or other machine-readable storage media for storing information.
  • the term "machine-readable storage medium" may include, but is not limited to, wireless channels and various other media capable of storing, containing, and/or carrying instructions and/or data.
  • Some embodiments of the present disclosure also provide a computer program product.
  • the computer program product includes computer program instructions, and when the computer program instructions are executed on a computer, the computer program instructions cause the computer to execute one or more steps in the target detection method described in the above-mentioned embodiments.
  • Some embodiments of the present disclosure also provide a computer program.
  • the computer program When the computer program is executed on the computer, the computer program causes the computer to execute one or more steps in the target detection method described in the above-mentioned embodiments.

Abstract

一种目标检测方法，包括：根据待检测图像，确定与待检测图像上的多个区域对应的多个检测点，以及各检测点对应的区域存在目标物的概率值；从所有检测点中筛选出概率值最大的第一检测点，以及概率值小于第一检测点的概率值、且大于或等于概率阈值的至少一个第二检测点；其中，第一检测点对应的区域存在目标物；判断每个第二检测点与第一检测点之间的第一距离是否大于或等于距离阈值，当所述第一距离大于或等于所述距离阈值时，则将相应的第二检测点的原概率值进行更新，得到更新后的概率值；将更新后的概率值与概率阈值进行比较，得到比较结果，并根据比较结果确定第二检测点对应的区域是否存在新的目标物。

