Multi-classification SVM human body action recognition method based on D-S evidence theory

Abstract

The invention relates to a multi-classification SVM (Support Vector Machine) human body action recognition method based on a D-S evidence theory. The method comprises the following steps: acquiring electromyographic signal feature information and acceleration signal feature information of a target part; after feature extraction is conducted on the electromyographic signal feature information and the acceleration signal feature information, the electromyographic signal feature information and the acceleration signal feature information are input into a first support vector machine SVM and a second support vector machine SVM respectively, and probability output of first action recognition and probability output of second action recognition are obtained, the first support vector machine SVM is obtained through historical surface electromyogram signals and corresponding action training, and the second support vector machine is obtained through historical acceleration characteristic signals and corresponding action training; and performing D-S evidence theory fusion on the probability output of the first action recognition and the probability output of the second action recognition to obtain a final action classification result. According to the method, the SVM posterior probability output of the surface electromyogram signals and the acceleration signals is fused through the D-S evidence theory, and the human body action recognition rate is increased.

Description

Technical Field

[0001] The present invention relates to the technical field of human motion recognition of biomedical signals, and particularly to a multi-class SVM human motion recognition method based on D-S evidence theory. Background Art

[0002] Human motion recognition refers to using some technical means to classify people's motion patterns and types. Under the background of the digital age, with the rapid development of advanced technologies such as artificial intelligence, human motion recognition has become one of the research hotspots in the field of biomedical engineering. It has important research significance and extensive social application value in the fields of rehabilitation medicine, sports science, disease treatment, security monitoring, safe driving, game entertainment, etc. The research on motion recognition algorithms has promoted the application of artificial intelligence technologies such as machine learning and deep learning in biomedical engineering. By combining the characteristics of biomedical signals, more efficient and robust algorithm models can be developed. In the future, with the continuous progress of technology, motion recognition algorithms will play an important role in more scenarios and contribute to the improvement of human health and quality of life. By combining the characteristics of biomedical signals, more efficient and robust algorithm models can be developed.

[0003] Currently, the main research methods for human motion recognition include computer vision methods using video monitors or infrared pair-beam sensors, wireless signal methods based on Wi-Fi signals and millimeter-wave radars, and sensor methods based on accelerometers and gyroscopes.

[0004] Among them, the computer vision method based on images and videos has the disadvantages of strong environmental dependence for the motion recognition monitoring system. Problems such as lighting conditions, background complexity, and occlusion will all affect the recognition accuracy; it requires high computing resources, and processing data requires relatively high computing power, especially when using complex deep learning models, which is a challenge for applications with high real-time requirements; and this technology involves image acquisition and is prone to privacy leakage; a single camera can only cover a limited viewing range, and multi-angle motion recognition may require multiple cameras to work together, increasing the complexity and cost of the system.

[0005] For the motion recognition detection technology based on wireless signals, compared with the methods based on sensors or computer vision, the method based on wireless signals may be slightly inferior in accuracy, especially in the recognition of complex motions; other electronic devices or obstacles in the environment may interfere with the wireless signals, thus affecting the accuracy of motion recognition; complex algorithms are required to analyze the changes in wireless signals and convert them into understandable motion information, which increases the difficulty and cost of technology research and development; the applicable range of this technology is currently relatively limited and not as wide as traditional methods.

[0006] The sensor-based action recognition monitoring system has the advantages of high precision, good real-time performance, environmental independence, and privacy protection. However, the disadvantage of a single-sensor action recognition system is its low recognition rate for multiple actions.

[0007] Identifying human actions based on sensor data is essentially a pattern classification problem. Since sensor data often contains noise, it is first necessary to denoise it. Furthermore, the sensor data can be segmented into small time windows, and the data in each window reflects the current action situation. To perform pattern classification, relevant features need to be extracted from the data within the window and combined into feature vectors. If the dimension of the vectors is too high, dimensionality reduction processing is required. Finally, these feature vectors are input into a classifier. After training the classifier algorithm for these models, unlabeled actions are recognized.

