CN108717548B - Behavior recognition model updating method and system for dynamic increase of sensors - Google Patents
Behavior recognition model updating method and system for dynamic increase of sensors Download PDFInfo
- Publication number
- CN108717548B CN108717548B CN201810315805.3A CN201810315805A CN108717548B CN 108717548 B CN108717548 B CN 108717548B CN 201810315805 A CN201810315805 A CN 201810315805A CN 108717548 B CN108717548 B CN 108717548B
- Authority
- CN
- China
- Prior art keywords
- recognition model
- behavior recognition
- incremental
- data
- updating
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/24323—Tree-organised classifiers
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
- G06V40/23—Recognition of whole body movements, e.g. for sport training
Abstract
The invention relates to a method and a system for updating a behavior recognition model facing dynamic addition of a sensor, wherein the method comprises the following steps: a model construction step, in which initial data of user behaviors are obtained through an initial sensor, and initial characteristic data are extracted to construct a behavior recognition model; an incremental characteristic data acquisition step, namely acquiring incremental data of user behaviors through the initial sensor and the newly added sensor, defining incremental characteristics and extracting incremental characteristic data; a step of model updating decision-making, which is to obtain diversity scores of each decision tree of the behavior recognition model according to average mutual information between prediction results of the decision trees in the behavior recognition model and mutual information between the prediction results and actual behaviors of users corresponding to the prediction results, and take the decision trees with the diversity scores smaller than an updating threshold value as the decision trees to be updated; and a step of dynamically updating the model, namely updating all decision trees to be updated by the incremental characteristic data so as to update the behavior recognition model.
Description
Technical Field
The invention relates to the fields of pervasive computing, incremental learning technology and behavior identification, in particular to a method and a system for updating a behavior identification model facing dynamic increase of a sensor.
Background
In recent years, a great deal of research has shown that the ability to perform daily activities is an important indicator of human physical health. Real-time and accurate behavior recognition can effectively monitor daily movement of people, and provides early warning for falling down for the old. The birth and development of the miniature wearable sensing device provides a new way for daily behavior monitoring. The characteristics of small volume, low power consumption and the like enable long-time and real-time behavior identification to be possible. With the increasing popularity of wearable devices, the number of wearable devices for users may increase resulting in an increasing number of sensors available for behavior recognition. How to improve the recognition performance of the original model by using a small amount of data with new sensors is an important challenge in wearable behavior recognition.
A traditional behavior recognition method based on a sensor establishes a classification model for calibration data collected off line. These models cannot accommodate variations in actual applications. With the advent of new wearable devices, more sensors can be used to improve performance of behavior recognition. It is also mentioned in the literature [ Farseev A, Chua T S.TweetFit: functional Multiple social media and Sensor Data for well Profile Learning [ C ]// AAAI.2017:95-101.MLA ] that various sensors can achieve better behavior recognition performance. However, it is difficult to integrate a new sensor into a pre-trained behavior recognition model. The appearance of new sensors will result in a corresponding increase in the feature dimensionality, i.e., feature class, of the input data, resulting in the failure of the pre-trained behavior recognition model. In order to fully utilize the data collected by the new sensor, the traditional method can only abandon the existing model and acquire the data again to train a new model. This approach will result in the information stored in the original model and the old data being discarded. Data annotation is a well-known and time-consuming task. Retraining the behavior recognition model would greatly waste time and labor. Therefore, it is a great challenge to adapt existing behavior recognition models to the advent of new sensors, while using minimal time and space costs.
In response to the above problems, many scholars and researchers have conducted related studies. Patent cn201610182598.x uses forward sequence selection algorithm to obtain the best feature vector in feature selection, and uses Relief-F to perform feature enhancement processing. And training in a model construction stage to obtain a base classifier and 3 weak classifiers, and integrating 4 classifiers to make decisions on human behaviors. Patent CN201710292408.4 uses LSTM model to model data obtained by sensors, trains neural network parameters to obtain recognition model, and can deal with modeling of relatively less data, but cannot deal with the situation of dynamically increasing number of sensors. Patent WO2015123373-a1 discloses a behavior recognition device. The device collects data using inertial sensors in a resource-constrained environment, improving the accuracy of behavior recognition by providing additional analysis based on a variety of factors.
