WO2022237902A1 - 对象检测方法、装置、设备及计算机存储介质 - Google Patents

对象检测方法、装置、设备及计算机存储介质 Download PDF

Info

Publication number
WO2022237902A1
WO2022237902A1 PCT/CN2022/092842 CN2022092842W WO2022237902A1 WO 2022237902 A1 WO2022237902 A1 WO 2022237902A1 CN 2022092842 W CN2022092842 W CN 2022092842W WO 2022237902 A1 WO2022237902 A1 WO 2022237902A1
Authority
WO
WIPO (PCT)
Prior art keywords
detection
detection frame
confidence
frame
degree
Prior art date
Application number
PCT/CN2022/092842
Other languages
English (en)
French (fr)
Inventor
谢伟
宇军
李逸
胡荣东
黄倩倩
Original Assignee
长沙智能驾驶研究院有限公司
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 长沙智能驾驶研究院有限公司 filed Critical 长沙智能驾驶研究院有限公司
Publication of WO2022237902A1 publication Critical patent/WO2022237902A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features

Definitions

  • the present application relates to the technical field of computer vision, and in particular to an object detection method, device, equipment and computer storage medium.
  • Computer vision recognition technology mainly detects objects through deep learning detection algorithms, and then uses tracking algorithms for object statistics. Compared with traditional object detection and statistical methods, computer vision recognition technology has the advantages of good stability and less damage.
  • Embodiments of the present application provide an object detection method, device, equipment, and computer storage medium, which can solve the technical problems of low object detection accuracy and easy occurrence of missed detection.
  • the embodiment of the present application provides an object detection method, and the object detection method includes:
  • the acquisition time of the second frame image is before the acquisition time of the first frame image, the first detection frame and the second detection frame
  • Both detection frames carry characteristic parameters, and the characteristic parameters include at least confidence
  • the detection object is determined according to the first detection frame after the confidence level is updated.
  • the embodiment of the present application provides an object detection device, and the object detection device includes:
  • the information acquisition module is used to obtain the first detection frame of the detected object in the first frame image and the second detection frame of the detected object in the second frame image; the acquisition time of the second frame image is before the acquisition time of the first frame image, Both the first detection frame and the second detection frame carry feature parameters, and the feature parameters include at least confidence;
  • a similarity determination module configured to determine at least one type of similarity information between the first detection frame and the second detection frame according to the characteristic parameters of the first detection frame and the characteristic parameters of the second detection frame;
  • An information fusion module configured to fuse at least one similarity parameter and feature parameter to obtain a fusion result
  • An information update module configured to update the confidence of the first detection frame according to the fusion result
  • the target determination module is configured to determine the detection object according to the first detection frame after the confidence level is updated.
  • the embodiment of the present application provides an object detection device, and the object detection device includes: a processor, and a memory storing computer program instructions;
  • the processor reads and executes computer program instructions to implement the object detection method provided in the first aspect of the embodiments of the present application.
  • the embodiments of the present application provide a computer storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the object detection method provided in the first aspect of the embodiments of the present application is implemented.
  • two frames of images before and after are taken into consideration to realize object detection.
  • the obtained Fusion of at least one kind of similarity information and feature parameters can realize the fusion of multiple information of two frame image detection frames, based on the fusion result, the confidence of the first detection frame can be updated, and the object can be determined according to the updated first detection frame ;
  • the technical solution of the embodiment of the present application can simultaneously refer to the feature parameters of the detection frames of the two frames of images before and after and the detection frames of the two frames of images when detecting the object
  • the referenced information is more comprehensive, when the object detection is performed based on the fusion result after the more comprehensive information is fused, the accuracy of object detection can be improved and missed detection can be reduced.
  • FIG. 1 is a schematic flow diagram of an object detection method provided by an embodiment of the present application.
  • FIG. 2 is a schematic flowchart of a part of an object detection method provided by another embodiment of the present application.
  • Fig. 3 is a schematic diagram of the principle of calculating the overlapping degree of different detection frames in an embodiment of the present application
  • Fig. 4 is a schematic flowchart of a part of an object detection method provided by another embodiment of the present application.
  • Fig. 5 is a schematic diagram of a detection frame matching situation involved in the related art
  • Fig. 6 is a schematic diagram of another detection frame matching situation involved in the related art.
  • Fig. 7 is a schematic diagram of the tracking sequence of tracking the same detection object and a schematic diagram of the principle of updating the confidence degree of the first detection frame in an embodiment of the present application;
  • Fig. 8 is a schematic structural frame diagram of an object detection device provided by an embodiment of the present application.
  • Fig. 9 is a schematic diagram of a structural framework of an object detection device provided by an embodiment of the present application.
  • the object detection method may include:
  • the acquisition time of the second frame of image is before the acquisition time of the first frame of image, and both the first detection frame and the second detection frame carry feature parameters, and the feature parameters include at least confidence.
  • S120 Determine at least one type of similarity information between the first detection frame and the second detection frame according to the characteristic parameters of the first detection frame and the characteristic parameters of the second detection frame.
  • two frames of images before and after are taken into consideration to realize object detection.
  • the obtained Fusion of at least one kind of similarity information and feature parameters can realize the fusion of multiple information of two frame image detection frames, based on the fusion result, the confidence degree of the first detection frame can be updated, and the detection can be determined according to the updated first detection frame Object; compared to the situation where only one frame of image or only the confidence of the detection frame is considered, the technical solution of the embodiment of the present application can simultaneously refer to the characteristic parameters of the detection frame of the two frames of images before and after and the detection of the two frames of images when detecting the object At least one kind of similarity information between frames, the referenced information is more comprehensive, when the object detection is performed based on the fusion result after the more comprehensive information is fused, the accuracy of object detection can be improved and missed detection can be reduced.
  • the frame number between the first frame image and the second frame image in the embodiment of the present application may be less than the preset frame number, and the preset frame number can be set according to actual needs and experience values .
  • the preset frame number in order to make the second frame image and the first frame image closer in time sequence, so that the second detection frame in the second frame image is more referential, the preset frame number can be set to a smaller The value of , for example, 2 frames, at this time, there is a frame interval between the second frame image and the first frame image, that is, the second frame image is the previous frame image of the first frame image.
  • step S110 before obtaining the first detection frame of the detected object in the first frame image and the second detection frame of the detected object in the second frame image, it may include Steps S111-S113 are as follows.
  • the target detection model of this application can be any existing neural network that can realize target detection.
  • the target detection model can be any one of a two-stage detection algorithm and a single-stage detection algorithm
  • the two-stage detection algorithm can be R-CNN (Region-Convolutional Neural Network, regional convolutional neural network), Fast R-CNN (fast regional convolutional neural network) series and any other convolutional neural network detection algorithm
  • the single-stage detection algorithm can be SSD (Single Shot MultiBox Detector, single-stage multi-frame prediction algorithm), YOLO (You Only Look Once) series and other convolutional neural network detection algorithms
  • the initial detection frame output by the target detection model carries position information and confidence.
  • the detection object in this embodiment of the present application may be any specified object, for example, a human head, an animal, or other objects in a certain place.
  • the first confidence threshold is smaller than a preset value, and the preset value can be set according to actual needs or experience values, and the first confidence threshold can be set according to actual needs or experience values on the basis of being smaller than the preset value.
  • the preset value can be set to a small value, such as 1 ⁇ 10 -4 or 1 ⁇ 10 -1 , so that the first
  • the confidence threshold is a small value, such as 0.5 ⁇ 10 -4 ; at this time, most of the initial detection frames can meet the condition of being greater than the first confidence threshold, so that more initial detection frames can be reserved for subsequent processing. .
  • the initial detection frame in the first frame image before obtaining the first detection frame of the detected object in the first frame image, can be preliminarily screened based on a small confidence threshold, that is, the first confidence threshold; Compared with the existing method of preliminary screening with a larger confidence threshold, the method of the embodiment of the present application can increase the number of first detection frames screened out, ensuring that as many first detection frames as possible can be detected in subsequent steps. Boxes are processed to reduce missed detections.
  • a local maximum search can be performed on the initial detection frame larger than the first confidence threshold.
  • the initial detection frames greater than the first confidence threshold are sorted in descending order of confidence to form a list of sorted initial detection frames, and the initial detection frame with the highest confidence is selected and added to the output list, and delete the initial detection frame with the highest confidence in the initial detection frame list; determine the degree of overlap or intersection-over-union ratio (IOU) between the initial detection frame with the highest confidence and other initial detection frames, in the initial detection frame list Delete the initial detection frame whose overlap degree is greater than the overlap threshold; repeat the previous steps until the initial detection frame list is empty, and the initial detection frame in the final output list is the initial detection frame selected according to the non-maximum suppression algorithm .
  • IOU intersection-over-union ratio
  • the second frame of images may be detected and screened through the above steps S111 to S313.
  • the initial detection frame of the second frame image is initially screened based on the first confidence threshold in step S112
  • the first confidence threshold can be set to a larger value, such as 0.5, so as to filter out the second detection frame with higher confidence and provide a data basis with higher confidence for object detection in subsequent images.