[0008] Support Vector Machine (SVM) is a supervised classifier based on statistical theory. Compared with traditional machine learning methods, it can solve problems such as the lack of standard mathematical guidance for the neural network structure, long update and optimization time, and local minimum points. It has strong advantages in the recognition of small samples and non-linear recognition. However, the discriminant output of the standard SVM belongs to hard decision-making, and the output is generally {-1, +1}, and it mainly targets two-class problems. This hard decision-making limits the subsequent data processing. Aiming at the traditional method of using only the "one-versus-one" recognition strategy in human motion recognition based on Support Vector Machine (SVM) and outputting the recognition result, when ignoring the large number of action types, it leads to low recognition efficiency and low recognition accuracy. Summary of the Invention

[0009] The purpose of the present invention is to provide a multi-class SVM human action recognition method based on D-S evidence theory, which improves the human action recognition rate by fusing the SVM posterior probability outputs of surface electromyogram signals and acceleration signals through D-S evidence theory.

[0010] To achieve the above purpose, the present invention provides the following solution:

[0011] A multi-class SVM human action recognition method based on D-S evidence theory, comprising:

[0012] Obtaining the electromyogram signal feature information and acceleration signal feature information of the target part;

[0013] After extracting the feature information of the myoelectric signal and the acceleration signal feature information, they are respectively input into the first support vector machine (SVM) and the second support vector machine (SVM) to obtain the probability output of the first action recognition and the probability output of the second action recognition. Among them, the first support vector machine (SVM) is trained through historical surface myoelectric signals and corresponding actions, and the second support vector machine is trained through historical acceleration feature signals and corresponding actions;

[0014] Fuse the probability output of the first action recognition and the probability output of the second action recognition by using the D-S evidence theory to obtain the final action classification result.

[0015] Optionally, obtaining the myoelectric signal feature information and the acceleration signal feature information of the target part includes:

[0016] Collect surface myoelectric signals and acceleration information;

[0017] Extract the action feature information from the surface myoelectric signal through permutation entropy to obtain the myoelectric signal feature information;

[0018] Filter the acceleration signal through a low-pass filter, and obtain the signal amplitude vector and the signal amplitude domain peak value of the acceleration signal as the acceleration signal feature information.

[0019] Optionally, obtaining the probability output of the first action recognition and the probability output of the second action recognition includes: respectively mapping the output of the first support vector machine (SVM) and the second support vector machine (SVM) to the target interval through the sigmoid function for the myoelectric signal feature information, the signal amplitude vector, and the signal amplitude domain peak value to obtain the probability output of the first action recognition and the probability output of the second action recognition.

[0020] Optionally, the kernel functions of the first support vector machine (SVM) and the second support vector machine (SVM) are radial basis kernel functions.

[0021] Optionally, the method for obtaining the probability output of the first recognized action and the probability output of the second recognized action is:

[0022] m(A i )=p i (1 - E(P(error)), i = 1, 2, 3,..., M

[0023] m(Θ)=E(P(error))

[0024] Among them, m(A i ) is the basic probability assignment of proposition A i 's, p iP is the posterior probability of the sample in each category, M is the number of sample points at the time point, E(P(error)) is the uncertainty, and m(Θ) is the basic probability assignment of the identification framework.

[0025] Optionally, the D-S evidence theory fusion method for the probability outputs of the first action recognition and the second action recognition is as follows:

[0026] m(φ) = 0

[0027]

[0028] Among them, m(φ) is the set function where m is the empty set, m(A) represents the degree of belief in proposition A, n is the number of probability assignment functions, j is the jth probability assignment function, and m j is the probability assignment function on the same identification framework Θ, and k represents the conflict probability between evidences.

[0029] The beneficial effects of the present invention are as follows: The present invention uses the permutation entropy of the EMG signal to extract the characteristic information of each action, which is applicable to non-linear data, has good robustness, and has a good inhibitory effect on noise. The present invention uses the D-S evidence theory for decision fusion after the soft output of the EMG characteristic signal and the acceleration characteristic signal by the multi-class support vector machine, overcomes the shortcomings of insufficient information of a single sensor and low recognition rate, and effectively solves the problem of multi-classification of human actions. The present invention overcomes the problem of hard classification of small samples by combining the support vector machine SVM and the D-S evidence theory. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0031] Figure 1 It is the biomedical signal acquisition diagram of the embodiment of the present invention;

