Wild animal trajectory tracking method and system based on trajectory perception and motion correction

Abstract

The invention discloses a wild animal trajectory tracking method and system based on trajectory perception and motion correction, and relates to the technical field of wild animal monitoring and computer vision, and the method comprises the steps: carrying out the preprocessing of the collected GPS trajectory data of corresponding wild animals; mapping the preprocessed GPS trajectory data to a high-dimensional feature space by using a feature encoder; converting the feature coding vector into a representation sequence for capturing time sequence dependence and context information through a constructed track encoder; inputting the representation sequence into the constructed trajectory sensing network for processing to obtain scores of candidate positions in a preset candidate list of a next time step; performing real-time track correction on the candidate position with the highest score in the preset candidate list, and taking the corrected candidate position as the track position of the wild animal in the next time step; sensor noise and environmental interference are effectively suppressed, and the precision and robustness of wild animal trajectory tracking are improved.

Description

Technical Field

[0001] This invention relates to the fields of wildlife monitoring and computer vision technology, and more specifically, to a method and system for wildlife trajectory tracking based on trajectory perception and motion correction. Background Technology

[0002] With the increasing demand for ecological protection and animal behavior research, the use of intelligent infrared camera terminals or similar monitoring equipment to conduct long-term and continuous observation of wild animals and obtain their movement trajectory data has become one of the important technical means in current scientific research and conservation work. Such equipment is usually deployed in complex and ever-changing field environments to continuously collect a large amount of trajectory data, providing a key data foundation for animal behavior analysis, habitat assessment and ecological environment protection.

[0003] Existing wildlife tracking methods are mostly based on linear models or heuristic algorithms, typically assuming that animal movement has relatively stable speed and direction over short timescales. However, wildlife movement in real-world scenarios often exhibits significant nonlinear characteristics and irregularities, and is easily affected by various external factors such as terrain, climate, vegetation obstruction, and human activities. This makes traditional tracking and prediction methods based on linear assumptions significantly less accurate and adaptable in complex environments. To overcome the limitations of linear models, some studies have introduced probabilistic graphical models such as Markov models and potential field models for trajectory prediction, which can, to some extent, uncover statistical patterns in movement sequences. However, these methods often still face problems such as insufficient characterization of nonlinear dynamic behavior and limited model generalization ability under long-term prediction, drastic environmental changes, or high-noise observation conditions.

[0004] In recent years, with the development of machine learning technologies such as deep learning, trajectory prediction methods based on regression models, sequence models, or attention mechanisms have been gradually applied to target tracking and trajectory inference tasks, improving the accuracy of trajectory prediction in some scenarios. However, these methods generally suffer from the following shortcomings: they are not robust enough to sparse data and incomplete observations in the field; they are sensitive to sensor noise, occlusion, and abnormal observation points, which can easily lead to deviations in prediction results; and their complex model structures and high computational costs result in high deployment costs in edge devices or real-time monitoring scenarios. In addition, existing methods mostly focus on unidirectional prediction of trajectory patterns, with insufficient consideration for the joint optimization of trajectory modeling and motion state estimation, and lack mechanisms for dynamic correction and uncertainty control of prediction results, making it difficult to maintain stable tracking performance in high-noise and highly uncertain field environments.

[0005] Therefore, this application is hereby submitted. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for tracking wildlife trajectories based on trajectory perception and motion correction, which can effectively suppress sensor noise and environmental interference, and improve the accuracy and robustness of wildlife trajectory tracking, while taking into account both complex nonlinear motion feature modeling and real-time requirements.

[0007] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0008] Firstly, this application provides a method for tracking wildlife trajectories based on trajectory perception and motion correction, including the following specific steps:

[0009] The collected GPS trajectory data of the corresponding wild animals are preprocessed to obtain the longitude and latitude information of the wild animals at each time step;

[0010] The preprocessed GPS trajectory data is mapped to a high-dimensional feature space using a feature encoder to obtain the feature encoding vectors corresponding to each time step.

[0011] The constructed trajectory encoder transforms the feature encoding vector into a representation sequence that captures temporal dependencies and contextual information.