  • the initial detection frame of the second frame image is preliminarily
  • the first confidence threshold can still be set to a small value, such as 0.5 ⁇ 10 ⁇ 4 .
  • the characteristic parameter in the embodiment of the present application may further include position information, and at least one kind of similarity information includes at least one of overlapping degree and distance.
  • determining at least one kind of similarity information between the first detection frame and the second detection frame includes: according to the position information of the first detection frame and the second detection frame The position information of the detection frame determines at least one of the overlapping degree and the distance between the first detection frame and the second detection frame.
  • the degree of overlap and the distance can be determined, and the degree of overlap and the distance can intuitively reflect the relative positions of the first detection frame and the second detection frame relationship, and then reflect the similarity between the first detection frame and the second detection frame, so that the overlap and distance can be used as similarity information and fused with the confidence.
  • the fusion result of the embodiment of the present application can take into account multiple situations, so as to more accurately match the first detection frame and the second detection frame, thereby improving the accuracy of object detection.
  • the location information in this embodiment of the present application may include coordinate information of the outline of the detection frame.
  • determining the degree of overlap between the first detection frame and the second detection frame according to the position information of the first detection frame and the position information of the second detection frame includes:
  • I ij is the degree of overlap between the i-th first detection frame and the j-th second detection frame; A is the area of the i-th first detection frame, that is, the first area; B is the area of the jth second detection frame, that is, the second area; both i and j are positive integers.
  • determining the distance between the first detection frame and the second detection frame according to the position information of the first detection frame and the position information of the second detection frame includes:
  • a boundary length, a second boundary length, a first center coordinate and a second center coordinate determine the distance between the first detection frame and the second detection frame.
  • the distance between the first detection frame and the second detection frame can be determined as follows:
  • G ij is the distance between the i-th first detection frame and the j-th second detection frame; w 1 is along the first direction in the i-th first detection frame (as shown in Figure 3 The length of the boundary extending along the x direction in ), that is, the length of the first boundary; h 1 is the length of the boundary extending along the second direction (like the y direction in Figure 3 ) in the ith first detection frame, that is, the second Boundary length; x 1 and y 1 are the coordinate values of the center point of the i-th first detection frame in the first direction and the second direction respectively, (x 1 , y 1 ) is the first center coordinate; x 2 and y 2 are the coordinate values of the center point of the j-th second detection frame in the first direction and the second direction respectively, and (x 2 , y 2 ) are the second center coordinates; both i and j are positive integers.
  • the distance between the first detection frame and the second detection frame can also be determined directly based on the difference between the center coordinates of the first detection frame and the second detection frame.
  • the first detection frame of the two detection frames The difference between the length of the first boundary and the length of the second boundary is large, and the distance determined directly by using the center coordinates is affected by the length of the boundary, which may not be enough.
  • Determine the distance between the first detection frame and the second detection frame based on the method shown in expression (2).
  • the difference between the center coordinates is divided by the boundary length, which can reduce the influence of the boundary length difference and make the obtained distance more accurate.
  • step S130 at least one kind of similarity information and feature parameters are fused to obtain a fusion result, including: the overlap degree and distance between the first detection frame and the second detection frame Fusion of at least one kind of information, the confidence degree of the first detection frame, and the confidence degree of the second detection frame to obtain the fusion result.
  • the distance is normalized; the weighted average of the overlapping degree and the normalized distance is determined; the weighted average, the confidence of the first detection frame, and the second detection frame are determined The product of the confidences of the boxes, as the fusion result.
  • the fused result can be obtained by:
  • I ij is the degree of overlap between the i-th first detection frame and the j-th second detection frame
  • G ij is the difference between the i-th first detection frame and the j-th second detection frame the distance between is the distance between the i-th first detection frame and the j-th second detection frame after normalization
  • s i is the confidence degree of the i-th first detection frame
  • s j is the j-th second detection The confidence of the box
  • ⁇ is the weight value, and the value range is (0,1); for I ij and The weighted average of ;
  • V ij is the fusion result of I ij , G ij , s i and s j , and both i and j are positive integers.
  • the normalization of the distance G ij can be realized, so that the normalization distance after processing Consistent with the degree of overlap I ij , they are all values between (0,1), so as to facilitate the fusion of the distance G ij and the degree of overlap I ij .
  • the distance G ij and the overlap degree I ij can be further fused with the confidence s i and s j to obtain the fusion result V ij .
  • the fusion result V ij can represent the i-th
  • the matching degree between a detection frame and the jth second detection frame indicates that the ith first detection frame and the jth second detection frame have a preliminary matching relationship.
  • step S140 updating the confidence level of the first detection frame according to the fusion result may include the following steps S141-S143.
  • the detection frame group includes a first detection frame and a second detection frame that match each other.
  • each first detection frame and each second detection frame can be determined by expressing (3) Carry out the fusion result of information fusion to obtain a fusion result matrix with M rows and N columns, and one-to-one matching between M first detection frames and N second detection frames can be realized according to the Hungarian matching algorithm and the fusion result matrix.
  • the Hungarian matching algorithm can perform an optimal pairing solution for the first detection frame and the second detection frame based on the matching degree between the first detection frame and the second detection frame represented by each element of the fusion result matrix.
  • the detection frame matching is based on the fusion result of at least one similarity information and confidence, and multiple information can be considered to improve the accuracy of detection frame matching.
  • the detection frame A and the detection frame B of the detection area are obviously from different detection objects, but when using single information such as a single
  • the detection frame A and the detection frame B are considered to be matching detection frames, and then the detection frame A and the detection frame B are considered to be from the same detection object, resulting in detection errors.
  • S142 Determine the confidence of the detection frame group according to the confidence of the first detection frame and the confidence of the second detection frame in the detection frame group.
  • the tracking sequence in the embodiment of the present application is a sequence set of detection results of tracking and detection of the same detection object based on images collected at different times.
  • the confidence of any detection frame group can be determined through the following timing model:
  • s' is the confidence degree of the detection frame group
  • s i is the confidence degree of the first detection frame in the detection frame group
  • s j is the confidence degree of the second detection frame in the detection frame group
  • n is The number of times the detection object of the first detection frame in the detection frame appears in the tracking sequence
  • q is a regularization factor, which can be set according to actual needs, for example, it can be set to 5
  • q/(n+q) is the first detection in the detection frame group
  • the weight of the confidence degree of the frame that is, the first weight
  • n/(n+q) is the weight of the confidence degree of the second detection frame in the detection frame group, that is, the second weight.
  • the timing model of (4) introduces the confidence degree of the second detection frame, which can update the confidence degree of the first detection frame more accurately.
  • Fig. 7 shows a schematic diagram of a tracking sequence for tracking the same detected object and a schematic diagram of a principle of updating the confidence of the first detection frame.
  • t-n to t in Figure 7 represent the n+1 moments in the tracking sequence, and the t moment is the acquisition moment of the first frame image, and the numbers above each moment represent the confidence of the detection frame in the image at each moment, before updating
  • the number 0.1 above the time t represents the confidence of the first detection frame in the first frame image before the update
  • the number 0.67 above the time t represents the first frame after the update Confidence of the first detection box in the image.
  • step S150 determining the detection object according to the first detection frame after the confidence update includes: determining the first detection frame greater than the second confidence threshold in the first detection frame after the confidence update A detection frame is used as a target detection frame, and an object corresponding to the target detection frame is used as a detection object.
  • the second confidence threshold in the embodiment of the present application is greater than a preset value, the preset value can be set according to actual needs or experience values, and the second confidence threshold can be set according to actual needs or experience values on the basis of being smaller than the preset value.
  • the second confidence threshold may be greater than the first confidence threshold, for example, the second confidence threshold may be set to 0.5, since in the previous step a smaller first confidence threshold is used as the screening threshold to preliminarily screen out the initial For the detection frame, more invalid first detection frames may be introduced.
  • the first detection frame after the confidence update and the first detection frame without the confidence update are performed again Screening can eliminate invalid detection frames with low confidence, and obtain a more accurate detection frame set, so as to more accurately determine the detection object.
  • step S150 determining the detection object according to the updated first detection frame of the confidence includes: determining a value greater than The first detection frame with the second confidence threshold is used as a target detection frame, and the object corresponding to the target detection frame is used as a detection object.
  • the image to be detected can be used as the new first frame of image
  • the current first frame image can be used as a new second frame image
  • the target detection frame determined through the above steps S110 to S150 can be used as the second detection frame of the new second frame image, that is, step S150
  • step S110 can be re-executed, so that S110 to S150 can be executed cyclically for each new first frame image to perform operations such as information fusion and confidence update, so as to improve the accuracy of object detection for each frame image.
  • the object detection apparatus 800 may include: an information acquisition module 801 , a similarity determination module 802 , an information fusion module 803 , an information update module 804 and an object determination module 805 .