Description

目标检测方法、装置、计算机设备及可读存储介质
本申请要求于2020年05月15日提交的、申请号为202010414578.7的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本公开实施例涉及一种目标检测方法、目标检测装置、计算机设备及计算机可读存储介质。
背景技术
在计算机视觉领域，目标检测是基础且具有实际应用意义的研究方向。通过目标检测可以识别图像中存在的目标物(例如人物、动物、植物、交通工具等)，并对识别出的目标物进行标记。
发明内容
一方面，提供一种目标检测方法。所述目标检测方法包括：根据待检测图像，确定与所述待检测图像上的多个区域对应的多个检测点，以及各检测点对应的区域存在目标物的概率值；从所有检测点中筛选出概率值最大的第一检测点，以及概率值小于所述第一检测点的概率值、且大于或等于概率阈值的至少一个第二检测点；其中，所述第一检测点对应的区域存在目标物；判断每个第二检测点与所述第一检测点之间的第一距离是否大于或等于距离阈值，当所述第一距离大于或等于所述距离阈值时，则将相应的第二检测点的概率值进行更新，得到更新后的概率值；将更新后的概率值与所述概率阈值进行比较，得到比较结果，并根据所述比较结果确定所述第二检测点对应的区域是否存在新的目标物。
在一些实施例中，所述将相应的第二检测点的原概率值进行更新，得到更新后的概率值，包括：根据各所述第一距离确定相应的第二检测点的概率减小量，其中，所述概率减小量与所述第一距离呈正相关；将所述第二检测点的原概率值与所述概率减小量的差值作为所述第二检测点更新后的概率值。
在一些实施例中，所述将相应的第二检测点的原概率值进行更新，得到更新后的概率值，包括：将所述第二检测点的概率值作为因变量输入预设函数，以得到第一参数；其中，所述预设函数为单调递减函数，且位于平面直角坐标系的第一象限内；将所述第二检测点到所述第一检测点的所述第一距离与所述第一参数相加得到第二参数；将所述第二参数作为自变量输入所述预设函数，以得到所述第二检测点更新后的概率值。
在一些实施例中，所述预设函数为以下高斯函数位于平面直角坐标系的第一象限内的部分：
Figure PCTCN2021086077-appb-000001
其中，a、b与c为实数常数，且a>0，x为所述第二参数，f(x)为所述第二检测点更新后的概率值。
在一些实施例中，所述预设函数为以下一次函数位于平面直角坐标系的第一象限内的部分：
f(x)＝kx+b；
其中，k与b为实数常数，且k＜0，x为所述第二参数，f(x)为所述第二检测点更新后的概率值。
在一些实施例中，所述将更新后的概率值与所述概率阈值进行比较，得到比较结果，并根据所述比较结果确定所述第二检测点对应的区域是否存在新的目标物，包括：判断更新后的概率值是否大于或等于所述概率阈值；若是，则确定所述第二检测点对应的区域存在新的目标物；若否，则确定所述第二检测点对应的区域不存在新的目标物。
在一些实施例中，还包括：当所述第一距离小于所述距离阈值时，则检测出所述第二检测点对应的区域和所述第一检测点对应的区域存在同一目标物。
在一些实施例中，所述根据待检测图像，确定与所述待检测图像上的多个区域对应的多个检测点，以及各检测点对应的区域存在目标物的概率值，包括：将所述待检测图像输入训练好的目标网络模型进行特征提取，得到相应的特征信息，所述特征信息包括特征图以及该特征图上各检测点存在目标物的概率值；所述特征图的分辨率为所述待检测图像的分辨率的1/n倍，n＞1。
在一些实施例中，所述特征信息包括各检测点的位置补偿精度；所述目标检测方法还包括：根据所述第一检测点及其位置补偿精度，确定该第一检测点对应的目标物在所述待检测图像上的中心点；以及，根据对应的区域存在新目标物的所述第二检测点及其位置补偿精度，确定该第二检测点对应的新目标物在所述待检测图像上的第二中心点。
在一些实施例中，所述特征信息包括各目标物的回归尺寸；所述目标检测方法还包括：根据各目标物的所述回归尺寸，确定所检测到的各目标物在所述待检测图像上所覆盖的区域。
在一些实施例中，所述目标网络模型包括下采样模块、上采样模块和残 差模块；所述下采样模块配置为对所述待检测图像执行下采样操作，以得到所述特征图；所述上采样模块配置为对所述特征图执行上采样操作；所述残差模块配置为从所述待检测图像中提取多个输入图像特征，以使所述上采样模块能够结合所述多个输入图像特征对所述特征图执行所述上采样操作。
在一些实施例中，所述概率阈值为0.4～0.6。
另一方面，提供一种目标检测装置。所述目标检测装置包括检测点确定模块、筛选模块、判断更新模块和比较确定模块。检测点确定模块被配置为根据待检测图像，确定与所述待检测图像上的多个区域对应的多个检测点，以及各检测点对应的区域存在目标物的概率值。筛选模块被配置为从所有检测点中筛选出概率值最大的第一检测点，以及概率值小于所述第一检测点的概率值、且大于或等于概率阈值的至少一个第二检测点；其中，所述第一检测点对应的区域存在目标物。判断更新模块被配置为判断每个第二检测点与所述第一检测点之间的第一距离是否大于或等于距离阈值，当所述第一距离大于或等于所述距离阈值时，则将相应的第二检测点的原概率值进行更新，得到更新后的概率值。比较确定模块被配置为将更新后的概率值与所述概率阈值进行比较，得到比较结果，并根据所述比较结果确定所述第二检测点对应的区域是否存在新的目标物。
再一方面，提供一种计算机设备。所述计算机设备包括：存储器；处理器；以及，储存在所述存储器上并可在所述处理器上运行的计算机程序指令，所述处理器执行所述计算机程序指令时实现如上述任一项实施例所述的目标检测方法中的。