[0012] The sequence is input into the constructed trajectory-aware network for processing to obtain the scores of each candidate position in the preset candidate list for the next time step;

[0013] The trajectory of the candidate position with the highest score in the preset candidate list is corrected in real time, and the corrected position is used as the trajectory position of the wild animal in the next time step.

[0014] Based on the above technical solution, the present invention can be further improved as follows.

[0015] Furthermore, the feature encoding vectors corresponding to each of the above time steps are as follows: ; in: ; In the formula, For time steps The feature encoding vector, As a learnable scalar, reflecting the first The contribution of class features to the final representation; Indicates the first Class features at time step The representation after projection. This represents the total number of feature types used. To make the first Each feature from its original dimension Mapping to higher dimensions The weight matrix, For the first Class features at time step The original representation, This means for all except the first Aggregate other features outside the class. For learnable contextual parameters, represent the first... Class features for the first The influence strength of class feature encoding; Represents the projection function. This is the corresponding bias vector.

[0016] Furthermore, the above representation sequence is specifically as follows: ; In the formula, This is the representation sequence output by the trajectory encoder. For multi-head self-attention quantum layer, This represents a sublayer of a feedforward neural network. Presentation layer normalization operation.

[0017] Furthermore, the specific processing results of the aforementioned multi-head self-attention sub-layer are as follows: , ; In the formula, Represents the normalized exponential function, This indicates aggregation at different scales. Scaling factor This is used to shield future information and ensure time steps. The output depends only on the masking matrix of the previous input. Represents the first in the matrix line, number Column elements; , , These represent the query vector, key vector, and value vector generated by the corresponding projection matrices, respectively.

[0018] Furthermore, the scores for each candidate position in the preset candidate list for the next time step are obtained in the following way:

[0019] Each candidate position in the preset candidate list is processed by a trajectory encoder to obtain the features of each candidate position;

[0020] Each candidate position feature is concatenated with its representation sequence to obtain the candidate key vector corresponding to each candidate position.

[0021] The sequence is represented by a linear transformation to obtain the corresponding query vector. Based on the query vector and each candidate key vector, the conditional probability of each candidate position being selected is calculated, and each conditional probability is used as the score of the candidate position.

[0022] Furthermore, the conditional probability of each candidate position being selected, calculated above, is as follows: ,in: , ; In the formula, Indicates that, given the current representation sequence Under the condition of, candidate position The conditional probability of being selected. Represents the query vector With candidate key vectors dot product, Indicates the scaling factor. It is an exponential function. This represents normalizing all candidate positions so that the sum of the probabilities of all candidates is 1, where This represents the total number of candidate positions. Indicates candidate position Embedded vector after feature encoding layer For the current moment The representation sequence, Indicates candidate position The candidate position features and their related features are obtained through the trajectory encoder. For candidate encoders, For the generated query vector, This is the projection matrix obtained through learning.

[0023] Furthermore, the above methods also include:

[0024] A trajectory tracking model is constructed using a feature encoder, a trajectory encoder, and a trajectory awareness network. The trajectory tracking model is trained in the following manner:

[0025] Each training sample is continuously input into the trajectory tracking model for processing to obtain the predicted position corresponding to each training sample. The training samples include at least the longitude and latitude information of the corresponding wild animals at each time step.

[0026] Based on the preset candidate list and predicted position for the next time step, calculate the loss function value corresponding to each training sample;

[0027] Based on the loss function value and backpropagation algorithm, at least the feature encoding layer, weight matrix, bias vector and projection matrix in the trajectory tracking model are updated;

[0028] The process continues until the loss function value reaches a preset condition, and then the trajectory tracking model with the loss function value reaching the preset condition is used to predict the trajectory position of wild animals in the next time step.

[0029] Furthermore, the above loss function value is obtained in the following way: ; In the formula, The value of the loss function. This represents the total number of time steps in the trajectory. Indicates the number of candidates Logarithm with base 0, Indicates that, given the current representation sequence Under the given conditions, the conditional probability of the actual position being selected from the candidate list.

[0030] Furthermore, the above-mentioned real-time trajectory correction of the candidate position with the highest score in the preset candidate list is as follows: ; In the formula, This indicates the position predicted based on the previous moment's state, control input, and process noise, i.e., the position after real-time trajectory correction. It is the state transition matrix. It is a control input matrix. It is a control vector. It is process noise.

