Motion state monitoring method of target object

Abstract

The invention discloses a method for monitoring the motion state of a target object, belongs to the technical field of target object monitoring, and can solve the problem that an existing method is difficult to meet monitoring requirements under complex conditions. The method comprises the following steps: S1, determining a motion type of a target object according to a body type, a task working condition and motion data of the target object; s2, determining a plurality of evaluation indexes of the target object according to the motion type and the motion data; and S3, determining the motion state of the target object according to the plurality of evaluation indexes, and performing early warning when the motion state accords with a preset abnormal state. The method is used for monitoring the motion state of the target object.

Description

Technical Field

[0001] The invention relates to a method for monitoring the motion state of a target object, and belongs to the technical field of target object monitoring. Background Art

[0002] After being domesticated by humans, working dogs, working mice, working horses and other animals can perform work that serves humans, and they play an important role in many areas of human society. In order to scientifically evaluate the health status, training effect and work efficiency of these animals, it is necessary to monitor their movement status in order to provide a reasonable basis for the evaluation process.

[0003] At present, the movement status of animals is mainly monitored by human movement status monitoring methods. However, due to the significant differences in body structure and movement patterns between humans and animals, the use of human monitoring methods for monitoring animals will lead to large monitoring errors and low monitoring accuracy. At the same time, different species of animals often have huge differences in body shape, movement posture, etc., and existing methods cannot perform differentiated monitoring for different species of animals, so it is difficult to meet the monitoring needs under complex conditions. Summary of the invention

[0004] The present invention provides a method for monitoring the motion state of a target object, which can solve the problem that existing methods are difficult to meet monitoring requirements under complex conditions.

[0005] The present invention provides a method for monitoring the motion state of a target object, the method comprising:

[0006] S1. determining the movement type of the target object according to the body shape, task working condition and movement data of the target object;

[0007] S2. determining a plurality of evaluation indicators of the target object according to the movement type and the movement data;

[0008] S3. Determine the motion state of the target object according to multiple evaluation indicators, and issue an early warning when the motion state meets a preset abnormal state.

[0009] Optionally, the S1 specifically includes:

[0010] Determine the data thresholds corresponding to different types of motion for target objects of different body sizes under different task conditions;

[0011] The movement type of the target object is determined using the data threshold according to the body shape, task conditions and movement data of the target object.

[0012] Optionally, the S1 specifically includes:

[0013] Constructing a type data set according to the motion data corresponding to different motion types of target objects of different body sizes under different task conditions, and constructing a type recognition model according to the type data set;

[0014] The type recognition model is used to determine the movement type of the target object according to the body shape, task conditions and movement data of the target object.

[0015] Optionally, construct a type dataset, including:

[0016] Use data augmentation methods to construct type datasets.

[0017] Optionally, the data augmentation method includes at least one of a time warping method, a noise injection method and a signal flipping method.

[0018] Optionally, the multiple evaluation indicators include number of steps, movement distance, movement trajectory and energy consumption.

[0019] Optionally, determining the number of steps of the target object according to the exercise type and the exercise data in S2 specifically includes:

[0020] Determining a normal cadence range of the target subject according to the exercise type;

[0021] According to the motion data, a hidden Markov model is used to determine the number of steps and the average step frequency of the target object, and the number of steps is adjusted when the average step frequency exceeds the normal step frequency range.

[0022] Optionally, determining the movement distance of the target object according to the movement type and the movement data in S2 specifically includes:

[0023] Determining the step length of the target object according to the body shape, task condition and exercise type of the target object;

[0024] The movement distance of the target object is determined according to the step length and the number of steps.

[0025] Optionally, the motion data includes acceleration and motion time; S2 determines the energy consumption of the target object according to the motion type and the motion data, specifically including:

[0026] determining a moving speed of the target object according to the acceleration;

[0027] The energy consumption of the target object is determined according to the movement speed, the movement time, the body shape and the movement type.

[0028] Optionally, the movement types include stationary, walking and running.

[0029] The beneficial effects that the present invention can produce include:

[0030] The present invention identifies the operation type of the target object according to its body shape, task conditions and motion data, and then determines its motion state according to the motion type, thereby realizing motion state monitoring of target objects of different body shapes under different task conditions and different motion types, improving the accuracy and scope of application of monitoring, being able to meet the monitoring needs under complex conditions, and providing a scientific basis for the health management and work efficiency evaluation of the target objects. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flow chart of a method for monitoring the motion state of a target object provided by an embodiment of the present invention;

[0032] Figure 2 A schematic diagram of a network structure for classifying motion types using a deep neural network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The present invention is described in detail below in conjunction with embodiments, but the present invention is not limited to these embodiments.

[0034] The embodiment of the present invention provides a method for monitoring the motion state of a target object. Figure 1 As shown, the method includes:

[0035] S1. Determine the movement type of the target object according to the body shape, task conditions and movement data of the target object.

[0036] Specifically, the target object may be a working dog, a working mouse, a working horse or other animals that can be trained. This embodiment is described using a working dog as an example, but is not limited to a working dog.

