Deep learning-based automatic recognition system for musk deer and sympatric ungulate footprint images

Abstract

The application discloses a deep learning-based musk deer and sympatric ungulate footprint image automatic identification system and relates to the technical field of animal footprint image identification; the system comprises a behavior rhythm extraction module which performs footprint instance segmentation on an original image, constructs a continuous footprint chain and outputs a locking footprint in a discrete jumping type footprint chain; a shape segmentation processing module which performs depth estimation on the locking footprint, obtains a normalized relative treading depth profile and a main depth extreme point; a depth dynamic feature confirmation module which extracts a forward centroid offset, a puncture shear steepness and a hoof suspension deviation degree based on the profile, generates a musk deer footprint confirmation index and obtains a classification result; and a structured archiving module which generates a structured identification record or performs shunt archiving according to the classification result.

Description

Technical Field

[0001] This invention relates to the field of animal footprint image recognition technology, and more specifically, to an automatic recognition system for musk deer and co-located even-toed ungulate footprint images based on deep learning. Background Technology

[0002] Musk deer are mostly distributed in forests, shrublands, and alpine meadows, characterized by their elusive activities, complex distribution areas, and high difficulty in direct observation. Footprint images are a commonly used indirect identification tool in field surveys, population monitoring, and habitat assessment. Existing animal footprint identification methods typically include manual on-site interpretation, expert review after image acquisition, and image classification based on machine vision or deep learning models. Among these, image-based automatic identification methods generally extract information such as footprint outline, size ratio, texture morphology, footprint center position, or footprint sequence distribution to classify and judge the target animal's footprints, thereby improving the efficiency of field sample processing and the consistency of identification results. However, the current problem is that musk deer footprints share certain similarities with those of other sympatric even-toed ungulates such as the Chinese serow, blue sheep, and domestic sheep in terms of outer contour, hoof structure, and local indentation morphology. Furthermore, field footprint images are easily affected by factors such as substrate moisture, shooting angle, lighting and shadows, footprint fragmentation, overlapping trampling, and lack of scale information. Existing identification methods, which primarily rely on the planar morphology of individual footprints or human experience, struggle to reliably distinguish the subtle differences between musk deer footprints and those of sympatric even-toed ungulates, and also find it difficult to standardize low-quality samples, atypical indentation samples, and samples requiring verification.

[0003] To address the above problems, this invention proposes a solution. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an automatic recognition system for footprint images of musk deer and co-located even-toed ungulates based on deep learning, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The system for automatic recognition of footprint images of musk deer and unilaterally distributed even-toed ungulates based on deep learning includes the following modules: a behavior rhythm extraction module is used to perform footprint instance segmentation on the original image, construct a continuous footprint chain, and output the footprints in the footprint chain with the rhythm label of discrete jump as locked footprints. The topography segmentation module is used to estimate the depth of the locked footprints to obtain a normalized relative trampling depth profile, and obtain the normalized longitudinal coordinates of the main depth extremum points in the search window at the front of the profile. The depth dynamic feature verification module is used to obtain the equivalent force center based on the weighted normalized relative trampling depth profile, and to obtain the positive centroid offset from the directional distance of the equivalent force center relative to the main depth extreme point. When the positive centroid offset meets the preset front-end concentrated force screening condition, the effective force interval is determined along the hoof and heel direction based on the positive centroid offset, starting from the normalized longitudinal coordinate of the main depth extreme point, and the puncture shear steepness is obtained within this effective force interval. The attenuation scale is determined based on the relative relationship between the positive centroid offset and the puncture shear steepness, and the hoof and heel suspension transition boundary is searched based on the attenuation scale. The hoof and heel suspension deviation is obtained based on the positional relationship between the hoof and heel suspension transition boundary relative to the main depth extreme point and the attenuation scale. The musk deer footprint verification index is generated based on the positive centroid offset, puncture shear steepness, and hoof and heel suspension deviation, and the classification result is obtained based on the musk deer footprint verification index. The classification result includes one of the following: musk deer candidate verification, non-musk candidate, and pending verification. The structured archiving module is used to generate structured identification records or perform split archiving based on classification results.

[0006] In a preferred embodiment, the behavior rhythm extraction module calculates footprint density, step length dynamic variance, and trajectory curvature for each continuous footprint chain. The footprint density is the ratio of the number of footprints belonging to the continuous footprint chain to the area benchmark, and the area benchmark includes at least the ground area converted according to the ground scale conversion factor. The step length dynamic variance is the sample variance of the step length sequence formed by the center distance between adjacent footprints in the footprint chain. The trajectory curvature is the mean of the ratio of the absolute value of the change in yaw angle between adjacent steps in the footprint chain to the corresponding step length. Based on footprint density threshold, step length variance threshold, and curvature threshold, rhythmic labels are assigned: footprint chains with footprint density higher than the footprint density threshold are labeled as high-density trampling type labels; footprint chains with footprint density not higher than the footprint density threshold and step length dynamic variance not higher than the step length variance threshold are labeled as low-density uniform type labels; footprint chains with footprint density not higher than the footprint density threshold, step length dynamic variance higher than the step length variance threshold, and trajectory curvature not lower than the curvature threshold are labeled as discrete jumping type labels; the remaining footprint chains are labeled as labels to be reviewed.

[0007] In a preferred embodiment, the footprint instance segmentation model used by the behavior rhythm extraction module is an instance segmentation network with a convolutional neural network or a visual Transformer as its backbone, and it is trained using a training set containing musk deer footprint samples and sympathetic even-toed ungulate footprint samples. The sympathetic even-toed ungulates include at least the Chinese serow, blue sheep, and domestic sheep. After segmentation, a binary mask, bounding box, and center point coordinates are obtained for each footprint, wherein the center point coordinates are used as input data for calculating the distance and direction between adjacent footprints when constructing a continuous footprint chain.

[0008] In a preferred embodiment, the depth estimation model used by the shape segmentation processing module is a deep learning network with an encoder-decoder structure. The encoder is used to extract multi-scale features of footprint texture and shadow, and the decoder restores spatial resolution through multi-level upsampling and skip connections, and outputs a single-channel depth matrix corresponding to the working image of the locked footprint. Each element in the single-channel depth matrix represents the relative trampling depth of the corresponding pixel position normalized to the interval [0, 1]. The normalized relative trampling depth profile is obtained by sampling the single-channel depth matrix along the longitudinal main axis of the footprint.

[0009] In a preferred embodiment, after obtaining the normalized longitudinal coordinates of the main depth extreme point, the topography segmentation processing module performs four-zone adaptive segmentation on the normalized relative trampling depth profile based on the first depth attenuation coefficient and the second depth attenuation coefficient, and the first depth attenuation coefficient is greater than the second depth attenuation coefficient. When searching from the normalized longitudinal coordinate of the principal depth extreme point towards the hoof and heel direction, the first boundary is obtained when the normalized relative trampling depth profile first drops to a position below the product of the principal depth extreme value and the first depth attenuation coefficient. When continuing the search from the first boundary toward the hoof and heel, the normalized relative trampling depth profile first drops to a position below the product of the principal depth extreme value and the second depth attenuation coefficient, thus obtaining the second boundary; The section from the hoof tip to the first boundary is the sharp-compression core zone. Centered on the normalized longitudinal coordinate of the main depth extreme point, the local section is determined based on the local morphology of the normalized relative trampling depth profile near the main depth extreme point, resulting in the double-lobe crack signature zone. The section from the first boundary to the second boundary is the main load attenuation zone, and the section from the second boundary to the hoof heel is the hoof heel light load wake zone. The longitudinal width of the sharp-compression core zone is used as a morphological feature parameter input into the species subdivision model along with the positive centroid offset and the puncture shear steepness during species subdivision. The transverse depth profile extracted along the vertical main axis within the double-lobe crack signature zone is used as an auxiliary feature parameter characterizing the morphology of the left and right hoof lobes.