[0031] Secondly, this application provides a wildlife trajectory tracking system based on trajectory perception and motion correction, applicable to any of the wildlife trajectory tracking methods based on trajectory perception and motion correction in the first aspect, including:

[0032] The preprocessing module is used to preprocess the collected GPS trajectory data of the corresponding wild animals to obtain the longitude and latitude information of the wild animals at each time step;

[0033] The feature encoder module is used to map the preprocessed GPS trajectory data to a high-dimensional feature space using the feature encoder, so as to obtain the feature encoding vector corresponding to each time step.

[0034] The trajectory encoder module is used to transform the feature encoding vector into a representation sequence that captures temporal dependencies and contextual information using a pre-constructed trajectory encoder.

[0035] The trajectory awareness network module is used to input the represented sequence into the constructed trajectory awareness network for processing, and obtain the scores of each candidate position in the preset candidate list for the next time step;

[0036] The trajectory location determination module is used to perform real-time trajectory correction on the candidate location with the highest score in the preset candidate list, and use the corrected location as the trajectory location of the wild animal in the next time step.

[0037] Thirdly, this application provides an electronic device, including: at least one processor, at least one memory, and a data bus;

[0038] In this system, the processor and memory communicate with each other via a data bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the wildlife trajectory tracking method based on trajectory perception and motion correction, as described in any of the first aspects.

[0039] Fourthly, this application provides a non-transitory computer-readable storage medium storing computer instructions that cause a computer to execute the wildlife trajectory tracking method based on trajectory perception and motion correction according to any one of the first aspects.

[0040] Compared with the prior art, the present invention has at least the following beneficial effects:

[0041] 1. First, this invention employs a contextual relationship mechanism in a feature encoder to capture the correlation between features such as timestamps and latitude / longitude during the encoding stage. This allows different features to influence each other during encoding, providing a richer and more accurate input representation for subsequent processing. Second, it utilizes a trajectory perceptron based on a multi-scale self-attention mechanism and a forward masking mechanism. By processing multiple spatial scales in parallel, it can simultaneously focus on features at different spatial scales of the trajectory sequence, such as local and global spatial distribution features. This effectively captures complex temporal dependencies and nonlinear motion patterns in animal trajectories, solving the problem of capturing complex temporal dependencies in historical trajectories and ensuring that the output depends only on historical input features. To prevent future information leakage, the system significantly improves the accuracy and effectiveness of future position prediction. Finally, through the dynamic system model and Kalman filter algorithm of the trajectory real-time correction module, the candidate positions output by the trajectory inference layer are dynamically corrected in real time. By modeling process noise and observation noise, the uncertainty of the system model and the uncertainty of sensor measurement are described respectively. Real-time measurement data are fused, and the weights of prediction information and observation information are adaptively balanced. This overcomes the problems of sensor noise processing, environmental interference and nonlinear motion mode processing in traditional trajectory tracking algorithms, reduces the error caused by sensor noise and external environmental fluctuations, and improves the accuracy and robustness of trajectory tracking.

[0042] 2. This invention adaptively adjusts the candidate position evaluation criteria through a trajectory sensor and uses a cross-entropy loss function for training and optimization, thereby improving the adaptability to trajectory changes and partial data loss in complex field environments and enhancing robustness in complex environments. The overall algorithm structure is clear, the parameters are flexible and controllable, and it is easy to optimize and tailor. It has good adaptability and practical value in intelligent infrared cameras or other resource-constrained field monitoring equipment and is suitable for embedded device deployment.

[0043] 3. In this application, a trajectory perception network is used to perform in-depth modeling of the temporal dependence and spatial patterns of historical trajectories, and a motion correction module is used to perform real-time dynamic correction of the prediction results. This enables high-precision and robust tracking of wildlife trajectories in complex field environments, providing more reliable data support for wildlife behavior analysis and ecological protection. Attached Figure Description

[0044] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0045] Figure 1 This is a flowchart of the tracking method in an embodiment of the present invention;

[0046] Figure 2 This is a flowchart of the algorithm in the tracking method of this invention.