  • the information acquisition module 801 is used to obtain the first detection frame of the detected object in the first frame image and the second detection frame of the detected object in the second frame image; the acquisition time of the second frame image is before the acquisition time of the first frame image, Both the first detection frame and the second detection frame carry feature parameters, and the feature parameters include at least confidence.
  • the similarity determination module 802 is configured to determine at least one type of similarity information between the first detection frame and the second detection frame according to the characteristic parameters of the first detection frame and the characteristic parameters of the second detection frame.
  • the information fusion module 803 is used to fuse at least one similarity parameter and feature parameter to obtain a fusion result.
  • the information update module 804 is configured to update the confidence of the first detection frame according to the fusion result.
  • the object determination module 805 is configured to determine the detection object according to the first detection frame after the confidence level is updated.
  • the object detection apparatus 800 provided in the embodiment of the present application may further include a detection module.
  • the detection module is used to obtain the first detection frame of the detection object in the first frame of image and the second detection frame of the detection object in the second frame of image, and detect the detection object in the first frame of image according to the target detection model, and obtain A plurality of initial detection frames; determine an initial detection frame greater than a first confidence threshold in a plurality of initial detection frames; the first confidence threshold is less than a preset value; The initial detection frame is screened, and the screened initial detection frame is determined as the first detection frame.
  • the characteristic parameter in the embodiment of the present application further includes location information, and at least one kind of similarity information includes at least one of overlapping degree and distance.
  • the similarity determination module 802 in the embodiment of the present application is specifically configured to determine the first detection frame and the second detection frame according to the position information of the first detection frame and the position information of the second detection frame At least one of the overlapping degree and the distance between them.
  • the information fusion module 803 in the embodiment of the present application is specifically used for at least one of the overlapping degree and distance information, the confidence of the first detection frame, and the confidence of the second detection frame Perform fusion to obtain the fusion result.
  • the location information includes coordinate information of the outline of the detection frame.
  • the similarity determination module 802 in the embodiment of the present application is specifically configured to determine the first boundary length, the second boundary length and the first center of the first detection frame according to the coordinate information of the first detection frame coordinates; determine the second center coordinate of the second detection frame according to the coordinate information of the second detection frame; determine the first detection frame and the second center coordinate according to the first boundary length, the second boundary length, the first center coordinate and the second center coordinate Distance between detection boxes.
  • the information fusion module 803 in the embodiment of the present application is specifically used to normalize the distance; determine the weighted average of the overlapping degree and the normalized distance; determine the weighted average The product of the value, the confidence of the first detection frame and the confidence of the second detection frame is used as the fusion result.
  • the information update module 804 in the embodiment of the present application is specifically configured to: determine multiple detection frame groups according to the fusion results; Determine the confidence of the detection frame group; update the confidence of the first detection frame in the detection group according to the confidence of the detection frame group; the detection frame group includes the first detection frame and the second detection frame that match each other.
  • the information update module 804 in the embodiment of the present application when determining the confidence of the detection frame group according to the confidence of the first detection frame and the confidence of the second detection frame in the detection frame group , specifically for: determining the number of occurrences of the detection object corresponding to the first detection frame in the tracking sequence; determining the first weight of the confidence of the first detection frame and the second weight of the confidence of the second detection frame according to the number of times; According to the first weight and the second weight, a weighted average of the confidence of the first detection frame and the confidence of the second detection frame is determined as the confidence of the detection frame group.
  • Tracking timing is a sequence collection of detection results of tracking and detecting the same detection object based on images collected at different times.
  • the object determination module 805 in the embodiment of the present application is specifically configured to determine the first detection frame greater than the second confidence threshold in the first detection frame after the confidence update, as the target detection frame , the object corresponding to the target detection frame is used as the detection object; the second confidence threshold is greater than a preset value.
  • the object determination module 805 in the embodiment of the present application is specifically configured to: determine the object greater than the second confidence in the first detection frame after the confidence update and the first detection frame without the confidence update
  • the first detection frame of the degree threshold is used as the target detection frame, and the object corresponding to the target detection frame is used as the detection object; the second confidence threshold is greater than a preset value.
  • Each module/unit in the object detection device 800 provided in the embodiment of the present application has the function of realizing each step in the object detection method provided in the embodiment of the present application, and can achieve its corresponding technical effects. For the sake of concise description, it will not be repeated here repeat.
  • the embodiment of the present application also provides an object detection device.
  • the object detection device 900 includes a processor 901 and a memory 902 storing computer program instructions.
  • the above-mentioned processor 901 may include a central processing unit (Central Processing Unit, CPU), or a specific integrated circuit (Application Specific Integrated Circuit, ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application .
  • CPU Central Processing Unit
  • ASIC Application Specific Integrated Circuit
  • Memory 902 may include mass storage for data or instructions.
  • memory 902 may include a hard disk drive (Hard Disk Drive, HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a Universal Serial Bus (Universal Serial Bus, USB) drive or two or more Combinations of multiple of the above.
  • memory 802 may include removable or non-removable (or fixed) media, or memory 802 may be a non-volatile solid-state memory.
  • Memory 902 may be internal or external to the object detection device.
  • memory 902 may include read only memory (ROM), random access memory (RAM), magnetic disk storage media devices, optical storage media devices, flash memory devices, electrical, optical, or other physical/tangible memory storage devices.
  • ROM read only memory
  • RAM random access memory
  • magnetic disk storage media devices e.g., magnetic disks
  • optical storage media devices e.g., magnetic disks
  • flash memory devices e.g., electrical, optical, or other physical/tangible memory storage devices.
  • memory 902 includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software comprising computer-executable instructions, and when the software is executed (e.g., by a or a plurality of processors) operable to perform the operations described with reference to the method according to an aspect of the present application.
  • the processor 901 reads and executes the computer program instructions stored in the memory 902 to implement the object detection method provided by the embodiment of the present application, and the corresponding technical effects achieved by the object detection method will not be repeated here for brevity.
  • the object detection device 900 provided in the embodiment of the present application may further include a communication interface 903 and a bus 904 .
  • a processor 901 , a memory 902 , and a communication interface 903 are connected through a bus 904 to complete mutual communication.
  • the communication interface 903 is mainly used to realize the communication between various modules, devices, units and/or devices in the embodiments of the present application.
  • the bus 904 includes hardware, software, or both, and couples the components of the object detection device 900 to each other.
  • a bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Super Transmission (Hyper Transport, HT) interconnect, Industry Standard Architecture (Industry Standard Architecture, ISA) bus, InfiniBand interconnect, Low Pin Count (LPC) bus, memory bus, Micro Channel Architecture (MCA) bus, peripheral component interconnect PCI bus, PCI-Express (PCI-X) bus, Serial Advanced Technology Attachment (SATA) bus, Video Electronics Standards Association Local (VLB) bus, or other suitable bus or a combination of two or more of these combination.
  • Bus 904 may comprise one or more buses, where appropriate. Although the embodiments of this application describe and illustrate a particular bus, this application contemplates any suitable bus or interconnect.
  • the object detection device 900 provided in the embodiment of the present application can execute the object detection method in the embodiment of the present application, and realize the corresponding technical effect of the object detection method. For the sake of brevity, the details will not be repeated here.
  • an embodiment of the present application also provides a computer storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, any object detection method in the above-mentioned embodiments is implemented. .
  • the computer storage medium may include a large-capacity memory for data or instructions.
  • the computer storage medium may include a hard disk drive (Hard Disk Drive, HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a Universal Serial Bus (Universal Serial Bus, USB) drive or two or a combination of more of the above.
  • computer storage media may include removable or non-removable (or fixed) media, or the computer storage media may be non-volatile solid-state memory.
  • the computer storage medium can be internal or external to the object detection device.
  • the functional blocks shown in the structural block diagrams described above may be implemented as hardware, software, firmware, or a combination thereof.
  • hardware When implemented in hardware, it can be, for example, an electronic circuit, an Application Specific Integrated Circuit (ASIC), appropriate firmware, a plug-in, a function card, and the like.
  • ASIC Application Specific Integrated Circuit
  • the elements of the present application are the programs or code segments employed to perform the required tasks.
  • Programs or code segments can be stored in machine-readable media, or transmitted over transmission media or communication links by data signals carried in carrier waves.
  • "Machine-readable medium" may include any medium that can store or transmit information.
  • machine-readable media examples include electronic circuits, semiconductor memory devices, read-only memory, flash memory, erasable read-only memory (EROM), floppy disks, compact disks-read-only (CD-ROMs), optical disks, hard disks, fiber optic media, radio frequency ( Radio Frequency, RF) link, etc.
  • Code segments may be downloaded via a computer network such as the Internet, an Intranet, or the like.