[0010] In a preferred embodiment, when the positive centroid offset is greater than a preset centroid offset threshold, the depth dynamic feature verification module determines that the locked footprint meets the preset front-end concentrated force screening condition and continues to perform the calculation of the puncture shear steepness; when the positive centroid offset is not greater than the preset centroid offset threshold, the depth dynamic feature verification module determines that the locked footprint does not meet the preset front-end concentrated force screening condition, triggers the first-order chain break feedback and stops the subsequent verification process, and outputs the locked footprint and its first-order chain break mark to the structured archiving module; wherein, the situation where the positive centroid offset is not greater than the preset centroid offset threshold includes: the equivalent force center is not located on the hoof heel side of the main depth extreme point, or although the equivalent force center is located on the hoof heel side of the main depth extreme point, its offset relative to the main depth extreme point in the hoof heel direction does not exceed the preset centroid offset threshold.

[0011] In a preferred embodiment, when the puncture shear steepness is not lower than a preset shear steepness threshold, the depth dynamic feature verification module determines that the locked footprint has passed the front-end puncture shear screening and continues to perform the determination of the attenuation scale and the search for the hoof and heel suspension transition boundary; when the puncture shear steepness is lower than the preset shear steepness threshold, the depth dynamic feature verification module triggers the second-order chain break feedback and stops the subsequent verification process, and outputs the locked footprint and its second-order chain break mark to the structured archiving module.

[0012] In a preferred embodiment, the depth dynamic feature verification module calculates the local descent rate within the scanning window, where the local descent rate is an estimate of the discrete derivative of the normalized relative trampling depth profile along the normalized longitudinal coordinate within the scanning window. When the local descent rate within the scanning window changes from a state not lower than the puncture shear steepness to a state lower than the product of the puncture shear steepness and a preset local rate attenuation ratio, the location where this change first occurs is taken as the hoof-heel suspension transition boundary, and the calculation of the hoof-heel suspension deviation continues. If there is no location within the scanning window that meets the above-mentioned change condition, the depth dynamic feature verification module determines that the hoof-heel suspension transition boundary is missing and triggers the third-order chain break feedback, outputting the locked footprint and its third-order chain break mark to the structured archiving module. The preset local rate attenuation ratio is greater than 0 and less than 1.

[0013] In a preferred embodiment, when the classification result is a candidate for musk deer, the deep dynamic feature confirmation module inputs the feature parameters in the morphological feature vector along with the positive centroid offset, puncture shear steepness, and hoof-heel overhang deviation into the species subdivision model to obtain the species candidate label. The species subdivision model is one of the following: interval template matching model, support vector machine, random forest, gradient boosting tree, or lightweight neural network, and the species template interval obtained from the training samples is used as the discrimination criterion of the model.

[0014] In a preferred embodiment, the chain break markers output by the deep dynamic feature verification module include first-order chain break markers, second-order chain break markers, and third-order chain break markers. The structured archiving module performs triage archiving based on the chain break markers output by the deep dynamic feature verification module: the locked footprints corresponding to the first-order chain break markers are assigned to the first-order chain break candidate sample sub-library, the locked footprints corresponding to the second-order chain break markers are assigned to the second-order chain break candidate sample sub-library, and the locked footprints corresponding to the third-order chain break markers are assigned to the third-order chain break candidate sample sub-library. The structured archiving module generates structured recognition records when the classification result is a musk deer candidate confirmation, generates non-musk deer candidate records when it is a non-musk deer candidate, and transfers the samples to the manual review queue when the sample chain length is less than a preset chain length threshold, the single-channel depth matrix variance is lower than a preset depth variance threshold, or the normalized relative trampling depth profile energy is lower than a preset effective indentation energy threshold, suspends the current judgment process, and transfers the samples to the manual review queue.