[0037] Specifically, working dogs include small, medium and large sizes, and can be divided according to key physiological parameters such as their weight, shoulder height and body length.

[0038] For example, working dogs weighing less than 10 kg, with a shoulder height of less than 40 cm and a body length of more than 50 cm can be classified as small dogs; working dogs weighing more than 30 kg, with a shoulder height of more than 60 cm and a body length of more than 80 cm can be classified as large dogs; and working dogs of other sizes can be classified as medium dogs.

[0039] Specifically, the task conditions may include a task type and a task environment.

[0040] More specifically, mission types may include search and rescue, patrol, and tracking; mission environments may include complex terrain, severe weather, and environments with dense buildings and crowds.

[0041] Specifically, the motion data may include acceleration, angular velocity, motion time, and the like.

[0042] In this embodiment, an acceleration sensor and a gyroscope sensor are worn on the body of the working dog, and then the acceleration sensor is used to collect the acceleration of the working dog along the x, y, and z directions at time t, which are respectively recorded as a and x (t), a y (t) and a z (t), and use the gyro sensor to collect the angular velocity ω(t) of the working dog at time t.

[0043] Specifically, this embodiment selects an accelerometer and a gyroscope sensor that have excellent performance in terms of accuracy, stability, sampling frequency, and waterproof performance. For example, the sampling frequency of the accelerometer and the gyroscope sensor can be greater than or equal to 100 Hz, and the waterproof level can be greater than or equal to IP67. The accelerometer and the gyroscope sensor can be installed on the collar of the working dog to ensure that they can stably collect data during the working dog's exercise.

[0044] Furthermore, in order to improve the accuracy and reliability of the data, the present embodiment may also pre-process the collected acceleration and angular velocity.

[0045] Specifically, this embodiment uses digital filtering algorithms such as low-pass filters and Kalman filters to pre-process the collected raw data.

[0046] The use of a low-pass filter can effectively remove high-frequency noise, making the data smoother and easier to process and analyze. For example, a low-pass filter with a cutoff frequency of 20 Hz can filter out high-frequency noise components caused by sensor jitter or environmental interference.

[0047] The prediction and update functions of Kalman filtering can effectively remove noise interference in the data and improve the accuracy and reliability of the data. x (t), a y (t) and a z (t) are respectively taken as the original acceleration a(t), and the acceleration after Kalman filtering is assumed to be The state equation of the Kalman filter is The measurement equation is Where v(t) is the measurement noise at time t, which is generally assumed to be zero-mean Gaussian noise with a specific covariance matrix R. This can provide more accurate accelerations in the x, y, and z directions:

[0048] After the preprocessing is completed, the present embodiment may also perform normalization processing on the preprocessed data, converting the data of different parameters into the same dimensional range to make them comparable, so as to facilitate subsequent analysis and processing.

[0049] This embodiment determines the motion type of the target object according to the normalized acceleration and angular velocity.

[0050] Specifically, the movement type may include stationary, walking, running, and other types.

[0051] In this embodiment, a threshold method or a machine learning method may be used to determine the motion type of the target object. The threshold method is preferably used when computing power is limited, while the machine learning method may be used when computing power is sufficient.

[0052] When the threshold method is adopted, S1 may specifically include:

[0053] Determine the data thresholds corresponding to different types of motion for target objects of different body sizes under different task conditions;

[0054] According to the target object's body shape, task conditions and motion data, the motion type of the target object is determined using data thresholds.

[0055] Specifically, the data threshold includes an acceleration threshold and an angular velocity threshold.

[0056] Specifically, the data threshold is set based on the individual characteristics of the working dog, the task type and the task environment. The impact of various aspects on the performance of working dogs is as follows:

[0057] 1. Individual characteristics

[0058] The breed, age, gender, body shape and other factors of different working dogs will lead to differences in their athletic performance. For example, the acceleration, gait frequency and other parameters of a German Shepherd and a Labrador Retriever when running may be different. By monitoring and analyzing the movement data of individual working dogs over a long period of time, data thresholds that meet their own characteristics can be set for each dog. For example, the acceleration threshold of a young working dog with better physical fitness during normal running can be set higher than that of an older working dog.

[0059] The health of working dogs is also an important factor affecting the setting of data thresholds. If a working dog suffers from a disease or is unwell, its athletic performance will change. For example, a working dog with arthritis may have an abnormal gait when walking, and the corresponding data threshold needs to be adjusted according to its health status. Since working dogs with obviously abnormal health conditions generally do not continue to participate in training and work, data thresholds do not need to be set for health conditions.

[0060] 2. Task Type

[0061] The type of task a working dog performs has a direct impact on its performance. For example, when a working dog is performing a search task, it may frequently change its direction and speed, and its data threshold needs to be set to adapt to this complex performance. On the other hand, the performance of a working dog performing a patrol task is relatively regular, and the data threshold can be set to focus more on stability and continuity.

[0062] 3. Task environment

[0063] Terrain and weather can also affect the performance of working dogs. In rugged mountain environments, working dogs may move slower and walk more cautiously, so the data threshold needs to be adjusted based on the terrain conditions. In severe weather conditions, such as strong winds and heavy rain, working dogs' performance can also be affected, so the data threshold needs to be adjusted based on the weather conditions.