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

本申请提供了一种对象检测方法、装置、设备及计算机存储介质。对象检测方法包括：获取第一帧图像中检测对象的第一检测框和第二帧图像中检测对象的第二检测框；第二帧图像的采集时刻在第一帧图像的采集时刻之前，第一检测框和第二检测框均携带有特征参数，特征参数至少包括置信度；根据第一检测框的特征参数和第二检测框的特征参数，确定第一检测框和第二检测框之间的至少一种相似度信息；对至少一种相似度信息和特征参数进行融合，得到融合结果；根据融合结果更新第一检测框的置信度；根据置信度更新后的第一检测框确定检测对象。

Description

对象检测方法、装置、设备及计算机存储介质
相关申请的交叉引用
本申请主张2021年05月14日在中国提交的中国专利申请号202110525549.2的优先权，其全部内容通过引用包含于此。
技术领域
本申请涉及计算机视觉技术领域，尤其涉及一种对象检测方法、装置、设备及计算机存储介质。
背景技术
随着计算机视觉的发展，计算机视觉识别技术广泛应用到对象密集场所以实现对象检测或统计，例如客流统计。
计算机视觉识别技术主要是通过深度学习检测算法对对象进行检测，然后利用跟踪算法进行对象统计，相比传统的对象检测和统计方法，计算机视觉识别技术有稳定性好、不易损坏等优点。
然而，在对图像或者视频进行对象检测的时候，若图像或视频中因对象排布比较紧密而产生对象的重叠，现有检测算法在检测这些重叠对象时准确率较低，很容易出现漏检，难以得到准确的统计结果。
发明内容
本申请实施例提供了一种对象检测方法、装置、设备及计算机存储介质，能够解决对象检测准确率较低、容易出现漏检的技术问题。
第一方面，本申请实施例提供了一种对象检测方法，对象检测方法包括：
获取第一帧图像中检测对象的第一检测框和第二帧图像中检测对象的第二检测框；第二帧图像的采集时刻在第一帧图像的采集时刻之前，第一 检测框和第二检测框均携带有特征参数，特征参数至少包括置信度；
根据第一检测框的特征参数和第二检测框的特征参数，确定第一检测框和第二检测框之间的至少一种相似度信息；
对至少一种相似度信息和特征参数进行融合，得到融合结果；
根据融合结果更新第一检测框的置信度；
根据置信度更新后的第一检测框确定检测对象。
第二方面，本申请实施例提供了一种对象检测装置，对象检测装置包括：
信息获取模块，用于获取第一帧图像中检测对象的第一检测框和第二帧图像中检测对象的第二检测框；第二帧图像的采集时刻在第一帧图像的采集时刻之前，第一检测框和第二检测框均携带有特征参数，特征参数至少包括置信度；
相似度确定模块，用于根据第一检测框的特征参数和第二检测框的特征参数，确定第一检测框和第二检测框之间的至少一种相似度信息；
信息融合模块，用于对至少一种相似度参数和特征参数进行融合，得到融合结果；
信息更新模块，用于根据融合结果更新第一检测框的置信度；
目标确定模块，用于根据置信度更新后的第一检测框确定检测对象。
第三方面，本申请实施例提供了一种对象检测设备，对象检测设备包括：处理器，以及存储有计算机程序指令的存储器；
处理器读取并执行计算机程序指令，以实现本申请实施例第一方面提供的对象检测方法。
第四方面，本申请实施例提供了一种计算机存储介质，该计算机存储介质上存储有计算机程序指令，计算机程序指令被处理器执行时实现本申请实施例第一方面提供的对象检测方法。
本申请实施例的技术，至少能够实现如下有益效果：
本申请实施例同时考虑前后两帧图像来实现对象的检测，在基于该两帧图像的检测框的特征参数确定两帧图像的检测框之间的至少一种相似度信息后，进一步将得到的至少一种相似度信息与特征参数进行融合，可实 现两帧图像检测框的多重信息的融合，基于融合结果可更新第一检测框的置信度，根据更新后的第一检测框可确定出对象；相比于只考虑一帧图像或只考虑检测框置信度的情形，本申请实施例的技术方案在检测对象时可同时参考前后两帧图像的检测框的特征参数以及两帧图像的检测框之间的至少一种相似度信息，所参考的信息更加全面，在将更加全面的信息进行融合后基于融合结果进行对象检测时，可提高对象检测的准确性，减少漏检。
附图说明
为了更清楚地说明本申请实施例的技术方案，下面将对本申请实施例中所需要使用的附图作简单的介绍，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1是本申请一实施例提供的对象检测方法的流程示意图；
图2是本申请另一实施例提供的对象检测方法的部分流程示意图；
图3是本申请一实施例中不同检测框重叠度计算的原理示意图；
图4是本申请又一实施例提供的对象检测方法的部分流程示意图；
图5是相关技术中涉及的一种检测框匹配情形示意图；
图6是相关技术中涉及的另一种检测框匹配情形示意图；
图7是本申请一实施例中对同一检测对象进行跟踪的跟踪时序示意图和更新第一检测框置信度的原理示意图；
图8是本申请一实施例提供的对象检测装置的结构框架示意图；
图9是本申请一实施例提供的对象检测设备的结构框架示意图。
具体实施方式
下面将详细描述本申请的各个方面的特征和示例性实施例，为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及具体实施例，对本申请进行进一步详细描述。应理解，此处所描述的具体实施例仅意在解释本申请，而不是限定本申请。对于本领域技术人员来说，本申请可以在不需要这些具体细节中的一些细节的情况下实施。下面对实施例的描述 仅仅是为了通过示出本申请的示例来提供对本申请更好的理解。
需要说明的是，在本文中，诸如第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。
下面对本申请的技术方案进行详细介绍。
本申请实施例提供了一种对象检测方法，如图1所示，该对象检测方法可以包括：
S110，获取第一帧图像中检测对象的第一检测框和第二帧图像中检测对象的第二检测框。
第二帧图像的采集时刻在第一帧图像的采集时刻之前，第一检测框和第二检测框均携带有特征参数，特征参数至少包括置信度。
S120，根据第一检测框的特征参数和第二检测框的特征参数，确定第一检测框和第二检测框之间的至少一种相似度信息。
S130，对至少一种相似度信息和特征参数进行融合，得到融合结果。
S140，根据融合结果更新第一检测框的置信度。
S150，根据置信度更新后的第一检测框确定检测对象。
本申请实施例提供的对象检测方法，至少能够实现如下有益效果：
本申请实施例同时考虑前后两帧图像来实现对象的检测，在基于该两帧图像的检测框的特征参数确定两帧图像的检测框之间的至少一种相似度信息后，进一步将得到的至少一种相似度信息与特征参数进行融合，可实现两帧图像检测框的多重信息的融合，基于融合结果可更新第一检测框的置信度，根据更新后的第一检测框可确定出检测对象；相比于只考虑一帧图像或只考虑检测框置信度的情形，本申请实施例的技术方案在检测对象 时可同时参考前后两帧图像的检测框的特征参数以及两帧图像的检测框之间的至少一种相似度信息，所参考的信息更加全面，在将更加全面的信息进行融合后基于融合结果进行对象检测时，可提高对象检测的准确性，减少漏检。
在一个可选的实施方式中，本申请实施例中的第一帧图像和第二帧图像之间间隔的帧数可小于预设帧数，该预设帧数可根据实际需求和经验值设置。在一个示例中，为了使第二帧图像和第一帧图像在时序上更加接近，从而使第二帧图像中的第二检测框更具参考性，可将预设帧数设置为一个较小的值，例如2帧，此时第二帧图像和第一帧图像之间间隔一帧，即第二帧图像是第一帧图像的前一帧图像。
在一个可选的实施方式中，参照图2所示，在步骤S110中，获取第一帧图像中检测对象的第一检测框和第二帧图像中检测对象的第二检测框之前，可以包括如下步骤S111-S113。
S111，根据目标检测模型对第一帧图像中的检测对象进行检测，得到多个初始检测框。
本申请目标检测模型可以是现有的任意一种可实现目标检测的神经网络，在一个示例中，目标检测模型可以是两阶段检测算法和单阶段检测算法的任意一种，两阶段检测算法可以R-CNN(Region-Convolutional Neural Network，区域卷积神经网络)、Fast R-CNN(快速区域卷积神经网络)系列等任意一种卷积神经网络检测算法，单阶段检测算法可以是SSD(Single Shot MultiBox Detector，单阶段多框预测算法)、YOLO(You Only Look Once)系列等任意一种卷积神经网络检测算法，目标检测模型输出的初始检测框携带有位置信息和置信度。
本申请实施例的检测对象可以是任意指定对象，例如某场所的人头、动物或其它物体。
S112，在多个初始检测框中确定大于第一置信度阈值的初始检测框。
第一置信度阈值小于预设数值，该预设数值可根据实际需求或经验值设置，第一置信度阈值可在小于预设数值的基础上根据实际需求或经验值设置。
为了减少漏检，通常不希望在初始阶段过滤掉过多的检测框，此时可将预设数值设置为一个较小的数值，例如1×10 -4或1×10 -1，从而使第一置信度阈值为一个较小的数值，例如0.5×10 -4；此时，大部分初始检测框可以满足大于第一置信度阈值这个条件，从而可以保留较多的初始检测框供后续处理过程使用。
本申请实施例通过获取第一帧图像中检测对象的第一检测框之前，可对基于一个较小的置信度阈值即第一置信度阈值对第一帧图像中的初始检测框进行初步筛选；相比于现有的以一个较大的置信度阈值进行初步筛选的方式，本申请实施例的方式可增加筛选出的第一检测框数量，保证后续步骤中能够对尽可能多的第一检测框进行处理，以减少漏检。
S113，根据非极大抑制算法对大于第一置信度阈值的初始检测框进行筛选，将筛选出的初始检测框确定为第一检测框。
根据非极大抑制算法的原理，可对大于第一置信度阈值的初始检测框进行局部最大搜索。
在一个示例中，将大于第一置信度阈值的各初始检测框中按照置信度从大到小的顺序进行排序，形成排序后的初始检测框列表，选择置信度最高的初始检测框添加到输出列表中，并在初始检测框列表中删除该置信度最高的初始检测框；确定置信度最高的初始检测框和其它初始检测框的重叠度或称交并比(IOU)，在初始检测框列表中删除重叠度大于重叠度阈值的初始检测框；重复前面的步骤，直到初始检测框列表为空，最终得到的输出列表中的初始检测框即为根据非极大抑制算法筛选出的初始检测框。
在一个可选的实施方式中，可通过上述步骤S111至S313对第二帧图像进行检测和筛选。
在一个示例中，若本申请实施例中的第二帧图像为当前图像检测中的起始帧图像，在基于步骤S112中的第一置信度阈值对第二帧图像的初始检测框进行初步筛选时，第一置信度阈值可以设置为一个较大的数值，例如0.5，从而筛选出较为置信度较高的第二检测框，为后续图像的对象检测提供置信度较高数据基础。