[0047] Figure 3 This is a schematic diagram of the feature encoder in an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of the trajectory encoder structure in an embodiment of the present invention;

[0049] Figure 5 This is a schematic diagram of the trajectory sensing network in an embodiment of the present invention;

[0050] Figure 6 This is a schematic diagram of the connection of the tracking system in an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0052] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0053] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0054] In the description of the embodiments of the present invention, "multiple" means at least two.

[0055] Example 1: Existing wildlife tracking methods often focus on unidirectional prediction of trajectory patterns, with insufficient consideration for the joint optimization of trajectory modeling and motion state estimation. They lack mechanisms for dynamic correction and uncertainty control of prediction results, making it difficult to maintain stable tracking performance in noisy and highly uncertain wild environments. This example provides a wildlife trajectory tracking method based on trajectory perception and motion correction, such as... Figure 1 and Figure 2 As shown, the specific steps include the following:

[0056] S1, preprocess the collected GPS trajectory data of the corresponding wild animals to obtain the longitude and latitude information of the wild animals at each time step.

[0057] The process involves using intelligent infrared camera terminals or similar devices to collect GPS trajectory data of wild animals. The preprocessing includes extracting the timestamp (time step), longitude, and latitude information from the GPS trajectory data and saving it to different files according to the animal's identification number. Each individual animal's data is saved separately to facilitate subsequent tracking experiments on individual animal trajectories. In addition, bilinear interpolation can be used to fill in missing GPS trajectory data to ensure the continuity of the trajectory, and an adaptive filter can be used to remove random noise from the data to further improve the quality of the trajectory data.

[0058] S2 uses a feature encoder to map the preprocessed GPS trajectory data to a high-dimensional feature space, obtaining the feature encoding vector corresponding to each time step.

[0059] Specifically, a feature encoder can be used to convert preprocessed GPS trajectory data into a unified high-dimensional vector representation. The structure of the feature encoder is as follows: Figure 3 As shown, the specific steps of this feature encoder are as follows:

[0060] 1. Each feature at time step The primitive representation of can be written as Its original dimensions vary depending on the data type; in order to combine the various Projecting onto a unified embedding space, the weight matrix and bias vector are defined as follows: ; in, To make the first Each feature from its original dimension Mapped to The weight matrix, For learnable contextual parameters, represent the first... Class features for the first The influence strength of class feature encoding This is the corresponding bias vector.

[0061] 2. To achieve dimensional alignment and create conditions for subsequent feature fusion, the feature encoder maps each feature to a unified space through linear transformation, as shown in the following formula. Considering the correlation between features such as timestamps and latitude and longitude in GPS trajectory data (e.g., the relationship between time interval and distance / speed), the feature encoder introduces contextual relationships between features during the mapping process, enabling different features to influence each other during encoding, thereby capturing trajectory features more accurately.

[0062] ; in, Indicates the first Class features in timestamps The representation after projection; To make the first Each feature from its original dimension Mapped to The weight matrix; For the first Class features at time step The original representation; For learnable contextual parameters, represent the first... Class features for the first The influence strength of class feature encoding; For the projection function, the first... Class features projected to the first A dimensional space compatible with class features ensures dimensional matching; This is the corresponding bias vector; This means for all except the first Aggregate other features outside the class.

[0063] 3. By introducing contextual relationships between features, the feature encoder can capture the correlation between timestamps and latitude / longitude during the encoding stage, such as the relationship between time intervals and distance traveled, and the relationship between timestamp changes and the rate of location change. This provides a richer and more accurate input representation for the subsequent trajectory encoder. For the differences in the importance of each feature in the task, the feature encoder uses learnable scalar weights to automatically adjust the contribution of each feature during training. All projected and weighted features are fused to obtain the feature encoding vector for each time step, providing rich and unified information for the subsequent trajectory encoder. Its construction formula is shown below: ; In the formula, The feature encoding vector for each time step, This indicates the total number of feature types used. As a learnable scalar, reflecting the first The contribution of class features to the final representation.

[0064] S3 transforms the feature encoding vector into a representation sequence that captures temporal dependencies and contextual information through a well-constructed trajectory encoder.