Although various machine learning methods have been successfully applied in the field of behavior recognition, they have some drawbacks in plasticity and dynamic adaptability:
1. in the face of a newly added sensor, the traditional behavior identification method is not suitable for a model due to inconsistent input dimensions, and the data acquired by newly added equipment cannot be fully utilized, so that the resource is greatly wasted.
2. In order to fully utilize the data of the newly added sensor, the traditional behavior identification method needs to abandon the existing model and collect the data again to construct a new model. Although this method can fully utilize all wearable devices to detect user behavior, it will cause waste of existing model knowledge and result in a lot of calibration work.
Therefore, it is urgently needed to design a robust behavior identification method facing dynamic increase of the number of sensors, and when the number of available sensing devices is dynamically increased, the accuracy of an existing model can be improved by using data from a new sensor at a small space-time cost.
Disclosure of Invention
In order to solve the problems, the invention relates to a behavior recognition model updating method facing dynamic addition of sensors, which comprises the following steps: a model construction step, in which initial data of user behaviors are obtained through an initial sensor, and initial characteristic data are extracted to construct a behavior recognition model; an incremental characteristic data acquisition step, namely acquiring incremental data of user behaviors through the initial sensor and the newly added sensor, defining incremental characteristics and extracting incremental characteristic data; a step of model updating decision-making, which is to obtain diversity scores of each decision tree of the behavior recognition model according to average mutual information between prediction results of the decision trees in the behavior recognition model and mutual information between the prediction results and actual behaviors of users corresponding to the prediction results, and take the decision trees with the diversity scores smaller than an updating threshold value as the decision trees to be updated; and a step of dynamically updating the model, namely updating all decision trees to be updated by the incremental characteristic data so as to update the behavior recognition model.
The invention relates to a behavior recognition model updating method, wherein the model construction step specifically comprises the following steps: processing the initial data by a sliding window method to obtain the initial characteristic data; and constructing a random forest classifier by using the initial characteristic data to obtain the behavior recognition model.
The behavior recognition model updating method provided by the invention comprises the following steps of: acquiring first incremental data of user behavior through the initial sensor; acquiring second incremental data of user behaviors through the newly added sensor; processing the first incremental data by a sliding window method to obtain first characteristic data corresponding to the initial sensor; processing the second incremental data by a sliding window method to extract the incremental features and obtain second feature data corresponding to the newly added sensor; the first characteristic data and the second characteristic data are combined to obtain the incremental characteristic data.
The behavior recognition model updating method provided by the invention obtains diversity scores by the following method:
The invention relates to a behavior recognition model updating method, wherein the step of dynamically updating the model specifically comprises the following steps: a sub-tree modification step, wherein a new sub-tree is constructed under a node by taking a certain increment characteristic of the node falling into a certain decision tree to be updated as a split attribute; and a leaf node splitting step, namely splitting the leaf node when the type of the user behavior in the leaf node of a certain decision tree to be updated is more than 1 and the number of samples falling into the leaf node is more than a splitting threshold.
The invention also relates to a behavior recognition model updating system facing dynamic addition of the sensor, which comprises the following steps: the model building module is used for obtaining initial data of user behaviors through an initial sensor and extracting initial characteristic data to build a behavior recognition model; the incremental characteristic data acquisition module is used for acquiring incremental data of user behaviors through the initial sensor and the newly added sensor, defining incremental characteristics and extracting incremental characteristic data; the model updating decision module is used for determining a decision tree to be updated in the updating process of the behavior recognition model; and the model dynamic updating module is used for updating all the decision trees to be updated by the incremental characteristic data so as to update the behavior recognition model.
The behavior recognition model updating system of the invention, wherein the initial model building module comprises: the initial data acquisition module is used for processing the initial data by a sliding window method to obtain the initial characteristic data; and the model construction training module is used for constructing a random forest classifier by using the initial characteristic data to obtain the behavior recognition model.