[0015] The technical effects and advantages of this invention's deep learning-based automatic recognition system for musk deer and unilaterally distributed even-toed ungulate footprint images are as follows: This invention, through the coordinated operation of a behavioral rhythm extraction module, a morphology segmentation module, a depth dynamic feature verification module, and a structured archiving module, forms a complete automatic recognition process from original footprint image input, continuous footprint chain screening, footprint depth and morphology analysis, depth dynamic feature verification, to archiving of recognition results. First, by segmenting footprint instances and constructing continuous footprint chains, the rhythmic features of the footprint chains are incorporated into the recognition process, and footprints in discrete, jumping footprint chains are used as locked footprints. This reduces the impact of high-density trampling, low-density uniform gait, and other interfering footprints on subsequent recognition, improving the relevance of input samples. Second, by estimating the depth of the locked footprints and obtaining a normalized relative trampling depth profile, the system searches for the principal depth extremum at the front of the footprint, enabling the system to not only rely on the footprint's outer contour or planar texture but also further... This invention utilizes the morphological information reflected by variations in footprint depth. Secondly, a depth dynamic feature verification module determines the equivalent force center, positive centroid offset, puncture shear steepness, attenuation scale, hoof-heel suspension transition boundary, and hoof-heel suspension deviation based on a normalized relative trampling depth profile. This generates a musk deer footprint verification index, enabling the quantification of features such as concentrated force at the front end, attenuation at the rear end, and hoof-heel suspension in musk deer footprints, thereby improving the ability to distinguish musk deer footprints from those of other cloven-hoofed animals in the same region. Finally, a structured archiving module generates structured recognition records or performs triage archiving based on the classification results, and can differentiate between non-musk candidates and samples awaiting verification, facilitating subsequent manual verification, sample tracking, and model updates. Therefore, this application can improve the automation, consistency, and efficiency of musk deer footprint identification processes, even in cases where field footprint images exhibit scale variations, substrate differences, incomplete indentations, or interference from other cloven-hoofed animals. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the automatic recognition system for footprint images of musk deer and co-located even-toed ungulates based on deep learning according to the present invention. Figure 2 This is a schematic diagram illustrating the construction and locking of the footprint chain in this invention; Figure 3 This is a schematic diagram of depth estimation and longitudinal profile extraction according to the present invention; Figure 4 This is a schematic diagram of the four-zone segmentation of the depth profile; Figure 5 This is a schematic diagram of the deep dynamic feature verification and chain breakage feedback of the present invention; Figure 6 This is a schematic diagram comparing the candidate musk deer footprints with the interfering footprints in the same domain according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example The system of this invention can acquire original images of footprints based on ordinary visible light imaging devices, including smartphones, digital cameras, infrared trigger cameras, patrol terminals, or other image acquisition devices that can obtain footprint images; during image acquisition, the main body of the footprint is located in the center of the image field of view, and the outer contour of the footprint, the hoof tip area, the hoof heel area, and the direction of the footprint chain are identifiable. Preferably, during data acquisition, the angle between the camera's optical axis and the ground normal can be controlled to not exceed a preset tilt angle threshold. The vertical distance between the lens and the center of the footprint is within the preset shooting distance range. Within, the image resolution is not lower than the preset resolution. As an example, Desirable , Desirable to , Desirable Pixels; The above values ​​are only for initialization or recommended acquisition conditions, and do not limit the present invention to be implemented under these numerical conditions; When the system needs to convert pixel distance into actual ground distance, a ground scale conversion coefficient can be established using the focal length, shooting height, shooting angle, and sensor size of the imaging device. , This indicates the ground length corresponding to a unit pixel; it can also be obtained through the device's built-in ranging module, the shooting distance recorded by the patrol terminal, or the footprint scale reference value input by the operator. If a metric scale cannot be obtained on-site, the system can use a normalized scale based on the average width of the footprint or the size of the footprint bounding box for identification, and mark it as a normalized scale result when archiving. Please see Figure 1 As shown, this invention discloses an automatic recognition system for footprint images of musk deer and co-located even-toed ungulates based on deep learning, comprising the following modules: a behavioral rhythm extraction module, a morphological segmentation processing module, a deep dynamic feature verification module, and a structured archiving module; The behavioral rhythm extraction module is used for instance segmentation, continuous footprint chain construction, and rhythm label assignment of raw footprint images. This module extracts footprint contours, bounding boxes, and center points using a deep learning instance segmentation model, forming continuous footprint chains based on spatial distance and directional continuity constraints. It calculates three rhythm parameters: footprint density, step length dynamic variance, and trajectory curvature. Based on these parameters, the footprint chains are classified into high-density trampling, low-density uniform, discrete jumping, and unreviewed types. Only footprints in the discrete jumping type are output as locked footprints, achieving preliminary screening of typical jumping behaviors in musk deer and providing effective samples for subsequent morphological and dynamic analysis. First, the footprint instance segmentation model is called on the original image. Perform pixel-level segmentation to obtain a binary mask for each footprint. , outer frame and center point coordinates Where i represents the footprint instance number, used to distinguish different footprints detected in the same original image; the footprint instance segmentation model can employ MaskR-CNN, YOLO segmentation network, U-Net variant, or other deep learning models with footprint instance segmentation capabilities; preferably, an instance segmentation network with a convolutional neural network or visual Transformer as its backbone can be used; as an example, MaskR-CNN with ResNet-50 as its backbone network can be used, with a training set containing no less than 5000 labeled footprints, covering four species: musk deer, Chinese serow, blue sheep, and domestic sheep. The annotation information includes footprint mask, bounding box, center point, and species label; a stochastic gradient descent optimizer is used, with an initial learning rate of 0.001 and 30 training epochs. The model can be used only if the segmentation mAP on the validation set is ≥0.85; the instance segmentation model only outputs detection results with a confidence level higher than a preset segmentation confidence threshold, which can be determined based on the balance between segmentation accuracy and recall on the validation set; The training set includes at least target musk deer footprint samples and sympatric even-toed ungulate footprint samples; sympatric even-toed ungulates may include Chinese serow, blue sheep, tufted deer, domestic sheep, yak calves, or other ungulates whose habitats overlap with those of musk deer; each footprint in the training samples is marked with at least the outer contour mask; preferably, the hoof tip, hoof heel, left and right hoof lobes dividing line, footprint chain direction, and species category may also be marked; the number of training samples, source regions, and species ratios can be expanded or adjusted according to the actual deployment area, as long as the model can learn the distinguishable morphological features between musk deer and sympatric even-toed ungulate footprints; Obtain the set of coordinates of the center of the footprint Then, this module converts the pixel coordinates to ground coordinates; if the ground scale conversion factor is... If available, then the first The ground coordinates of the center of each footprint are as follows ;like If unavailable, the image coordinates are normalized to a preset scale space, and the normalized step size is used in subsequent processing; Before constructing a continuous footprint chain, scale correction or perspective correction is performed on the original image; when scale reference, ground calibration point or shooting posture information is detected, the image coordinates are mapped to ground coordinates according to the homography matrix; when ground coordinates cannot be obtained, normalization is performed using the footprint bounding box size or average footprint width. Please see Figure 2 As shown, this module then constructs a footprint chain based on the set of center points; specifically, using spatial distance, directional continuity, and footprint orientation consistency as constraints, footprint centers that satisfy the adjacent relationship are connected into a continuous footprint chain; for any two candidate footprint centers... and When the distance between the two Connection threshold less than preset step size Furthermore, the angle between the direction of the line connecting the two footprints and the direction of the local main axis is less than a preset direction threshold. When the two are in the same footprint chain, they are considered as adjacent nodes. and All parameters are adjustable and can be adaptively calibrated based on animal stride length statistics within the target area or training samples; as an example, The initial value can be set as an extension of the upper limit of the jumping step length for musk deer and homing even-toed ungulates within the target area, for example... ; The initial value can be taken as to The values ​​between; For each consecutive footprint chain, this module calculates at least the following rhythmic statistics: footprint density. Step size dynamic variance and trajectory curvature ; Among them, footprint density It is the ratio of the number of footprints belonging to the continuous footprint chain to the area reference; the area reference is the ground area determined according to the ground scale conversion factor; if the ground scale conversion factor is not available, the area reference is the normalized area obtained by normalizing the footprint bounding box size or the average width of the footprint. Specifically, when the original image can provide actual scale information, the area where the footprint chain is located is converted into the actual ground area, and the footprint density is obtained by dividing the number of footprints belonging to the continuous footprint chain by the actual ground area. When the original image cannot provide actual scale information, the size of the footprint bounding box or the average width of the footprint is used as the normalized scale reference. The area containing the footprint chain is normalized, and the footprint density is obtained by dividing the number of footprints belonging to the continuous footprint chain by the normalized area. Footprint density is calculated using normalized area. At that time, the structured archiving module marks the structured recognition record as a normalized scaling result.

