Method for evaluating work performance of working dog

Abstract

The invention discloses a working dog work performance evaluation method, belongs to the technical field of working dog evaluation, and can solve the problem that an existing evaluation method is low in accuracy and efficiency. The method comprises the following steps: S1, determining a working state, a motion state and a health state of a working dog at each moment according to motion data of the working dog at each moment in a plurality of moments; s2, when the combination of the motion state and the health state meets a preset condition, determining a plurality of evaluation indexes of the working dog according to the motion data or the working state; and S3, evaluating the work performance of the working dog according to the plurality of evaluation indexes. The method is used for evaluating the working performance of the working dog.

Description

Technical Field

[0001] The present invention relates to a method for evaluating the working performance of working dogs, belonging to the technical field of working dog evaluation. Background Art

[0002] Working dogs play an irreplaceable and important role in many fields such as public security, fire fighting, and search and rescue. Their working performance is directly related to the success or failure of tasks. However, at present, the evaluation of the working performance of working dogs mostly relies on manual observation and subjective judgment. This evaluation method is not only inefficient but also easily affected by the subjective factors of evaluators, resulting in inaccurate evaluation results, and thus unable to comprehensively, objectively and efficiently reflect the actual performance and contributions of working dogs. Summary of the Invention

[0003] The present invention provides a method for evaluating the working performance of working dogs, which can solve the problems of low accuracy and efficiency of existing evaluation methods.

[0004] The present invention provides a method for evaluating the working performance of working dogs, and the method includes:

[0005] S1. Determine the working state, movement state and health state of the working dog at each moment according to the movement data of the working dog at multiple moments;

[0006] S2. When the combination of the movement state and the health state meets a preset condition, determine multiple evaluation indexes of the working dog according to the movement data or the working state;

[0007] S3. Evaluate the working performance of the working dog according to multiple evaluation indexes.

[0008] Optionally, the S1 specifically includes:

[0009] S11. Construct a state recognition model based on a feature extraction method, a long short-term memory network and an attention mechanism;

[0010] S12. According to the movement data of the working dog at multiple moments, use the state recognition model to determine the working state, movement state and health state of the working dog at each moment.

[0011] Optionally, the S11 specifically includes:

[0012] Construct a feature extraction unit based on a feature extraction method; the feature extraction unit is used to extract data features from the movement data of the working dog;

[0013] Construct a state recognition unit based on a long short-term memory network and an attention mechanism; the state recognition unit is used to recognize the working state, movement state and health state of the working dog according to the data features;

[0014] Construct a state recognition model based on the feature extraction unit and the state recognition unit.

[0015] Optionally, the feature extraction method is a one-dimensional convolutional feature extraction method.

[0016] Optionally, construct a state recognition unit based on a long short-term memory network and an attention mechanism, specifically including:

[0017] Construct a first loss function, a second loss function, and a third loss function for classifying the working state, motion state, and health state of the working dog respectively;

[0018] Based on the first loss function, the second loss function, and the third loss function, construct a state recognition unit based on a long short-term memory network and an attention mechanism.

[0019] Optionally, based on the first loss function, the second loss function, and the third loss function, construct a state recognition unit based on a long short-term memory network and an attention mechanism, specifically including:

[0020] According to the correlation relationship between the working state, motion state, and health state of the working dog, use the first loss function, the second loss function, and the third loss function to construct a total loss function;

[0021] Based on the total loss function, construct a state recognition unit based on a long short-term memory network and an attention mechanism.

[0022] Optionally, the multiple evaluation indicators include working duration, exercise intensity, path rationality, and coverage.

[0023] Optionally, the working state includes patrolling, searching, tracking, and resting; determining the working duration of the working dog according to the working state specifically includes:

[0024] According to the working state at each moment and the time interval between adjacent moments, determine the patrolling duration, searching duration, tracking duration, and resting duration of the working dog;

[0025] Determine the working duration of the working dog according to the patrolling duration, the searching duration, the tracking duration, and the resting duration.

[0026] Optionally, the motion data includes acceleration; determining the exercise intensity of the working dog according to the motion data specifically includes:

[0027] Determine the acceleration change rate and speed change rate of the working dog according to the acceleration;

[0028] Determine the exercise intensity of the working dog according to the acceleration change rate and the velocity change rate.