[0064] According to the influence of the above-mentioned aspects on the working dog's sports performance, this embodiment sets acceleration thresholds and angular velocity thresholds corresponding to different sports types according to the sports conditions of working dogs of different sizes under different task conditions, so as to accurately distinguish the working dog's sports type.

[0065] For example, the specific setting scheme of the data threshold may be as follows:

[0066] 1. Considering the normal activities of dogs in general environment, the data threshold setting scheme for working dogs of different sizes is as follows:

[0067] For medium-sized dogs, a running threshold and a static threshold can be set, where the running threshold includes the acceleration threshold A corresponding to running. th1 The angular velocity threshold ω corresponding to running th1 , the static threshold includes the acceleration threshold A corresponding to static th2 The angular velocity threshold ω corresponding to rest th2 , A th1 >A th2 And ω th1 >ω th2 .when And the duration exceeds the preset time, and |ω(t)|>ω th1 When , the working dog's movement type is determined to be running; when And |ω(t)|<ω th2 When the working dog is detected, the movement type is judged as stationary, and the rest are classified as walking or other movement types.

[0068] The movement range of small dogs is relatively small, and the corresponding data threshold should be appropriately lower than that of medium-sized dogs; the movement range of large dogs is relatively large, and the corresponding data threshold should be appropriately increased compared with medium-sized dogs.

[0069] 2. The data threshold correction scheme for different task types is as follows:

[0070] 1) When performing search and rescue missions, working dogs need to frequently change their movement direction and speed, with large movement amplitudes and varied movement types.

[0071] For running threshold: To improve the flexibility and agility of identifying the type of movement, the acceleration threshold can be increased to adapt to the characteristics of rapid acceleration and deceleration during running. Set threshold A th ′=A th +α·dA th , where A th ′ is the corrected acceleration threshold; A th is the original acceleration threshold; dA th is the value of the acceleration threshold increase, which is a non-negative number; α is the coefficient of working dogs of different sizes, medium-sized dogs can take α = 2, large dogs can take α = 4, and small dogs can take α = 1. For example, the acceleration threshold of a medium-sized dog after correction is ± 2.5 m / s 2 , small dogs: ±2.0m / s 2 , large dogs: ±3.0m / s 2 , where the positive and negative signs represent different directions of acceleration.

[0072] At the same time, the angular velocity threshold can be increased to adapt to the characteristics of rapid turning during running, and the threshold ω′ can be set. th =ω th +β·dω th , where ω th ′ is the corrected angular velocity threshold; ω th is the original angular velocity threshold; dω th is the value of the angular velocity threshold increase, which is a non-negative number; β is the coefficient of working dogs of different sizes, β = 1 for medium-sized dogs, β = 2 for large dogs, and β = 0.5 for small dogs. For example, the angular velocity threshold of a medium-sized dog after correction is ±150° / s, that of a small dog is ±120° / s, and that of a large dog is ±180° / s, where the positive and negative signs represent different directions of the angular velocity.

[0073] In addition, the duration for judging running can be shortened, for example, the duration can be set to 1 second, so as to more quickly identify the movement type of the working dog.

[0074] For the static threshold: Since the working dog may be static for a short time to sniff the scent during the search process, the static threshold should be set more strictly. For example, the acceleration threshold can be set to less than 0.5m / s 2 , the angular velocity threshold is less than 10° / s.

[0075] 2) When performing patrol tasks, the working dogs' movement performance is relatively regular, mainly walking or jogging at a steady speed, with a small range of movement and less change in movement type.

[0076] For the running threshold: To avoid misjudgment, the acceleration threshold can be appropriately lowered and the threshold A can be set. th ′=A th -α·dA th , where A th ′ is the corrected acceleration threshold; A th is the original acceleration threshold; dA t x is the value of the acceleration threshold reduction, which is a non-negative number; α is the coefficient for working dogs of different sizes, where α=1 for medium-sized dogs, α=2 for large dogs, and α=0.5 for small dogs. For example, the acceleration threshold for medium-sized dogs after correction is ±1.5m / s 2 , small dogs: ±1.0m / s 2 , large dogs: ±2.0m / s 2 .

[0077] At the same time, the angular velocity threshold can be lowered and the threshold ω′ can be set th =ω th -β·dω th , where ω th ′ is the corrected angular velocity threshold; ω th is the original angular velocity threshold; dω th is the value by which the angular velocity threshold is reduced, which is a non-negative number; β is the coefficient for working dogs of different sizes, where β = 1 for medium-sized dogs, β = 2 for large dogs, and β = 0.5 for small dogs. For example, the angular velocity threshold for medium-sized dogs after correction is ±80° / s, for small dogs is ±60° / s, and for large dogs is ±100° / s.

[0078] In addition, the duration for judging running can be extended, for example, the duration can be set to 5 seconds to ensure the stability of the exercise type.

[0079] For the static threshold: To adapt to the fact that working dogs are less static during patrol, the acceleration threshold and angular velocity threshold can be appropriately lowered. For example, the acceleration threshold can be set to less than 0.3m / s 2 , the angular velocity threshold is less than 5° / s.

[0080] 2) When performing tracking tasks, working dogs need to maintain a certain speed and directional stability, but may quickly adjust the type of movement due to sudden changes in the tracking target.

[0081] For the running threshold: the acceleration threshold can be left unchanged, but the angular velocity threshold can be appropriately corrected. For example, the angular velocity threshold for a medium-sized dog is ±120° / s, for a small dog is ±100° / s, and for a large dog is ±150° / s. At the same time, according to the characteristics of the tracking task, the duration of judging running can be set to 4 seconds to balance fast response and stability.

[0082] For the stillness threshold: no correction is required and the stillness threshold under the general environment and normal activities of dogs mentioned above can be directly used.

[0083] 3. The threshold correction schemes for different task environment types are as follows:

[0084] 1) When working on complex terrain, working dogs need to walk more carefully, slow down their speed, and change their gait frequently.

[0085] For the running threshold: the acceleration threshold can be appropriately lowered to adapt to the characteristics of small speed changes, and the threshold A is set th ′=A th -α·dA th , where A th ′ is the corrected acceleration threshold; A th is the original acceleration threshold; dA th is the value by which the acceleration threshold is reduced, which is a non-negative number; α is the coefficient of working dogs of different sizes, where α=1 for medium-sized dogs, α=2 for large dogs, and α=0.5 for small dogs.

[0086] At the same time, the angular velocity threshold can be increased to adapt to the characteristics of frequent turning, and the threshold ω′ can be set th =ω th +β·dω th , where ω th ′ is the corrected angular velocity threshold; ω th is the original angular velocity threshold; dω th is the value by which the angular velocity threshold is increased, which is a non-negative number; β is the coefficient of working dogs of different sizes, where β = 1 for medium-sized dogs, β = 2 for large dogs, and β = 0.5 for small dogs.

[0087] For the immobility threshold: Since working dogs may frequently stop to observe the terrain, the immobility threshold should be set more leniently. For example, the acceleration threshold can be set to less than 0.4m / s 2 , the angular velocity threshold is less than 15° / s.

[0088] 2) In bad weather, working dogs’ athletic performance will be affected by wind and slippery ground, causing their speed to slow down and their range of motion to decrease.

[0089] For running threshold: you can lower the acceleration threshold and set threshold A th′=A th -α·dA th , where A th ′ is the corrected acceleration threshold; A th is the original acceleration threshold; dA th is the value of the acceleration threshold reduction, which is a non-negative number; α is the coefficient of working dogs of different sizes, α = 1 for medium-sized dogs, α = 2 for large dogs, and α = 0.5 for small dogs. For example, the acceleration threshold of a medium-sized dog after correction is ±1.2m / s 2 , small dogs: ±0.9m / s 2 , large dogs: ±1.5m / s 2 .

[0090] At the same time, the angular velocity threshold can be lowered and the threshold ω′ can be set th =ω th -β·dω th , where ω th ′ is the corrected angular velocity threshold; ω th is the original angular velocity threshold; dω th is the value by which the angular velocity threshold is reduced, which is a non-negative number; β is the coefficient for working dogs of different sizes, where β = 1 for medium-sized dogs, β = 2 for large dogs, and β = 0.5 for small dogs. For example, the angular velocity threshold for medium-sized dogs after correction is ±70° / s, for small dogs is ±50° / s, and for large dogs is ±90° / s.

[0091] For the static threshold: Since working dogs may stop more frequently to avoid wind and rain in bad weather, the static threshold should be set more leniently. For example, the acceleration threshold can be set to less than 0.4m / s 2 , the angular velocity threshold is less than 10° / s.

[0092] 2) In environments with dense buildings and crowds, the types of movement of working dogs change frequently due to increased interference from the environment.

[0093] For the running threshold: the acceleration threshold can be appropriately increased, threshold A th ′=A th +α·dA th , where A th ′ is the corrected acceleration threshold; A th is the original acceleration threshold; dA th is the value of the acceleration threshold increase, which is a non-negative number; α is the coefficient of working dogs of different sizes, medium-sized dogs can take α = 2, large dogs can take α = 4, and small dogs can take α = 1. For example, the acceleration threshold of a medium-sized dog after correction is ±1.8m / s 2 , small dogs: ±1.3m / s 2 , large dogs: ±2.2m / s2 .

[0094] At the same time, the angular velocity threshold can be appropriately increased to adapt to the characteristics of fast turning movements. The threshold ω′ th =ω th +β·dω th , where ω th ′ is the corrected angular velocity threshold; ω th is the original angular velocity threshold; dω th is the value of the angular velocity threshold increase, which is a non-negative number; β is the coefficient of working dogs of different sizes, β = 1 for medium-sized dogs, β = 2 for large dogs, and β = 0.5 for small dogs. For example, the angular velocity threshold of a medium-sized dog after correction is ±130° / s, that of a small dog is ±110° / s, and that of a large dog is ±160° / s.

[0095] For the static threshold: Since there may be a short stop, the static threshold should be set moderately. For example, the acceleration threshold can be set to less than 0.3m / s 2 , the angular velocity threshold is less than 8° / s.

[0096] This embodiment determines the running threshold and the stationary threshold based on the body shape and task conditions of the working dog, and then compares the working dog's motion data with the running threshold and the stationary threshold, respectively, to determine the motion type of the target object. When the motion data exceeds the corresponding threshold range, it is determined that the working dog's motion type has changed.

[0097] Using machine learning methods, S1 can specifically include:

[0098] A type data set is constructed based on the motion data corresponding to different motion types of target objects of different body sizes under different task conditions, and a type recognition model is constructed based on the type data set;

[0099] The type of motion of the target object is determined using a type recognition model based on the target object's body shape, task conditions, and motion data.

[0100] Furthermore, the above-mentioned construction of a type recognition model based on the type data set specifically includes:

[0101] Based on the type data set, a type recognition model is built using a machine learning algorithm.

[0102] The machine learning method converts the collected acceleration time series data and angular velocity time series into the frequency domain and extracts the time domain and frequency domain features, making full use of the statistical information and frequency information of the time series data to provide a richer feature representation for motion type classification. This method can not only effectively handle complex motion patterns, but also adapt to the needs of motion type recognition in different scenarios. The following are the specific implementation steps of this method:

[0103] 1) Data segmentation: Based on the aforementioned data preprocessing, data segmentation is performed to segment the continuous acceleration time series and angular velocity time series into time windows of fixed length, such as 2 seconds / time window, with an overlap of 50%.

[0104] 2) Fourier transform: Use Fast Fourier Transform (FFT) to convert time series signals into frequency domain. FFT can efficiently calculate the spectrum of the signal.

[0105] 3) Time domain feature extraction: Extract time domain features, including acceleration and angular velocity, as well as the mean and statistical variance of acceleration and angular velocity.

[0106] 4) Frequency domain feature extraction: extract frequency domain features, such as spectrum energy, main frequency components, etc.

[0107] 5) Feature fusion: Fusion of time domain features and frequency domain features to form a type data set as input for the machine learning model.

[0108] 6) Training and classification of machine learning models: Use type data sets to perform classification training on traditional machine learning algorithms such as SVM and random forest or deep learning algorithms such as CNN and LSTM to obtain a type recognition model, and then use the type recognition model to identify the movement type of working dogs.

[0109] In order to identify the movement types of working dogs of different sizes, it is necessary to collect movement data of working dogs of various sizes in different task types and different task environments.

[0110] On the one hand, all these motion data can be directly divided into three types: running, walking, and stillness for training, without distinguishing between body shape, task type, and task environment.

[0111] On the other hand, the body size, mission type and mission environment can be mapped as features and used as inputs to the type recognition model. For example, for each data, body size features (0: small dog, 1: medium dog, 2: large dog, etc.), mission type features (0: search and rescue mission, 1: patrol mission, 2: tracking mission, etc.), and mission environment features (0: complex terrain, 1: bad weather, 2: buildings and crowded environment, etc.) can be added. Here are just a few typical features, and other specific features can also be added.

[0112] For example, when using a shallow deep neural network to build a type recognition model, an embedding layer can be used as a feature mapping method to input the working dog's body shape, task type, task environment, etc. as features into the deep neural network for classification. The network structure is as follows: Figure 2 As shown in the figure. The embedding layer is a common method for mapping discrete inputs (such as numeric indices) into a continuous vector space. Although it is commonly used for word embedding in natural language processing, it can also be used to map numbers into feature vectors with the same length as the time domain signal and the frequency domain signal, and concatenate them together using the concat method.

[0113] Furthermore, the above-mentioned construction type data set may specifically include:

[0114] Use data augmentation methods to construct type datasets.

[0115] Specifically, the data augmentation method may include at least one of a time warping method, a noise injection method, and a signal flipping method.

[0116] Machine learning methods usually require a large amount of motion data for model training. Generally, motion data is collected by putting a collar device equipped with a sensor on the working dog, and the data is manually labeled according to the working dog's body shape, task type, task environment and movement type to generate a certain amount of training data for model training. However, this traditional training method has two main problems: first, the data collection and labeling process consumes a lot of human and dog resources, and is inefficient; second, the amount of data that can be collected is still limited, resulting in low recognition accuracy of the model when applied to the classification of a large amount of actual combat data.

[0117] To solve these two problems, this embodiment uses data augmentation methods such as time warping, noise injection, and signal flipping to expand the type data set, increase the diversity and quantity of data, and improve the generalization ability of the model.

[0118] 1. Time Warp Method

[0119] Time Warping is a method of generating new time series signals by nonlinearly stretching or compressing the time axis of a time series signal. FastDTW is an optimized version of the dynamic time warping algorithm (DTW algorithm), which can significantly reduce the computational complexity while maintaining the alignment effect of the DTW algorithm. FastDTW can be used to efficiently time warp time series signals and generate enhanced data with different rhythms, thereby increasing data diversity and improving the generalization ability of machine learning models. The specific implementation steps of the time warping method are:

[0120] First, input a multidimensional time series signal X, whose shape is (n samples , n features ), where n samples represents the time step, n featuresRepresents the feature dimension of each time step. For example, a signal containing 100 time steps and 2 features in each time step can be represented as a 100×2 matrix. Next, define a time warping function TimeWarp(X, γ), where γ is the warping factor used to control the degree of time warping. The value of the warping factor γ is usually between 0 and 1. For example, γ=0.1 means the maximum degree of warping is 10%. For each time step t (from 0 to n samples -1), randomly generates a distorted time index t′. The distorted time index is achieved by adding a random offset Δt to the original time index t, where Δt follows a uniform distribution Uniform(-γn samples ,γn samples Then, the warped time index t′ is clipped to ensure that its value is within the valid range (i.e., 0<t′<n samples ).

[0121] Based on the distorted time index t', the corresponding value is extracted from the original signal X to generate the distorted signal X'. Specifically, for each distorted time index t', the feature value of the corresponding time step is extracted from the original signal X to form a new signal X'. In order to increase the diversity of the data, the above process can be repeated multiple times to generate multiple distorted versions of the signal X 1 ′,X 2 ′,X 3 ′…X k ′, where k represents the number of enhanced samples. These distorted signals can be used as training data to improve the generalization ability of machine learning models.

[0122] If you need to align the distorted signal with the original signal, you can use the FastDTW algorithm. FastDTW uses dynamic programming to find the optimal time alignment path and calculate the similarity between the distorted signal and the original signal. This step can further optimize the training effect of the model in some application scenarios. Finally, the generated distorted signal X 1 ′,X 2 ′,X 3 ′…X k ′ is input into the machine learning model together with the original signal X as training data. In this way, the model can learn richer timing patterns, thereby improving the robustness to different rhythm changes.

[0123] 2. Noise injection method

[0124] Noise injection refers to adding random noise to the original time series signal to generate new enhanced samples. The noise usually follows a Gaussian distribution or other specific distribution, and its intensity can be controlled by the variance of the noise.

[0125] Input timing signal X=(x 1 , x 2 , ..., x n ), where n is the time step. Generate noise Y with the same shape as the original signal according to a specified noise distribution (such as Gaussian distribution). The strength of noise Y is controlled by the variance, which is usually set to a certain proportion of the signal standard deviation. The generated noise Y is added to the original signal X to generate an enhanced signal X′=X+Y.

[0126] 3. Signal inversion method

[0127] Signal flipping is a method used to enhance the diversity of multidimensional time series signal data. Assume that the input multidimensional time series signal is X, and its shape is (n samples , n features ), where n samples represents the time step, n features Represents the feature dimension of each time step. The core idea of ​​signal flipping is to completely reverse the signal X on the time axis to generate an enhanced signal X′. The specific implementation steps of the signal flipping method are:

[0128] For each feature dimension, its value is completely flipped in the order of time steps. The mathematical expression of the flip operation is X′i, j = Xn samples -i-1,j. Among them, i=0, 1,...,n samples -1; j = 0, 1, ..., n feactures -1. In this way, the signal is flipped on the time axis column by column.

[0129] S2. Determine multiple evaluation indicators of the target object according to the movement type and movement data.

[0130] Specifically, the multiple evaluation indicators may include the number of steps, movement distance, movement trajectory and energy consumption.

[0131] In S2, the number of steps of the target object is determined according to the exercise type and the exercise data, which may specifically include:

[0132] Determine the normal cadence range for the target subject based on the type of exercise;

[0133] Based on the motion data, the hidden Markov model is used to determine the number of steps and average cadence of the target subject, and the number of steps is adjusted when the average cadence exceeds the normal cadence range.

[0134] Specifically, in this embodiment, the normal cadence range can be pre-set according to the working dog's body shape, task type, task environment, and exercise type.

[0135] Specifically, the method of determining the number of steps using the hidden Markov model is as follows:

[0136] 1. Data preprocessing: Wavelet transform is used to preprocess the raw acceleration data to remove noise and extract multi-scale features.

[0137] 2. Step statistics:

[0138] 1) Use the hidden Markov model to preprocess the vertical acceleration Conduct analysis.

[0139] 2) The gait process is modeled as a series of hidden state sequences, each of which represents a gait event, such as "foot touching the ground", "foot leaving the ground", etc.

[0140] 3) Describe how acceleration data is mapped to gait stages by defining the transition probabilities between different gait events and the conditional probability distribution between hidden states and observed values.

[0141] 4) Apply the Viterbi algorithm to decode the most likely hidden state sequence and record the time interval T between two adjacent gait events. step .

[0142] 5) Set a suitable time window (30 seconds or 1 minute), and use the hidden Markov model to identify all "foot touchdown" events in each time window. Assuming that n "foot touchdown" events are detected in a time window, and each gait cycle "foot touchdown" occurs twice, the number of steps in the time window is By adding up the number of steps in all time windows, we can get the total number of steps of the target object.