在另一个示例中，若本申请实施例中的第二帧图像为当前图像检测中 的起始帧图像，在基于步骤S112中的第一置信度阈值对第二帧图像的初始检测框进行初步筛选时，仍可将第一置信度阈值设置为一个较小的数值，例如0.5×10 -4。
本领域技术人员可以理解，在基于步骤S112中的第一置信度阈值对第二帧图像的初始检测框进行初步筛选后，还需要执行S113进行基于非极大抑制的筛选，具体筛选原理可参照前面的内容，此处不再赘述。
在一个可选的实施方式中，在步骤S120中，本申请实施例中的特征参数还可以包括位置信息，至少一种相似度信息包括重叠度和距离中的至少一种信息。
根据第一检测框的特征参数和第二检测框的特征参数，确定第一检测框和第二检测框之间的至少一种相似度信息，包括：根据第一检测框的位置信息和第二检测框的位置信息，确定第一检测框和第二检测框之间的重叠度和距离中的至少一种信息。
本申请实施例可基于第一检测框和第二检测框的位置信息，确定重叠度和距离中的至少一种，重叠度和距离可以直观地反映第一检测框和第二检测框的相对位置关系，进而反映第一检测框和第二检测框的相似度，从而重叠度和距离可作为相似度信息，与置信度进行融合。相比于单一信息的检测框匹配，本申请实施例的融合结果可兼顾多种情况，从而更准确地对第一检测框和第二检测框进行匹配，进而提高对象检测的准确率。
本申请实施例中的位置信息可以包括检测框轮廓的坐标信息。
在一个可选的实施方式中，根据第一检测框的位置信息和第二检测框的位置信息，确定第一检测框和第二检测框之间的重叠度，包括：
根据第一检测框的坐标信息确定第一检测框的第一面积；根据第二检测框的坐标信息确定第二检测框的第二面积；根据第一面积和第二面积确定第一检测框和第二检测框之间的重叠度。
参照图3所示的检测框重叠示意图，上述重叠度的确定方式可表达为：
Figure PCTCN2022092842-appb-000001
在表达式(1)中，I ij为第i个第一检测框和第j个第二检测框之间的重叠度；A为第i个第一检测框的面积，即第一面积；B为第j个第二检测 框的面积，即第二面积；i和j均为正整数。
在另一个可选的实施方式中，根据第一检测框的位置信息和第二检测框的位置信息，确定第一检测框和第二检测框之间的距离，包括：
根据第一检测框的坐标信息确定第一检测框的第一边界长度、第二边界长度和第一中心坐标；根据第二检测框的坐标信息确定第二检测框的第二中心坐标；根据第一边界长度、第二边界长度、第一中心坐标和第二中心坐标，确定第一检测框和第二检测框之间的距离。
在一个示例中，可通过如下方式确定第一检测框和第二检测框之间的距离：
Figure PCTCN2022092842-appb-000002
在表达式(2)中，G ij为第i个第一检测框和第j个第二检测框之间的距离；w 1为第i个第一检测框中沿第一方向(如图3中的x方向)延伸的边界的长度，即第一边界长度；h 1为第i个第一检测框中沿第二方向(如图3中的y方向)延伸的边界的长度，即第二边界长度；x 1和y 1分别为第i个第一检测框的中心点在第一方向和第二方向的坐标值，(x 1,y 1)即为第一中心坐标；x 2和y 2分别为第j个第二检测框的中心点在第一方向和第二方向的坐标值，(x 2,y 2)即为第二中心坐标；i和j均为正整数。
在一个示例中，还可直接基于第一检测框和第二检测框的中心坐标之差确定第一检测框和第二检测框之间距离，在该种情况下，若两个检测框的第一边界长度或第二边界长度相差较大，直接采用中心坐标确定出的距离受边界长度的影响，可能不够，基于表达式(2)所示的方式确定第一检测框和第二检测框之间的距离时，对中心坐标之差进行了除以边界长度的操作，可减小边界长度差异带来影响，使得到的距离更加准确。
在一个可选的实施方式中，在步骤S130中，对至少一种相似度信息和特征参数进行融合，得到融合结果，包括：对第一检测框和第二检测框之间的重叠度和距离中的至少一种信息、第一检测框的置信度、第二检测框的置信度进行融合，得到融合结果。
在一个可选的实施方式中，对距离进行归一化处理；确定重叠度和归一化处理后的距离的加权平均值；确定该加权平均值、第一检测框的置信 度和第二检测框的置信度的乘积，作为融合结果。
在一个示例中，可通过以下方式得到融合结果：
Figure PCTCN2022092842-appb-000003
在表达式(3)中，I ij为第i个第一检测框和第j个第二检测框之间的重叠度G ij为第i个第一检测框和第j个第二检测框之间的距离；
Figure PCTCN2022092842-appb-000004
为归一化处理后的第i个第一检测框和第j个第二检测框之间的距离；s i为第i个第一检测框的置信度；s j为第j个第二检测框的置信度；θ为权重值，取值范围为(0,1)；
Figure PCTCN2022092842-appb-000005
为I ij和
Figure PCTCN2022092842-appb-000006
的加权平均值；V ij为I ij、G ij、s i和s j的融合结果，i和j均为正整数。
在表达式(3)，通过对第i个第一检测框和第j个第二检测框之间的距离G ij进行指数化运算，可实现对距离G ij的归一化处理，使归一化处理后的距离
Figure PCTCN2022092842-appb-000007
与重叠度I ij一致，均为(0,1)之间的数值，以便于进行距离G ij和重叠度I ij的融合。在此基础上，通过表达式(3)，距离G ij和重叠度I ij还可进一步与置信度s i和s j进行融合，得到融合结果V ij，融合结果V ij可表征第i个第一检测框和第j个第二检测框的匹配度，表示第i个第一检测框和第j个第二检测框具有初步匹配关系。
在一个可选的实施方式中，参照图4所示，在步骤S140中，根据融合结果更新第一检测框的置信度，可以包括如下步骤S141-S143。
S141，根据融合结果确定多个检测框组。
检测框组包括相互匹配的第一检测框和第二检测框。
在一个示例中，对于M个第一检测框和N个第二检测框(M和N均为正整数)，通过表达(3)可确定出每个第一检测框和每个第二检测框进行信息融合的融合结果，得到M行N列融合结果矩阵，可根据匈牙利匹配算法和该融合结果矩阵实现M个第一检测框和N个第二检测框之间的一一匹配。具体地，匈牙利匹配算法可基于该融合结果矩阵各元素表征的第一检测框和第二检测框的匹配度，对第一检测框和第二检测框进行最优配对求解。
相对于基于单一信息进行检测框匹配的情况，基于至少一个相似度信息和置信度的融合结果进行检测框的匹配，可考虑多重信息以提高检测框 匹配的准确率。
在一个示例中，参照图5，若检测区域的检测框A和检测框B的中心相同但大小不同，则检测框A和检测框B明显来自不同的检测对象，但在采用单一信息例如单一的距离信息来进行检测框的匹配时，会认为检测框A和检测框B是可以相互匹配的检测框，进而认为检测框A和检测框B来自同一检测对象，从而发生检测错误。
在另一个示例中，参照图6，对于检测区域的检测框A、B和C，若采用单一信息例如单一的重叠度信息来进行检测框的匹配，由于检测框A与检测框B重叠度、以及检测框A与检测框C的重叠度均为0，此时无法判断检测框A与检测框B、检测框C中的哪个检测框更匹配。
S142，根据检测框组中第一检测框的置信度和第二检测框的置信度，确定检测框组的置信度。
在一个可选的实施方式中，确定第一检测框对应的检测对象在跟踪时序中出现的次数；根据该次数确定第一检测框的置信度的第一权重、以及第二检测框的置信度的第二权重；根据第一权重和第二权重，确定第一检测框的置信度和第二检测框的置信度的加权平均值，作为检测框组的置信度。
本申请实施例中的跟踪时序是基于不同时刻采集的图像对同一检测对象进行跟踪检测的检测结果的序列集合。
在一个示例中，可通过如下时序模型确定任意一个检测框组的置信度：
Figure PCTCN2022092842-appb-000008
在表达式(4)中，s′为检测框组的置信度，s i为检测框组中第一检测框的置信度，s j为检测框组中第二检测框的置信度，n为检测框中第一检测框的检测对象在跟踪时序中出现的次数；q为正则因子，可根据实际需求设置，例如可设置为5；q/(n+q)为检测框组中第一检测框的置信度的权重，即第一权重；n/(n+q)为检测框组中第二检测框的置信度的权重，即第二权重。
S143，根据检测框组的置信度更新检测组中第一检测框的置信度。
根据表达式(4)所示的时序模型可知，若某个第一检测框在跟踪时 序中出现的次数越多，也就是n值越大，则第一权重q/(n+q)越小，第二权重n/(n+q)越大，则s′值更受与该第一检测框匹配的第二检测框的置信度s j的影响。也即某一第一检测框对应的对象在跟踪时序中出现的越久，则该第一检测框是真实检测对象的概率越大，由于目标检测模型检测的第一检测框的置信度可能会出现波动(例如某些第一检测框的置信度过小)从而导致漏检，此时认为第二检测框的置信度相对于第一检测框的置信度的可靠性更高，因此可以通过表达式(4)的时序模型引入第二检测框的置信度，可较为准确地更新第一检测框的置信度。
图7示出了对同一检测对象进行跟踪的跟踪时序示意图和更新第一检测框置信度的原理示意图。图7中的t-n至t表示跟踪时序中的n+1个时刻，t时刻为第一帧图像的采集时刻，各时刻上方的数字表示各时刻的图像中检测框的置信度，在更新前的跟踪时序示意图中，t时刻上方的数字0.1表示更新前的第一帧图像中第一检测框的置信度，在更新后的跟踪时序示意图中，t时刻上方的数字0.67表示更新后的第一帧图像中第一检测框的置信度。
由图7可以看出，可以看出如若某个对象的检测框在t时刻出现置信度很低的情况，若按照正常置信度阈值(一般为0.5)进行筛选，该检测框就会被剔除，从而造成漏检；在本申请实施例的一种实施方式中，假设该检测框的对象在跟踪时序中出现的次数为10，即n＝10，则在利用如表达式(4)所示的时序模型进行置信度的更新后，该检测框的置信度由0.1更新为0.67，则按照正常置信度阈值筛选时该检测框不会被剔除，从而避免了漏检情况。
在一个可选的实施方式中，在步骤S150中，根据置信度更新后的第一检测框确定检测对象，包括：在置信度更新后的第一检测框中确定大于第二置信度阈值的第一检测框，作为目标检测框，目标检测框对应的对象作为检测对象。
本申请实施例中的第二置信度阈值大于预设数值，预设数值可根据实际需求或经验值设置，第二置信度阈值可在小于预设数值的基础上根据实际需求或经验值设置。
在一个示例中，第二置信度阈值可以大于第一置信度阈值，例如第二置信度可设置为0.5，由于前序步骤中是以较小的第一置信度阈值作为筛选阈值初步筛选出初始检测框的，可能引入了较多的无效的第一检测框，在本步骤中，基于第二置信度阈值对置信度更新后的第一检测框和置信度未更新的第一检测框进行再次筛选，可剔除置信度较低的无效检测框，得到更准确的检测框集合，从而更精确地确定出检测对象。