又一方面，提供一种计算机可读存储介质。所述计算机可读存储介质存储有计算机程序指令，所述计算机程序指令在处理器上运行时，使得所述处理器执行如上述任一实施例所述的目标检测方法中的。
又一方面，提供一种计算机程序产品。所述计算机程序产品包括计算机程序指令，在计算机上执行所述计算机程序指令时，所述计算机程序指令使计算机执行如上述任一实施例所述的目标检测方法。
又一方面，提供一种计算机程序。当所述计算机程序在计算机上执行时，所述计算机程序使计算机执行如上述任一实施例所述的目标检测方法。
附图说明
为了更清楚地说明本公开中的技术方案，下面将对本公开一些实施例中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本公开的一些实施例的附图，对于本领域普通技术人员来讲，还 可以根据这些附图获得其他的附图。此外，以下描述中的附图可以视作示意图，并非对本公开实施例所涉及的产品的实际尺寸、方法的实际流程、信号的实际时序等的限制。
图1为根据本公开一些实施例的一种目标检测方法的流程图；
图2为根据本公开一些实施例的一种目标网络模型；
图3为根据本公开一些实施例的待检测图像及其对应的特征图；
图4为根据本公开一些实施例的另一种目标网络模型；
图5为根据本公开一些实施例的目标网络模型中的残差模块；
图6为根据本公开一些实施例的一种特征图；
图7为根据本公开一些实施例的另一种目标检测方法的流程图；
图8为根据本公开一些实施例的再一种目标检测方法的流程图；
图9为根据本公开一些实施例的又一种目标检测方法的流程图；
图10为根据本公开一些实施例的又一种目标检测方法的流程图；
图11为根据本公开一些实施例的一种目标检测装置的框图；
图12为根据本公开一些实施例的另一种目标检测装置的框图；
图13为根据本公开一些实施例的一种计算机设备的框图。
具体实施方式
下面将结合附图，对本公开一些实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本公开一部分实施例，而不是全部的实施例。基于本公开所提供的实施例，本领域普通技术人员所获得的所有其他实施例，都属于本公开保护的范围。
除非上下文另有要求，否则，在整个说明书和权利要求书中，术语“包括(comprise)”及其其他形式例如第三人称单数形式“包括(comprises)”和现在分词形式“包括(comprising)”被解释为开放、包含的意思，即为“包含，但不限于”。在说明书的描述中，术语“一个实施例(one embodiment)”、“一些实施例(some embodiments)”、“示例性实施例(exemplary embodiments)”、“示例(example)”、“特定示例(specific example)”或“一些示例(some examples)”等旨在表明与该实施例或示例相关的特定特征、结构、材料或特性包括在本公开的至少一个实施例或示例中。上述术语的示意性表示不一定是指同一实施例或示例。此外，所述的特定特征、结构、材料或特点可以以任何适当方式包括在任何一个或多个实施例或示例中。
以下，术语“第一”、“第二”仅用于描述目的，而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量。由此，限定有“第 一”、“第二”的特征可以明示或者隐含地包括一个或者更多个该特征。在本公开实施例的描述中，除非另有说明，“多个”的含义是两个或两个以上。
在描述一些实施例时，可能使用了“耦接”和“连接”及其衍伸的表达。例如，描述一些实施例时可能使用了术语“连接”以表明两个或两个以上部件彼此间有直接物理接触或电接触。又如，描述一些实施例时可能使用了术语“耦接”以表明两个或两个以上部件有直接物理接触或电接触。然而，术语“耦接”或“通信耦合(communicatively coupled)”也可能指两个或两个以上部件彼此间并无直接接触，但仍彼此协作或相互作用。这里所公开的实施例并不必然限制于本文内容。
本文中“适用于”或“被配置为”的使用意味着开放和包容性的语言，其不排除适用于或被配置为执行额外任务或步骤的设备。
本文参照作为理想化示例性附图的剖视图和/或平面图描述了示例性实施方式。在附图中，为了清楚，放大了层和区域的厚度。因此，可设想到由于例如制造技术和/或公差引起的相对于附图的形状的变动。因此，示例性实施方式不应解释为局限于本文示出的区域的形状，而是包括因例如制造而引起的形状偏差。例如，示为矩形的蚀刻区域通常将具有弯曲的特征。因此，附图中所示的区域本质上是示意性的，且它们的形状并非旨在示出设备的区域的实际形状，并且并非旨在限制示例性实施方式的范围。
图像的目标检测是计算机视觉领域最为常见和具有实际应用意义的一个研究方向。近年来，卷积神经网络在图像处理领域中，尤其是目标的检测识别方面，表现出了极大的优势。然而，相关技术中的算法都需要在算法中预先设定锚框。这种方法需要待检测目标的先验知识，增加了算法中需要设定的超参数，并且会产生大量的无效待检锚框，浪费计算资源。同时，在对未知的有重叠的目标物进行检测时，相关技术中的方式是直接去除目标物附近概率值较大的潜在目标点，这使得在含有密集目标的场景中的检测能力较差。
基于此，本公开一些实施例提供一种目标检测方法，参见图1，该目标检测方法包括步骤101～步骤105。
步骤101、根据待检测图像，确定与待检测图像上的多个区域对应的多个检测点，以及各检测点对应的区域存在目标物的概率值。