The behavior recognition model updating system of the invention, wherein the incremental characteristic data acquisition module comprises: the incremental data acquisition module is used for acquiring first incremental data of user behaviors through the initial sensor and acquiring second incremental data of the user behaviors through the newly added sensor; an incremental data processing module, configured to obtain the incremental feature data, where the first incremental data is processed by a sliding window method to obtain first feature data corresponding to the initial sensor, and the second incremental data is processed by a sliding window method to extract the incremental feature and obtain second feature data corresponding to the newly added sensor; the first characteristic data and the second characteristic data are combined to obtain the incremental characteristic data.
The behavior recognition model updating system provided by the invention is characterized in that the model updating decision module obtains diversity scores through the following formula:
The behavior recognition model updating system of the invention, wherein the model dynamic updating module comprises: the sub-tree modification module is used for constructing a new sub-tree under the decision tree to be updated; taking a certain increment characteristic of a node falling into a certain decision tree to be updated as a split attribute, and constructing a new sub-tree under the node; the leaf node splitting module is used for splitting leaf nodes of the decision tree to be updated; when the type of the user behavior in a leaf node of a certain decision tree to be updated is larger than 1 and the number of samples falling into the leaf node is larger than a splitting threshold value, splitting the leaf node.
Drawings
Fig. 1 is a schematic workflow diagram of a behavior recognition model updating method according to an embodiment of the present invention.
Fig. 2 is a flowchart of a behavior recognition model updating method according to an embodiment of the present invention.
FIG. 3 is a flow chart of building a behavior recognition model according to an embodiment of the present invention.
FIG. 4 is a flow chart of acquiring incremental feature data according to an embodiment of the present invention.
FIG. 5A is a schematic diagram of a single decision tree in the behavior recognition model according to an embodiment of the present invention.
Fig. 5B and 5C are schematic diagrams of a decision tree growing mechanism of a behavior recognition model according to an embodiment of the present invention.
FIG. 6 is a comparative illustration of test accuracy tests of an embodiment of the present invention.
FIG. 7 is a comparative graph of training time trials for an embodiment of the present invention.
FIG. 8 is a comparative graph of test time trials for an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more clearly understood, the following describes in detail a behavior recognition model updating method and system for sensor dynamic addition according to the present invention with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Fig. 1 is a schematic workflow diagram of a behavior recognition model updating method for dynamic addition of sensors according to an embodiment of the present invention. Fig. 2 is a flowchart of a behavior recognition model updating method according to an embodiment of the present invention. Referring to fig. 1, firstly, behavior data of a user or a user group is collected by an existing sensor (initial sensor) as initial data to construct a behavior recognition model for recognizing user behaviors; in the subsequent use process of the behavior recognition model, behavior data of a user or a user group is obtained through an existing initial sensor and a newly added sensor (a newly added sensor) to serve as incremental data, incremental features are defined, and incremental feature data are extracted; judging which decision trees in the behavior recognition model need to be updated according to a diversity generation strategy based on mutual information so as to determine the decision trees to be updated; and updating the decision tree to be updated by using a characteristic increment tree growth mechanism so as to update the behavior recognition model. The specific steps are shown in fig. 2:
step S1, constructing a behavior recognition model of the user;
step S2, obtaining incremental characteristic data;
and step S3, selecting a decision tree to be updated, and updating the behavior recognition model by using the incremental characteristic data. When the number of user sensors increases, a jump back to step S2 continues with the updating of the behavior recognition model.
Firstly, constructing a behavior recognition model
In the process of constructing the behavior recognition model, the behavior recognition model is constructed by adopting data collected by an acceleration sensor, a gyroscope and the like which are arranged in the intelligent hardware equipment.