在另一个可选的实施方式中，在步骤S150中，根据置信度更新后的第一检测框确定检测对象，包括：在更新后的第一检测框和未更新的第一检测框中确定大于第二置信度阈值的第一检测框，作为目标检测框，目标检测框对应的对象作为检测对象。
在本申请实施例中，基于置信度更新后的第一检测框可确定出与前序图像匹配度高、持续出现的检测对象，基于置信度未更新的第一检测框可确定出首次出现的检测对象，从而保证对象检测的全面性，减少漏检，
在一个示例中，经过上述步骤S110至S150后，若需要对第一帧图像之后(指该图像的采集时刻之后)的某一帧图像进行对象检测时，该待检测的图像可作为新的第一帧图像，当前的第一帧图像可作为新的第二帧图像，经过上述步骤S110至S150确定出的目标检测框可作为新的第二帧图像的第二检测框，也即，步骤S150之后可以重新执行步骤S110，从而对于每一个新的第一帧图像可循环执行S110至S150，进行信息的融合、置信度的更新等操作，以提高针对每一帧图像进行对象检测的准确率。
基于同一发明构思，本申请实施例还提供了一种对象检测装置。如图8所示，该对象检测装置800可以包括：信息获取模块801、相似度确定模块802、信息融合模块803、信息更新模块804和对象确定模块805。
信息获取模块801用于获取第一帧图像中检测对象的第一检测框和第二帧图像中检测对象的第二检测框；第二帧图像的采集时刻在第一帧图像的采集时刻之前，第一检测框和第二检测框均携带有特征参数，特征参数至少包括置信度。
相似度确定模块802用于根据第一检测框的特征参数和第二检测框的特征参数，确定第一检测框和第二检测框之间的至少一种相似度信息。
信息融合模块803用于对至少一种相似度参数和特征参数进行融合，得到融合结果。
信息更新模块804用于根据融合结果更新第一检测框的置信度。
对象确定模块805用于根据置信度更新后的第一检测框确定检测对象。
在一个可选的实施方式中，本申请实施例提供的对象检测装置800还可以包括检测模块。
该检测模块用于在获取第一帧图像中检测对象的第一检测框和第二帧图像中检测对象的第二检测框，根据目标检测模型对第一帧图像中的检测对象进行检测，得到多个初始检测框；在多个初始检测框中确定大于第一置信度阈值的初始检测框；第一置信度阈值为小于预设数值；根据非极大抑制算法对大于第一置信度阈值的初始检测框进行筛选，将筛选出的初始检测框确定为第一检测框。
在一个可选的实施方式中，本申请实施例中的特征参数还包括位置信息，至少一种相似度信息包括重叠度和距离中的至少一种信息。
在一个可选的实施方式中，本申请实施例中的相似度确定模块802具体用于根据第一检测框的位置信息和第二检测框的位置信息，确定第一检测框和第二检测框之间的重叠度和距离中的至少一种信息。
在一个可选的实施方式中，本申请实施例中的信息融合模块803具体用于对重叠度和距离中的至少一种信息、第一检测框的置信度、以及第二检测框的置信度进行融合，得到融合结果。
在一个可选的实施方式中，位置信息包括检测框轮廓的坐标信息。
在一个可选的实施方式中，本申请实施例中的相似度确定模块802具体用于根据第一检测框的坐标信息确定第一检测框的第一边界长度、第二边界长度和第一中心坐标；根据第二检测框的坐标信息确定第二检测框的第二中心坐标；根据第一边界长度、第二边界长度、第一中心坐标和第二中心坐标，确定第一检测框和第二检测框之间的距离。
在一个可选的实施方式中，本申请实施例中的信息融合模块803具体用于对距离进行归一化处理；确定重叠度和归一化处理后的距离的加权平均值；确定该加权平均值、第一检测框的置信度和第二检测框的置信度的 乘积，作为融合结果。
在一个可选的实施方式中，本申请实施例中的信息更新模块804具体用于：根据融合结果确定多个检测框组；根据检测框组中第一检测框的置信度和第二检测框的置信度，确定检测框组的置信度；根据检测框组的置信度更新检测组中第一检测框的置信度；检测框组包括相互匹配的第一检测框和第二检测框。
在一个可选的实施方式中，本申请实施例中的信息更新模块804，在根据检测框组中第一检测框的置信度和第二检测框的置信度，确定检测框组的置信度时，具体用于：确定第一检测框对应的检测对象在跟踪时序中出现的次数；根据次数确定第一检测框的置信度的第一权重、以及第二检测框的置信度的第二权重；根据第一权重和第二权重，确定第一检测框的置信度和第二检测框的置信度的加权平均值，作为检测框组的置信度。跟踪时序是基于不同时刻采集的图像对同一检测对象进行跟踪检测的检测结果的序列集合。
在一个可选的实施方式中，本申请实施例中的对象确定模块805具体用于在置信度更新后的第一检测框中确定大于第二置信度阈值的第一检测框，作为目标检测框，目标检测框对应的对象作为检测对象；第二置信度阈值大于预设数值。
在另一个可选的实施方式中，本申请实施例中的对象确定模块805具体用于：在置信度更新后的第一检测框和置信度未更新的第一检测框中确定大于第二置信度阈值的第一检测框，作为目标检测框，目标检测框对应的对象作为检测对象；第二置信度阈值大于预设数值。
本申请实施例提供的对象检测装置800中的各个模块/单元具有实现本申请实施例提供的对象检测方法中各个步骤的功能，并能达到其相应的技术效果，为简洁描述，在此不再赘述。
基于同一发明构思，本申请实施例还提供了一种对象检测设备。如图9所示，该对象检测设备900包括处理器901以及存储有计算机程序指令的存储器902。
具体地，上述处理器901可以包括中央处理器(Central Processing  Unit，CPU)，或者特定集成电路(Application Specific Integrated Circuit，ASIC)，或者可以被配置成实施本申请实施例的一个或多个集成电路。
存储器902可以包括用于数据或指令的大容量存储器。举例来说而非限制，存储器902可包括硬盘驱动器(Hard Disk Drive，HDD)、软盘驱动器、闪存、光盘、磁光盘、磁带或通用串行总线(Universal Serial Bus，USB)驱动器或者两个或更多个以上这些的组合。在一个实例中，存储器802可以包括可移除或不可移除(或固定)的介质，或者存储器802是非易失性固态存储器。存储器902可在对象检测设备的内部或外部。
在一个实例中，存储器902可包括只读存储器(ROM)，随机存取存储器(RAM)，磁盘存储介质设备，光存储介质设备，闪存设备，电气、光学或其他物理/有形的存储器存储设备。因此，通常，存储器902包括一个或多个编码有包括计算机可执行指令的软件的有形(非暂态)计算机可读存储介质(例如，存储器设备)，并且当该软件被执行(例如，由一个或多个处理器)时，其可操作来执行参考根据本申请的一方面的方法所描述的操作。
处理器901通过读取并执行存储器902中存储的计算机程序指令，以实现本申请实施例提供的对象检测方法，并该对象检测方法达到的相应技术效果，为简洁描述在此不再赘述。
在一个示例中，本申请实施例提供的对象检测设备900还可包括通信接口903和总线904。其中，如图9所示，处理器901、存储器902、通信接口903通过总线904连接并完成相互间的通信。
通信接口903，主要用于实现本申请实施例中各模块、装置、单元和/或设备之间的通信。
总线904包括硬件、软件或两者，将对象检测设备900的部件彼此耦接在一起。举例来说而非限制，总线可包括加速图形端口(Accelerated Graphics Port，AGP)或其他图形总线、增强工业标准架构(Extended Industry Standard Architecture，EISA)总线、前端总线(Front Side Bus，FSB)、超传输(Hyper Transport，HT)互连、工业标准架构(Industry Standard Architecture，ISA)总线、无限带宽互连、低引脚数(LPC)总线、 存储器总线、微信道架构(MCA)总线、外围组件互连(PCI)总线、PCI-Express(PCI-X)总线、串行高级技术附件(SATA)总线、视频电子标准协会局部(VLB)总线或其他合适的总线或者两个或更多个以上这些的组合。在合适的情况下，总线904可包括一个或多个总线。尽管本申请实施例描述和示出了特定的总线，但本申请考虑任何合适的总线或互连。
本申请实施例提供的对象检测设备900可以执行本申请实施例中的对象检测方法，并实现该对象检测方法的相应技术效果，为简洁描述，在此不再赘述。
基于同一发明构思，本申请实施例还提供了一种计算机存储介质，该计算机存储介质上存储有计算机程序指令，该计算机程序指令被处理器执行时实现上述实施例中的任意一种对象检测方法。
本申请实施例提供的计算机存储介质可以包括用于数据或指令的大容量存储器。举例来说而非限制，该计算机存储介质可包括硬盘驱动器(Hard Disk Drive，HDD)、软盘驱动器、闪存、光盘、磁光盘、磁带或通用串行总线(Universal Serial Bus，USB)驱动器或者两个或更多个以上这些的组合。在一个实例中，计算机存储介质可以包括可移除或不可移除(或固定)的介质，或者计算机存储介质是非易失性固态存储器。计算机存储介质可在对象检测设备的内部或外部。
需要明确的是，本申请并不局限于上文所描述并在图中示出的特定配置和处理。为了简明起见，这里省略了对已知方法的详细描述。在上述实施例中，描述和示出了若干具体的步骤作为示例。但是，本申请的方法过程并不限于所描述和示出的具体步骤，本领域的技术人员可以在领会本申请的精神后，作出各种改变、修改和添加，或者改变步骤之间的顺序。
以上所述的结构框图中所示的功能块可以实现为硬件、软件、固件或者它们的组合。当以硬件方式实现时，其可以例如是电子电路、专用集成电路(Application Specific Integrated Circuit，ASIC)、适当的固件、插件、功能卡等等。当以软件方式实现时，本申请的元素是被用于执行所需任务的程序或者代码段。程序或者代码段可以存储在机器可读介质中，或者通过载波中携带的数据信号在传输介质或者通信链路上传送。“机器可读介 质”可以包括能够存储或传输信息的任何介质。机器可读介质的例子包括电子电路、半导体存储器设备、只读存储器、闪存、可擦除只读存储器(EROM)、软盘、只读光盘(CD-ROM)、光盘、硬盘、光纤介质、射频(Radio Frequency，RF)链路，等等。代码段可以经由诸如因特网、内联网等的计算机网络被下载。
以上所述，仅为本申请的具体实施方式，所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，上述描述的系统、模块和单元的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。应理解，本申请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申请揭露的技术范围内，可轻易想到各种等效的修改或替换，这些修改或替换都应涵盖在本申请的保护范围之内。