例如，可以将待检测图像输入训练好的目标网络模型中进行特征提取，得到相应的特征信息，该特征信息包括目标特征图以及该目标特征图上各检测点存在目标物的概率值。其中，该目标特征图的分辨率为该待检测图像的分辨率的1/n倍，n＞1。
例如，该目标网络模型为hourglass网络模型。
示例性的，如图2所示，该目标网络模型包括下采样模块11和上采样模块12，其中，一个下采样模块11和一个上采样模块12组成一个沙漏结构1。
该下采样模块11可以配置成对该待检测图像A进行下采样操作，以得到特征图。在一些示例中，该下采样操作可以是以均匀的间隔从该待检测图像A中抽取数据，从而降低该待检测图像A的尺寸。例如，如果执行2倍的下采样操作，可以将该待检测图像A的尺寸从4×4减小为2×2(即此时得到的特征图的分辨率为待检测图像A的1/2倍)。如果执行4倍的下采样，可以将该待检测图像A的尺寸从8×8减小为2×2(即此时得到的特征图的分辨率为待检测图像A的1/4倍)。如图3所示，以2倍下采样为例，下采样操作可以将该待检测图像A中2×2的区域01内的所有像素映射成特征图A'中的一个像素02，例如可以将该区域内的所有像素的均值作为下采样后图像中该一个像素的像素值。而在特征图A′中，每个像素02对应一个检测点。
该上采样模块12可以配置成对所得到的特征图执行上采样操作。在一些示例中，该上采样操作可以是以均匀地间隔对该特征图进行插值。例如，如果执行2倍的上采样操作，可以将图像特征的尺寸从2×2增加为4×4(即此时可以将特征图较好地还原为待检测图像A)。如果执行4倍的上采样操作，可以将该特征图的尺寸从2×2增加为8×8(即此时可以将特征图较好地还原为待检测图像A)。即，上采样操作可以包括对图像执行插值操作，例如，邻插值(如双线性插值、双三次插值、样条插值等)、基于边缘的插值和/或基于区域的插值。
在此基础上，示例性的，该目标网络模型还包括残差模块13。例如图4示出了该目标网络模型的另一种结构形式，图5示出了图4中的一个残差模块。通过该残差模块13可以从待检测图像中提取多个输入图像特征，而上采样模块12能够结合所述多个输入图像特征对上述特征图执行上采样操作，从而可以将特征图更好的还原为待检测图像A。
需要说明的是，上述示出的尺寸可以不是其实际尺寸，而仅用于代表待检测图像A与特征图之间的尺寸的比例关系。例如，在输入的待检测图像A的尺寸为1024×1024时，执行2倍下采样操作后得到的特征图的尺寸可以是512×512(即此时得到的特征图的分辨率为待检测图像A的1/2倍)，执行4倍下采样操作后得到的特征图的尺寸可以是256×256(即此时得到的特征图的分辨率为待检测图像A的1/4倍)。即，通过该下采样模块11后得到的特征图的分辨率可以是该待检测图像A的分辨率的1/n倍，n＞1。
参见图2，将待检测图像A输入该目标网络模型后，可以通过多个沙漏结构1可以执行多次下采样操作和上采样操作，而且由于每次执行下采样操作得到特征图后，都通过上采样操作将该特征图较好的还原为待检测图像A，因此，可以得到多种尺寸不同的特征图。在此基础上，该目标网络模型可以对所得到的一个或多个特征图进行处理，进而实现输出上述特征信息(该特征信息包括目标特征图以及该目标特征图上各检测点存在目标物的概率值)。
该目标特征图的分辨率为待检测图像的分辨率的1/n倍，n＞1，因此，该目标特征图上每个检测点可以对应待检测图像上的一块区域，例如一块2×2区域，或者，一块4×4区域。
该目标网络模型可以输出该目标特征图上(i，j)位置处的检测点(也即待检测图像上与该检测点对应的区域)出现第c类目标的概率值P (i，j)。即，当P (i，j)＝1时，则表示(i，j)位置处的检测点对应的区域存在第c类目标；当P (i，j)＝0时，则表示(i，j)位置处的检测点对应的区域不存在第c类目标；当0＜P (i，j)＜1时，则表示(i，j)位置处的检测点对应的区域可能存在第c类目标。
步骤102、从所有检测点中筛选出概率值最大的第一检测点，以及概率值小于第一检测点的概率值、且大于或等于概率阈值的至少一个第二检测点；其中，第一检测点对应的区域存在目标物。
图6示出了一种目标特征图。在一些示例中，上述特征信息包括该目标特征图的宽度和高度信息，并且可以显示该目标特征图中目标类别的数量。
如图6所示，第一检测点T1存在目标物(即北极熊)的概率值为1；第二检测点为T2存在目标物(即北极熊)的概率值大于0且小于1。即，通过将待检测图像输入训练好的目标网络模型，可以直接检测出第一检测点T1对应的区域存在目标物(例如图6中的两个第一检测点T1对应的区域分别存在一只北极熊)。并且，在一些示例中，上述特征信息还包括各检测点的位置补偿精度。通过该训练好的目标网络模型，可以根据该第一检测点T1及其位置补偿精度，确定该第一检测点T1对应的目标物在所述待检测图像A上的中心点。在此基础上，示例性的，上述特征信息还包括各目标物的回归尺寸。首先可以由该第一检测点T1周围的图像特征(例如周围检测点的概率值)得到目标物在目标特征图上的宽度(例如左侧的第一检测点T1对应的目标物的宽度为W)和高度(例如左侧的第一检测点T1对应的目标物的高度为H)，然后根据该回归尺寸，可以确定该目标物在待检测图像上所覆盖的区域，从而可以准确的检测出目标物的位置。此外，由于本实施例中，直接通过目标 特征图上的多个检测点检测出目标物，因此，采用本实施例的目标检测方法，无需预先设置锚框即可实现检测图像中的目标物，从而不需要设置待检测目标物的先验知识，不需要在算法中设定的超参数，不会产生大量的无效待检锚框，节省了计算资源。