FIG. 3 is a flow chart of building a behavior recognition model according to an embodiment of the present invention. As shown in fig. 3, the step of constructing the behavior recognition model specifically includes:
step S11, reading behavior data of an initial sensor (an accelerometer, a gyroscope and the like) containing a user behavior category as initial data;
step S12, sliding the intercepted initial data with fixed time length (such as 5 seconds) by using a sliding window method, and processing each window data through a preprocessing algorithm, wherein the preprocessing algorithm includes but is not limited to data filtering and the like;
step S13, intercepting the data obtained by preprocessing by using a sliding window, where adjacent windows overlap by 50%, and the features extracted by each window include but are not limited to: mean, standard deviation, minimum, maximum, mode, interval, mean point number, direct current component, peak value after fast Fourier transform, mean, standard deviation energy entropy and the like;
step S14, combining the extracted features and the user behavior categories into initial feature data for constructing a random forest classifier;
and step S15, training an initial random forest model, setting trees as 100, and dividing each tree into at least 2 nodes to obtain a behavior recognition model.
Secondly, acquiring incremental characteristic data
FIG. 4 is a flow chart of acquiring incremental feature data according to an embodiment of the present invention. As shown in fig. 4, the step of constructing the behavior recognition model specifically includes:
step S21, reading behavior data of an initial sensor (an accelerometer, a gyroscope and the like) containing a user behavior category as first incremental data;
step S22, reading the behavior data of the newly added sensor (accelerometer, gyroscope, etc.) and taking the behavior data of the user behavior category as second incremental data;
step S23, sliding the intercepted first incremental data with fixed time length (such as 5 seconds) by using a sliding window method, and processing each window data through a preprocessing algorithm, wherein the preprocessing algorithm includes but is not limited to data filtering and the like; intercepting the data obtained by preprocessing by using a sliding window, wherein the adjacent windows are overlapped by 50%, and the extracted features of each window include but are not limited to: mean, standard deviation, minimum, maximum, mode, interval, mean point number, direct current component, peak value after fast Fourier transform, mean, standard deviation energy entropy and the like; combining the extracted features with user behavior categories into first feature data
Step S24, processing the second incremental data by the method of step S23 to obtain second characteristic data and obtain the incremental characteristics of the second incremental data;
and step S25, combining the first characteristic data and the second characteristic data to obtain incremental characteristic data for updating the behavior recognition model.
Third, behavior recognition model updating
FIG. 5A is a schematic diagram of a single decision tree of a behavior recognition model according to an embodiment of the present invention. As shown in fig. 5A, the behavior recognition model may recognize user behavior data of an initial sensor, but after incremental behavior data is obtained by a new sensor, the behavior recognition model cannot improve behavior recognition performance of an existing model by using data of the new sensor, and needs to perform dynamic update of the model, where the dynamic update of the model mainly includes a diversity generation strategy and a feature incremental tree growth mechanism based on mutual information.
(one) mutual information-based diversity generation strategy
In order to improve the diversity of decision trees in the integrated classifier, the invention provides a diversity generation strategy based on mutual information.
How to measure the effect of individual learners on the diversity of the integrated classifier is an important challenge in the diversity generation process. The invention provides a new diversity scoring rule-a diversity scoring rule based on mutual information. Which can take into account both the accuracy and the redundancy of a single classifier.
In the field of information theory, mutual information is proposed to measure the correlation between two variables:
I(X1,X2) A larger value of (A) means X1And X2The correlation of (a) is higher.
Likewise, the diversity measure, I (h), can be performed using mutual informationiY) may be used to measure the correlation between the predicted outcome of the ith decision tree and the actual behavior of the user, I (h)i,hk) Can be used to measure the redundancy between the prediction of the ith decision tree and the prediction of the kth decision tree. The mutual information based diversity scoring rules may be formalized as follows:
wherein S (h)i) A diversity score for the ith decision tree in the behavior recognition model,
hiis the predicted result of the ith decision tree, hkAnd M is the total number of decision trees in the behavior recognition model, and y is the user behavior corresponding to the ith decision tree.