[0019] Step size dynamic variance ,in This represents the distance between the centers of adjacent footprints within the same footprint chain. This represents the sample variance of the step length sequence of the footprint chain; trajectory curvature. , represents the statistical value of the ratio of the change in yaw angle between adjacent steps in the footprint chain to the corresponding step length. In this embodiment, it is preferably the average value of the ratio of the absolute value of the change in yaw angle between adjacent steps in the footprint chain to the corresponding step length. This module assigns rhythm labels to the footprint chain based on the aforementioned rhythm statistical parameters; specifically, it introduces a footprint density threshold. Step size variance threshold and curvature threshold ;in, , , All are preset adjustable thresholds, which respectively represent the footprint density limit for distinguishing between dense trampling scenarios, the stride length dispersion limit for distinguishing between uniform and non-uniform gait, and the trajectory curvature limit for distinguishing between straight-line walking and jump-avoidance walking; as an example, Can be initialized to , Can be initialized to , Can be initialized to ; , , All of these are preset adjustable thresholds, which can be adaptively updated through training set statistics, cross-validation, or historical samples. If footprint density Above the footprint density threshold Then the footprint chain is marked as high-density trampling type; if the footprint density Not higher than the footprint density threshold And step size dynamic variance Not higher than the step size variance threshold If so, it will be marked as a low-density uniform speed label; If footprint density Not higher than the footprint density threshold Step size dynamic variance Above the step size variance threshold And trajectory curvature Not lower than the curvature threshold If the condition is met, it is marked as a discrete jump type label; otherwise, it is marked as a label to be reviewed. For footprint chains tagged as discrete jump types, this module generates downstream operator parameter packages. The parameter package includes at least: the front-end search window ratio. Crack detection switch Number of profile samples Horizontal filter width And the set of depth attenuation coefficients required for four-band segmentation All of the above parameters can be preset or adaptively updated based on training sample statistics; as an example, Can be initialized to , used to indicate the normalized range of positions where the depth extremum should be located at the front of the footprint, and the calibration basis is the 95th quantile of the statistical distribution of the longitudinal coordinates of the extreme points of the musk deer sample; Can be initialized to ; It can be initialized to a preset ratio of several pixels width or footprint width; the narrow baseline gap detector switch is set to on; This module also outputs a list of precise pixel coordinates of all locked footprints belonging to the footprint chain and passes it to the shape segmentation processing module.

[0020] The morphology segmentation module obtains the normalized relative trampling depth matrix of footprints through a lightweight depth estimation network, and extracts a one-dimensional relative trampling depth profile along the longitudinal main axis of the footprints. It locates the main depth extreme points within the front-end search window, and performs four-zone adaptive segmentation on the footprint profile based on two-level depth attenuation coefficients, dividing it into a sharp pressure core zone, a double-lobed fissure signature zone, a main load attenuation zone, and a hoof and heel light load wake zone. It extracts key morphological features such as longitudinal width and fissure morphology, providing standardized structural features for subsequent depth dynamic feature confirmation and species subdivision. This module crops the original image based on the bounding box of the footprint mask; the cropping area is centered on the center of the footprint and extends outward from the bounding box by a preset cropping ratio. This ensures that the hoof tip, hoof heel, edge indentation, and a small background area are all included in the working drawing; This is an adjustable parameter; as an example, it can be initialized to... Then, scale the cropped image to the preset working size. As an example, it can be scaled to Pixels; if the computing power of the deployment end is high or higher precision is required, a larger working size can also be used; Furthermore, principal axis alignment is performed on the working image; specifically: principal component analysis is performed on the foreground pixels in the footprint mask, and the direction of the first principal component is taken as the longitudinal principal axis of the footprint; the hoof tip and hoof heel end are determined based on the sharpness of the footprint front end, the initial depth estimate, the shape of the mask endpoints, or the forward direction of the footprint chain, so that the normalized longitudinal coordinates are... Corresponding to the direction of the hoof tip, Corresponding to the direction of the hoof and heel; If the direction of the hoof tip conflicts with the direction of the footprint chain, the joint determination result of the direction of the footprint chain and the shape of the hoof tip shall be given priority; if it still cannot be confirmed, it shall be marked as a sample with uncertain direction and enter the manual review or low confidence output process. Next, monocular relative depth estimation is performed on the aligned working image; the depth estimation model can be a lightweight convolutional neural network, encoder-decoder network, lightweight visual Transformer network, or other models that can infer relative depth from footprint texture, shadow and indentation boundaries; Preferably, in this embodiment, a lightweight convolutional neural network is used to adapt to deployment on portable terminals in the field; as a non-limiting example, MobileNetV3-Small can be used as the encoder backbone, combined with a multi-level upsampling decoder and skip connections to output a single-channel depth matrix. Single-channel depth matrix Each element in the expression represents the relative stepping depth at the corresponding pixel position; where It has been normalized to The interval is defined as follows: the larger the value, the deeper the relative depth. If the training samples include measured 3D depth maps, the supervision signal can be obtained through structured light scanning, photogrammetry, laser scanning, or other 3D reconstruction methods. If only 2D images are available, a semi-supervised or self-supervised method can be used to train the relative depth estimation model. The training loss includes one or more of the following: scale-invariant depth loss, gradient consistency loss, boundary preservation loss, and mask region reconstruction loss. The number of training rounds, learning rate, number of samples, and number of network parameters can be determined based on the computing power and sample size of the deployment end. As long as the relative depth matrix corresponding to the footprint working map can be output, the specific details are not considered as limiting conditions of this invention. It should be noted that during the training phase, MobileNetV3-Small uses footprint images with measured 3D depth information as supervised samples. The measured depth is obtained through structured light scanning, photogrammetry, or laser scanning. For samples without measured depth, semi-supervised training is performed using the consistency constraint of multiple images of the same footprint under different lighting angles. Before entering the depth estimation model, the input image undergoes white balance correction, brightness normalization, footprint mask cropping, and principal axis alignment. The single-channel depth matrix output by the depth estimation model is normalized to [0, 1] and isolated noise points are removed by median filtering. Please see Figure 3 As shown, the single-channel depth matrix is ​​obtained. Subsequently, a one-dimensional normalized relative trampling depth profile was extracted along the longitudinal principal axis of the footprint. Specifically: normalize the longitudinal principal axis of the footprint to Set in this direction 1 sampling point; at each sampling point, a preset horizontal window is taken along the vertical principal axis. The depth statistics within the point are used as the profile depth; the depth statistics can be the median, mean, weighted mean, or mean after removing extreme values; in this embodiment, the median is preferred to reduce the influence of soil particles, leaf shading, and local reflections on the profile; subsequently, the normalized relative trampling depth profile is... Perform smoothing to obtain a smooth profile. Smoothing methods can include moving average, Gaussian filtering, median filtering, or spline smoothing. This module is in the front-end search window. Searching inside The principal depth extremum point is normalized and its longitudinal coordinate is marked as... , This refers to the normalized coordinates of the deepest relative indentation location within the front region of the footprint along the longitudinal principal axis. It can be determined by comparing the smooth profiles of each sampling point within the front search window. Once the size is determined, its mathematical expression can be found as follows: ;in, The ratio of the front search window is used to limit the depth extreme value to be located in the pressure area from the hoof tip to the forefoot. Normalized longitudinal coordinates for determining the principal depth extreme points Then, four-band adaptive segmentation is performed; the four-band boundary is defined as the boundary dynamically determined by a preset depth attenuation coefficient. Specifically, a first depth attenuation coefficient is introduced. Second depth attenuation coefficient And satisfy ; The relative depth control coefficient representing the transition from the sharply compressed core zone to the main load attenuation zone. This represents the relative depth control factor for the transition from the main load attenuation zone to the heel light load wake zone; as an example, Can be initialized to , Can be initialized to These values ​​are only initial or recommended values ​​obtained from sample statistics. Please see Figure 4 As shown, the boundaries of the four zones are determined as follows: First Boundary Defined as the normalized longitudinal coordinate of the autonomous depth extreme point When searching towards the hoof and heel... First drop to The following locations; if no precise isocenters exist among the sampling points, they can be determined through linear interpolation or spline interpolation. If the location is not found within the search range, then... Mark as missing and proceed to the exception branch for handling; second boundary Defined as self When searching towards the hoof and heel... First drop to The following locations; if no exact isopoints exist, they are also determined by interpolation; if If missing, it is marked as a sample with abnormal main load attenuation; Normalized vertical coordinates based on the main depth extreme points First boundary Second boundary The footprints are divided longitudinally into the following four functional zones: Sharp, pressure-bearing core zone To cover the hoof tip to the first boundary The area is used to characterize the degree of concentrated force on the hoof tip or forefoot area; Bilobed slit signature band For A local region centered on L0 and whose range is automatically determined by the local morphology of the normalized relative trampling depth profile near L0 along the longitudinal direction is used to characterize the gap morphology and narrow baseline features between the left and right hoof flaps. Main load attenuation zone To cover to The region is used to characterize the attenuation process of the footprint as it transitions from concentrated force at the front end to weaker force at the rear end; hoof light load trail To cover to The area is used to characterize whether there are obvious dragging, flat treading, or loose treading trails in the hoof and heel area; Signature band in the double-lobed slit In the middle, extract the lateral depth profile. ,in Normalized horizontal coordinates in the direction perpendicular to the principal axis; based on Calculate the depth of the fracture valley Double-petal peak distance Depth difference between left and right lobes Isofracture characteristics; as an example, fracture valley depth Defined as the difference between the average peak depth of the left and right hoof lobes and the depth of the central fissure; the peak distance between the two lobes. Defined as the lateral distance between the local peak values ​​of the left and right hoof lobes; if If enabled, the above-mentioned crack features are incorporated into the morphological feature vector; if the cracks in the field sample are unclear or the mask quality is insufficient, the crack features are marked as missing values ​​and the confidence of the sample is reduced. To illustrate how this invention handles the subtle differences between candidate musk deer footprints and sympatric interference footprints, please refer to [link to relevant documentation]. Figure 6 Musk deer candidate footprints and sympatric even-toed ungulate footprints share certain similarities in outer contour, double hoof flap structure, and local indentation morphology, which can easily lead to confusion when relying solely on planar contours or human experience for judgment. Musk deer candidate footprints are generally characterized by a relatively narrow and long overall shape, a sharp hoof tip, a small fork, relatively concentrated pressure at the front end, and a lighter indentation in the heel area. Sympatric interference footprints, on the other hand, may be characterized by a relatively wide overall shape, a relatively blunt hoof tip, a heavier indentation at the rear end, or a more gradual decrease in depth. Figure 6 This is only used to illustrate the differences in morphology and depth mentioned above, and does not imply that the footprints of different species have exactly the same appearance under all substrate, posture and collection conditions; Based on the above differences, this invention does not directly use the outer contour of the footprint as the only criterion for discrimination. Instead, it further utilizes the normalized relative trampling depth profile of the locked footprint, combined with the principal depth extreme point, equivalent force center, positive centroid offset, puncture shear steepness and hoof and heel suspension deviation, to conduct dynamic confirmation analysis on musk deer candidate footprints, thereby improving the ability to distinguish musk deer footprints from interfering footprints in the same domain. Ultimately, this module generates morphological feature vectors. in, Indicates the width of the sharply compressed core band. Indicates the width of the main load attenuation band. express to The slope of the depth decrease between them express to The slope of the depth descent between these parameters; the above features can be added or removed according to the actual training situation.

