Multi-view animal identity recognition method and device

Abstract

The present invention relates to a multi-view animal identity recognition method and device. The multi-view animal identity recognition method comprises: acquiring single-animal behavior videos at the same moment in at least two views and multi-animal behavior videos at the same moment in at least two views. A deep learning model reuse strategy is used to successively and respectively train a deep learning instance segmentation model and a deep learning identity recognition model, an animal instance segmentation task and an animal identity recognition task are designed as complementary tasks, in light of video acquisition in multiple views, the problems of needing to manually label identity data, having difficulty in labeling an animal identity set and recognizing an animal identity, and having significant errors are directly solved, and the defects of having difficulty in obtaining a robust identity feature acquisition quantity and low identity recognition accuracy are further solved. Compared with the prior art, the multi-view animal identity recognition method and device disclosed in the present application can achieve the objectives of easily obtaining the robust identity feature acquisition quantity and improving the identity recognition accuracy.

Description

A multi-perspective animal identification method and identification device Technical Field

[0001] The present invention relates to the field of visual recognition technology, and in particular to a multi-viewing animal identification method and an identification device thereof. Background Art

[0002] In the development of drugs for social disorders such as autism, anxiety, and phobias, identifying behavioral differences in animals before and after drug administration is an important indicator for judging drug efficacy.

[0003] However, due to limitations in animal tracking technology, existing methods are unable to identify each animal during their natural social interactions. They only allow one animal to move freely, and the others must be restrained to minimize interference with their identities, as in the three-box social experiment. This limitation stems from the animals' similar appearance, making it difficult for traditional algorithms to reduce errors in identifying animals. This makes it difficult to label animal identity sets, making animal identification difficult and error-prone. This translates to difficulties in obtaining robust identity features and low identification accuracy.

[0004] Summary of the Invention

[0005] In order to solve the problems of difficulty in labeling animal identity sets, difficulty in animal identity recognition and large errors, which lead to difficulty in obtaining robust identity features and low identity recognition accuracy, the present invention proposes a multi-view animal identity recognition method and its recognition device.

[0006] The technical solution adopted by the present invention is a multi-view animal identification method, comprising:

[0007] Obtaining single-animal behavior videos at the same moment from at least two viewing angles, and multi-animal behavior videos at the same moment from at least two viewing angles;

[0008] Extract animal instance outlines from single-animal behavior videos and multi-animal behavior videos, and use the animal instance outlines to train deep learning instance segmentation models;

[0009] Perform instance segmentation on single-animal behavior videos and multi-animal behavior videos based on a deep learning instance segmentation model, obtain single-animal segmentation instances from single-animal behavior videos, and obtain multi-animal segmentation instances from multi-animal behavior videos;

[0010] The single animal segmentation instances from different perspectives are stitched together, and the stitched images are used as input patterns, which are then used to train a deep learning identity recognition model.

[0011] Multiple animal segmentation instances from different perspectives are spliced ​​together to obtain partial or complete patterns of a single animal. Based on the deep learning identity recognition model, the partial or complete patterns of a single animal are identified and matched to obtain the identity information of the single animal.

[0012] Preferably, the step of extracting animal instance outlines from the single-animal behavior video and the multi-animal behavior video includes:

[0013] The overlapping parts of animals occluded at different times in the single-animal behavior video are extracted as single-animal background video, and the overlapping parts of animals occluded at different times in the multi-animal behavior video are extracted as multi-animal background video;

[0014] In single animal behavior videos and multi-animal behavior videos, single animal background videos and multi-animal background videos are eliminated respectively to obtain animal instance outlines.

[0015] Preferably, after obtaining the animal instance outline, the method further includes synthesizing the animal instance outline with different backgrounds to generate an image containing the animal instance outline annotation.

[0016] Preferably, the deep learning instance segmentation model is trained based on YOLACT++.

[0017] Preferably, the deep learning identity recognition model is trained based on the deep learning image classification model of EfficientNet.

[0018] Preferably, the step of splicing multiple animal segmentation instances from different perspectives to obtain a partial or complete image of a single animal includes:

[0019] Partial or complete single-animal segmentation instances of a single animal are obtained from multiple animal segmentation instances, and then partial or complete single-animal segmentation instances of a single animal under different perspectives are spliced ​​to obtain a partial or complete pattern of a single animal.