值得指出的是，相关技术中，在检测重叠的目标物(比如检测图5中位于左侧的第一检测点T1和第二检测点T2对应的两只北极熊)时，由于第二检测点T2与该左侧的第一检测点T1之间的距离比较近，且第二检测点T2的概率值小于该左侧的第一检测点T1的概率值，导致第二检测点T2对应的目标物不能被检测到，即，相关技术中，对于该左侧的第一检测点T1和第二检测点T2，只能检测到一只北极熊，从而难以有效的对密集目标进行检测。
步骤103、判断每个第二检测点与第一检测点之间的第一距离是否大于或等于距离阈值，当所述第一距离大于或等于所述距离阈值时，则执行步骤104。
步骤104、将相应的第二检测点的原概率值进行更新，得到更新后的概率值。
步骤105、将更新后的概率值与概率阈值进行比较，得到比较结果，并根据比较结果确定第二检测点对应的区域是否存在新的目标物。
其中，该距离阈值可以根据应用场景进行设置，即不同的应用场景(例如十字路口对应的待检测图像，和学校门口对应的待检测图像)，对应的距离阈值也不同。本实施例对此不做限定。
本实施例中，若判断出第二检测点T2与第一检测点T1之间的距离大于或等于距离阈值，则认为在第二检测点T1对应的区域可能存在新的目标物。此时，通过执行步骤104和步骤105，将第二检测点T2的原概率值进行更新，得到更新后的概率值，并将更新后的概率值与概率阈值进行比较，得到比较结果，可以根据比较结果确定第二检测点对应的区域是否存在新的目标物。因此相较相关技术而言，本实施例可以更有效的对密集目标进行检测。
示例性的，该概率阈值为0.4～0.6。例如，该概率阈值可以为0.5。
在一些实施例中，如图7所示，上述步骤105包括步骤1051～步骤1053。
步骤1051、判断更新后的概率值是否大于或等于概率阈值；若是，则执行步骤1052、确定第二检测点对应的区域存在新的目标物；若否，则执行步骤1053、确定第二检测点对应的区域不存在新的目标物。
在此基础上，在一些实施例中，上述步骤104包括步骤104A和步骤104B。
如图8所示，步骤104A、根据各所述第一距离确定相应的第二检测点的概率减小量，其中，所述概率减小量与所述第一距离呈正相关。其中，每个 第一距离都有其对应的概率减小量。
步骤104B、将所述第二检测点的原概率值与所述概率减小量的差值作为所述第二检测点更新后的概率值。
本实施例中，由于概率减小量与第一距离呈正相关，可以使第二检测点T2更新后的概率值小于该第二检测点T2更新前的概率值。样有利于减小后续数据处理量，提高运算效率。同时由于第二检测点T2更新后的概率值小于该第二检测点T2更新前的概率值，能够筛选出更新前概率值较大的第二检测点，从而能够更准确的实现相邻目标的检测。同时，对于更新前的概率值相同、但到第一检测点的距离不同的第二检测点而言，到该第一检测点的距离越远的第二检测点，更新后的概率值相较更新前的概率值降低的幅度越大，这样能够有效的减小后续数据处理量，提高运算效率，同时能够更准确的实现相邻目标的检测识别。
在此基础上，在另一些实施例中，上述步骤104包括步骤1041～步骤1043。
如图9所示，步骤1041、将第二检测点的原概率值作为因变量输入预设函数，以得到第一参数；其中，预设函数为单调递减函数，且位于平面直角坐标系的第一象限内；
步骤1042、将第二检测点到第一检测点的第一距离与第一参数相加得到第二参数；
步骤1043、将第二参数作为自变量输入预设函数，以得到第二检测点更新后的概率值。
本实施例中，对于更新前的概率值相同，但距离第一检测点的距离不同的第二检测点而言，距离第一检测点的距离越远的第二检测点(即第一距离越大的第二检测点)，更新后的概率值相较更新前的概率值降低的幅度越大，这样能够有效的减小后续数据处理量，提高运算效率，同时能够更准确的实现相邻目标的检测识别。
需要说明的是，该单调递减函数的形式有多种，例如，本公开包括但不限于以下示出的一些示例。
在一些示例中，该预设函数为以下高斯函数位于平面直角坐标系的第一象限内的部分：
Figure PCTCN2021086077-appb-000002
其中，a、b与c为实数常数，且a>0，x为上述第二参数，f(x)为第二检测点更新后的概率值。
本示例中，高斯函数位于平面直角坐标系的第一象限内的部分为单调递减函数，而且根据高斯函数可知，第二参数越大，所得到的更新后的概率值则越小，也即，对于更新前的概率值相同，但距离第一检测点的距离不同的第二检测点而言，该第二检测点到该第一检测点的距离值越大，该第二检测点更新后的概率值相较更新前的概率值降低的幅度也就越大，因此可以更准确的实现相邻目标的检测识别。
在另一些示例中，该预设函数为以下一次函数位于平面直角坐标系的第一象限内的部分：
f(x)＝kx+b；
其中，k与b为实数常数，且k＜0，x为上述第二参数，f(x)为第二检测点更新后的概率值。
本示例中，该一次函数位于平面直角坐标系的第一象限内的部分为单调递减函数，而且根据该一次函数可知，第二参数越大，所得到的更新后的概率值则越小，也即第二检测点到第一检测点的距离值越大，所得到的更新后的概率值相较更新前的概率值降低的幅度越大，因此可以更准确的实现相邻目标的检测识别。
在一些实施例中，参见图10，该目标检测方法还包括：若判断出第二检测点与第一位检测点之间的距离小于距离阈值，则执行步骤106。
步骤106、确定第二检测点对应的区域和第一检测点对应的区域存在同一目标物。例如，参见图6，如果第二检测点T2落在第一检测点对应的覆盖左侧北极熊矩形区域内，则可以认为两者对应同一目标物。