[0021] Please see Figure 5As shown, the depth dynamic feature verification module performs quantitative calculations of force and motion features based on the footprint depth profile. This module obtains the equivalent force center through weighted integration, calculates the positive centroid offset to determine the concentrated force characteristics at the front end; calculates the puncture shear steepness within the effective force range to characterize the attenuation characteristics of the hoof tip puncture indentation; searches for the hoof and heel suspension transition boundary based on the attenuation scale and calculates the hoof and heel suspension deviation; integrates the above three features to generate a musk deer footprint verification index, enabling automatic differentiation between musk deer and sympatric even-toed ungulates; samples confirmed as musk deer are further input into a species subdivision model, outputting specific species candidate labels. This module is based on Determine the equivalent center of force for the entire footprint. , The average stress location is obtained by weighting the relative depths of each longitudinal sampling point, meaning that the deeper the location, the greater its contribution to the stress center; the calculation formula is as follows: In discrete implementations, sampling point weighted averaging, trapezoidal integrals, or Simpson's integrals can be used to obtain the results. If the total depth energy of the normalized relative trampling depth profile is lower than the preset effective indentation energy threshold... If the footprint depth information is insufficient, the system will terminate the depth dynamic feature verification and enter the degradation process; a preset effective indentation energy threshold is used. The initial value of is 0.05 in this embodiment, which means the lower limit of the sum of the depth values ​​of all sampling points in the normalized relative trampling depth profile divided by the number of sampling points. The calibration basis is the lower 5th percentile of the statistical distribution of the total depth energy of normal footprint samples. This module will define the equivalent center of force. The directed distance between the normalized longitudinal coordinate L0 of the principal depth extremum point and the centroid offset is defined as the positive centroid offset. This is used to characterize the degree to which the equivalent center of force shifts towards the hoof heel relative to the principal depth extreme point; since in this invention x=0 represents the hoof tip direction and x=1 represents the hoof heel direction, the positive centroid shift amount Defined as equivalent force center Subtract the normalized longitudinal coordinate L0 of the principal depth extremum point; To avoid minor offsets caused by depth estimation noise, local unevenness of the substrate, or atypical indentations being misidentified as valid features, this module presets a centroid offset threshold. Only when the positive centroid offset Greater than the preset centroid offset threshold When the footprint is determined to have front-end puncture force characteristics, it proceeds to the subsequent puncture shear characteristic assessment; when the positive centroid offset is... Less than or equal to the preset centroid offset threshold At that time, including as well as In both scenarios, the footprint is determined not to meet the front-end puncture force requirements, triggering the first-order chain break feedback, and the sample is then handed over to the structured archiving module for distribution; among them, As an adjustable parameter, it can be calibrated based on the statistical distribution of musk deer samples and co-located even-toed ungulate samples; as an example, It can be initialized to 0.08, and its calibration can be based on the lower 5th percentile of the statistical distribution of the positive centroid offset in musk deer samples. Since the footprint used for subsequent puncture and shear feature assessment already meets... Therefore, in subsequent calculations, Always positive; After the footprints are filtered by the concentrated force at the front end, this module normalizes the longitudinal coordinate L0 of the main depth extreme point to the tail end of the effective force interval. The interval is taken as the effective force-bearing interval, in which for The smaller of the two, i.e., the end of the effective force range, is denoted as . , for The smaller of 1; the effective force-bearing range represents the force attenuation region transitioning from the extreme depth point at the front end to the overall force-bearing center, its boundary being determined by the previous stage. Determine the correct length, rather than setting a fixed length separately; Within the effective stress zone, this module determines whether the depth profile exhibits a rapid descent characteristic. To characterize this characteristic, the puncture shear steepness K can be extracted, where K represents the average rate of descent from the peak depth at the front end to the tail end of the effective stress zone. It can be defined as: ;in, To prevent extremely small positive numbers with a denominator of zero; the puncture shear kurtosis K represents... to The average rate of descent between the two is used to quantify the shear steepness formed after the hoof tip or forefoot area enters the soil; If K is not lower than the preset shear steepness threshold This indicates that significant puncture shear attenuation occurred after the hoof tip or forefoot area entered the soil, and the system entered the hoof and heel suspension characteristic judgment; if Less than the preset shear steepness threshold This triggers a second-order chain break feedback, indicating that the footprint is more likely to correspond to a steady walk, flat tread, or blunt hoof imprint type; among which, This can be determined through training sample statistics, ROC curves, cross-validation, or Bayesian optimization; as an example... Can be initialized to ; After the footprint is determined by puncture and shear features, this module uses the positive centroid offset. The relative relationship with the puncture shear steepness K forms the attenuation scale τ, the calculation formula of which can be found below: ;in, To prevent extremely small positive numbers with a denominator of zero; This represents the normalized longitudinal distance scale corresponding to the transition from the front-end puncture state to the rear-end gentle stress state; when the shear steepness is large... A relatively small value indicates that the stress attenuation is more concentrated; when the shear kurtosis is small, A relatively large value indicates a more gradual decrease in force. Subsequently, in Starting from, with To find the overhanging transition boundary of the hoof and heel within the scanning window of the scale. The length of the scanning window can be adjusted by the scanning window magnification m. As an example, m is a constant greater than 1 and less than 5. The preferred initial value is m=3. This step focuses on the earliest position where the local descent rate changes from a rapid descent state to a smooth dragging or light load state. The judgment condition is the local slope change condition. The local descent rate is used to characterize the degree of local attenuation of the normalized relative trampling depth profile along the hoof tip to hoof heel direction within the scanning window. Specifically, adjacent sampling points are selected in the scanning window in ascending order of normalized longitudinal coordinates. The decrease in profile depth value of the previous sampling point relative to the profile depth value of the next sampling point is calculated, and the ratio of this decrease to the normalized longitudinal coordinate increment between the two sampling points is positively evaluated to obtain the local descent rate at the corresponding position. If the profile depth value of the next sampling point is not lower than the profile depth value of the previous sampling point, the local descent rate at that position is recorded as zero. Within the scanning window, the earliest position where the local descent rate changes from above the preset high slope condition to below the preset low slope condition is located is identified, and this position is recorded as... The low slope condition can be determined by the local rate decay proportionality coefficient. It means that the conditions are met. As an example, Can be initialized to If no location satisfying the local slope change condition is found within the scanning window, the third-order chain break feedback is triggered, indicating that the footprint does not show a clear hoof and heel suspension transition boundary, which may correspond to a blunt hoof slipping out, hoof and heel dragging, or excessive soil disturbance. Obtaining the hoof and heel suspension transition boundary Then, based on the transition boundary of the hoof and heel in the air. The proportional relationship between the distance between L0 and τ is used to obtain the hoof and heel suspension deviation η, and its calculation formula is shown below: ;in, To prevent extremely small positive numbers with a denominator of zero; It is used to characterize the geometric proximity of the heel suspension transition point to the front depth extreme point; if η is too large, it usually means that the transition point is far away from the front force area, and if η is too small, it may indicate local noise or abnormal peaks. Therefore, a reasonable range or confidence level can be set in combination with sample statistics. positive centroid offset The puncture shear steepness K and the hoof and heel suspension deviation η are combined to form a musk deer footprint confirmation index. As a feasible implementation method, the musk deer footprint confirmation index It can be calculated using the following formula: ;in, To prevent extremely small positive numbers with denominators that are zero or too small from having unstable values; It comprehensively reflects the forward centroid shift, puncture shear steepness, and hoof and heel suspension transition characteristics; If musk deer footprint confirmation index Not lower than the preset threshold for musk deer confirmation If the musk deer footprint confirmation index is 0, then output the candidate confirmed musk deer; when the musk deer footprint confirmation index is 0, the candidate confirmed musk deer footprint confirmation index is 0. Below the preset threshold for confirming musk deer If no downgrade conditions are triggered, output a non-musk candidate; if the sample triggers downgrade conditions, key boundaries are missing, species templates conflict, or confidence is insufficient, output "to be reviewed". This can be determined through validation set statistics of musk deer samples and sympatric even-toed ungulate samples; as an example, Can be initialized to ; For footprints confirmed as musk deer candidates, this module further performs species subdivision; species subdivision does not use a fixed unique value as an absolute boundary, but instead uses species template intervals or classifier models obtained from training samples; specifically, it can... , , , , , , , The species segmentation model is based on input features, and the model can be an interval template matching model, support vector machine, random forest, gradient boosting tree, lightweight neural network, or Bayesian classification model; if the interval template matching method is used, then each musk deer species... A corresponding set of statistical templates The statistical template includes at least the statistical intervals for the species in terms of frontal extreme position, core zone width, centroid offset, shear steepness, suspension deviation and crack characteristics; When the features of the sample to be judged fall into the template range of a certain species and the overlap with the templates of other species is less than the preset conflict threshold, the corresponding species label is output; if it falls into multiple templates or the matching degree with all templates is insufficient, the musk deer candidate is output and the species is to be verified. As an example, in a certain training sample library, musk deer samples... It can be mainly distributed in the more forward but slightly wider anterior pressure zone, in the musk deer samples. It can be mainly distributed in the area closer to the hoof tip and with a narrower core zone; therefore, the musk deer template can be set as an example. Located in the extreme range of the first species, Located within the width range of the first core zone; the musk deer template is Located in the extreme range of the second species, Located within the width range of the second core band; the specific range is determined by actual sample statistics and is not considered a fixed limitation of this invention.