Claims (16)

  1. 一种对象检测方法，包括：
    获取第一帧图像中检测对象的第一检测框和第二帧图像中检测对象的第二检测框；所述第二帧图像的采集时刻在所述第一帧图像的采集时刻之前，所述第一检测框和所述第二检测框均携带有特征参数，所述特征参数包括置信度和位置信息；
    根据所述第一检测框的所述位置信息和所述第二检测框的所述位置信息，确定所述第一检测框和所述第二检测框之间的重叠度和距离中的至少一种相似度信息；
    对所述重叠度和所述距离中的至少一种相似度信息、所述第一检测框的置信度、以及所述第二检测框的置信度进行融合，得到融合结果；
    根据所述融合结果更新所述第一检测框的置信度；
    根据置信度更新后的所述第一检测框确定检测对象。
  2. 根据权利要求1所述的对象检测方法，其中，所述获取第一帧图像中检测对象的第一检测框和第二帧图像中检测对象的第二检测框之前，还包括：
    根据目标检测模型对所述第一帧图像中的检测对象进行检测，得到多个初始检测框；
    在多个所述初始检测框中确定大于第一置信度阈值的所述初始检测框；所述第一置信度阈值为小于预设数值；
    根据非极大抑制算法对大于第一置信度阈值的所述初始检测框进行筛选，将筛选出的所述初始检测框确定为所述第一检测框。
  3. 根据权利要求1所述的对象检测方法，其中，所述位置信息包括检测框轮廓的坐标信息；
    根据所述第一检测框的所述位置信息和所述第二检测框的所述位置信息，确定所述第一检测框和所述第二检测框之间的距离，包括：
    根据所述第一检测框的所述坐标信息确定所述第一检测框的第一边界长度、第二边界长度和第一中心坐标；
    根据所述第二检测框的所述坐标信息确定所述第二检测框的第二中心 坐标；
    根据所述第一边界长度、所述第二边界长度、所述第一中心坐标和所述第二中心坐标，确定所述第一检测框和所述第二检测框之间的距离。
  4. 根据权利要求1所述的对象检测方法，其中，所述对所述重叠度和所述距离中的至少一种信息、所述第一检测框的置信度、以及所述第二检测框的置信度进行融合，得到融合结果，包括：
    对所述距离进行归一化处理；
    确定所述重叠度和归一化处理后的所述距离的加权平均值；
    确定所述加权平均值、所述第一检测框的置信度和所述第二检测框的置信度的乘积，作为融合结果。
  5. 根据权利要求1-4中任意一项所述的对象检测方法，其中，所述根据所述融合结果更新所述第一检测框的置信度，包括：
    根据所述融合结果确定多个检测框组；所述检测框组包括相互匹配的第一检测框和第二检测框；
    根据所述检测框组中所述第一检测框的置信度和所述第二检测框的置信度，确定所述检测框组的置信度；
    根据所述检测框组的置信度更新所述检测组中所述第一检测框的置信度。
  6. 根据权利要求5所述的对象检测方法，其中，所述根据所述检测框组中所述第一检测框的置信度和所述第二检测框的置信度，确定所述检测框组的置信度，包括：
    确定所述第一检测框对应的检测对象在跟踪时序中出现的次数；所述跟踪时序是基于不同时刻采集的图像对同一检测对象进行跟踪检测的检测结果的序列集合；
    根据所述次数确定所述第一检测框的置信度的第一权重、以及所述第二检测框的置信度的第二权重；
    根据所述第一权重和所述第二权重，确定所述第一检测框的置信度和所述第二检测框的置信度的加权平均值，作为所述检测框组的置信度。
  7. 根据权利要求1-4中任意一项所述的对象检测方法，其中，所述根 据置信度更新后的所述第一检测框确定所述检测对象，包括：
    在置信度更新后的所述第一检测框中确定大于第二置信度阈值的所述第一检测框，作为所述目标检测框，所述目标检测框对应的对象作为所述检测对象；
    或，在置信度更新后的所述第一检测框和置信度未更新的所述第一检测框中确定大于所述第二置信度阈值的所述第一检测框，作为所述目标检测框，所述目标检测框对应的对象作为所述检测对象；
    所述第二置信度阈值大于预设数值。
  8. 一种对象检测装置，包括：
    信息获取模块，用于获取第一帧图像中检测对象的第一检测框和第二帧图像中检测对象的第二检测框；所述第二帧图像的采集时刻在所述第一帧图像的采集时刻之前，所述第一检测框和所述第二检测框均携带有特征参数，所述特征参数包括置信度和位置信息；
    相似度确定模块，用于根据所述第一检测框的所述位置信息和所述第二检测框的所述位置信息，确定所述第一检测框和所述第二检测框之间的重叠度和距离中的至少一种相似度信息；
    信息融合模块，用于对所述重叠度和所述距离中的至少一种相似度信息、所述第一检测框的置信度、以及所述第二检测框的置信度进行融合，得到融合结果；
    信息更新模块，用于根据所述融合结果更新所述第一检测框的置信度；
    对象确定模块，用于根据置信度更新后的所述第一检测框确定所述检测对象。
  9. 根据权利要求8所述的对象检测装置，其中，所述对象检测装还包括：
    检测模块，用于根据目标检测模型对所述第一帧图像中的检测对象进行检测，得到多个初始检测框；
    所述检测模块，还用于在多个所述初始检测框中确定大于第一置信度阈值的所述初始检测框；所述第一置信度阈值为小于预设数值；
    所述检测模块，还用于根据非极大抑制算法对大于第一置信度阈值的 所述初始检测框进行筛选，将筛选出的所述初始检测框确定为所述第一检测框。
  10. 根据权利要求8所述的对象检测装置，其中，所述位置信息包括检测框轮廓的坐标信息；
    所述相似度确定模块，还用于根据所述第一检测框的所述坐标信息确定所述第一检测框的第一边界长度、第二边界长度和第一中心坐标；
    根据所述第二检测框的所述坐标信息确定所述第二检测框的第二中心坐标；
    根据所述第一边界长度、所述第二边界长度、所述第一中心坐标和所述第二中心坐标，确定所述第一检测框和所述第二检测框之间的距离。
  11. 根据权利要求8所述的对象检测装置，其中，
    所述信息融合模块，还用于对所述距离进行归一化处理；确定所述重叠度和归一化处理后的所述距离的加权平均值；确定所述加权平均值、所述第一检测框的置信度和所述第二检测框的置信度的乘积，作为融合结果。
  12. 根据权利要求8-11中任意一项所述的对象检测装置，其中，
    所述信息更新模块，还用于根据所述融合结果确定多个检测框组；所述检测框组包括相互匹配的第一检测框和第二检测框；根据所述检测框组中所述第一检测框的置信度和所述第二检测框的置信度，确定所述检测框组的置信度；根据所述检测框组的置信度更新所述检测组中所述第一检测框的置信度。
  13. 根据权利要求12所述的对象检测装置，其中，
    所述信息更新模块，还用于确定所述第一检测框对应的检测对象在跟踪时序中出现的次数；所述跟踪时序是基于不同时刻采集的图像对同一检测对象进行跟踪检测的检测结果的序列集合；根据所述次数确定所述第一检测框的置信度的第一权重、以及所述第二检测框的置信度的第二权重；根据所述第一权重和所述第二权重，确定所述第一检测框的置信度和所述第二检测框的置信度的加权平均值，作为所述检测框组的置信度。
  14. 根据权利要求8-11中任意一项所述的对象检测装置，其中，
    所述对象确定模块，还用于在置信度更新后的所述第一检测框中确定 大于第二置信度阈值的所述第一检测框，作为所述目标检测框，所述目标检测框对应的对象作为所述检测对象；
    所述对象确定模块，还用于在置信度更新后的所述第一检测框和置信度未更新的所述第一检测框中确定大于所述第二置信度阈值的所述第一检测框，作为所述目标检测框，所述目标检测框对应的对象作为所述检测对象；
    所述第二置信度阈值大于预设数值。
  15. 一种对象检测设备，包括：处理器，以及存储有计算机程序指令的存储器；
    所述处理器读取并执行所述计算机程序指令，以实现如权利要求1-7中任意一项所述的对象检测方法。
  16. 一种计算机存储介质，所述计算机存储介质上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现如权利要求1-7中任意一项所述的对象检测方法。
PCT/CN2022/092842 2021-05-14 2022-05-13 对象检测方法、装置、设备及计算机存储介质 WO2022237902A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110525549.2 2021-05-14
CN202110525549.2A CN112949785B (zh) 2021-05-14 2021-05-14 对象检测方法、装置、设备及计算机存储介质