[0029] Optionally, the motion data includes position information; determining the path rationality of the working dog according to the motion data specifically includes:

[0030] Use a path planning algorithm to determine the optimal path of the target task;

[0031] Determine the actual path of the working dog according to the position information, and determine the path rationality of the working dog according to the actual path and the optimal path.

[0032] The beneficial effects that the present invention can produce include:

[0033] The present invention can determine the working state, motion state and health state of a working dog according to its motion data, and then evaluate the working performance of the working dog based on the working state or motion data. It can accurately identify the working state of the working dog, and compared with manual observation and subjective judgment, it improves the accuracy and efficiency of the evaluation, and can comprehensively, objectively and efficiently reflect the actual performance and contribution of the working dog.

[0034] Based on the one-dimensional convolutional feature extraction method and the long short-term memory network, and combined with the attention mechanism, the present invention constructs a state recognition model for identifying the working state, motion state and health state of a working dog, realizing the fusion of multiple algorithms for the state recognition of a working dog, and can improve the robustness and recognition accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flowchart of the method for evaluating the working performance of a working dog provided by an embodiment of the present invention;

[0036] Figure 2 It is a schematic diagram of the network structure of the state recognition model based on the traditional feature extraction method provided by an embodiment of the present invention;

[0037] Figure 3 It is a schematic diagram of the network structure of the state recognition model based on the one-dimensional convolutional feature extraction method provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0038] The present invention will be described in detail below with reference to the embodiments, but the present invention is not limited to these embodiments.

[0039] An embodiment of the present invention provides a method for evaluating the working performance of a working dog, as Figure 1 shown, the method includes:

[0040] S1. Determine the working status, motion status, and health status of a working dog at each moment based on the motion data of the working dog at multiple moments.

[0041] Specifically, the working status of the working dog can include patrolling, searching, tracking, and resting; the motion status can include normal and abnormal; and the health status can include good, fatigued, and stressed.

[0042] Specifically, the motion data can include the position information of the working dog, the accelerations of the working dog along the three axes of x, y, and z, the angular velocities of the working dog around the three axes of x, y, and z, and the magnetic field intensities of the three axes of x, y, and z. Where x and y are two mutually perpendicular coordinate axes on the horizontal plane, and z is the vertical coordinate axis.

[0043] In this embodiment, a wearable instrument device integrating instruments such as an accelerometer, a gyroscope, a magnetometer, and a positioning module is stably worn on the neck of the working dog to collect relevant motion data.

[0044] The functions of each instrument are as follows:

[0045] 1. Accelerometer: Used to collect the linear accelerations of the working dog along the three axes of x, y, and z. The sampling frequency range can be 0.001 Hz - 100 Hz, and the acceleration can reflect the linear motion change of the working dog.

[0046] 2. Gyroscope: Used to collect the angular velocities of the working dog around the three axes of x, y, and z. The acceleration can reflect the rotational motion changes of the working dog such as turning and rolling. The range of the gyroscope can be ±2000 DPS, and the sampling frequency range can be between 0.001 Hz - 100 Hz.

[0047] 3. Magnetometer: Used to collect the magnetic field intensities on the three axes of x, y, and z. The magnetic field intensity can reflect the position and attitude changes of the working dog in the magnetic field environment. The range of the magnetometer on the x and y axes is ±1300 μT, and the range on the z axis is ±2500 μT. The sampling frequency range can be 0.001 Hz - 25 Hz.

[0048] 4. Positioning module: Used to collect the real-time position information of the working dog. The positioning module can be a GPS positioning module or a Beidou positioning module.

[0049] During the execution of various tasks by the working dog, the above instruments continuously collect relevant data. Whether in different scenarios such as complex urban street environments, rugged mountain terrains, or complex internal spaces of buildings, it can ensure the integrity and continuity of data collection to comprehensively obtain the motion data of the working dog under various working conditions.

[0050] After collecting the motion data, this embodiment can also preprocess the operation data. The preprocessing specifically includes:

[0051] 1. Noise removal

[0052] For the collected motion data, use a Butterworth low-pass filter for noise removal processing. Mean filtering, median filtering, Kalman filtering, extended Kalman filtering and other methods can also be used for noise removal processing. The Butterworth filter can effectively suppress high-frequency noise components. By setting an appropriate cut-off frequency, the biological signals reflecting the actual activities of the working dog are retained. For example, according to the frequency characteristics of each instrument, select a cut-off frequency of 10 Hz, so that the noise signals above this cut-off frequency are greatly attenuated, thereby improving the quality and reliability of the data.