[0065] The trajectory encoder utilizes a multi-scale self-attention mechanism and a forward masking mechanism to capture the temporal dependencies and contextual features of the trajectory, transforming the input vector sequence into a representation sequence that captures temporal dependencies and contextual information, thus providing contextual basis for subsequent trajectory inference. The structure of the trajectory encoder is as follows: Figure 4 As shown, the specific steps of the trajectory encoder are as follows:

[0066] 1. For a given input sequence, the input embedding is first mapped to a different vector space for self-attention computation through a linear transformation, enabling the algorithm to capture the correlation between time steps by calculating similarity in subsequent steps; the resulting query, key, and value vectors are shown in the following equation: ; in, Given an input sequence of embedding vectors, each element ; These generate projection matrices for the query, key, and value, respectively. These matrices can be learned through training. The dimension representing the internal representation of attention; These represent the query matrix, key matrix, and value matrix across all time steps, with each row vector corresponding to the representation of each time step.

[0067] 2. Employ a multi-scale self-attention (MSA) mechanism to enhance the model's ability to focus on multi-scale features. See [link to relevant documentation]. Figure 4 Multi-scale kernels can be 1×3, 1×5, and 1×7 convolutional kernels, respectively. Convolutional kernels of different scales capture the features of trajectory sequences at different spatial scales, thereby providing a more comprehensive understanding of the temporal dependencies and spatial distribution features of the trajectories. Multi-scale self-attention can be expressed as the following formula: ; Among them, Q, K and These represent query, key, and value, respectively. express Attention output at any given moment; d represents the dimension of K; represents time; n represents the scale, which can take values ​​of 1, 2, and 3. Value 1 corresponds to a 1×3 scale, value 2 corresponds to a 1×5 scale, and value 3 corresponds to a 7×7 scale. This indicates aggregation at different scales; Softmax(·) is the normalization exponential function. This is a scaling factor used to prevent the dot product value from becoming too large, which would cause the function gradient to become too small and ensure numerical stability.

[0068] 3. Multi-scale self-attention mechanisms, by processing multiple spatial scales in parallel, can simultaneously focus on the characteristics of trajectory sequences at different spatial scales, such as local and global spatial distribution features, thereby more comprehensively capturing the temporal dependencies and spatial patterns of the trajectory; to ensure that at computation time... When outputting, only time 1 to... The information is used to apply unidirectional masking to multi-scale attention to prevent future information leakage and ensure the effectiveness of real-time prediction, as shown in the following formula: ; In the formula, , A masking matrix, used to shield future information and ensure time step... The output depends only on the previous input; Represents the first in the matrix line, number The elements of the column, if It is set to negative infinity to shield future information; through this one-way masking, the algorithm ensures that it does not leak future information when generating the representation of the current moment, thus meeting the requirements of time series prediction.

[0069] 4. The trajectory encoder is further optimized using residual mechanisms and layer normalization, integrating multi-scale self-attention and feedforward neural networks to obtain the trajectory encoder output sequence, as shown in the following equation: ; in, The output of the trajectory encoder represents a sequence. Then it equals the initial input embedding sequence. ; For multi-head self-attention sublayers, implement query, key and value calculation and masking operations; It is a feedforward neural network sublayer containing two linear layers and a non-linear activation function, used to further transform the representation at each time step; This is a layer normalization operation used to stabilize the training process and accelerate convergence.

[0070] S4, the sequence is input into the constructed trajectory perception network for processing to obtain the scores of each candidate position in the preset candidate list for the next time step.

[0071] The trajectory encoder output sequence is input into the trajectory sensing network (PN) to score the next position against a set of candidate positions, predicting the animal's next position. The structure of the trajectory sensing network (PN) is as follows: Figure 5 As shown, the specific steps of trajectory perception are as follows:

[0072] 1. Assume the trajectory encoder is at time step The output is The query vector is obtained through linear transformation. As shown in the following formula: ; in, The learned projection matrix; For the trajectory encoder at time step The output sequence is represented by the given information. The generated query vector is used to calculate the similarity with the candidate key vector; This refers to the dimension used in the trajectory inference layer.