另一方面，本公开一些实施例可以根据上述方法示例对目标检测装置进行功能模块的划分，例如，可以对应各个功能划分各个功能模块，也可以将两个或两个以上的功能集成在一个处理模块中。上述集成的模块既可以采用硬件的形式实现，也可以采用软件功能部件的形式实现。需要说明的是，本公开一些实施例中对模块的划分是示意性的，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式。
在采用对应各个功能划分各个功能模块的情况下，图11示出了上述实施例中涉及的目标检测装置的一种可能的组成示意图，如图11所示，该目标检测装置200包括：检测点确定模块21、筛选模块22、判断更新模块23和比较确定模块24。
其中，检测点确定模块21，被配置为根据待检测图像，确定与待检测图像上的多个区域对应的多个检测点，以及各检测点对应的区域存在目标物的 概率值。也即，该检测点确定模块21可以支持该目标显示装置200执行上述步骤101。
筛选模块22，被配置为从所有检测点中筛选出概率值最大的第一检测点，以及概率值小于第一检测点的概率值、且大于或等于概率阈值的至少一个第二检测点；其中，第一检测点对应的区域存在目标物。也即，该筛选模块22可以支持该目标显示装置200执行上述步骤102。
判断更新模块23，被配置为判断每个第二检测点与第一检测点之间的距离是否大于或等于距离阈值，若是，则将相应的第二检测点的概率值进行更新，得到更新后的概率值。也即，该判断更新模块23可以支持该目标显示装置200执行上述步骤103和步骤104。
比较确定模块24，被配置为将更新后的概率值与概率阈值进行比较，得到比较结果，并根据比较结果确定第二检测点对应的区域是否存在新的目标物。也即，该比较确定模块24可以支持该目标显示装置200执行上述步骤105。
需要说明的是，上述方法实施例涉及的各步骤的所有相关内容均可以援引到对应功能部件的功能描述，在此不再赘述。
本公开一些实施例提供的目标检测装置，被配置为执行上述目标检测方法，因此能够达到与上述目标检测方法相同的效果。
在采用集成的控制部件的情况下，图12示出了上述实施例中所涉及的目标检测装置的另一种可能的组成示意图。如图12所示，该目标检测装置200包括：处理部件31、通信部件32和存储部件33。
处理部件31被配置为对目标检测装置的动作进行控制管理，例如，处理部件31被配置为支持目标检测装置执行图1中的步骤101～步骤105，和/或被配置为本文所描述的技术的其它过程。通信部件32被配置为支持目标检测装置与其他网络实体的通信。存储部件33，被配置为存储目标检测装置的程序代码和数据。
示例性的，处理部件31为处理器。其可以实现或执行结合本公开公开内容所描述的各种示例性的逻辑方框，部件和电路。处理器也可以是实现计算功能的组合，例如包含一个或多个微处理器组合，数字信号处理器(Digital Signal Processor，DSP)和微处理器的组合等等。通信部件32可以是通信接口。存储部件33可以是存储器。
在此基础上，本公开一些实施例还提供一种显示装置，该显示装置包括上述任一实施例所述的目标检测装置。
示例性的，该显示装置可以为AR头盔、AR眼镜、手机、平板电脑、电 视机、显示器、笔记本电脑、数码相框、导航仪等任何具有显示功能的产品或部件。
本公开一些实施例提供的显示装置，可以通过目标检测装置执行上述目标检测方法，因此能够达到与上述目标检测方法相同的效果。
本公开实施例还提供一种计算机设备，如图13所示，该计算机设备300包括：存储器42；处理器41；以及，储存在存储器42上并可在处理器41上运行的计算机程序指令43，处理器41执行计算机程序指令43时实现上述目标检测方法中的一个或多个步骤。
本公开的一些实施例提供了一种计算机可读存储介质(例如，非暂态计算机可读存储介质)，该计算机可读存储介质中存储有计算机程序指令，计算机程序指令在处理器上运行时，使得处理器执行如上述实施例中任一实施例所述的目标检测方法中的一个或多个步骤。
示例性的，上述计算机可读存储介质可以包括，但不限于：磁存储器件(例如，硬盘、软盘或磁带等)，光盘(例如，CD(Compact Disk，压缩盘)、DVD(Digital Versatile Disk，数字通用盘)等)，智能卡和闪存器件(例如，EPROM(Erasable Programmable Read-Only Memory，可擦写可编程只读存储器)、卡、棒或钥匙驱动器等)。本公开描述的各种计算机可读存储介质可代表用于存储信息的一个或多个设备和/或其它机器可读存储介质。术语“机器可读存储介质”可包括但不限于，无线信道和能够存储、包含和/或承载指令和/或数据的各种其它介质。
本公开的一些实施例还提供了一种计算机程序产品。该计算机程序产品包括计算机程序指令，在计算机上执行该计算机程序指令时，该计算机程序指令使计算机执行如上述实施例所述的目标检测方法中的一个或多个步骤。
本公开的一些实施例还提供了一种计算机程序。当该计算机程序在计算机上执行时，该计算机程序使计算机执行如上述实施例所述的目标检测方法中的一个或多个步骤。
上述计算机设备、计算机可读存储介质、计算机程序产品及计算机程序的有益效果和上述一些实施例所述的目标检测方法的有益效果相同，此处不再赘述。
以上所述，仅为本公开的具体实施方式，但本公开的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本公开揭露的技术范围内，想到变化或替换，都应涵盖在本公开的保护范围之内。因此，本公开的保护范围应以所述权利要求的保护范围为准。