[0053] 2. Data formatting and synchronization

[0054] Unify the formatting of data from different instruments to ensure that all data has a consistent format and accurate timestamps. Through a timestamp-based synchronization algorithm, ensure that the data collected by the accelerometer, gyroscope, magnetometer and positioning module are accurately aligned on the time axis, so that subsequent data fusion and analysis can be accurately performed.

[0055] S1 can specifically include:

[0056] S11. Construct a state recognition model based on the feature extraction method, long short-term memory network and attention mechanism.

[0057] S11 can specifically include:

[0058] S111. Construct a feature extraction unit based on the feature extraction method; the feature extraction unit is used to extract data features from the motion data of the working dog.

[0059] The feature extraction unit can further process the preprocessed motion data and extract data features that can be used to identify the working state, motion state and health state of the working dog. The construction of the feature extraction unit can adopt traditional feature extraction methods or one-dimensional convolutional neural network (1D-CNN) feature extraction methods.

[0060] Specifically, the process of the traditional feature extraction method includes:

[0061] 1. Extract feature values

[0062] Extract the following feature values from the data collected by the accelerometer, gyroscope, magnetometer and positioning module respectively:

[0063] Mean value: Used to measure the average level of the data and can reflect the average situation of the working dog's movement within a certain period of time.

[0064] Mean absolute deviation: It is used to measure the average deviation degree of data relative to the average value, and can reflect the fluctuation of the working dog's movement.

[0065] Standard deviation: It is used to measure the dispersion degree of data, and can reflect the stability and change degree of the working dog's movement.

[0066] Minimum value and maximum value: They are respectively the minimum and maximum values in the data, and can reflect the extreme state of the working dog's movement.

[0067] Interquartile range: It is used to measure the middle dispersion degree of data.

[0068] Skewness: It is used to measure the asymmetry of data distribution, and can reflect whether there is a bias in the working dog's movement.

[0069] Kurtosis: It is used to measure the peakedness of data distribution, and can reflect the central tendency in the working dog's movement.

[0070] Energy: It is used to evaluate the energy size of the working dog's movement, and is related to the movement intensity of the working dog. The calculation formula of energy is:

[0071]

[0072] In formula (1), E is the movement energy of the working dog, and x i is the i-th sample data in the preprocessed movement data, and N is the number of sample data.

[0073] 2. Windowing processing

[0074] Adopt windowing technology to divide the continuous data collected by each instrument into multiple time windows according to a fixed value n. Each time window contains 2n sample data and overlaps n sample data with the previous time window. For the sample data within each time window, calculate the above-extracted feature values, and label the working state, movement state, and health state with the highest frequency within this time window.

[0075] 3. Frequency domain conversion

[0076] Apply the Fast Fourier Transform (FFT for short) to the sample data after windowing processing to convert the time-domain data into frequency-domain data, so as to obtain the distribution information of the movement data on different frequency components, in order to further mine the deep features of the working dog's movement, such as the energy distribution at different frequencies, for subsequent distinguishing the working state, movement state, and health state of the working dog.

[0077] Specifically, the process of the one-dimensional convolutional feature extraction method includes:

[0078] 1D-CNN has a powerful ability to automatically extract highly relevant features and is particularly suitable for processing time series data. It can slide the convolutional kernels of the convolutional layer over the time series data to automatically learn the local features in the data, which are used for subsequent identification of the working state, motion state, and health state of working dogs. To achieve preliminary feature extraction, we can use the input layer, convolutional layer, and output layer in 1D-CNN to implement feature extraction.

[0079] Input layer: Receive the preprocessed motion data and determine the data dimension of the input layer according to the number of features to be extracted. For example, if m features are to be extracted, then the data dimension of the input layer is m. Then, perform the aforementioned windowing process on the time series data.

[0080] Convolutional layer: Use several convolutional layers, such as convolutional layers with 128, 128, 256, and 256 convolutional kernels respectively, and set the stride to 1. The calculation method of the convolutional layer is where is the output of the j-th feature map in the l-th layer, is the input of the i-th feature map in the (l - 1)-th layer, is the convolutional kernel of the l-th layer, * represents the convolution operation, is the j-th bias term in the l-th layer, f is the activation function, such as the ReLU function f(x) = max(0, x). By stacking multiple convolutional layers, complex patterns in the data are gradually extracted.

[0081] Output layer: Use the output of 1D-CNN as the input of the subsequent state recognition unit.

[0082] Specifically, 1D-CNN can be 1D-CNN based on the VGG structure (VNet), 1D-CNN based on the EfficienNet structure (ENet), or 1D-CNN based on the ResNet structure (RNet), etc.

[0083] S112. Construct a state recognition unit based on the long short-term memory network and the attention mechanism; the state recognition unit is used to identify the working state, motion state, and health state of working dogs according to data features.

[0084] Long Short-Term Memory (LSTM) is a special type of Recursive Neural Network (RNN), which is particularly suitable for processing and predicting time series-based data. In the field of action recognition, LSTM is widely adopted because of its ability to capture long-term dependencies. This ability is crucial for extracting action features from time series data, as the recognition of the working, movement, and health states of working dogs usually requires an in-depth understanding of the evolution of signal patterns over time.

[0085] LSTM stores long-term and short-term information through cell memory states and hidden states, and transfers information between time series data, ensuring that past observations have a significant impact on current predictions. This is particularly important for distinguishing subtle action differences. For example, when distinguishing between "patrolling" and "searching", the sequential change in position is crucial. In addition, LSTM shows flexibility in processing sequences of different lengths, which is particularly important for identifying variable speed information in the movement data of working dogs.

[0086] The Attention Mechanism is a resource allocation scheme. Its core idea is that when processing input data, it does not pay equal attention to all parts, but dynamically allocates "attention" according to the current context, thereby highlighting key information.

[0087] In this embodiment, by introducing an attention encoder into the state recognition unit, the attention encoder is used to capture the attributes of specific aspects, enhancing the ability of the state recognition unit to analyze the nuances of complex content. The combination of LSTM and the attention mechanism provides a powerful tool for processing time series data. By simulating the process of human attention, the state recognition unit can pay more attention to important parts when processing input data, thereby improving the performance and effectiveness of the state recognition unit.

[0088] Specifically, the attention mechanism can be a multi-head self-attention mechanism.

[0089] In S112, constructing a state recognition unit based on the long short-term memory network and the attention mechanism can specifically include:

[0090] Respectively constructing a first loss function, a second loss function, and a third loss function for classifying the working state, movement state, and health state of the working dog;

[0091] According to the first loss function, the second loss function, and the third loss function, constructing a state recognition unit based on the long short-term memory network and the attention mechanism.

[0092] Furthermore, constructing a state recognition unit based on the first loss function, the second loss function, and the third loss function, using a long short-term memory network and an attention mechanism, specifically includes:

[0093] Construct a total loss function using the first loss function, the second loss function, and the third loss function according to the correlation relationship among the working state, the motion state, and the health state of the working dog;

[0094] Construct a state recognition unit based on the total loss function, using a long short-term memory network and an attention mechanism.

[0095] Specifically, let the first loss function for the working state be L1, the second loss function for the motion state classification be L2, and the third loss function for the health state classification be L3. The total loss function can be obtained by weighted summation of each loss, and the total loss function can be expressed as:

[0096] L = α1L1 + α2L2 + α3L3 (2)

[0097] In formula (2), L is the total loss function; L1, L2, and L3 are the first loss function, the second loss function, and the third loss function respectively; α1, α2, and α2 are the weight coefficients of the first loss function, the second loss function, and the third loss function respectively, and can be determined according to the importance of the three classification tasks of the working state classification, the motion state classification, and the health state classification.

[0098] However, since there is a high correlation among the working state, the motion state, and the health state, there is also a high correlation among the three classification tasks. Each classification task is not carried out independently. Therefore, in this embodiment, the total loss function can also be determined based on the correlation relationship among the working state, the motion state, and the health state of the working dog, and the correlation between classification tasks can be strengthened through a regularization term to enhance the synergy effect between classification tasks and reduce the conflict between classification tasks. At this time, the total loss function can be expressed as:

[0099]

[0100] In formula (3), L is the total loss function; Corr(L i , L j ) is the loss correlation between classification task L i and classification task L j , i = 1, 2, 3 and j = 1, 2, 3 and i ≠ j; λ is the regularization coefficient.

[0101] Among them, Corr(L i , L j)It can be calculated according to the correlation relationship among the working state, motion state and health state of the working dog, and by using Pearson correlation coefficient, mutual information or Jensen-Shannon divergence (JS divergence), etc.

[0102] S113. Construct a state recognition model according to the feature extraction unit and the state recognition unit.

[0103] Specifically, in this embodiment, based on the aforementioned feature extraction unit and state recognition unit, combined with a fully connected layer, a dropout layer and a classifier, a state recognition model is constructed. The network structure of the state recognition model based on the traditional feature extraction method is as Figure 2 shown, and the network structure of the state recognition model based on the one-dimensional convolutional feature extraction method is as Figure 3 shown.

[0104] Figure 2 and Figure 3 the Attention in it represents the attention encoder, Dense represents the fully connected layer, Dropout represents the dropout layer, and Softmax represents the classifier.

[0105] Among them, each neuron in the fully connected layer is connected to all neurons in the previous layer, forming a fully connected topological structure. The parameters mainly include a weight matrix and a bias vector. The weight matrix is used to represent the connection strength between each input feature and the output feature, and its size is usually the number of input units multiplied by the number of output units; the bias vector is used to adjust the offset of the output so that the model can better fit the data. Its working principle is linear transformation and non-linear activation. Commonly used activation functions can be used, such as ReLU, sigmoid, tanh, etc.

[0106] The dropout layer randomly discards the outputs of a part of neurons during the model training process, so that the model does not overly rely on certain neurons, thereby improving the generalization ability and robustness of the model.

[0107] As the last layer of the network, the classifier serves as the output layer for multi-classification problems, corresponding to processing classification tasks such as working state classification, motion state classification and health state classification. Specifically, when processing the working state classification task, the classifier can output the probability distribution of the working state; when processing the motion state classification task, the classifier can output the probability of abnormal motion state; when processing the health state classification task, the classifier can use a regression layer to output the score of the health state.

[0108] After the state recognition model is constructed, in this embodiment, the collected historical motion data of the working dog is used to train the state recognition model. The training process is as follows:

[0109] 1. Division of training set and test set

[0110] During training, it is necessary to first prepare labeled data for each classification task. For the work status classification task, it is necessary to label the categories of each work status; for the motion status classification task, it is necessary to label the motion status abnormal events; for the health status classification task, it is necessary to label the scores corresponding to the good, fatigued, and stressed health statuses.

[0111] Then, randomly divide the historical motion data of the working dog into a training set and a test set according to a certain proportion, ensuring that the training set and the test set are representative in terms of data distribution. For example, the sample proportion of each category of work status in the training set and the test set is similar to its proportion in the overall data.

[0112] 2. Class Weight Adjustment

[0113] During the training process, calculate the weighted loss of samples of different categories according to the class weights, pay more attention to the minority class samples, prevent the model from biasing towards the majority class, and improve the accuracy of the model in classifying each status.

[0114] Exemplarily, the calculation formula for class weights is:

[0115]

[0116] In Equation (4), W i is the class weight of the samples of the i-th category; n_instances is the total number of samples in the dataset; n_classes is the number of categories; n_instances i is the number of samples of the i-th class.

[0117] 3. Hyperparameter Adjustment

[0118] Through multiple experiments and optimizations, determine the hyperparameters of the model. For example, the learning rate is initially set to 0.0001, and during the training process, according to the change of the total loss function, adopt a learning rate decay strategy. For example, after a certain number of iterations, the learning rate is multiplied by a decay factor less than 1 to ensure that the model can converge quickly in the initial stage of training and be stably optimized in the later stage. The number of iterations is determined according to the training effect. An optimizer such as Stochastic Gradient Descent (SGD) can be used to train the model to minimize the classification cross-entropy loss function.

[0119] S12. According to the motion data of the working dog at each of multiple moments, use the status recognition model to determine the work status, motion status, and health status of the working dog at each moment.

[0120] Specifically, the trained status recognition model can determine the current work status, motion status, and health status of the working dog according to the real-time motion data of the working dog.

[0121] Taking the determination of the working state as an example, the state recognition model can calculate the probabilities of the working dog being in different categories of motion states. When the probability of a certain category is the highest and its probability is greater than the preset threshold, it is determined that the working dog is currently in the motion state of that category. For example, the state recognition model calculates that the probability of the working dog being in the patrol state is 0.8, the probability of being in the search state is 0.15, and the probability of being in other states is 0.05, etc. Suppose the preset threshold is 0.7. Since the probability of the patrol state is the highest and its probability is greater than the preset threshold of 0.8, it is determined that the working dog is currently in the patrol state. Among them, the preset threshold can be reasonably set according to the relevant experimental results of the working dog.

[0122] S2. When the combination of the motion state and the health state meets the preset conditions, determine multiple evaluation indicators of the working dog according to the motion data or the working state.

[0123] Specifically, the preset condition is that the motion state is normal and the health state is good. When the combination of the motion state and the health state meets this preset condition, it means that the working dog is in normal work and its work performance can be evaluated. If the preset condition is not met, it means that the motion state of the working dog is abnormal or the health state is poor. At this time, the training or work of the working dog should be suspended, and the evaluation of its work performance should also be stopped.

[0124] Specifically, the multiple evaluation indicators can include working duration, exercise intensity, path rationality, and coverage.

[0125] In this embodiment, the above-mentioned determination of the working duration of the working dog according to the working state can specifically include:

[0126] Determine the patrol duration, search duration, tracking duration, and rest duration of the working dog according to the working state at each moment and the time interval between adjacent moments;

[0127] Determine the working duration of the working dog according to the patrol duration, search duration, tracking duration, and rest duration.

[0128] Specifically, the process of determining the working duration specifically includes:

[0129] 1. Statistic the working duration

[0130] To comprehensively evaluate the degree of engagement of working dogs in different working states, the duration of working dogs in different types of working states can be calculated based on the recognition results of the working states. Specifically, during the working process of working dogs, the working state of the working dog can be automatically judged at fixed time intervals (such as every minute or every five minutes). If the working dog continuously remains in the patrol state within a certain continuous period, then this period will be counted into the patrol duration. Similarly, the durations of other types of working states such as searching and tracking are statistically calculated. Through this cumulative calculation method, we can accurately obtain the specific time that working dogs invest in different working states, such as patrol duration, search duration, tracking duration, rest duration, etc. These data not only provide a data basis for quantitatively evaluating the working performance of working dogs, but also provide a scientific basis for the training optimization of working dogs and the rational utilization of canine power.

[0131] 2. Determine the effective duration

[0132] To ensure the accuracy of the data, when situations such as interruptions, rests, and task conversions occur during the work of working dogs, the durations corresponding to these situations should be subtracted from the working duration. For example, the rest duration is subtracted from the working duration to obtain the effective duration. The effective duration can reflect the actual time that working dogs invest in work and provide a data basis for evaluating the concentration and execution ability of working dogs.

[0133] In this embodiment, determining the exercise intensity of the working dog based on the motion data specifically includes:

[0134] Determine the acceleration change rate and speed change rate of the working dog according to the acceleration;

[0135] Determine the exercise intensity of the working dog according to the acceleration change rate and speed change rate.

[0136] Specifically, the process of determining the exercise intensity specifically includes:

[0137] Calculate the acceleration change rate according to the acceleration of the working dog. The calculation formula is:

[0138]

[0139] In formula (5), a rate is the acceleration change rate of the working dog; Δa is the acceleration change amount of the working dog between adjacent moments; Δt is the time interval between adjacent moments.

[0140] The acceleration change rate can reflect the acceleration and deceleration conditions of the working dog. Frequent and large-amplitude acceleration and deceleration indicate a high exercise intensity.

[0141] Meanwhile, the speed of the working dog can be calculated by integrating the acceleration and other methods. Based on the speed, the rate of change of speed can be calculated, and the calculation formula is as follows:

[0142]

[0143] In Equation (6), v rate is the rate of change of speed of the working dog; Δv is the change in speed of the working dog between adjacent time instants; Δt is the time interval between adjacent time instants.

[0144] The rate of change of speed can reflect the speed fluctuation of the working dog. A large speed fluctuation means that the activity intensity of the working dog changes greatly.

[0145] By comprehensively analyzing the acceleration rate of change and the speed rate of change, the analysis result of the exercise intensity of the working dog can be obtained to evaluate the working effort and working effect of the working dog.

[0146] In this embodiment, determining the path rationality of the working dog based on the motion data specifically includes:

[0147] Using a path planning algorithm to determine the optimal path for the target task;

[0148] Determining the actual path of the working dog based on the position information, and determining the path rationality of the working dog based on the actual path and the optimal path.

[0149] Specifically, the process of determining the path rationality specifically includes:

[0150] 1. Determine the optimal path

[0151] Select a suitable path planning algorithm according to the type of the target task of the working dog to generate the optimal path. For example, for search tasks, a coverage path planning algorithm such as spiral search, grid search, etc. can be used to ensure that the optimal path fully covers the entire search area; for tracking tasks, a prediction algorithm based on the motion model of the tracking target can be considered, and the optimal path can be planned in combination with the motion ability of the working dog and environmental factors.

[0152] Set constraint conditions: In the actual scenario, environmental factors have an important impact on the actual path of the working dog. Therefore, when planning the optimal path, it is necessary to fully consider the environmental constraint conditions. For example, terrain and landforms (such as mountains, rivers, buildings, etc.) will affect the movement speed and accessibility of the working dog; obstacles (such as trees, electric poles, etc.) will limit the passage path of the working dog. Convert these environmental factors into corresponding constraint conditions and incorporate them into the path planning algorithm to make the generated optimal path more in line with the actual situation.

[0153] Generate the optimal path: Using the selected path planning algorithm and constraints, traverse the search area or track the target on the computer according to a certain strategy, and finally simulate and generate the optimal path of the working dog. At the same time, record the coordinates and order of each key point on the optimal path as the benchmark for subsequent comparison with the actual path.

[0154] 2. Analyze the actual path

[0155] Fit the actual path and discretize the key points: Fit the position information collected by the positioning module to obtain a continuous curve, which is the actual path of the working dog. Then, discretize the actual path into a series of key points at a certain time interval or distance interval, and these key points should accurately reflect the actual movement trajectory of the working dog.

[0156] Extract path features: Analyze the actual path and the discretized key points to extract their path features. For example, calculate indicators such as the total length of the actual path, the coverage rate of the search area, and the distance change from the tracked target. These path features will be used to measure the deviation between the actual path and the optimal path.

[0157] 3. Calculate and analyze the path deviation

[0158] Select the deviation calculation method: According to the extracted path features, select a suitable deviation calculation method to measure the deviation between the actual path and the optimal path. Common deviation calculation methods include the Euclidean distance method, the Dynamic Time Warping (DTW) algorithm, etc. The Euclidean distance method is suitable for calculating the straight-line distance difference between two paths in space; the DTW algorithm takes into account the time series information of the path and can better handle the non-linear deformation and time asynchronization problems in the path.

[0159] Calculate the deviation: Use the selected deviation calculation method to calculate the deviation values between the actual path and the optimal path at each key point respectively. Then conduct statistical analysis on all deviation values, such as calculating statistical quantities such as the average value and standard deviation, to quantify the overall deviation degree. The smaller the deviation value, the closer the actual path is to the optimal path, indicating that the actual path of the working dog is more reasonable.

[0160] Visualize and analyze the deviation results: Visualize the calculated deviation values, such as drawing a deviation distribution histogram, a line chart of the deviation changing with time or the path, etc. Through visual analysis, the distribution and change trend of the deviation can be intuitively observed, and the rationality of the actual path planning of the working dog can be further understood. At the same time, combined with the type of target task and environmental factors, in-depth analysis can be carried out on the reasons for the deviation to provide a reference basis for the optimization of the target task.

[0161] In this embodiment, determining the range coverage of a working dog based on the motion data may specifically include:

[0162] By calculating the ratio of the actual coverage area of the working dog in the search area to the preset coverage area of the search task, the range coverage of the working dog is obtained, which measures the comprehensiveness of the working dog's search of the search area. Specifically, the motion trajectory of the working dog can be determined based on the position data of the working dog, so as to accurately calculate the actual coverage area of the working dog in the search area, and then calculate the ratio of the actual coverage area to the preset coverage area, that is, the range coverage. The range coverage can intuitively reflect whether the working dog has completed the search of the search area comprehensively and without omission. If the range coverage is close to or reaches 100%, it indicates that the working dog has a high search comprehensiveness; on the contrary, if the range coverage is low, it indicates that there are search omissions or the search is not thorough enough.

[0163] S3. Evaluate the working performance of the working dog according to multiple evaluation indicators.

[0164] In this embodiment, the above multiple evaluation indicators can be scored respectively to obtain the scores of the multiple evaluation indicators. Then, the scores of the multiple evaluation indicators are weighted and summed to obtain the comprehensive score of the working dog. The calculation formula of the comprehensive score is:

[0165] S = w1T + w2I + w3Q + w4F (7)

[0166] In formula (7), S is the comprehensive score; T, I, Q, and F are the scores of working duration, exercise intensity, task path rationality, and coverage range respectively; w1, w2, w3, and w4 are the weight coefficients corresponding to working duration, exercise intensity, path rationality, and coverage range respectively, and w1 + w2 + w3 + w4 = 1.

[0167] The higher the comprehensive score, the better the working performance of the working dog. Based on the comprehensive score, the working performance of the working dog can be evaluated comprehensively and objectively.

[0168] This embodiment can determine the working state, motion state, and health state of the working dog based on the motion data of the working dog, and then evaluate the working performance of the working dog based on the working state or motion data, which can achieve accurate identification of the working state of the working dog. Compared with manual observation and subjective judgment, it improves the accuracy and efficiency of the evaluation, and can comprehensively, objectively, and efficiently reflect the actual performance and contribution of the working dog.

[0169] This embodiment constructs a state recognition model for identifying the working state, motion state, and health state of a working dog based on a one-dimensional convolutional feature extraction method and a long short-term memory network, and combines an attention mechanism, realizing the fusion of multiple algorithms for the state recognition of a working dog, which can improve the robustness and recognition accuracy of the model.

[0170] As described above, these are only several embodiments of the present application and do not impose any form of limitation on the present application. Although the present application is disclosed above with preferred embodiments, it is not intended to limit the present application. Any person skilled in the relevant art can make some changes or modifications within the scope of the technical solution of the present application by using the disclosed technical content, which are equivalent to equivalent implementation cases and all fall within the scope of the technical solution.

Claims

1. An evaluation method for the working performance of working dogs, characterized in that, The method includes: S1. Determine the working state, motion state, and health state of the working dog at each of multiple moments based on the motion data of the working dog at each moment; S2. When the combination of the motion state and the health state meets a preset condition, determine multiple evaluation indicators of the working dog based on the motion data or the working state; S3. Evaluate the working performance of the working dog according to multiple evaluation indicators.

2. The method according to claim 1, wherein The specific content of S1 includes: S11. Construct a state recognition model based on a feature extraction method, a long short-term memory network, and an attention mechanism; S12. Based on the motion data of the working dog at each of multiple moments, use the state recognition model to determine the working state, motion state, and health state of the working dog at each moment.

3. The method according to claim 2, wherein The specific content of S11 includes: Construct a feature extraction unit based on a feature extraction method; the feature extraction unit is used to extract data features from the motion data of the working dog; Construct a state recognition unit based on a long short-term memory network and an attention mechanism; the state recognition unit is used to recognize the working state, motion state, and health state of the working dog according to the data features; Construct a state recognition model according to the feature extraction unit and the state recognition unit.

4. The method according to claim 2 or 3, characterized in that, The feature extraction method is a one-dimensional convolutional feature extraction method.

5. The method according to claim 3, characterized in that, Constructing a state recognition unit based on a long short-term memory network and an attention mechanism specifically includes: Construct a first loss function, a second loss function, and a third loss function for classifying the working state, motion state, and health state of the working dog respectively; According to the first loss function, the second loss function, and the third loss function, construct a state recognition unit based on a long short-term memory network and an attention mechanism.

6. The method according to claim 5, wherein According to the first loss function, the second loss function, and the third loss function, constructing a state recognition unit based on a long short-term memory network and an attention mechanism specifically includes: According to the correlation relationship among the working state, motion state, and health state of the working dog, construct a total loss function using the first loss function, the second loss function, and the third loss function; According to the total loss function, construct a state recognition unit based on a long short-term memory network and an attention mechanism.

7. The method according to claim 1, wherein The multiple evaluation indicators include working duration, motion intensity, path rationality, and coverage.

8. The method according to claim 7, characterized in that, The working state includes patrolling, searching, tracking, and resting; determining the working duration of the working dog according to the working state specifically includes: Determine the patrolling duration, searching duration, tracking duration, and resting duration of the working dog according to the working state at each moment and the time interval between adjacent moments; Determine the working duration of the working dog according to the patrolling duration, the searching duration, the tracking duration, and the resting duration.

9. The method according to claim 7, characterized in that The motion data includes acceleration; determining the motion intensity of the working dog according to the motion data specifically includes: Determine the acceleration change rate and the speed change rate of the working dog according to the acceleration; Determine the motion intensity of the working dog according to the acceleration change rate and the speed change rate.

10. The method according to claim 7, characterized in that The motion data includes position information; determining the path rationality of the working dog according to the motion data specifically includes: Using a path planning algorithm to determine the optimal path of the target task; Determining the actual path of the working dog according to the position information, and determining the path rationality of the working dog according to the actual path and the optimal path.

Images