[0032] Figure 2 It is the flowchart of a multi-class SVM human action recognition method based on the D-S evidence theory according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0034] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] As Figure 2 shown, this embodiment provides a multi-class SVM human action recognition method based on the D-S evidence theory, including:

[0036] Obtain the electromyogram signal feature information of the target part (leg) and the acceleration signal feature information of the (trunk);

[0037] After feature extraction of the electromyogram signal feature information and the acceleration signal feature information, input them into the first support vector machine SVM and the second support vector machine SVM respectively to obtain the probability output of the first action recognition and the probability output of the second action recognition. Among them, the first support vector machine SVM is obtained by training with historical surface electromyogram signals and corresponding actions, and the second support vector machine is obtained by training with historical acceleration feature signals and corresponding actions;

[0038] Fuse the probability output of the first action recognition and the probability output of the second action recognition by the D-S evidence theory to obtain the final action classification result.

[0039] Further, obtaining the electromyogram signal feature information and the acceleration signal feature information of the target part includes:

[0040] Collect surface electromyogram signals and acceleration information;

[0041] Extract action feature information from the surface electromyogram signal through permutation entropy to obtain electromyogram signal feature information;

[0042] Filter the acceleration signal through a low-pass filter, and obtain the signal amplitude vector and the peak value in the signal amplitude domain of the acceleration signal as the acceleration signal feature information.

[0043] Specifically, extracting action feature information from the surface electromyogram signal through permutation entropy includes: Permutation entropy is a non-linear dynamic parameter based on complexity measure. It is similar to the Lyapunov exponent in reflecting the complexity of one-dimensional time series, is especially suitable for non-linear data, has good robustness, and has a good inhibitory effect on noise. The basic principle is as follows.

[0044] Given any one-dimensional time series {x i}, where i = 1, 2,..., T. In the time series arrangement of length T, every other sample point is taken to form a continuous n-sample point to compose an n-dimensional vector X i = [x(i), x(i + 1),... x(i + n - 1)]. Sort X i in ascending order to obtain the arrangement order [x(i + j1 - 1) ≤ x(i + j2 - 1) ≤,..., ≤ x(i + j n - 1)] (j1, j2,..., j n ) are the corresponding serial numbers of the sample points in the original sequence. In this way, the permutation and combination method of the vector X i is one of the permutation and combination methods in n!. Then, perform probability statistics on the permutation orders of each permutation and combination in the entire time series,

[0045]

[0046] where (x i , x i+1 ,... x i+n-1 ) are n consecutive sample points in the sequence, and # represents the number of times the permutation situation π appears in the sequence. After obtaining the probabilities of various permutation situations, calculate the permutation and combination entropy:

[0047]

[0048] Acceleration signal preprocessing and feature estimation include: Since the frequencies of acceleration signals generated by almost all human activities are below 20 Hz, the sampling frequency of the acceleration sensor is set to 50 Hz according to the characteristics of human movements under the condition of satisfying the Shannon theorem, and 1.5 s is set as the completion time of an action. Since the collected acceleration signal is the combined acceleration signal A B of the acceleration A G generated by human activities and the gravitational acceleration A T , in order to obtain the human acceleration A B , in this embodiment, a low-pass filter with a cut-off frequency of 0.25 Hz (passband 0.01 dB, stopband -100 dB) is used to filter the acceleration signal after eliminating abnormal noise peaks by third-order median filtering to obtain the gravitational acceleration A G , and then the approximate value of A B = A T - A G is obtained to get A B .

[0049] The Signal Magnitude Area (SMA) characterizes the intensity of body movement per unit time, and its formula is:

[0050]

[0051] Where T is the movement time of a human body movement, that is, the time of the observation window. According to the Mathie experiment and the actual situation of actions such as falling, in this embodiment, T = 1.5 s is selected, that is, 75 sampling points, which is the same as the action time of surface electromyogram signal acquisition.

[0052] At the same time, a Signal Magnitude Vector (SMV) is introduced. In essence, it provides a measure of the intensity of movement, and its definition is as follows:

[0053]

[0054] Where x i , y i , z i are the body acceleration A B sampling point values on the x-axis, y-axis, and z-axis at the i-th moment, respectively.