Claims (15)

  1. 一种目标检测方法，包括：
    根据待检测图像，确定与所述待检测图像上的多个区域对应的多个检测点，以及各检测点对应的区域存在目标物的概率值；
    从所有检测点中筛选出概率值最大的第一检测点，以及概率值小于所述第一检测点的概率值、且大于或等于概率阈值的至少一个第二检测点；其中，所述第一检测点对应的区域存在目标物；
    判断每个第二检测点与所述第一检测点之间的第一距离是否大于或等于距离阈值，当所述第一距离大于或等于所述距离阈值时，将相应的第二检测点的原概率值进行更新，得到更新后的概率值；
    将更新后的概率值与所述概率阈值进行比较，得到比较结果，并根据所述比较结果确定所述第二检测点对应的区域是否存在新的目标物。
  2. 根据权利要求1所述的目标检测方法，其中，所述将相应的第二检测点的原概率值进行更新，得到更新后的概率值，包括：
    根据各所述第一距离确定相应的第二检测点的概率减小量，其中，所述概率减小量与所述第一距离呈正相关；
    将所述第二检测点的原概率值与所述概率减小量的差值作为所述第二检测点更新后的概率值。
  3. 根据权利要求1所述的目标检测方法，其中，所述将相应的第二检测点的原概率值进行更新，得到更新后的概率值，包括：
    将所述第二检测点的原概率值作为因变量输入预设函数，以得到第一参数；其中，所述预设函数为单调递减函数，且位于平面直角坐标系的第一象限内；
    将所述第二检测点到所述第一检测点的所述第一距离与所述第一参数相加得到第二参数；
    将所述第二参数作为自变量输入所述预设函数，以得到所述第二检测点更新后的概率值。
  4. 根据权利要求3所述的目标检测方法，其中，
    所述预设函数为以下高斯函数位于平面直角坐标系的第一象限内的部分：
    Figure PCTCN2021086077-appb-100001
    其中，a、b与c为实数常数，且a>0，x为所述第二参数，f(x)为所述第二检测点更新后的概率值。
  5. 根据权利要求3所述的目标检测方法，其中，
    所述预设函数为以下一次函数位于平面直角坐标系的第一象限内的部分：
    f(x)＝kx+b；
    其中，k与b为实数常数，且k＜0，x为所述第二参数，f(x)为所述第二检测点更新后的概率值。
  6. 根据权利要求1～5中任一项所述的目标检测方法，其中，所述将更新后的概率值与所述概率阈值进行比较，得到比较结果，并根据所述比较结果确定所述第二检测点对应的区域是否存在新的目标物，包括：
    判断更新后的概率值是否大于或等于所述概率阈值；若是，则确定所述第二检测点对应的区域存在新的目标物；若否，则确定所述第二检测点对应的区域不存在新的目标物。
  7. 根据权利要求1～6中任一项所述的目标检测方法，还包括：
    当所述第一距离小于所述距离阈值时，则检测出所述第二检测点对应的区域和所述第一检测点对应的区域存在同一目标物。
  8. 根据权利要求1～7中任一项所述的目标检测方法，其中，所述根据待检测图像，确定与所述待检测图像上的多个区域对应的多个检测点，以及各检测点对应的区域存在目标物的概率值，包括：
    将所述待检测图像输入训练好的目标网络模型进行特征提取，得到相应的特征信息，所述特征信息包括目标特征图以及该目标特征图上各检测点存在目标物的概率值；
    所述目标特征图的分辨率为所述待检测图像的分辨率的1/n倍，n＞1。
  9. 根据权利要求8所述的目标检测方法，其中，所述特征信息包括各检测点的位置补偿精度；
    所述目标检测方法还包括：
    根据所述第一检测点及其位置补偿精度，确定该第一检测点对应的目标物在所述待检测图像上的中心点；以及，
    根据对应的区域存在新目标物的所述第二检测点及其位置补偿精度，确定该第二检测点对应的新目标物在所述待检测图像上的第二中心点。
  10. 根据权利要求9所述的目标检测方法，其中，所述特征信息包括各目标物的回归尺寸；
    所述目标检测方法还包括：
    根据各目标物的所述回归尺寸，确定所检测到的各目标物在所述待检测 图像上所覆盖的区域。
  11. 根据权利要求8～10中任一项所述的目标检测方法，其中，所述目标网络模型包括下采样模块、上采样模块和残差模块；
    所述下采样模块配置为对所述待检测图像执行下采样操作，以得到特征图；
    所述上采样模块配置为对所述特征图执行上采样操作；
    所述残差模块配置为从所述待检测图像中提取多个输入图像特征，以使所述上采样模块能够结合所述多个输入图像特征对所述特征图执行所述上采样操作。
  12. 根据权利要求1～11中任一项所述的目标检测方法，其中，
    所述概率阈值为0.4～0.6。
  13. 一种目标检测装置，包括：
    检测点确定模块，被配置为根据待检测图像，确定与所述待检测图像上的多个区域对应的多个检测点，以及各检测点对应的区域存在目标物的概率值；
    筛选模块，被配置为从所有检测点中筛选出概率值最大的第一检测点，以及概率值小于所述第一检测点的概率值、且大于或等于概率阈值的至少一个第二检测点；其中，所述第一检测点对应的区域存在目标物；
    判断更新模块，被配置为判断每个第二检测点与所述第一检测点之间的第一距离是否大于或等于距离阈值，当所述第一距离大于或等于所述距离阈值时，则将相应的第二检测点的原概率值进行更新，得到更新后的概率值；
    比较确定模块，被配置为将更新后的概率值与所述概率阈值进行比较，得到比较结果，并根据所述比较结果确定所述第二检测点对应的区域是否存在新的目标物。
  14. 一种计算机设备，包括：
    存储器；
    处理器；以及，
    储存在所述存储器上并可在所述处理器上运行的计算机程序指令，所述处理器执行所述计算机程序指令时实现如权利要求1～12中任一项所述的目标检测方法。
  15. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序指令，所述计算机程序指令在处理器上运行时，使得所述处理器执行如权利要求1～12中任一项所述的目标检测方法。
PCT/CN2021/086077 2020-05-15 2021-04-09 目标检测方法、装置、计算机设备及可读存储介质 WO2021227723A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/765,366 US20220351413A1 (en) 2020-05-15 2021-04-09 Target detection method, computer device and non-transitory readable storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010414578.7A CN111598088B (zh) 2020-05-15 2020-05-15 目标检测方法、装置、计算机设备及可读存储介质
CN202010414578.7 2020-05-15