Publications (1)

Publication Number Publication Date
WO2022237902A1 true WO2022237902A1 (zh) 2022-11-17

Family

ID=76233835

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/092842 WO2022237902A1 (zh) 2021-05-14 2022-05-13 对象检测方法、装置、设备及计算机存储介质

Country Status (2)

Country Link
CN (1) CN112949785B (zh)
WO (1) WO2022237902A1 (zh)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112949785B (zh) * 2021-05-14 2021-08-20 长沙智能驾驶研究院有限公司 对象检测方法、装置、设备及计算机存储介质
CN113695058B (zh) * 2021-10-28 2022-03-15 南通金驰机电有限公司 换热器生产用智能废料破碎装置自保护方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108021848A (zh) * 2016-11-03 2018-05-11 浙江宇视科技有限公司 客流量统计方法及装置
WO2020056903A1 (zh) * 2018-09-21 2020-03-26 北京字节跳动网络技术有限公司 用于生成信息的方法和装置
US20200265591A1 (en) * 2019-04-23 2020-08-20 Jiangnan University Visual Multi-Object Tracking based on Multi-Bernoulli Filter with YOLOv3 Detection
CN112150505A (zh) * 2020-09-11 2020-12-29 浙江大华技术股份有限公司 目标对象跟踪器的更新方法和装置、存储介质及电子装置
CN112749590A (zh) * 2019-10-30 2021-05-04 上海高德威智能交通系统有限公司 目标检测方法、装置、计算机设备和计算机可读存储介质
CN112949785A (zh) * 2021-05-14 2021-06-11 长沙智能驾驶研究院有限公司 对象检测方法、装置、设备及计算机存储介质

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108629354B (zh) * 2017-03-17 2020-08-04 杭州海康威视数字技术股份有限公司 目标检测方法及装置
CN108038452B (zh) * 2017-12-15 2020-11-03 厦门瑞为信息技术有限公司 一种基于局部图像增强的家电手势快速检测识别方法
US10748035B2 (en) * 2018-07-05 2020-08-18 Mitsubishi Electric Research Laboratories, Inc. Visually aided active learning for training object detector
CN111199165B (zh) * 2018-10-31 2024-02-06 浙江宇视科技有限公司 图像处理方法及装置
CN111783655A (zh) * 2020-06-30 2020-10-16 Oppo广东移动通信有限公司 一种图像处理方法、装置、电子设备和存储介质

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108021848A (zh) * 2016-11-03 2018-05-11 浙江宇视科技有限公司 客流量统计方法及装置
WO2020056903A1 (zh) * 2018-09-21 2020-03-26 北京字节跳动网络技术有限公司 用于生成信息的方法和装置
US20200265591A1 (en) * 2019-04-23 2020-08-20 Jiangnan University Visual Multi-Object Tracking based on Multi-Bernoulli Filter with YOLOv3 Detection
CN112749590A (zh) * 2019-10-30 2021-05-04 上海高德威智能交通系统有限公司 目标检测方法、装置、计算机设备和计算机可读存储介质
CN112150505A (zh) * 2020-09-11 2020-12-29 浙江大华技术股份有限公司 目标对象跟踪器的更新方法和装置、存储介质及电子装置
CN112949785A (zh) * 2021-05-14 2021-06-11 长沙智能驾驶研究院有限公司 对象检测方法、装置、设备及计算机存储介质

Also Published As

Publication number Publication date
CN112949785A (zh) 2021-06-11
CN112949785B (zh) 2021-08-20

Similar Documents

Publication Publication Date Title
WO2022237902A1 (zh) 对象检测方法、装置、设备及计算机存储介质
EP3295424B1 (en) Systems and methods for reducing a plurality of bounding regions
CN108388879B (zh) 目标的检测方法、装置和存储介质
JP2022548438A (ja) 欠陥検出方法及び関連装置、機器、記憶媒体、並びにコンピュータプログラム製品
CN109190446A (zh) 基于三元组聚焦损失函数的行人再识别方法
CN112633276B (zh) 训练方法、识别方法、装置、设备、介质
US8948522B2 (en) Adaptive threshold for object detection
CN111178161B (zh) 一种基于fcos的车辆追踪方法及系统
CN112016401A (zh) 基于跨模态行人重识别方法及装置
CN111597933B (zh) 人脸识别方法和装置
CN112801008A (zh) 行人重识别方法、装置、电子设备及可读存储介质
TWI789128B (zh) 人臉識別方法、裝置、設備及存儲介質
WO2021143865A1 (zh) 定位方法及装置、电子设备、计算机可读存储介质
CN112766218B (zh) 基于非对称联合教学网络的跨域行人重识别方法和装置
CN113780145A (zh) 精子形态检测方法、装置、计算机设备和存储介质
CN114692778B (zh) 用于智能巡检的多模态样本集生成方法、训练方法及装置
CN106778731A (zh) 一种车牌定位方法及终端
CN115471476A (zh) 一种部件缺陷检测方法、装置、设备及介质
CN112149601A (zh) 兼容遮挡的面部属性识别方法、装置和电子设备
CN113989721A (zh) 目标检测方法和目标检测模型的训练方法、装置
US20220392192A1 (en) Target re-recognition method, device and electronic device
CN112819953B (zh) 三维重建方法、网络模型训练方法、装置及电子设备
CN111368625A (zh) 一种基于级联优化的行人目标检测方法
CN116343099B (zh) 一种基于机器视觉的计算机屏幕信息防窃照系统
CN110942008A (zh) 一种基于深度学习的面单信息定位方法和系统

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: 22806864

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: 22806864

Country of ref document: EP

Kind code of ref document: A1