[0055] Furthermore, obtaining the probability outputs of the first action recognition and the second action recognition. The probability of recognizing an action includes: respectively passing the electromyogram signal feature information, the signal magnitude vector, and the signal magnitude domain peak through the sigmoid function and the radial basis kernel function to map the outputs of the first support vector machine (SVM) and the second support vector machine (SVM) to the target interval, and obtaining the probability outputs of the first action recognition and the second action recognition.

[0056] Furthermore, the kernel functions of the first support vector machine (SVM) and the second support vector machine (SVM) are radial basis kernel functions.

[0057] Obtaining the probability outputs of the first action recognition and the second action recognition, that is, the method for obtaining the BPA of each action signal is given by formulas (8)-(9).

[0058] Specifically, the commonly used method at present is to use the method proposed by Platt to map the output f(x) value of the SVM to [0, 1] through the sigmoid function. The output accuracy is similar to the original SVM accuracy, so as to realize the posterior probability output of the SVM. The posterior probability output is:

[0059]

[0060] When A < 0, the monotonicity of the above formula can be guaranteed.

[0061] A and B can be obtained by the cross-entropy error function:

[0062]

[0063] Among them, is the original training set of the possible maximum estimated values of parameters A and B (f i , y i ), and the transformed training set (f i , t i ), y i is the label category.

[0064] According to the support vector machine theory, the upper bound of the classification error rate of the test sample is the ratio of the average number of support vectors in the training sample to the total number of training samples, that is:

[0065]

[0066] It reflects the uncertainty of the SVM for the training sample.

[0067] For multi-class SVM, because the "one-versus-one" method is adopted in this embodiment, the BPA distribution output by each unit is:

[0068] m(A i ) = p i (1 - E(P(error)), i = 1, 2, 3,..., M (8)

[0069] m(Θ) = E(P(error)) (9)

[0070] Among them, m(A i ) is the basic probability assignment of proposition A i , p i is the posterior probability of the sample in each class, M is the number of sample points at the time point, E(P(error) is the uncertainty, and m(Θ) is the basic probability assignment of the recognition frame Θ.

[0071] Furthermore, the method of fusing the probability output of the first action recognition and the probability of the second action recognition by the D-S evidence theory is formulas (12)-(13).

[0072] Specifically, D-S evidence reasoning is an imprecise reasoning theory, and its core task is to construct a correct evidence model and determine a suitable decision-making.

[0073] In a recognition frame Θ of proposition A, the set function satisfies m: 2 Θ → [0, 1], and:

[0074] m(φ) = 0 (10)

[0075]

[0076] Among them, the degree of belief in proposition A is m(A), where m represents the basic probability assignment of the recognition frame Θ.

[0077] If A ∈ Θ and m(A) > 0, then A is called a focal element. Let m1, m2,..., m n be credibility assignments on Θ, and their orthogonal sum is defined as:

[0078] m(φ) = 0 (12)

[0079]

[0080] Among them, represents the conflict probability between evidences. m(φ) is the set function where m is the empty set, m(A) represents the degree of belief in proposition A, n is the number of probability assignment functions, j is the j-th probability assignment function, and m j is a probability assignment function on the same recognition frame Θ.

[0081] To test the effectiveness of the support vector machine model, Table 1 lists the SVM posterior probability outputs of the permutation entropy [x1, x2, x3, x4] test samples of 4-channel surface electromyography signals for 16 test samples with 2 test samples for each action. Table 2 lists the SVM posterior probability outputs of the signal amplitude vector and the peak value in the signal amplitude domain [y1, y2] test samples of the acceleration signals for 16 test samples with 2 test samples for each action. Among them, P1 to P8 represent the probabilities of forward fall, walk, sit, up stairs, down stairs, sit-to-stand, backward fall, and run actions respectively. The action corresponding to the largest probability is considered the corresponding action. The recognition situation and the actual situation are the classification labels for each action. 1 to 8 are the classification labels for forward fall, walk, sit, up stairs, down stairs, sit-to-stand, backward fall, and run actions respectively. The shaded marks indicate the misrecognition results.