[0020] Preferably, partial or complete single animal segmentation instances of a single animal are obtained from multiple animal segmentation instances, and the projection relationship in the multiple animal segmentation instances is obtained through video shooting parameters. After adjusting the size of the projection, the projections from different perspectives are spliced ​​to obtain partial or complete patterns of a single animal.

[0021] Preferably, the projection relationship in the multi-animal segmentation instance is obtained, the projections under different viewing angles are compared and matched with the single animal segmentation instance in the database, the size of the matched projections is adjusted, and then the projections under different viewing angles are spliced.

[0022] Preferably, the single animal segmentation instances under different viewing angles are spliced ​​together, and the spliced ​​images are one picture.

[0023] The present invention also proposes a multi-view animal identification device, comprising:

[0024] A multi-camera array behavior acquisition unit that acquires single-animal behavior videos at the same time from at least two viewing angles, and multi-animal behavior videos at the same time from at least two viewing angles;

[0025] An instance segmentation model training unit extracts animal instance outlines from single-animal behavior videos and multi-animal behavior videos, and uses the animal instance outlines to train a deep learning instance segmentation model;

[0026] The instance segmentation unit performs instance segmentation on single-animal behavior videos and multi-animal behavior videos based on a deep learning instance segmentation model, obtaining single-animal segmentation instances from single-animal behavior videos and multi-animal segmentation instances from multi-animal behavior videos;

[0027] The single animal identification unit stitches together the single animal segmentation instances from different perspectives and uses the stitched images as input patterns to train the deep learning identification model.

[0028] The multi-animal identification unit splices multiple animal segmentation instances from different perspectives to obtain partial or complete patterns of a single animal, and identifies and matches the partial or complete patterns of a single animal based on a deep learning identification model to obtain the identity information of a single animal.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The present application discloses a multi-perspective animal identification method, comprising: obtaining a single-animal behavior video at the same moment from at least two perspectives, and a multi-animal behavior video at the same moment from at least two perspectives; extracting animal instance contours from the single-animal behavior video and the multi-animal behavior video, and using the animal instance contours to train a deep learning instance segmentation model; performing instance segmentation on the single-animal behavior video and the multi-animal behavior video based on the deep learning instance segmentation model, obtaining a single-animal segmentation instance from the single-animal behavior video, and obtaining a multi-animal segmentation instance from the multi-animal behavior video; splicing the single-animal segmentation instances from different perspectives, using the spliced ​​image as an input pattern, and using the input pattern to train a deep learning identification model; splicing the multi-animal segmentation instances from different perspectives to obtain a partial or full pattern of a single animal, and identifying and matching the partial or full pattern of the single animal based on the deep learning identification model, thereby obtaining the identity information of the single animal. This application uses a deep learning model reuse strategy to train the deep learning instance segmentation model and the deep learning identity recognition model respectively, designing the animal instance segmentation task and the animal identity recognition task as complementary tasks, and combining video acquisition from multiple perspectives to directly solve the problems of manually labeling identity data, difficulty in labeling animal identity sets, and difficulty and large error in animal identity recognition, and further solve the defects of difficulty in obtaining robust identity features and low identity recognition accuracy.

[0031] The present application also discloses a multi-perspective animal identification device, comprising: a multi-camera array behavior acquisition unit, which obtains single-animal behavior videos at the same time from at least two perspectives, and multi-animal behavior videos at the same time from at least two perspectives; an instance segmentation model training unit, which extracts animal instance contours from single-animal behavior videos and multi-animal behavior videos, and uses the animal instance contours to train a deep learning instance segmentation model; an instance segmentation unit, which performs instance segmentation on single-animal behavior videos and multi-animal behavior videos based on a deep learning instance segmentation model, obtains single-animal segmentation instances from single-animal behavior videos, and obtains multi-animal segmentation instances from multi-animal behavior videos; a single-animal identification unit, which splices single-animal segmentation instances from different perspectives, uses the spliced ​​image as an input pattern, and uses the input pattern to train a deep learning identification model; a multi-animal identification unit, which splices multi-animal segmentation instances from different perspectives to obtain partial or full patterns of a single animal, and identifies and matches partial or full patterns of a single animal based on a deep learning identification model, thereby obtaining identity information of the single animal. It directly solves the problems of needing to manually label identity data, difficulty in labeling animal identity sets, difficulty in animal identity recognition and large errors, and further solves the defects of difficulty in obtaining robust identity features and low identity recognition accuracy.