[0022] The structured archiving module, based on the classification results and chain break markers from the deep dynamic feature verification module, assigns samples to the corresponding chain break sub-library or generates structured identification records for musk deer. It performs downgrade control on abnormal samples such as those with excessively short footprint chains or insufficient depth information, transferring them to the manual review queue. It also supports closed-loop sample updates, triggering incremental model training when the number of newly added review samples reaches a threshold, thereby improving the system's long-term recognition accuracy and generalization ability. When the deep dynamic feature confirmation module outputs Not less than At that time, this module generates a structured recognition record; the record can be in JSON, XML, database table entries or other structured formats; the record content includes at least: original image path, acquisition time, acquisition location, imaging device information, footprint chain number, footprint mask, center point coordinates, rhythm label, etc. , , , , , , , The species candidate labels, confidence levels, and whether a normalized scale is used are considered. If there are manual verification results, the manually verified labels are used as subsequent training or validation data. When a link break occurs in the deep dynamic feature verification module, this module performs a traffic splitting process based on the broken link node; if the first-order link break occurs, i.e. If the centroid shift requirement is not met, the sample is classified into the first-order chain break candidate sample sub-library; if it is a second-order chain break, i.e. If the shear kurtosis requirement is not met, it is classified into the second-order chain break candidate sample sub-library; if it is a third-order chain break, i.e. If missing, it will be classified into the third-order broken chain candidate sample sub-library; the above diversion categories can be adjusted according to the species of symbiotic animals in the actual target area, and are not limited to specific animal names; The naming of the sub-library is directly anchored to the chain break stage marker output by the deep dynamic feature confirmation module. In different deployment areas, the aforementioned sub-library may correspond to different morphological candidate names of syn-dominant ungulates, but the naming identifier of the sub-library is uniformly named using the aforementioned stage number. When the behavior rhythm extraction module fails to form a valid footprint chain, the morphology segmentation module fails to obtain a valid depth profile, or the depth dynamic feature verification module experiences parameter out-of-bounds errors, this module executes degradation control. Degradation conditions include, but are not limited to: footprint chain length being less than a preset chain length threshold. The instance segmentation confidence level is lower than the preset segmentation confidence level threshold. The depth matrix variance is lower than the preset depth variance threshold. The normalized relative trampling depth profile energy is lower than the preset effective indentation energy threshold. The normalized longitudinal coordinates of the principal depth extreme point, the first boundary, or the second boundary are missing; the principal axis direction cannot be determined; the effective force range is out of bounds. As an example, a preset chain length threshold is used. In this embodiment, the initial value is 3, which means the minimum number of footprints that should be included in the footprint chain. The calibration basis is that at least 3 consecutive footprints are required to calculate the step size variance and trajectory curvature. In this embodiment, the initial value is This is used to determine whether the depth estimation result is approximately an invalid plane output; For locked footprints that meet any of the above downgrade conditions, this module will suspend the current judgment process and write the original image of the sample, the footprint chain number to which it belongs, and the intermediate results of the completed modules into the manual review queue for operator review; the manual review queue can be implemented by local database table entries, message queues or file directories; This module can also perform closed-loop update functionality; when the number of new samples in a certain category reaches a preset feedback threshold... After manual review or high-confidence automatic labeling confirmation, the system can start the incremental training process; incremental training may include fine-tuning of instance segmentation model, fine-tuning of depth estimation model, rhythm threshold update, four-band segmentation coefficient update, and species template interval update; to avoid model drift, the new model after incremental training must meet the preset performance conditions on the independent validation set, such as no decrease in recall, improvement in overall accuracy, or improvement in the F1 value of the target category, before it can replace the online model; This is an adjustable parameter; as an example, it can be initialized to... One new sample.