(2) The first term I (h) in the formulaiAnd y) represents mutual information of the output of the ith decision tree and the sample real category, and is used for measuring the correlation, and the higher the value of the mutual information is, the higher the precision of the ith decision tree is proved.
(2) Second term in the formula
Is the average mutual information between the ith decision tree and other decision trees. The larger the second term, the stronger the correlation of the ith decision tree with other decision trees.
Then, S (h) can be obtained from the formula (2)i) The larger the value, the higher the precision of the ith decision tree and the lower the redundancy with other decision trees. Thus, the present invention selects S (h)i) The decision trees with small values are used as the decision trees needing to be updated, and the rest decision trees are kept unchanged. By the strategy, the difference between decision trees can be maximized on the basis of ensuring the identification precision of the integrated classifier. This integrated diversity metric strategy is a mutual information based diversity generation strategy (MIDGS).
To maximize the diversity of the ensemble learner, it is important to select a suitable proportion of decision trees for further updating. The invention provides two methods for determining the proportion of a decision tree to be further updated. The method comprises the steps of sequencing decision trees according to the value of S, and then selecting a certain proportion of decision trees for further growth. Another approach is to define a specified update threshold. The decision tree (to-be-updated decision tree) with the value of S smaller than the update threshold is selected for updating. And after the decision tree to be updated is updated, combining the updated decision tree with the non-updated decision tree to obtain an updated behavior recognition model.
Feature increment tree growth mechanism
Fig. 5B and 5C are schematic diagrams of a decision tree growing mechanism of a behavior recognition model according to an embodiment of the present invention. The higher the accuracy of the decision tree, the better the performance of the integrated classifier. Therefore, the invention provides an incremental tree growing mechanism for a decision tree which is obtained by an MIDGS strategy and needs to be updated, as shown in FIG. 5A, the existing random forest behavior recognition model, namely a characteristic incremental tree growing mechanism (FITGM), is promoted by using newly added sensor data, and the incremental tree growing mechanism comprises two tree structure operations. Fig. 5B and 5C give diagrams of two operations of FITGM.
1. Modifying a sub-tree
With the advent of new sensors, new features are correspondingly obtained. When a new feature gets a better score at a decision tree node, we discard the current partition and the child nodes of the current node and reconstruct a new sub-tree with the new feature as the splitting attribute, as shown in fig. 5B. The common scoring method comprises information gain and a kini coefficient, wherein the information gain is an important index in feature selection and is defined as how much information can be brought to a classification system by one feature, and the more information is brought, the more important the feature is; the kini index is used to determine the degree of disorder of the sample, and the more the classes contained in the population are disordered, the larger the kini index is. Although updating sub-trees of the decision tree with feature data (incremental features) obtained by newly added sensors has the potential to improve the performance of the decision tree, it may also result in more time consumption and loss of information stored in the existing tree structure. Therefore, we propose two constraints to limit the modification to the subtree:
(1) the number of newly added samples meets formula 3 or formula 4;
(2) the score of the new feature is better than the score of the current partition, score (f)new)＞scorenode。
(1) And (2) to ensure that modifications to a subtree yield more revenue than the cost of discarding the current subtree.
Wherein n isINCIs the number of samples contained in the data block of the incremental feature, and d is the node depth (assuming that the root node depth is 1 and its child node depth is 2).
n≥nnode(4)
Wherein n isnodeIs the number of samples that have fallen into the current node.
2. Splitting leaf nodes
In the characteristic increment learning stage, the data distribution changes correspondingly with the arrival of the increment data. In order to adapt to the dynamic change of data distribution, the invention adopts a decision tree online growth strategy similar to ORF-Saffari [ Saffari A, Leistner C, Santner J, et al.on-line random concerns for [ C ]// Computer Vision Workshos (ICCVWorkshos), 2009IEEE 12th International Conference on.IEEE,2009: 1393-. The threshold may be self-adjusting for specific uses. The attribute with the best split score is selected to split the leaf node as shown in figure 5C.