[0143] In order to improve the accuracy of the model step count statistics, this embodiment also uses the average step frequency to correct the step count. The specific steps of the correction are as follows:

[0144] 1. Use the formula Calculate the average step frequency, where f avg is the average step frequency, T w is the length of the time window, N steps is the time window T w The number of steps within.

[0145] 2. If the average step frequency f is calculated avg If the calculated average step frequency f does not exceed the normal step frequency range of the target object under the corresponding task conditions, the calculated total number of steps will be used as the final step calculation result; if the calculated average step frequency f avg If it exceeds the normal cadence range of the target subject under the corresponding task conditions, the peak detection method is used to determine the average cadence again.

[0146] The method for determining the average cadence using the peak detection method is as follows:

[0147] 1) Acceleration data after filtering in the vertical direction Conduct in-depth analysis, when When it changes from less than zero to greater than zero (i.e., crossing the zero axis upward), it is recorded as a peak point, which represents a gait event, such as "foot touching the ground", "foot leaving the ground", etc. At the same time, the time interval between two adjacent peak points is recorded, which is the time interval between two adjacent gait events T step .

[0148] 2) Set a suitable time window (30 seconds or 1 minute), and use the peak detection method to identify all "foot touchdown" events in each time window. Assuming that n "foot touchdown" events are detected in a time window, and each gait cycle "foot touchdown" occurs twice, the number of steps in the time window is By adding up the number of steps in all time windows, we can get the total number of steps of the target object.

[0149] 3) Calculate the average cadence based on the total number of steps using the above average cadence formula.

[0150] 3. Determine whether the average step frequency calculated by the peak detection method also exceeds the normal step frequency range. If so, it means that the step count calculation result of the Markov model is correct or the sensor and other equipment have failed, and further confirmation is required; if not, the total number of steps calculated by the peak detection method is used as the final step count calculation result.

[0151] In S2, the movement distance of the target object is determined according to the movement type and the movement data, which may specifically include:

[0152] Determine the target object's step length based on the target object's body size, task conditions, and exercise type;

[0153] Determine the movement distance of the target object based on the step length and number of steps.

[0154] Specifically, this embodiment establishes a relationship model between the stride length and body parameters (weight, shoulder height) of working dogs of different body sizes through extensive experiments and in-depth data analysis.

[0155] Specifically, the stride length of a medium-sized dog is where k 1 , k 2 , k 3 is the coefficient obtained by fitting a large amount of experimental data, m and h are the weight and shoulder height of a medium-sized dog respectively; the stride length of a small dog is where k 6 , k 7 , k 8 is the coefficient obtained by fitting a large amount of experimental data, ms and h s are the weight and shoulder height of small dogs; the stride length of large dogs is where k 12 , k 13 , k 14 is the coefficient obtained by fitting a large amount of experimental data, m l and h l They are the weight and shoulder height of large dogs respectively.

[0156] Then, this embodiment takes into account the change in the step length of the working dog when it moves in different terrain environments and corrects the step length.

[0157] Specifically, when a medium-sized dog moves on complex and rugged terrain such as mountains, let the terrain slope be θ, and the corrected step length when going uphill is L′=L×(1+k 4 × sinθ), the corrected step length when going downhill is L′=L×(1-k 4 × sinθ), where k 4 is the coefficient of influence of slope on medium-sized dogs; when small dogs exercise indoors, considering factors such as ground friction, the friction coefficient is set to μ, and the corrected step length is L′ s =L s ×(1-k 9 ×μ), where k 9 is the friction coefficient; large dogs are on uphill and downhill sections, assuming the terrain slope is θ l , the corrected step length after uphill slope is L′ l =L l ×(1+k 15 ×sinθ l ), the step length after downhill correction is L′ i =L l ×(1-k 15 ×sinθ l ), where k 15 is the coefficient of influence of slope on large dogs.

[0158] Specifically, this embodiment accurately calculates the movement distance of the working dog based on the number of steps calculated in real time and the corrected step length. Assume that the number of steps of a medium-sized dog, a small-sized dog, and a large-sized dog are N, N respectively. s and N l , then the movement distance of a medium-sized dog is D = N × L, and the movement distance of a small dog is D s =N s ×L s , the movement distance of large dogs is D l =N l ×L l .

[0159] In S2, the energy consumption of the target object is determined according to the exercise type and the exercise data, which may specifically include:

[0160] Determine the moving speed of the target object based on the acceleration;

[0161] Determine the target subject's energy expenditure based on exercise speed, exercise duration, body size, and exercise type.

[0162] Specifically, in this embodiment, the movement speed of the working dog can be calculated by integrating the acceleration or the like.

[0163] Specifically, this embodiment establishes an energy consumption estimation model based on key data (such as acceleration, speed, exercise time, etc.) during the movement of working dogs through extensive experiments and in-depth data analysis.

[0164] Specifically, the energy expenditure of a medium-sized dog is E m =k m ×v 2 ×t+BMR m , where k m is the energy consumption coefficient of a medium-sized dog, which is related to factors such as the dog's body size and muscle content; v is the movement speed of a medium-sized dog, and t is the exercise time of a medium-sized dog. In actual use, it is necessary to fully consider the physiological characteristics of medium-sized dogs (such as basal metabolic rate, muscle content, etc.) and their effects on k m Make reasonable adjustments, BMR m is the basal metabolic rate of a medium-sized dog.