[0023] The following examples are only used to illustrate the operation of the present invention and do not constitute a limitation on the scope of the present invention; the parameter values ​​in the examples are all initial values ​​or example results under a certain training sample library and a certain deployment scenario, and those skilled in the art can adjust them according to actual sample statistics; Using a musk deer sample as an example, in an image of moist soil in an alpine shrubland, the behavioral rhythm extraction module identified a continuous footprint chain consisting of multiple footprints; after scale conversion, the footprint density of this footprint chain was... Below the footprint density threshold Step size dynamic variance Above the step size variance threshold trajectory curvature Above the curvature threshold Therefore, it is marked as a discrete jump type and enters the topography segmentation processing module; the topography segmentation processing module performs principal axis alignment and depth estimation on one of the locked footprints, and obtains the depth extremum point within the front search window. ;by , When used as an example initial value, the system can obtain the first boundary. Second boundary For example, Can be , Can be , Can be This indicates that the footprint has a relatively obvious front-end concentrated pressure zone, main load attenuation zone, and rear-end light load wake; the depth dynamic characteristic confirmation module further determines whether the equivalent force center of the footprint is located on the hoof heel side of the depth extreme point. If the positive offset δ between the two meets the preset centroid offset requirement, it continues to determine whether there is obvious shear attenuation within the effective force range; if the puncture shear steepness K meets the preset requirement, it continues to search for the hoof heel suspended transition boundary; if all three judgments are true, and Not less than At that time, the system outputs a candidate confirmation of musk deer; if the sample in the species subdivision template... , , , and If the sample is closer to the musk deer template, output a musk deer candidate label; if there is a template conflict or insufficient confidence, output a musk deer candidate or species to be verified. Taking a musk deer sample as an example, in another image of moist humus soil at the edge of a temperate mixed coniferous and broadleaf forest, the behavioral rhythm extraction module also identified a low-density, high-step-variance, and high-curvature footprint chain, thus labeling it as a discrete skipping type; after the morphology segmentation processing module estimates the depth of the locked footprints, it obtains the depth extremum points closer to the hoof tips within the front search window; for example, Can be , Can be , Can be This is characterized by a narrow, sharply compressed core zone and rapid front-end depth decay; the depth dynamic characteristic verification module calculates... , , , and All meet the candidate criteria for musk deer; if the sample's , If the fissure features are closer to the musk deer template, then output the musk deer candidate label; The two examples above are only used to illustrate that the present invention can process different musk deer footprints through the same process and output corresponding species candidate results through templates or classifiers.

[0024] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0025] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0026] Those skilled in the art will recognize that the modules and algorithm modules of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0027] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0028] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0029] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based automatic recognition system for musk deer and co-located even-toed ungulate footprint images, characterized in that, The module includes the following: The behavior rhythm extraction module is used to perform footprint instance segmentation on the original image, construct a continuous footprint chain, and output the footprints in the footprint chain with the rhythm label of discrete jump as locked footprints; The topography segmentation module is used to estimate the depth of the locked footprints to obtain a normalized relative trampling depth profile, and obtain the normalized longitudinal coordinates of the main depth extremum points in the search window at the front of the profile. The depth dynamic feature verification module is used to obtain the equivalent force center based on the weighted normalized relative trampling depth profile, and to obtain the positive centroid offset from the directional distance of the equivalent force center relative to the main depth extreme point. When the positive centroid offset meets the preset front-end concentrated force screening condition, the effective force interval is determined along the hoof and heel direction based on the positive centroid offset, starting from the normalized longitudinal coordinate of the main depth extreme point, and the puncture shear steepness is obtained within this effective force interval. The attenuation scale is determined based on the relative relationship between the positive centroid offset and the puncture shear steepness, and the hoof and heel suspension transition boundary is searched based on the attenuation scale. The hoof and heel suspension deviation is obtained based on the positional relationship between the hoof and heel suspension transition boundary relative to the main depth extreme point and the attenuation scale. The musk deer footprint verification index is generated based on the positive centroid offset, puncture shear steepness, and hoof and heel suspension deviation, and the classification result is obtained based on the musk deer footprint verification index. The classification result includes one of the following: musk deer candidate verification, non-musk candidate, and pending verification. The structured archiving module is used to generate structured identification records or perform split archiving based on classification results.

2. The automatic recognition system for musk deer and co-located even-toed ungulate footprint images based on deep learning according to claim 1, characterized in that, The behavior rhythm extraction module calculates footprint density, step length dynamic variance, and trajectory curvature for each continuous footprint chain. Footprint density is the ratio of the number of footprints belonging to the continuous footprint chain to the area benchmark. The area benchmark includes at least the ground area converted according to the ground scale conversion factor. Step length dynamic variance is the sample variance of the step length sequence formed by the center distance between adjacent footprints in the footprint chain. Trajectory curvature is the mean of the ratio of the absolute value of the change in yaw angle between adjacent steps in the footprint chain to the corresponding step length. Based on footprint density threshold, step length variance threshold, and curvature threshold, rhythmic labels are assigned: footprint chains with footprint density higher than the footprint density threshold are labeled as high-density trampling type labels; footprint chains with footprint density not higher than the footprint density threshold and step length dynamic variance not higher than the step length variance threshold are labeled as low-density uniform type labels; footprint chains with footprint density not higher than the footprint density threshold, step length dynamic variance higher than the step length variance threshold, and trajectory curvature not lower than the curvature threshold are labeled as discrete jumping type labels; the remaining footprint chains are labeled as labels to be reviewed.