[0073] 2. For candidate positions in the preset candidate list Candidates are obtained through a trajectory encoder, along with their related features. In order to merge the current representation sequence The features of the candidate positions are combined with those of the candidate positions, and the two are concatenated to obtain the candidate key vector through the candidate encoder, as shown in the following formula: ; in, Candidate position Embedded vector after processing by the feature encoding layer; For the current moment The representation sequence; The concatenation operation between the candidate embedding and the representation sequence generates a joint feature vector; This is a candidate encoder used to map the concatenated vector to the same space as the query vector. Let be the candidate key vector, representing the features of the candidate positions and their relationship in the current representation sequence.

[0074] 3. Utilizing query vectors and each candidate key vector The similarity between the pairs is calculated using dot product, and then normalized using softmax to obtain candidate positions. The conditional probability of being selected is shown in the following formula: ; in, Represents the query vector With candidate key vectors The dot product of two values ​​measures the similarity or correlation between them. This is a scaling factor used to balance the dot product values ​​and prevent the softmax gradient from vanishing due to excessively large values. It is an exponential function, used to amplify the impact of larger scores; This represents normalizing all candidate positions so that the sum of the probabilities of all candidates is 1, where This represents the total number of candidate positions. This indicates that, given the current representation sequence Under the condition of, candidate position The probability of being selected. Indicates including And other relevant information.

[0075] In the above, the core idea of ​​the trajectory-aware network is to transform the representation sequence output by the trajectory encoder... Convert to query vector Meanwhile, the original features of the candidate positions are processed to obtain candidate key vectors. This allows the algorithm to adaptively adjust the evaluation criteria for candidate positions in different contexts.

[0076] S5 performs real-time trajectory correction on the candidate position with the highest score in the preset candidate list, and uses the corrected position as the trajectory position of the wild animal in the next time step. By continuously obtaining the trajectory position of the animal, the trajectory tracking process is completed.

[0077] Specifically, the real-time trajectory correction module can correct the predicted candidate trajectory position in real time, calculate the Kalman gain based on real-time observation data to correct the prediction error, and update the trajectory position to obtain a more accurate trajectory output, reducing errors caused by sensor noise or environmental changes. The input of the real-time trajectory correction module is the candidate position output by the trajectory sensor. By using a Kalman filter to reduce the error between the tracked trajectory and the true trajectory, real-time trajectory tracking correction is achieved, as shown in the following equation: ; in, It is the state transition matrix. It is a control input matrix. It is a control vector. It is process noise (system noise), which represents the uncertainty of the system model, such as environmental disturbances and model approximation errors; The position is predicted based on the previous moment's state, control input, and process noise. , process noise covariance matrix Calculated through initialization using the real-time trajectory correction module; process noise. It follows a pattern with a mean of 0 and a covariance matrix of . The Gaussian distribution, i.e. .

[0078] S6 can construct a trajectory tracking model through a feature encoder, a trajectory encoder, and a trajectory perception network. Then, the prediction results of the trajectory inference layer are optimized using the cross-entropy loss function. The parameters of the feature encoder, trajectory encoder, and trajectory inference layer are updated through the backpropagation algorithm, enabling the model to better learn trajectory patterns and improve prediction accuracy.

[0079] S61, continuously input each training sample into the trajectory tracking model for processing, and obtain the predicted position corresponding to each training sample. The training samples include at least the longitude and latitude information of the corresponding wild animals at each time step.

[0080] S62, based on the preset candidate list and predicted position of the next time step, calculate the loss function value corresponding to each training sample;

[0081] In this context, the first candidate in the preset candidate list can be set as the actual location, i.e., at each time step. The real label is The loss function is defined as the mean of the negative log-likelihood over all time steps, as shown in the following formula: ; in, The overall loss value reflects the algorithm's performance across the entire training sample. The closer the value is to 0, the higher the probability of obtaining a true candidate and the smaller the loss. Let be the total number of time steps in the trajectory. Since the first time step has no prior information, the summation in the formula ranges from 1 to 1. ; This means averaging the loss over all time steps, so that the loss has a stable scale and is not affected by changes in trajectory length; Indicates the number of candidates The base logarithm is used to normalize the scale, making the losses under different numbers of candidates comparable; This indicates that, given the current representation sequence Under the given conditions, the probability that the first candidate position (the actual position) is selected.

[0082] S63, based on the loss function value and backpropagation algorithm, at least the feature encoding layer, weight matrix, bias vector and projection matrix in the trajectory tracking model are updated; until the loss function value reaches a preset condition, and the trajectory tracking model with the loss function value reaching the preset condition is used to predict the trajectory position of wild animals in the next time step; wherein, the preset condition may be that the loss function value is less than a threshold.

[0083] Example 2: This application provides a wildlife trajectory tracking system based on trajectory perception and motion correction, applied to the wildlife trajectory tracking method based on trajectory perception and motion correction in Example 1, such as... Figure 6 As shown, it includes:

[0084] The preprocessing module is used to preprocess the collected GPS trajectory data of the corresponding wild animals to obtain the longitude and latitude information of the wild animals at each time step;

[0085] The feature encoder module is used to map the preprocessed GPS trajectory data to a high-dimensional feature space using the feature encoder, so as to obtain the feature encoding vector corresponding to each time step.

[0086] The trajectory encoder module is used to transform the feature encoding vector into a representation sequence that captures temporal dependencies and contextual information using a pre-constructed trajectory encoder.

[0087] The trajectory awareness network module is used to input the represented sequence into the constructed trajectory awareness network for processing, and obtain the scores of each candidate position in the preset candidate list for the next time step;

[0088] The trajectory location determination module is used to perform real-time trajectory correction on the candidate location with the highest score in the preset candidate list, and use the corrected location as the trajectory location of the wild animal in the next time step.

[0089] Example 3: This application provides an electronic device, including: at least one processor, at least one memory, and a data bus;

[0090] The processor and memory communicate with each other via a data bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the wildlife trajectory tracking method based on trajectory perception and motion correction as described in Example 1.

[0091] Example 4: This application provides a non-transitory computer-readable storage medium that stores computer instructions that cause a computer to execute the wildlife trajectory tracking method based on trajectory perception and motion correction described in Example 1.

[0092] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0093] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0094] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0095] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0096] Those skilled in the art will understand that all or part of the steps in the above facts and methods can be implemented by a program instructing related hardware. The program or the program described therein can be stored in a computer-readable storage medium. When the program is executed, it includes the following steps: at this time, the corresponding method steps are introduced. The storage medium can be ROM / RAM, magnetic disk, optical disk, etc.

[0097] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for tracking wildlife trajectories based on trajectory perception and motion correction, characterized in that, The specific steps include the following: The collected GPS trajectory data of the corresponding wild animals are preprocessed to obtain the longitude and latitude information of the wild animals at each time step; The preprocessed GPS trajectory data is mapped to a high-dimensional feature space using a feature encoder to obtain the feature encoding vectors corresponding to each time step. The constructed trajectory encoder transforms the feature encoding vector into a representation sequence that captures temporal dependencies and contextual information. The represented sequence is input into the constructed trajectory perception network for processing to obtain the scores of each candidate position in the preset candidate list for the next time step; The candidate position with the highest score in the preset candidate list is corrected in real time, and the corrected position is used as the trajectory position of the wild animal in the next time step.

2. The wildlife trajectory tracking method based on trajectory perception and motion correction according to claim 1, characterized in that, The feature encoding vectors corresponding to each time step are specifically as follows: ； in: ； In the formula, For time steps The feature encoding vector, As a learnable scalar, reflecting the first The contribution of class features to the final representation; Indicates the first Class features at time step The representation after projection. This represents the total number of feature types used. To make the first Each feature from its original dimension Mapping to higher dimensions The weight matrix, For the first Class features at time step The original representation, This means for all except the first Aggregate other features outside the class. For learnable contextual parameters, represent the first... Class features for the first The influence strength of class feature encoding; Represents the projection function. This is the corresponding bias vector.

3. The wildlife trajectory tracking method based on trajectory perception and motion correction according to claim 1, characterized in that, The representation sequence is specifically as follows: ； In the formula, This is the representation sequence output by the trajectory encoder. For multi-head self-attention quantum layer, This represents a sublayer of a feedforward neural network. Presentation layer normalization operation.

4. The wildlife trajectory tracking method based on trajectory perception and motion correction according to claim 3, characterized in that, The specific processing result of the multi-head self-attention sublayer is as follows: ， ； In the formula, Represents the normalized exponential function, This indicates aggregation at different scales. Scaling factor This is used to shield future information and ensure time steps. The output depends only on the masking matrix of the previous input. Represents the first in the matrix line, number Column elements; , , These represent the query vector, key vector, and value vector generated by the corresponding projection matrices, respectively.

5. The wildlife trajectory tracking method based on trajectory perception and motion correction according to claim 1, characterized in that, The scores for each candidate position in the preset candidate list for the next time step are obtained in the following way: Each candidate position in the preset candidate list is processed by a trajectory encoder to obtain the features of each candidate position; Each candidate position feature is concatenated with the representation sequence to obtain the candidate key vector corresponding to each candidate position; The representation sequence is transformed linearly to obtain the corresponding query vector. Based on the query vector and each candidate key vector, the conditional probability of each candidate position being selected is calculated, and each conditional probability is used as the score of the candidate position.

6. The wildlife trajectory tracking method based on trajectory perception and motion correction according to claim 5, characterized in that, The conditional probability of each candidate position being selected is calculated as follows: ,in: ， ； In the formula, Indicates that, given the current representation sequence Under the condition of, candidate position The conditional probability of being selected. Represents the query vector With candidate key vectors dot product, Indicates the scaling factor. It is an exponential function. This represents normalizing all candidate positions so that the sum of the probabilities of all candidates is 1, where This represents the total number of candidate positions. Indicates candidate position Embedded vector after feature encoding layer For the current moment The representation sequence, Indicates candidate position The candidate position features and their related features are obtained through the trajectory encoder. For candidate encoders, For the generated query vector, This is the projection matrix obtained through learning.

7. The wildlife trajectory tracking method based on trajectory perception and motion correction according to claim 1, characterized in that, The method further includes: A trajectory tracking model is constructed using the feature encoder, the trajectory encoder, and the trajectory awareness network. The trajectory tracking model is trained in the following manner: Each training sample is continuously input into the trajectory tracking model for processing to obtain the predicted position corresponding to each training sample. The training sample includes at least the longitude and latitude information of the corresponding wild animal at each time step. Based on the preset candidate list and the predicted position at the next time step, calculate the loss function value corresponding to each training sample; Based on the loss function value and backpropagation algorithm, at least the feature encoding layer, weight matrix, bias vector and projection matrix in the trajectory tracking model are updated; The trajectory tracking model is used to predict the trajectory position of wild animals in the next time step until the loss function value reaches a preset condition.

8. The wildlife trajectory tracking method based on trajectory perception and motion correction according to claim 7, characterized in that, The loss function value is obtained in the following way: ； In the formula, The value of the loss function. This represents the total number of time steps in the trajectory. Indicates the number of candidates Logarithm with base 0, Indicates that, given the current representation sequence Under the given conditions, the conditional probability of the actual position being selected from the candidate list.

9. The wildlife trajectory tracking method based on trajectory perception and motion correction according to claim 1, characterized in that, The trajectory of the candidate position with the highest score in the preset candidate list is corrected in real time, specifically as follows: ； In the formula, This indicates the position predicted based on the previous moment's state, control input, and process noise, i.e., the position after real-time trajectory correction. It is the state transition matrix. It is a control input matrix. It is a control vector. It is process noise.

10. A wildlife trajectory tracking system based on trajectory perception and motion correction, characterized in that, include: The preprocessing module is used to preprocess the collected GPS trajectory data of the corresponding wild animals to obtain the longitude and latitude information of the wild animals at each time step; The feature encoder module is used to map the preprocessed GPS trajectory data to a high-dimensional feature space using the feature encoder, so as to obtain the feature encoding vector corresponding to each time step. The trajectory encoder module is used to transform the feature encoding vector into a representation sequence that captures temporal dependencies and contextual information through a constructed trajectory encoder. The trajectory sensing network module is used to input the representation sequence into the constructed trajectory sensing network for processing, and obtain the scores of each candidate position in the preset candidate list for the next time step; The trajectory location determination module is used to perform real-time trajectory correction on the candidate location with the highest score in the preset candidate list, and use the corrected location as the trajectory location of the wild animal in the next time step.

Images