[0032] Compared with the prior art, the multi-view animal identification method and identification device disclosed in the present application can achieve the purpose of easily obtaining robust identity features and improving the accuracy of identity identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The present invention is described in detail below with reference to the embodiments and accompanying drawings, in which:

[0034] FIG1 shows a schematic flow chart of a multi-view animal identification method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] To make the objectives, technical solutions, and advantages of the present invention more apparent, embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar components or components having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0036] The present invention discloses a multi-view animal identification method, as shown in FIG1 , comprising:

[0037] S10: Acquire a single animal behavior video at the same moment from at least two viewing angles, and a multi-animal behavior video at the same moment from at least two viewing angles;

[0038] S20: Extract animal instance outlines from single-animal behavior videos and multi-animal behavior videos, and use the animal instance outlines to train a deep learning instance segmentation model;

[0039] S30: Perform instance segmentation on single-animal behavior videos and multi-animal behavior videos based on a deep learning instance segmentation model, obtain single-animal segmentation instances from single-animal behavior videos, and obtain multi-animal segmentation instances from multi-animal behavior videos;

[0040] S40: Splicing single animal segmentation instances from different perspectives, using the spliced ​​images as input patterns, and using the input patterns to train a deep learning identity recognition model;

[0041] S50: Splicing multiple animal segmentation instances from different perspectives to obtain a partial or complete pattern of a single animal, and identifying and matching the partial or complete pattern of a single animal based on a deep learning identity recognition model to obtain the identity information of the single animal.

[0042] This application uses a deep learning model reuse strategy to train the deep learning instance segmentation model and the deep learning identity recognition model respectively, designing the animal instance segmentation task and the animal identity recognition task as complementary tasks, and combining video acquisition from multiple perspectives to directly solve the problems of manually labeling identity data, difficulty in labeling animal identity sets, and difficulty and large error in animal identity recognition, and further solve the defects of difficulty in obtaining robust identity features and low identity recognition accuracy.

[0043] It's important to note that attaching physical tags to animals is an effective method for identifying their identities. However, these tags can affect the behavior of animals that are only social, and increasing the number of animals can also increase overall equipment costs. Furthermore, directly identifying animals through images is an alternative to physical tags. Using biometric features like nose prints to identify animals is a method that has only been validated in certain species, such as cats and dogs, and is not applicable to mice, which are commonly used in drug development.

[0044] Different from the above-mentioned method of identifying animal identities, the present application identifies animal identities through deep learning, and trains the deep learning instance segmentation model through a large number of animal instance contours obtained from single-animal behavior videos and multi-animal behavior videos, ensuring that the deep learning instance segmentation model has a large number of identified animal instance contours, and by splicing single animal segmentation instances under multiple perspectives, a large number of input patterns are obtained, and then the deep learning identity recognition model is trained with a large number of input patterns, so that the deep learning identity recognition model has a large number of identified input patterns. Therefore, when the appearance of animals is extremely similar, they can also be identified according to the deep learning identity recognition model.

[0045] This application also addresses the difficulty of obtaining sufficient training data when deep learning is used to identify multiple animals. Furthermore, when adding a new animal, only single-animal behavior videos of the new animal, captured from the same moment in time from at least two different perspectives, are required to automatically identify the new animal.

[0046] It should also be noted that acquiring videos of the same moment from at least two perspectives (single-animal behavior videos and multi-animal behavior videos) is intended to address the issues of feature loss and perspective deviation that occur in image-based animal identification, and to comprehensively capture stable and complete animal appearance features from multiple perspectives. In this application, at least two videos of the same moment from the same perspective are required. Obviously, videos from more perspectives can be captured based on actual needs. Correspondingly, the recognition results are more accurate, the time required to collect animal instance outlines is faster, and the computational complexity is greater.

[0047] Among them, the animal instance contour is a graphic obtained by cutting out the animal part in the video. The same animal has different animal instance contours in different postures and at different distances from the camera. In order to train the deep learning instance segmentation model in the subsequent steps, it is necessary to obtain the animal instance contour in advance to serve the subsequent instance segmentation.

[0048] When performing instance segmentation, instances are segmented in the video according to the trained deep learning instance segmentation model. The contour area is obtained after computational recognition in the computer vision task. Specifically, the contour area can be obtained by computationally recognizing pixel points.