Third, related experiments of the invention
In order to further verify the effectiveness of the behavior recognition model updating method and system facing the dynamic increase of the sensor and to illustrate the use method of the invention, the inventor also takes the motion behavior recognition as an example to perform experiments. The experiment adopts the Daily exercise behavior Data Set Daily and sports activities Data Set of University of California Irvine (University of California Irvine) used for a machine learning database, hereinafter referred to as DSADS, which comprises 19 types of Daily exercise behaviors collected by 8 participants in total of 4 males and 4 females.
(one) data acquisition
Three types of sensors were used in the experiment: the 3-axis accelerometer, the 3-axis gyroscope and the 3-axis magnetometer are respectively fixed on five parts of a body: the torso, left and right arms, and left and right legs, i.e., 15 sensors in total. The athletic performance mainly collects 19 kinds, and the 19 kinds of performance mainly include: sitting, standing, lying on its side, going upstairs and downstairs, standing in an elevator, walking in a parking lot, walking on a treadmill at a speed of 4 km/h, walking on a treadmill at 15 inclined positions at a speed of 4 km/h, running on a treadmill at a speed of 8 km, stepping exercise, training on a cross trainer, riding a bicycle in a horizontal position, riding a bicycle in a vertical position, rowing, jumping, basketball.
(II) feature extraction
And extracting features from the acquired athletic performance data. These features include two broad categories: (1) time domain features, i.e. features that have a time dependence during the time-varying sequence, including mean, standard deviation, minimum, maximum, mode, interval, number of over-mean points (2) frequency domain features, are typically used to find the periodicity information in the signal: direct current component, peak value, average value, standard deviation, energy entropy and the like after fast Fourier transform. A single sensor extracts 27-dimensional features, for a total of 15 sensors, 405-dimensional features.
(III) classification
And in order to illustrate the effectiveness of the method, a random forest retraining model of a conventional machine learning method is used as a comparison experiment, and the test precision and the training time are used as performance test standards. The test precision refers to the proportion of the correctly classified samples in all the samples, and the training time refers to the training time required for constructing a behavior recognition model capable of recognizing a new class. In the experimental process, a single sensor data extraction characteristic initial classifier is selected, and the method or the random forest is respectively adopted for retraining when new sensors continuously appear. The results of the experiments are shown in FIGS. 6, 7 and 8. As can be seen from FIG. 6, the behavior recognition model updating method facing dynamic increase of sensors, Feature associative Random Forest, hereinafter referred to as FIRF, of the invention, the test accuracy on the DSADS test set is always higher than Random Forest retraining, Random Forest-Retrain, hereinafter referred to as RF-Retrain, and the method is proved to be an effective method for solving dynamic increase of the number of sensors. From FIG. 7, it can be seen that the training time required for FIRF, which is a method of the present invention, is much less than that of RF-Retrain, which has a significant advantage in time consumption. As can be seen from FIG. 8, the FIRF testing time obtained by the present invention is substantially consistent with that of RF-Retrain, even better, and the redundancy of the model constructed by the FIRF testing time is low. Experiments prove that the characteristic increment random forest provided by the invention can update the existing behavior recognition model by using the newly added sensor data, and obtains the behavior recognition model with higher precision and lower redundancy by using less training time.
Claims (10)
1. A behavior recognition model updating method facing dynamic addition of sensors is characterized by comprising the following steps:
a model construction step, in which initial data of user behaviors are obtained through an initial sensor, and initial characteristic data are extracted to construct a behavior recognition model;
an incremental characteristic data acquisition step, namely acquiring incremental data of user behaviors through the initial sensor and the newly added sensor, defining incremental characteristics and extracting incremental characteristic data;
a step of model updating decision-making, which is to obtain diversity scores of each decision tree of the behavior recognition model according to average mutual information between prediction results of the decision trees in the behavior recognition model and mutual information between the prediction results and actual behaviors of users corresponding to the prediction results, and take the decision trees with the diversity scores smaller than an updating threshold value as the decision trees to be updated;
and a step of dynamically updating the model, namely updating all decision trees to be updated by the incremental characteristic data so as to update the behavior recognition model.