[0082] Table 1

[0083]

[0084] Table 2

[0085]

[0086]

[0087] Then, according to Equation (7), the uncertainties of the sEMG signal and the acceleration signal in the framework due to ignorance are E1(P(error)) = 0.1046 and E2(P(error)) = 0.0692 respectively. Then, according to Equations (8) and (9), the posterior probability outputs of the sEMG signal and the acceleration signal are respectively converted into BPA. Then, Equations (12) and (13) are used to perform D-S evidence theory fusion on the BPA of the sEMG signal and the acceleration signal of each action. Since the test sample data volume is large, with a total of 160 samples, Table 3 only lists the D-S evidence fusion results of the sEMG signal in Table 1 and the acceleration signal in Table 2.

[0088] As can be seen from the shaded markings in Table 1, Table 2, and Table 3, among the 16 test samples of the 4-channel sEMG signal, 3 samples are misclassified, and among the 16 test samples of the acceleration signal, 5 samples are misclassified. All the sitting and sitting-to-standing actions are misclassified. After the D-S evidence fusion of the sEMG signal and the acceleration signal, only 1 sample is misclassified among the 16 test samples. Moreover, after fusing the posterior probabilities of the sEMG signal and the acceleration signal, the credibility distribution is more inclined to the corresponding correct actions, overcoming the shortcoming of insufficient information obtained by a single sensor and improving the recognition rate of human actions.

[0089] Table 3

[0090]

[0091]

[0092] The present invention will be further described below in conjunction with specific experiments:

[0093] A multi-class SVM human action recognition method based on D-S evidence theory includes:

[0094] Step 1: Feature extraction of surface electromyogram signal and acceleration signal:

[0095] (1) Extract the feature information of 8 actions from the 4-channel electromyogram signals of the legs through permutation entropy. It is applicable to non-linear data, has good robustness, and has a good inhibitory effect on noise.

[0096] (2) After collecting the acceleration signals of 8 actions of the human torso, filter the acceleration signals after third-order median filtering through a low-pass filter to eliminate abnormal noise peaks.

[0097] (3) Use the acceleration signal amplitude domain and the signal amplitude vector as the acceleration signal feature information to provide the degree of body movement intensity per unit time and a measure of the degree of movement intensity.

[0098] Specifically, an mt400 electromyogram signal acquisition instrument from Noraxon Company in the United States was used in the experiment. Disposable electromyogram electrode patches were attached to four muscles of the subjects, namely the tibialis anterior, gastrocnemius, rectus femoris, and semitendinosus muscles, to collect four-channel surface electromyogram signals for a total of 8 actions (including forward fall, walk, sit, walk up stairs, walk down stairs, sit-to-stand, backward fall, run), with a sampling frequency of 1000 Hz. The acceleration signal was collected using an i4 Motion acceleration acquisition instrument from TRCHNO CONCEPT Company in France. The acceleration sensor was placed in the middle of the waist, with a sampling frequency of 50 Hz. During the fall experiment, the fall scenarios were completed by healthy adult male experimenters falling forward and backward onto the floor on a mat. After the fall, they did not return to the upright state, and ordinary running data of the same subjects were collected on a flat cement road surface. The placement positions of the surface electromyogram signals (sEMG, Surface Electromyography) and acceleration signals collected in the experiment are as shown in Figure 1 shown.

[0099] Among them, there are 40 samples for each action, a total of 320 samples. The training samples and test samples are evenly divided, with 20 samples for each action, a total of 160 samples. For sEMG, the energy threshold method was used to determine the start time of each action of the sEMG signal, and the subsequent 1500 sampling point data was taken as the effective part of the action.

[0100] Step 2: Human action recognition based on support vector machine and D-S evidence reasoning:

[0101] (1) Use the sigmoid function to map the output of the support vector machine in [0, 1], ensuring similar output accuracy while realizing the posterior probability output of the support vector machine.

[0102] (2) Through the soft output of the multi-class support vector machine for the electromyogram feature signal and acceleration feature signal, D-S evidence theory is used for decision fusion, overcoming the shortcomings of insufficient single-sensor information and low recognition rate, and improving the recognition rate of human actions.

[0103] Specifically, since the SVM toolbox libsvm is used, the multi-classification method of SVM is the "one-versus-one" algorithm. According to equations (8) and (9), the permutation entropy of the 4-channel surface electromyogram signals and the signal amplitude vector and signal amplitude domain peak value of the acceleration signal of 160 training samples of 8 actions are respectively mapped to [0, 1] through the sigmoid function to obtain the probability output of recognition.

[0104] To further illustrate the effectiveness of the D-S evidence theory-based human action recognition method using sEMG signals and acceleration signals, this paper compares four posture recognition methods: the D-S evidence theory-based action recognition method, the multi-feature SVM-based action recognition method (i.e., using [x1, x2, x3, x4, y1, y2] as the feature vector and inputting it into SVM for recognition), the EMG signal-based SVM action recognition method, and the acceleration signal-based SVM action recognition method. The statistical results of the human action recognition experiments for the four methods are shown in Table 4.

[0105] Table 4

[0106]

[0107]

[0108] As can be seen from Table 4, based on the recognition of 8 actions, the signals of a single sensor have a low recognition rate for multi-action recognition, and it is basically impossible to recognize individual actions. The overall recognition rate is 66.25%. Compared with the multi-feature SVM-based action recognition method, the D-S evidence theory-based human action recognition method proposed in this embodiment has a better recognition rate than the multi-feature SVM-based action recognition method in forward fall, sitting, going up stairs, backward fall, and running. The recognition rates of sitting to standing and going down stairs are slightly lower than those of the multi-feature SVM-based action recognition method. The overall recognition rate is 88.75%, which is higher than the overall recognition rate of 86.88% of the multi-feature SVM-based action recognition method.

[0109] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A multi-class SVM human action recognition method based on D-S evidence theory, characterized in that, Including: Obtaining the electromyography signal feature information and acceleration signal feature information of the target part; After feature extraction of the electromyography signal feature information and acceleration signal feature information, input them into the first support vector machine (SVM) and the second support vector machine (SVM) respectively to obtain the probability output of the first action recognition and the probability output of the second action recognition. Among them, the first support vector machine (SVM) is trained through historical surface electromyography signals and corresponding actions, and the second support vector machine is trained through historical acceleration feature signals and corresponding actions; Fusing the probability output of the first action recognition and the probability output of the second action recognition by D-S evidence theory to obtain the final action classification result.

2. The multi-class SVM human action recognition method based on D-S evidence theory according to claim 1, characterized in that Obtaining the electromyography signal feature information and acceleration signal feature information of the target part includes: Collecting surface electromyography signals and acceleration information; Extracting action feature information from the surface electromyography signals through permutation entropy to obtain the electromyography signal feature information; Filtering the acceleration signal through a low-pass filter to obtain the signal amplitude vector and the peak value in the signal amplitude domain of the acceleration signal as the acceleration signal feature information.

3. The multi-class SVM human action recognition method based on D-S evidence theory according to claim 2, characterized in that, Obtaining the probability output of the first action recognition and the probability output of the second action recognition includes: mapping the output of the first support vector machine (SVM) and the second support vector machine (SVM) to the target interval through the sigmoid function for the electromyography signal feature information, the signal amplitude vector, and the peak value in the signal amplitude domain respectively to obtain the probability output of the first action recognition and the probability output of the second action recognition.

4. The multi-class SVM human action recognition method based on D-S evidence theory according to claim 1, characterized in that, The kernel functions of the first support vector machine (SVM) and the second support vector machine (SVM) are radial basis kernel functions.

5. The multi-class SVM human action recognition method based on D-S evidence theory according to claim 1, characterized in that, The method for obtaining the probability output of the first recognized action and the probability output of the second recognized action is: m(A i ) = p i (1 - E(P(error)), i = 1, 2, 3,..., M m(Θ) = E(P(error)) Among them, m(A i ) is the basic probability assignment of proposition A i , p i is the posterior probability of the sample in each category, M is the number of sample points at the time point, E(P(error)) is the uncertainty, and m(Θ) is the basic probability assignment of the recognition framework.

6. The multi-class SVM human action recognition method based on D-S evidence theory according to claim 5, wherein The method for fusing the probability output of the first action recognition and the probability output of the second action recognition by D-S evidence theory is: m(φ) = 0 Among them, m(φ) is the set function where m is the empty set, m(A) represents the degree of belief in proposition A, n is the number of probability assignment functions, j is the j-th probability assignment function, and m j is the probability assignment function on the same recognition framework Θ, and k represents the conflict probability between evidences.

Images