[0165] The energy expenditure of a small dog is E s =k s ×k sr ×v 2 ×t+BMR s , where k s is the energy consumption coefficient of small dogs. Considering that the basal metabolic rate of small dogs is relatively high, k is set s >1;k sr is the exercise-related energy coefficient of small dogs, v is the exercise speed of small dogs, t is the exercise time of small dogs, BMR s is the basal metabolic rate of small dogs.

[0166] The energy expenditure of a large dog is E l =k l ×k lr ×v 2 ×t+BMR l , where k l is the energy consumption coefficient of large dogs. Considering the continuous characteristics of large dogs in guiding the blind and other tasks, 0<k s <1; k lris the exercise-related energy coefficient of large dogs, v is the exercise speed of large dogs, t is the exercise time of large dogs, BMR l is the basal metabolic rate of large dogs.

[0167] In addition, the present embodiment can also obtain the position data of the working dog by integrating the acceleration of the working dog, and then determine the motion trajectory of the working dog according to the position data.

[0168] S3. Determine the motion state of the target object based on multiple evaluation indicators, and issue an early warning when the motion state meets the preset abnormal state.

[0169] Specifically, the movement state includes a normal state and an abnormal state. This embodiment can flexibly set the value range of the evaluation index corresponding to the normal state and the abnormal state of the working dog according to the specific situation. For example, if the movement distance of a search and rescue dog is too short and the energy consumption is too high, it may indicate that the search and rescue dog is fatigued and should be set to an abnormal state; if the movement distance of a guide dog is abnormal, it may indicate that there is a problem with the guide route or that the guide dog is unwell, and should be set to an abnormal state.

[0170] This embodiment determines the evaluation index of the working dog by implementing data transmission and analysis, and visualizes the calculation results of each evaluation index so that the staff can understand the working dog's movement status at a glance. When an abnormal state occurs, this embodiment issues an early warning message so that the staff can further check and handle the abnormal state in time.

[0171] The present invention identifies the operation type of the target object according to its body shape, task conditions and motion data, and then determines its motion state according to the motion type, thereby realizing motion state monitoring of target objects of different body shapes under different task conditions and different motion types, improving the accuracy and scope of application of monitoring, being able to meet the monitoring needs under complex conditions, and providing a scientific basis for the health management and work efficiency evaluation of the target objects.

[0172] The above are only a few embodiments of the present application and do not constitute any form of limitation to the present application. Although the present application is disclosed as above with preferred embodiments, it is not intended to limit the present application. Any technician familiar with the profession, without departing from the scope of the technical solution of the present application, using the technical content disclosed above to make slight changes or modifications are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A method for monitoring the motion state of a target object, characterized in that: The method comprises: S1. Determine the movement type of the target object according to the body shape, task working condition and movement data of the target object; S2. determining a plurality of evaluation indicators of the target object according to the movement type and the movement data; S3. Determine the motion state of the target object according to multiple evaluation indicators, and issue an early warning when the motion state meets a preset abnormal state.

2. The method according to claim 1, characterized in that The S1 specifically includes: Determine the data thresholds corresponding to different types of motion for target objects of different body sizes under different task conditions; The movement type of the target object is determined using the data threshold according to the body shape, task conditions and movement data of the target object.

3. The method according to claim 1, characterized in that The S1 specifically includes: Constructing a type data set according to the motion data corresponding to different motion types of target objects of different body sizes under different task conditions, and constructing a type recognition model according to the type data set; The type recognition model is used to determine the movement type of the target object according to the body shape, task conditions and movement data of the target object.

4. The method according to claim 3, characterized in that Build type data sets, including: Use data augmentation methods to construct type datasets.

5. The method according to claim 4, characterized in that The data augmentation method includes at least one of a time warping method, a noise injection method and a signal flipping method.

6. The method according to claim 1, characterized in that Multiple evaluation indicators include number of steps, exercise distance, exercise trajectory and energy consumption.

7. The method according to claim 6, characterized in that S2 determines the number of steps of the target object according to the exercise type and the exercise data, specifically including: Determining a normal cadence range of the target subject according to the exercise type; According to the motion data, a hidden Markov model is used to determine the number of steps and the average step frequency of the target object, and the number of steps is adjusted when the average step frequency exceeds the normal step frequency range.

8. The method according to claim 7, characterized in that S2 determines the movement distance of the target object according to the movement type and the movement data, specifically including: Determining the step length of the target object according to the body shape, task condition and exercise type of the target object; The movement distance of the target object is determined according to the step length and the number of steps.

9. The method according to claim 6, characterized in that The motion data includes acceleration and motion time; S2 determines the energy consumption of the target object according to the motion type and the motion data, specifically including: determining a moving speed of the target object according to the acceleration; The energy consumption of the target object is determined according to the movement speed, the movement time, the body shape and the movement type.

10. The method according to any one of claims 1 to 9, characterized in that The types of movement include stillness, walking, and running.

Images