[0049] The stitched images serve as input patterns, and the stitched images can include not only static images but also continuous videos. For example, a continuous video can be used as input pattern to capture a certain animal movement or behavior, thereby achieving higher recognition accuracy.

[0050] In some embodiments, the step of extracting animal instance outlines from single-animal behavior videos and multi-animal behavior videos includes:

[0051] The overlapping parts of animals occluded at different times in the single-animal behavior video are extracted as single-animal background video, and the overlapping parts of animals occluded at different times in the multi-animal behavior video are extracted as multi-animal background video;

[0052] In single animal behavior videos and multi-animal behavior videos, single animal background videos and multi-animal background videos are eliminated respectively to obtain animal instance outlines.

[0053] Specifically, the step of extracting the outline of the animal instance in the video first extracts the single-animal background video and the multi-animal background video from the single-animal behavior video and the multi-animal behavior video. The purpose of extracting the single-animal background video and the multi-animal background video is to ensure that the obtained animal instance outline is not interfered with by the background video (single-animal background video and multi-animal background video). In addition, it also serves as a key step in subsequently eliminating the background video from the single-animal behavior video and the multi-animal behavior video. It should be noted that since the animals in the single-animal behavior video and the multi-animal behavior video are not in a motionless state at all times, when the animal walks away, the part of the background blocked by the animal can be identified. Therefore, the animal background can be obtained by extracting the overlapping parts blocked by the animal at different times.

[0054] In some specific embodiments, after obtaining the animal instance outline, the method further includes synthesizing the animal instance outline with different backgrounds to generate an image containing the animal instance outline annotation.

[0055] It should be noted that after extracting the animal instance outline, different animal instance outlines can be synthesized with different backgrounds. For example, if it is necessary to identify an animal in a certain background environment, the background environment can be synthesized with the animal instance outline and the synthesized image can be labeled. This can further improve the accuracy of subsequent recognition.

[0056] In some embodiments, a deep learning instance segmentation model is trained based on YOLACT++.

[0057] Specifically, YOLACT++ has near real-time dynamic object segmentation capabilities, with a frame rate of nearly 40 per second and extremely high accuracy. In addition, YOLACT++ is an improved version of YOLACT, which includes a deeper backbone network, a larger feature pyramid, and the introduction of the ProtoNet mechanism, which is used to better segment and represent the shape of the animal instance outline.

[0058] In some embodiments, a deep learning identity recognition model is trained based on a deep learning image classification model based on EfficientNet.

[0059] Specifically, EfficientNet is a new convolutional neural network architecture proposed in recent years. Its main goal is to improve the efficiency of the model, while maintaining or improving the performance of the model, while minimizing the complexity and computational requirements of the model. It effectively utilizes the model parameters by scaling the depth, width, and resolution simultaneously in compound scaling, thereby improving the model's performance.

[0060] In some embodiments, the step of splicing multiple animal segmentation instances from different perspectives to obtain a partial or complete image of a single animal includes:

[0061] Partial or complete single-animal segmentation instances of a single animal are obtained from multiple animal segmentation instances, and then partial or complete single-animal segmentation instances of a single animal under different perspectives are spliced ​​to obtain a partial or complete pattern of a single animal.

[0062] It should be noted that, in the above steps, multiple animal segmentation instances under different viewing angles can be spliced ​​first, and then a partial or full pattern of a single animal can be obtained from the spliced ​​partial or full patterns of multiple animals.

[0063] Specifically, first obtaining part or all of the single-animal segmentation instances of a single animal from the multiple-animal segmentation instances can reduce the amount of calculation. Since in some recognition scenarios, only certain specific animal individuals need to be identified, and not all animal individuals need to be identified, if the multiple animal segmentation instances under different perspectives are spliced ​​first, the calculation will be redundant during the splicing process, and the response speed will be slower.

[0064] In some specific embodiments, partial or complete single-animal segmentation instances of a single animal are obtained from multiple-animal segmentation instances, and the projection relationship in the multiple-animal segmentation instances is obtained through video shooting parameters. After adjusting the size of the projection, the projections from different perspectives are spliced ​​to obtain a partial or complete pattern of a single animal.

[0065] Specifically, the video shooting parameters include the distance between the individual animals and the camera, the spacing between different cameras, the viewing angle of the camera, the zoom length of the camera, and other parameters. The above parameter information is used to obtain the projection relationship in the multi-animal segmentation instance, and after splicing, a partial or complete image of a single animal is obtained. It should be noted that due to the obstruction of different animals or the environment, it is sometimes impossible to obtain a complete single animal segmentation instance from the multi-animal segmentation instance. More often, a partial single animal segmentation instance is obtained. Therefore, after splicing, only a partial image of a single animal can be obtained.

[0066] In some further embodiments, the projection relationship in multiple animal segmentation instances is obtained, the projections at different perspectives are compared and matched with the single animal segmentation instances in the database, the size of the matched projections is adjusted, and then the projections at different perspectives are spliced.

[0067] Specifically, after obtaining the projections and before stitching them together, the obtained projections are compared and matched with single animal segmentation instances in the database. The single animal segmentation instances in the database refer to the single animal segmentation instances in the database. This determines whether the projections of the single animal from different perspectives fall into the single animal segmentation instances in the database. If the comparison and match are successful, the size of the matched projections is adjusted before stitching. This improves recognition accuracy and avoids incorrect stitching processes.

[0068] In some embodiments, single animal segmentation instances from different perspectives are spliced ​​together, and the spliced ​​images are a single picture.

[0069] Specifically, in order to reduce the amount of calculation and increase the recognition speed, the spliced ​​images are output as one picture.

[0070] The present invention discloses a multi-view animal identification device, comprising:

[0071] A multi-camera array behavior acquisition unit that acquires single-animal behavior videos at the same time from at least two viewing angles, and multi-animal behavior videos at the same time from at least two viewing angles;

[0072] An instance segmentation model training unit extracts animal instance outlines from single-animal behavior videos and multi-animal behavior videos, and uses the animal instance outlines to train a deep learning instance segmentation model;

[0073] The instance segmentation unit performs instance segmentation on single-animal behavior videos and multi-animal behavior videos based on a deep learning instance segmentation model, obtaining single-animal segmentation instances from single-animal behavior videos and multi-animal segmentation instances from multi-animal behavior videos;

[0074] The single animal identification unit stitches together the single animal segmentation instances from different perspectives and uses the stitched images as input patterns to train the deep learning identification model.

[0075] The multi-animal identification unit splices multiple animal segmentation instances from different perspectives to obtain partial or complete patterns of a single animal, and identifies and matches the partial or complete patterns of a single animal based on a deep learning identification model to obtain the identity information of a single animal.

[0076] It directly solves the problems of needing to manually label identity data, difficulty in labeling animal identity sets, difficulty in animal identity recognition and large errors, and further solves the defects of difficulty in obtaining robust identity features and low identity recognition accuracy.

[0077] Specifically, the multi-view animal identification device includes a multi-camera array behavior acquisition unit for shooting single animal behavior videos and multi-animal behavior videos at different perspectives.

[0078] Preferably, the multi-camera array behavior acquisition unit is a multi-camera calibration system in the patent publication number CN112862900B and the patent name "Multi-view camera calibration device, calibration method and storage medium".

[0079] By calibrating the camera, we obtain its internal and external parameters. We calculate the projection relationship among multiple animal segmentation instances in multiple perspectives based on the camera parameters. We match single animal segmentation instances belonging to the same animal based on the projection distance, resize them, and stitch them together. The images are then input into a deep learning identity recognition model to infer the identity of each animal.

[0080] In this specification, the use of terms such as "Embodiment 1," "this embodiment," and "in one embodiment" indicates that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in the invention or at least one embodiment or example of the invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example; furthermore, the specific features, structures, materials, or characteristics described may be appropriately combined in any one or more embodiments or examples.

[0081] In the description of this specification, the terms "connect," "install," "fix," "dispose," and "have" are to be understood in a broad sense. For example, "connect" can mean a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in this application based on specific circumstances.

[0082] In the description of this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising the element.

[0083] The above description of the embodiments is to facilitate ordinary technicians in this technical field to understand and apply the technology of this case. People familiar with the technology in this field can obviously make various modifications to these examples easily and apply the general principles described here to other embodiments without having to go through creative work. Therefore, this case is not limited to the above embodiments. Modifications to the following situations should all be within the scope of protection of this case: ① A new technical solution implemented based on the technical solution of the present invention and combined with existing common knowledge, the technical effect produced by the new technical solution does not exceed the technical effect of the present invention; ② The equivalent replacement of some features of the technical solution of the present invention with the known technology, the technical effect produced is the same as the technical effect of the present invention; ③ The technical solution of the present invention is expandable, and the substantive content of the expanded technical solution does not exceed the technical solution of the present invention; ④ The equivalent transformation made by the content of the description and drawings of the present invention is directly or indirectly applied to other related technical fields.

Claims

1. A multi - perspective animal identity recognition method, characterized in that, it includes: Obtain single - animal behavior videos at the same time under at least two perspectives, and multi - animal behavior videos at the same time under at least two perspectives; Extract the animal instance contours in the single - animal behavior videos and multi - animal behavior videos, and use the animal instance contours to train a deep - learning instance segmentation model; Based on the deep - learning instance segmentation model, perform instance segmentation on the single - animal behavior videos and multi - animal behavior videos, obtain single - animal segmentation instances from the single - animal behavior videos, and obtain multi - animal segmentation instances from the multi - animal behavior videos; Stitch the single - animal segmentation instances under different perspectives, and use the stitched image as an input pattern, and use the input pattern to train a deep - learning identity recognition model; Stitch the multi - animal segmentation instances under different perspectives to obtain partial or all patterns of a single animal, and perform recognition and matching on the partial or all patterns of a single animal based on the deep - learning identity recognition model, so as to obtain the identity information of a single animal.

2. The multi - perspective animal identity recognition method according to claim 1, characterized in that, The step of extracting the animal instance contours in the single - animal behavior videos and multi - animal behavior videos includes: Extract the overlapping parts of the animals blocked at different times in the single - animal behavior video as a single - animal background video, and extract the overlapping parts of the animals blocked at different times in the multi - animal behavior video as a multi - animal background video; Eliminate the single - animal background video and multi - animal background video in the single - animal behavior video and multi - animal behavior video respectively, so as to obtain the animal instance contours.

3. The multi - perspective animal identity recognition method according to claim 2, characterized in that, After obtaining the animal instance contours, it further includes synthesizing the animal instance contours with different backgrounds to generate an image with animal instance contour annotations.

4. The multi - perspective animal identity recognition method according to claim 1, characterized in that, Train the deep - learning instance segmentation model based on YOLACT++.

5. The multi - perspective animal identity recognition method according to claim 1, characterized in that, Train the deep - learning identity recognition model based on the deep - learning image classification model of EfficientNet.

6. The multi - perspective animal identity recognition method according to claim 1, characterized in that, The step of stitching the multi - animal segmentation instances under different perspectives to obtain partial or all patterns of a single animal includes: Obtain partial or all single - animal segmentation instances of a single animal from the multi - animal segmentation instances, and then stitch the partial or all single - animal segmentation instances of a single animal under different perspectives to obtain partial or all patterns of a single animal.

7. The multi - perspective animal identity recognition method according to claim 6, characterized in that, Obtain partial or all single - animal segmentation instances of a single animal from the multi - animal segmentation instances, obtain the projection relationship in the multi - animal segmentation instances through video shooting parameters, adjust the size of the projection, and then stitch the projections under different perspectives to obtain partial or all patterns of a single animal.

8. A multi-view animal identification method according to claim 7, characterized in that, obtain the projection relationship in the multi-animal segmentation instances, compare and match the projections from different views with the single-animal segmentation instances in the database, adjust the size of the matched projections, and then splice the projections from different views.

9. A multi-view animal identification method according to claim 1, characterized in that, splice the single-animal segmentation instances from different views, and the spliced image is a single picture.

10. A multi-view animal identification device, characterized in that, comprising: a multi-camera array behavior acquisition unit, which acquires single-animal behavior videos at the same moment from at least two views, and multi-animal behavior videos at the same moment from at least two views; an instance segmentation model training unit, which extracts the animal instance contours in the single-animal behavior videos and multi-animal behavior videos, and uses the animal instance contours to train a deep learning instance segmentation model; an instance segmentation unit, which performs instance segmentation on the single-animal behavior videos and multi-animal behavior videos based on the deep learning instance segmentation model, obtains single-animal segmentation instances from the single-animal behavior videos, and obtains multi-animal segmentation instances from the multi-animal behavior videos; a single-animal identification unit, which splices the single-animal segmentation instances from different views, uses the spliced image as an input pattern, and uses the input pattern to train a deep learning identification model; a multi-animal identification unit, which splices the multi-animal segmentation instances from different views to obtain part or all of the patterns of a single animal, and performs identification and matching on part or all of the patterns of a single animal based on the deep learning identification model, so as to obtain the identity information of a single animal.

Images