3. The automatic recognition system for musk deer and co-located even-toed ungulate footprint images based on deep learning according to claim 2, characterized in that, The footprint instance segmentation model used in the behavior rhythm extraction module is an instance segmentation network with a convolutional neural network or visual Transformer as its backbone. It is trained using a training set containing musk deer footprint samples and unilaterally distributed even-toed ungulate footprint samples. The unilaterally distributed even-toed ungulates include at least the Chinese serow, blue sheep, and domestic sheep. After segmentation, a binary mask, bounding box, and center point coordinates are obtained for each footprint. The center point coordinates are used as input data to calculate the distance and direction between adjacent footprints when constructing a continuous footprint chain.

4. The automatic recognition system for musk deer and co-located even-toed ungulate footprint images based on deep learning according to claim 1, characterized in that, The depth estimation model used in the shape segmentation processing module is a deep learning network with an encoder-decoder structure. The encoder is used to extract multi-scale features of footprint texture and shadow. The decoder restores spatial resolution through multi-level upsampling and skip connections, and outputs a single-channel depth matrix corresponding to the working image of the locked footprint. Each element in the single-channel depth matrix represents the relative trampling depth of the corresponding pixel position normalized to the interval [0, 1]. The normalized relative trampling depth profile is obtained by sampling the single-channel depth matrix along the longitudinal main axis of the footprint.

5. The automatic recognition system for musk deer and co-located even-toed ungulate footprint images based on deep learning according to claim 4, characterized in that, After obtaining the normalized longitudinal coordinates of the main depth extreme point, the topography segmentation module performs four-zone adaptive segmentation on the normalized relative trampling depth profile based on the first depth attenuation coefficient and the second depth attenuation coefficient, and the first depth attenuation coefficient is greater than the second depth attenuation coefficient. When searching from the normalized longitudinal coordinate of the principal depth extreme point towards the hoof and heel direction, the first boundary is obtained when the normalized relative trampling depth profile first drops to a position below the product of the principal depth extreme value and the first depth attenuation coefficient. When continuing the search from the first boundary toward the hoof and heel, the normalized relative trampling depth profile first drops to a position below the product of the principal depth extreme value and the second depth attenuation coefficient, thus obtaining the second boundary; The section from the hoof tip to the first boundary is a sharp pressure core zone. Taking the normalized longitudinal coordinate of the main depth extreme point as the center, the local section is determined according to the local morphology of the normalized relative trampling depth profile near the main depth extreme point, and the double-lobed crack signature zone is obtained. The section from the first boundary to the second boundary is a main load attenuation zone, and the section from the second boundary to the hoof heel is a hoof heel light load wake zone. The longitudinal width of the sharp, compressed core zone is used as a morphological feature parameter input into the species subdivision model along with the positive centroid offset and the puncture shear steepness. The transverse depth profile extracted along the vertical principal axis within the double-lobed cleft signature zone is used as an auxiliary feature parameter characterizing the morphology of the left and right hoof lobes.

6. The automatic recognition system for musk deer and co-located even-toed ungulate footprint images based on deep learning according to claim 1, characterized in that, When the positive centroid offset is greater than the preset centroid offset threshold, the depth dynamic feature verification module determines that the locked footprint meets the preset front-end concentrated force screening condition and continues to calculate the puncture shear steepness. When the positive centroid offset is not greater than the preset centroid offset threshold, the depth dynamic feature verification module determines that the locked footprint does not meet the preset front-end concentrated force screening condition, triggers the first-order chain break feedback and stops subsequent verification processing, and outputs the locked footprint and its first-order chain break mark to the structured archiving module. Among them, the situation where the positive centroid offset is not greater than the preset centroid offset threshold includes: the equivalent force center is not located on the hoof heel side of the main depth extreme point, or although the equivalent force center is located on the hoof heel side of the main depth extreme point, its offset relative to the main depth extreme point in the hoof heel direction does not exceed the preset centroid offset threshold.

7. The automatic recognition system for musk deer and co-located even-toed ungulate footprint images based on deep learning according to claim 1, characterized in that, When the puncture shear steepness is not lower than the preset shear steepness threshold, the deep dynamic feature verification module determines that the locked footprint has passed the front-end puncture shear screening and continues to perform the determination of the attenuation scale and the search for the hoof and heel suspension transition boundary; when the puncture shear steepness is lower than the preset shear steepness threshold, the deep dynamic feature verification module triggers the second-order chain break feedback and stops the subsequent verification process, and outputs the locked footprint and its second-order chain break mark to the structured archiving module.

8. The automatic recognition system for musk deer and co-located even-toed ungulate footprint images based on deep learning according to claim 1, characterized in that, The depth dynamic feature verification module calculates the local descent rate within the scanning window, where the local descent rate is the estimated value of the discrete derivative of the normalized relative trampling depth profile along the normalized longitudinal coordinate within the scanning window. When the local descent rate changes from a state not lower than the puncture shear steepness to a state lower than the product of the puncture shear steepness and the preset local rate attenuation ratio within the scanning window, the location where this change first occurs is taken as the hoof and heel suspension transition boundary, and the calculation of the hoof and heel suspension deviation continues. If there is no location within the scanning window that meets the above transition conditions, the depth dynamic feature verification module determines that the hoof and heel suspension transition boundary is missing and triggers the third-order chain break feedback, outputting the locked footprint and its third-order chain break mark to the structured archiving module. The preset local rate attenuation ratio is greater than 0 and less than 1.

9. The automatic recognition system for musk deer and co-located even-toed ungulate footprint images based on deep learning according to claim 5, characterized in that, When the classification result is a candidate for musk deer, the deep dynamic feature confirmation module inputs the feature parameters in the morphological feature vector along with the positive centroid offset, puncture shear steepness, and hoof-heel overhang deviation into the species subdivision model to obtain the species candidate label. The species subdivision model is one of the following: interval template matching model, support vector machine, random forest, gradient boosting tree, or lightweight neural network, and the species template interval obtained from the training samples is used as the discrimination criterion of the model.

10. The automatic recognition system for musk deer and co-located even-toed ungulate footprint images based on deep learning according to claim 1, characterized in that, The chain break markers output by the deep dynamic feature verification module include first-order chain break markers, second-order chain break markers, and third-order chain break markers; the structured archiving module performs split archiving based on the chain break markers output by the deep dynamic feature verification module: the locking footprints corresponding to the first-order chain break markers are assigned to the first-order chain break candidate sample sub-library, the locking footprints corresponding to the second-order chain break markers are assigned to the second-order chain break candidate sample sub-library, and the locking footprints corresponding to the third-order chain break markers are assigned to the third-order chain break candidate sample sub-library. The structured archiving module generates a structured identification record when the classification result is a confirmed candidate of musk deer, generates a non-musk deer candidate record when the result is a non-musk deer candidate, and transfers the sample to the manual review queue when the result is pending review. The structured archiving module performs degradation control when the footprint chain length is less than the preset chain length threshold, the single-channel depth matrix variance is less than the preset depth variance threshold, or the normalized relative trampling depth profile energy is less than the preset effective indentation energy threshold. This suspends the current judgment process and transfers the sample to the manual review queue.

Images