Publications (1)

Publication Number Publication Date
WO2021227723A1 true WO2021227723A1 (zh) 2021-11-18

Family

ID=72191476

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/086077 WO2021227723A1 (zh) 2020-05-15 2021-04-09 目标检测方法、装置、计算机设备及可读存储介质

Country Status (3)

Country Link
US (1) US20220351413A1 (zh)
CN (1) CN111598088B (zh)
WO (1) WO2021227723A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111598088B (zh) * 2020-05-15 2023-12-29 京东方科技集团股份有限公司 目标检测方法、装置、计算机设备及可读存储介质
CN112505304B (zh) * 2020-12-11 2023-10-03 成都北方石油勘探开发技术有限公司 一种含水量的检测方法及系统
CN112817716B (zh) * 2021-01-28 2024-02-09 厦门树冠科技有限公司 一种视觉检测处理方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108764325A (zh) * 2018-05-23 2018-11-06 腾讯科技（深圳）有限公司 图像识别方法、装置、计算机设备和存储介质
CN109344789A (zh) * 2018-10-16 2019-02-15 北京旷视科技有限公司 人脸跟踪方法及装置
CN109815868A (zh) * 2019-01-15 2019-05-28 腾讯科技（深圳）有限公司 一种图像目标检测方法、装置及存储介质
CN111598088A (zh) * 2020-05-15 2020-08-28 京东方科技集团股份有限公司 目标检测方法、装置、计算机设备及可读存储介质

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10571560B2 (en) * 2015-09-21 2020-02-25 Saab Ab Detecting objects in images
US9727800B2 (en) * 2015-09-25 2017-08-08 Qualcomm Incorporated Optimized object detection
CN109784290B (zh) * 2019-01-23 2021-03-05 科大讯飞股份有限公司 一种目标检测方法、装置、设备及可读存储介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108764325A (zh) * 2018-05-23 2018-11-06 腾讯科技（深圳）有限公司 图像识别方法、装置、计算机设备和存储介质
CN109344789A (zh) * 2018-10-16 2019-02-15 北京旷视科技有限公司 人脸跟踪方法及装置
CN109815868A (zh) * 2019-01-15 2019-05-28 腾讯科技（深圳）有限公司 一种图像目标检测方法、装置及存储介质
CN111598088A (zh) * 2020-05-15 2020-08-28 京东方科技集团股份有限公司 目标检测方法、装置、计算机设备及可读存储介质

Also Published As

Publication number Publication date
US20220351413A1 (en) 2022-11-03
CN111598088A (zh) 2020-08-28
CN111598088B (zh) 2023-12-29

Similar Documents

Publication Publication Date Title
US9697416B2 (en) Object detection using cascaded convolutional neural networks
WO2021227723A1 (zh) 目标检测方法、装置、计算机设备及可读存储介质
US11321593B2 (en) Method and apparatus for detecting object, method and apparatus for training neural network, and electronic device
US10452893B2 (en) Method, terminal, and storage medium for tracking facial critical area
WO2018103608A1 (zh) 一种文字检测方法、装置及存储介质
WO2020252917A1 (zh) 一种模糊人脸图像识别方法、装置、终端设备及介质
TW202011733A (zh) 對影像進行目標取樣的方法及裝置
CN110866871A (zh) 文本图像矫正方法、装置、计算机设备及存储介质
US11790499B2 (en) Certificate image extraction method and terminal device
CN108230354B (zh) 目标跟踪、网络训练方法、装置、电子设备和存储介质
CN108986152B (zh) 一种基于差分图像的异物检测方法及装置
WO2017054442A1 (zh) 一种图像信息识别处理方法及装置、计算机存储介质
WO2020082731A1 (zh) 电子装置、证件识别方法及存储介质
Xiao et al. Defocus blur detection based on multiscale SVD fusion in gradient domain
CN112509003B (zh) 解决目标跟踪框漂移的方法及系统
US20150220804A1 (en) Image processor with edge selection functionality
CN114283343B (zh) 基于遥感卫星图像的地图更新方法、训练方法和设备
CN115660945A (zh) 一种坐标转换方法、装置、电子设备及存储介质
CN114898154A (zh) 一种增量式目标检测方法、装置、设备及介质
JP2016053763A (ja) 画像処理装置、画像処理方法及びプログラム
CN116486126A (zh) 一种模版确定方法、装置、设备及存储介质
JP4387889B2 (ja) テンプレート照合装置および方法
JP2016045538A (ja) 情報処理装置、画像判定方法、及びプログラム
CN115205202A (zh) 视频检测方法、装置、设备及存储介质
CN114359352A (zh) 图像处理方法、装置、设备、存储介质及计算机程序产品

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: 21804493

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21804493

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 21804493

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 29.06.2023)