2. The behavior recognition model updating method according to claim 1, wherein the model building step specifically includes:
processing the initial data by a sliding window method to obtain the initial characteristic data;
and constructing a random forest classifier by using the initial characteristic data to obtain the behavior recognition model.
3. The behavior recognition model updating method according to claim 1, wherein the incremental feature data obtaining step specifically includes:
acquiring first incremental data of user behavior through the initial sensor; acquiring second incremental data of user behaviors through the newly added sensor;
processing the first incremental data by a sliding window method to obtain first characteristic data corresponding to the initial sensor; processing the second incremental data by a sliding window method to extract the incremental features and obtain second feature data corresponding to the newly added sensor; the first characteristic data and the second characteristic data are combined to obtain the incremental characteristic data.
4. The behavior recognition model updating method according to claim 1, wherein the diversity score is obtained by:
5. The behavior recognition model updating method according to claim 1, wherein the model dynamic updating step specifically includes:
a sub-tree modification step, wherein a new sub-tree is constructed under a node by taking a certain increment characteristic of the node falling into a certain decision tree to be updated as a split attribute;
and a leaf node splitting step, namely splitting the leaf node when the type of the user behavior in the leaf node of a certain decision tree to be updated is more than 1 and the number of samples falling into the leaf node is more than a splitting threshold.
6. A system for updating a behavior recognition model for dynamic addition of sensors, comprising:
the model building module is used for obtaining initial data of user behaviors through an initial sensor and extracting initial characteristic data to build a behavior recognition model;
the incremental characteristic data acquisition module is used for acquiring incremental data of user behaviors through the initial sensor and the newly added sensor, defining incremental characteristics and extracting incremental characteristic data;
the model updating decision module is used for determining a decision tree to be updated in the updating process of the behavior recognition model; obtaining diversity scores of each decision tree of the behavior recognition model according to average mutual information between prediction results of the decision trees in the behavior recognition model and mutual information between the prediction results and actual behaviors of users corresponding to the prediction results, and taking the decision trees with the diversity scores smaller than an update threshold value as decision trees to be updated;
and the model dynamic updating module is used for updating all the decision trees to be updated by the incremental characteristic data so as to update the behavior recognition model.
7. The behavior recognition model update system of claim 6, wherein the model building module comprises:
the initial data acquisition module is used for processing the initial data by a sliding window method to obtain the initial characteristic data;
and the model construction training module is used for constructing a random forest classifier by using the initial characteristic data to obtain the behavior recognition model.
8. The behavior recognition model updating system of claim 6, wherein the incremental feature data acquisition module comprises:
the incremental data acquisition module is used for acquiring first incremental data of user behaviors through the initial sensor and acquiring second incremental data of the user behaviors through the newly added sensor;
an incremental data processing module, configured to obtain the incremental feature data, where the first incremental data is processed by a sliding window method to obtain first feature data corresponding to the initial sensor, and the second incremental data is processed by a sliding window method to extract the incremental feature and obtain second feature data corresponding to the newly added sensor; the first characteristic data and the second characteristic data are combined to obtain the incremental characteristic data.
9. The behavior recognition model update system of claim 6, wherein the model update decision module obtains the diversity score by:
10. The behavior recognition model update system of claim 6, wherein the model dynamic update module comprises:
the sub-tree modification module is used for constructing a new sub-tree under the decision tree to be updated; taking a certain increment characteristic of a node falling into a certain decision tree to be updated as a split attribute, and constructing a new sub-tree under the node;
the leaf node splitting module is used for splitting leaf nodes of the decision tree to be updated; when the type of the user behavior in a leaf node of a certain decision tree to be updated is larger than 1 and the number of samples falling into the leaf node is larger than a splitting threshold value, splitting the leaf node.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810315805.3A CN108717548B (en) | 2018-04-10 | 2018-04-10 | Behavior recognition model updating method and system for dynamic increase of sensors |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810315805.3A CN108717548B (en) | 2018-04-10 | 2018-04-10 | Behavior recognition model updating method and system for dynamic increase of sensors |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108717548A CN108717548A (en) | 2018-10-30 |
| CN108717548B true CN108717548B (en) | 2020-09-15 |
Family
ID=63898960
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810315805.3A Active CN108717548B (en) | 2018-04-10 | 2018-04-10 | Behavior recognition model updating method and system for dynamic increase of sensors |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108717548B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109726195B (en) * | 2018-11-26 | 2020-09-11 | 北京邮电大学 | Data enhancement method and device |
| CN111353600B (en) * | 2020-02-20 | 2023-12-12 | 第四范式（北京）技术有限公司 | Abnormal behavior detection method and device |
| CN111401507B (en) * | 2020-03-12 | 2021-01-26 | 大同公元三九八智慧养老服务有限公司 | Adaptive decision tree fall detection method and system |
| CN112861796B (en) * | 2021-03-12 | 2023-04-07 | 中国科学院计算技术研究所 | Feature adaptive motion recognition method |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8948501B1 (en) * | 2009-12-22 | 2015-02-03 | Hrl Laboratories, Llc | Three-dimensional (3D) object detection and multi-agent behavior recognition using 3D motion data |
| CN105718490A (en) * | 2014-12-04 | 2016-06-29 | 阿里巴巴集团控股有限公司 | Method and device for updating classifying model |
| CN106156809A (en) * | 2015-04-24 | 2016-11-23 | 阿里巴巴集团控股有限公司 | For updating the method and device of disaggregated model |
| CN106446964B (en) * | 2016-10-21 | 2018-04-03 | 河南大学 | A kind of gradient lifting decision tree update method of increment type |
-
2018
- 2018-04-10 CN CN201810315805.3A patent/CN108717548B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN108717548A (en) | 2018-10-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108717548B (en) | Behavior recognition model updating method and system for dynamic increase of sensors | |
| CN108446733A (en) | A kind of human body behavior monitoring and intelligent identification Method based on multi-sensor data | |
| CN110334573B (en) | Human motion state discrimination method based on dense connection convolutional neural network | |
| CN107798563B (en) | Internet advertisement effect evaluation method and system based on multi-mode characteristics | |
| CN104586402B (en) | A kind of feature extracting method of physical activity | |
| Ghazali et al. | Common sport activity recognition using inertial sensor | |
| CN106910314A (en) | A kind of personalized fall detection method based on the bodily form | |
| CN111274998A (en) | Parkinson's disease finger knocking action identification method and system, storage medium and terminal | |
| CN110490109A (en) | A kind of online human body recovery action identification method based on monocular vision | |
| CN112699785A (en) | Group emotion recognition and abnormal emotion detection method based on dimension emotion model | |
| Shen et al. | Feature analysis of mouse dynamics in identity authentication and monitoring | |
| CN115563484A (en) | Street greening quality detection method based on physiological awakening identification | |
| CN111931568A (en) | Human body falling detection method and system based on enhanced learning | |
| Youn et al. | Wearable sensor-based biometric gait classification algorithm using WEKA | |
| Du et al. | RETRACTED: Research on the intelligent model of progress in physical education training based on motion sensor | |
| Ning et al. | Fall detection algorithm based on gradient boosting decision tree | |
| CN109271889B (en) | Action recognition method based on double-layer LSTM neural network | |
| CN113345590B (en) | User mental health monitoring method and system based on heterogeneous graph | |
| Fañez et al. | Transfer learning and information retrieval applied to fall detection | |
| CN114881079A (en) | Human body movement intention abnormity detection method and system for wearable sensor | |
| WO2022100187A1 (en) | Mobile terminal-based method for identifying and monitoring emotions of user | |
| CN111210912A (en) | Parkinson prediction method and device | |
| CN107045624A (en) | A kind of EEG signals pretreatment rolled into a ball based on maximum weighted and sorting technique | |
| Kao et al. | GA-SVM applied to the fall detection system | |
| Kumari et al. | Effect of reduced dimensionality on deep learning for human activity recognition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |