CN115439934A - Self-adaptive step frequency detection method based on CNN-LSTM motion mode identification - Google Patents

Self-adaptive step frequency detection method based on CNN-LSTM motion mode identification Download PDF

Info

Publication number
CN115439934A
CN115439934A CN202211102627.9A CN202211102627A CN115439934A CN 115439934 A CN115439934 A CN 115439934A CN 202211102627 A CN202211102627 A CN 202211102627A CN 115439934 A CN115439934 A CN 115439934A
Authority
CN
China
Prior art keywords
data
lstm
cnn
motion
frequency detection
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.)
Pending
Application number
CN202211102627.9A
Other languages
Chinese (zh)
Inventor
杨运成
吴飞
朱润哲
杨明泽
朱海
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shanghai University of Engineering Science
Original Assignee
Shanghai University of Engineering Science
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Shanghai University of Engineering Science filed Critical Shanghai University of Engineering Science
Priority to CN202211102627.9A priority Critical patent/CN115439934A/en
Publication of CN115439934A publication Critical patent/CN115439934A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • 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
    • G06V40/25—Recognition of walking or running movements, e.g. gait recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/30—Noise filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/7715—Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks

Abstract

The invention relates to a self-adaptive step frequency detection method based on CNN-LSTM motion mode identification, which comprises the following steps: acquiring data acquired by a sensor when a user walks; inputting the collected data into a CNN-LSTM model for motion mode identification and classification; inputting the acquired data and the classification result into a step frequency detection algorithm adopting adaptive threshold crest detection to obtain a step frequency detection result; the step of adopting the CNN-LSTM model to carry out motion pattern recognition classification comprises the following steps: making the acquired data into two-dimensional slice data, and inputting the slice data into a CNN (continuous neural network); extracting motion characteristics of the slice data by using the convolution layer; performing dimensionality reduction and downsampling through a pooling layer; expanding the data into a one-dimensional array through the flattening layer and inputting the one-dimensional array into the LSTM layer; classifying different motion modes of the data by using an LSTM layer; and outputting the classification result through the full connection layer. Compared with the prior art, the pedestrian identification method has the advantages of self-adaption to different motion modes of pedestrians, no need of manual feature extraction, high identification accuracy, effective elimination of pseudo wave crests and the like.

Description

Self-adaptive step frequency detection method based on CNN-LSTM motion mode identification
Technical Field
The invention relates to the technical field of step frequency detection, in particular to a self-adaptive step frequency detection method based on CNN-LSTM motion mode identification.
Background
With the continuous development of information technology, indoor location services such as shopping navigation, parking navigation, industrial positioning, and the like have become an indispensable part of people's daily life and work. The method has the advantages that positioning equipment is not required to be deployed in advance, the independence is strong, the method is suitable for unknown environments, and a Pedestrian Dead Reckoning (PDR) indoor positioning method based on the smart phone is made out of a plurality of indoor position service technologies and is widely concerned and researched. PDR methods use the readings of inertial sensors built into most modern smartphones to calculate the user's stride frequency, step size, and heading in order to estimate position in real time.
An important ring of the PDR positioning technology at present is how to implement robust and high-precision step frequency detection. Usually, the position where a user carries a mobile phone while walking is not a single fixed position, and common carrying positions include holding the chest with a hand, swinging the arms, making a call, and putting in a pocket. Meanwhile, the false walking state caused by the operation of the user in the non-walking state also causes the problem of excessive counting, so that the method has important significance for the research of recognizing the pedestrian motion mode and carrying out step frequency detection aiming at different mode adaptive thresholds.
Chinese patent CN202110521228.5 discloses a dead reckoning positioning method based on pedestrian motion state identification, the method comprising: constructing a pedestrian motion state identification classification model: collecting three-axis acceleration data of the pedestrian in five motion states of walking, jogging, left striding, right striding and reversing, and constructing a pedestrian motion state recognition classification model. Recognizing the motion state of the pedestrian: collecting three-axis acceleration data and three-axis gyroscope data of the pedestrian, and identifying the pedestrian motion state by using a pedestrian motion state identification classification model. Step frequency detection is carried out to obtain single step frequency: and (4) performing step frequency detection on the acquired triaxial acceleration data to obtain single step frequency. Step length estimation: if the identified pedestrian motion state is walking, left stepping, right stepping or reversing, estimating a single step by adopting a linear step model and combining a single step frequency; and if the pedestrian motion state is identified as jogging, estimating a single step size by combining a Weinberg nonlinear step size model with the single step frequency. Course estimation: and carrying out integral solution on the acquired three-axis gyroscope data by adopting angular velocity converted based on a quaternion coordinate system to calculate a course angle, and finishing course angle correction by adopting a heuristic offset elimination algorithm. And (3) dead reckoning: and setting the initial position coordinates and the initial course angle of the pedestrians, and updating the positions of the pedestrians in the five motion states according to the motion state of the pedestrians, the single step length and the corrected course angle.
The prior art has the following defects:
in a conventional motion pattern recognition method, a machine learning method such as a decision tree, a random forest, a support vector machine, an artificial neural network, or the like is generally used. While these conventional classifiers can identify different motion activities, the accuracy of the identification is highly dependent on different types of features extracted manually from the source sensor data, such as statistical features (e.g., mean, variance, energy), time domain features (e.g., zero-crossing ratio), frequency domain features (e.g., fast fourier transform). And, the sensitivity of different types of features to different motion modes is different, and if a large amount of experience accumulation is needed for selecting proper features, other models are difficult to popularize.
Because the gravity center of the body fluctuates up and down in the vertical direction when a user walks, the combined acceleration data has the periodic oscillation characteristic of wave crests and wave troughs, and the walking steps are determined according to the number of the wave crests of the acceleration data in the traditional peak detection. Although the prior art can eliminate most of false peaks generated by factors such as motion mode and body shaking through filtering wave energy, the problem of excessive counting caused by partial false peaks still exists.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide an adaptive step frequency detection method based on the CNN-LSTM motion mode identification.
The purpose of the invention can be realized by the following technical scheme:
as an aspect of the present invention, an adaptive stride frequency detection method based on CNN-LSTM motion pattern recognition is provided, the detection method includes the following steps: acquiring data acquired by a sensor when a user walks; inputting the collected data into a CNN-LSTM model for motion mode identification and classification; inputting the acquired data and the classification result into a step frequency detection algorithm adopting adaptive threshold crest detection to obtain a step frequency detection result;
the steps of adopting the CNN-LSTM model to carry out motion pattern recognition classification comprise:
making the acquired data into two-dimensional slice data, and inputting the slice data into a CNN;
extracting motion characteristics of slice data by using the convolutional layer;
performing dimensionality reduction and downsampling through a pooling layer;
expanding the data into a one-dimensional array through the flattening layer and inputting the one-dimensional array into the LSTM layer;
classifying different motion modes of the data by using an LSTM layer;
and outputting a classification result through the full connection layer, wherein the classification result comprises a motion state and a mobile phone posture.
As a preferred technical solution, the sensor data includes three-axis acceleration and three-axis angular velocity.
As a preferred technical solution, the step of generating the acquired data into the two-dimensional slice data is to sequentially generate 30 pieces of data into the two-dimensional slice data with the size of 30 × 6.
As a preferred technical scheme, the activation function of the full connection layer is Soft-max.
As a preferred technical scheme, the motion state comprises normal walking and fast walking, and the mobile phone posture comprises hand holding, communication, swing arm and pocket.
As a preferred technical solution, the step frequency detection algorithm includes the following steps:
filtering the acquired data by using a filter;
and screening the filtered data according to the following formula to obtain a Peak value Peak meeting the conditions t ；
Peak t ＝(a t ≥(a t-k ∶a t-1 )&&a t ≤(a t+1 ∶a t+k ))
Wherein, a t The method comprises the steps that a sum acceleration value is obtained under a t sample, k is a sample quantity threshold value on the left side and the right side of the t, and the k is set according to a motion state classification result of collected data;
the dynamic threshold window size is calculated according to the following equation,
W＝(T s *F s )-1
wherein, T s For each step of time, F s Sampling frequency for the current mobile phone;
then, the dynamic threshold value TH under the current window is calculated according to the formula peak ；
Figure BDA0003839991940000031
Wherein, W max Is the maximum value of the current window, W min D is the approaching peak value degree;
and comparing the peak value after the minimum peak distance filtering with the dynamic threshold value in the window, and if the current peak value is greater than or equal to the dynamic threshold value in the window, judging that the current peak value is a real step.
As a preferable technical scheme, the filter comprises a low-pass filter and a convolution smoothing filter, and the cut-off frequency of the low-pass filter is 3Hz.
As a preferred technical solution, the threshold k of the number of samples is set to 20 in the normal walking mode of the sport mode, and is set to 10 in the fast walking mode.
As a preferred technical scheme, each step of time T s Set to 0.507s in the sport mode in normal walking mode and 0.469 seconds in fast walking mode.
As a preferred technical solution, the approach peak degree d is set to 1.05 when the mobile phone is in a handheld state and in a telephone state, and is set to 0.9 when the mobile phone is in a pocket state and in a swing arm state.
Compared with the prior art, the invention has the following beneficial effects:
1) The invention designs and uses a CNN-LSTM network model, adopts CNN to extract features, avoids the secondary artificial extraction of different types of features from source sensor data by the traditional classification method, and improves the accuracy of motion mode classification by mining the correlation among data features through LSTM.
2) The invention provides a peak detection algorithm of a self-adaptive threshold value aiming at the problem of traditional peak detection, which reduces data noise and eliminates obvious false peaks by identifying a user motion mode and utilizing combined filtering, and self-adaptively detects whether the current step counting meets the conditions or not by utilizing selection of minimum peak distance and a dynamic threshold value method.
Drawings
FIG. 1 is a schematic flow diagram of the present invention;
FIG. 2 is a structural diagram of a CNN-LSTM network model according to the present invention.
Detailed Description
The invention is described in detail below with reference to the figures and the specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation manner and a specific operation process are given, but the scope of the present invention is not limited to the following embodiments.
As an embodiment of the invention, a CNN-LSTM deep web framework model is adopted that does not need secondary feature extraction from source data. The model firstly utilizes CNN to automatically extract and reduce the data characteristics of the dimension sensor, and inputs the characteristics into the LSTM network to identify the correlation among the characteristics, and finally has more accurate classification results on the motion modes. The wave crests are screened by the minimum peak distance selection and dynamic threshold method self-adaptive detection method, so that the step frequency detection accuracy is higher, and the step frequency detection under different motion modes is more robust.
As shown in fig. 1, in a normal case, the walking state (normal walking, fast walking) of the pedestrian has a step frequency lower than 3Hz, and three-axis acceleration (acc _ x, acc _ y, acc _ z) and three-axis angular velocity (ang _ x, ang _ y, ang _ z) data of a sensor with a mobile phone output frequency of 50Hz are sequentially made into two-dimensional slice data with a size of 30 × 6 every 30 (about 2-3 steps), so that the model can learn data information with time sequence information and complete gait more comprehensively. The slice data is input into the CNN, the convolutional layer is used for extracting the motion characteristics of the data, and dimension reduction down-sampling is carried out on the slice data through the pooling layer with the size of 2 to select the characteristics with larger weight and remove noise information interference, so that the problem that the classification precision of the traditional classifier is influenced by a mode of manually extracting the characteristics in advance is solved. And then, the data is expanded into a one-dimensional array through the flattening layer and is input into the LSTM layer, different motion modes used in walking of pedestrians are analyzed by utilizing the analysis capability of the LSTM for the correlation of time series data, the model classification accuracy is improved, and finally, a classification result is output through a full-connection layer with an activation function of Soft-max.
Under the condition of only inputting original acceleration angle speed data without using artificial secondary extraction features, the accuracy of the CNN-LSTM network model identification is far higher than that of a traditional decision tree and a random forest method, and compared with the CNN-LSTM network model identification only using a single CNN model and an LSTM model, the accuracy of the CNN-LSTM network model is respectively improved by 2.77% and 2.57%.
Because the cheap inertial sensor used in the mobile phone has lower measurement accuracy and the sensor is influenced by the body shake of the user in the relative movement process during data acquisition, a large amount of noise exists in the acquired original data, and meanwhile, a pseudo peak can be caused to influence the step frequency detection step counting accuracy. Therefore, a set of two combinations of filtering is provided before the step frequency detection is performed: low pass filtering and convolution smoothing filtering. Firstly, low-pass filtering is utilized to reduce noise in original data and eliminate obvious pseudo wave peaks; and then, the convolution smoothing filter can be used for smoothing the data for the second time to obtain a motion rule more beneficial to step frequency detection. Firstly, a low-pass filter with the cut-off frequency of 3Hz is utilized to filter noise and eliminate a large number of pseudo wave crests, and the low-pass filtering formula is as follows:
Figure BDA0003839991940000051
and then, convolution smoothing filtering is utilized, namely, a continuous subset of adjacent data points is fitted with a low-order polynomial through a linear least square method, the data precision is improved under the condition that the signal trend is not changed, and the smooth data obtains a motion rule which is more beneficial to step frequency detection. The formula is as follows:
Figure BDA0003839991940000052
wherein Y is j Is the filtered value, C i Is the convolution coefficient, and m is the polynomial fitting order.
Even if a large amount of noise can be filtered after data preprocessing, part of unfiltered false peaks still exist between two peaks, and the method based on the traditional peak detection can bring the false peaks into normal step counting, so that the error is increased. Therefore, according to the condition that the movement periods are consistent under the same movement mode and the Peak value is the local maximum value under the current step, the minimum Peak distance is used for removing other wave Peak numbers in the adjacent peaks, and the Peak value Peak after filtering is calculated t The formula is as follows:
Peak t ＝(a t ≥(a t-k ∶a t-1 )&&a t ≤(a t+1 ∶a t+k ))
wherein, a t Refers to the sum of acceleration values under t samples, and k is the sample number threshold on the left and right sides of t. Because the frequency of the movement cycle of the user is different when the user walks normally and quickly, an adaptive threshold algorithm is designed, the current movement state is detected according to the classification result, and the minimum peak distance k value is adaptively adjusted. K =20 in the normal walking mode and k =10 in the fast walking mode are set.
The user can generate different periodic characteristics by using sensor data under different mobile phone gestures in the walking process. In the handheld mode, a plurality of small wave peaks appear first and a larger wave peak appears in each actual one-step process. These large peaks correspond to true steps, while the influence step-counting wavelet peaks need to be eliminated as false peaks. In the pocket mode, a large peak appears first in each actual step, then a few small peaks appear, and then a middle peak appears, and the rule is that the peak needing real step counting is a large-middle rule. In the traditional fixed threshold setting method, a smaller fixed threshold is set to introduce a pseudo peak to add in step counting, and a larger fixed threshold is set to take a middle peak in a pocket mode as a pseudo peak to remove the step counting. Therefore, the step counting precision under different motion modes can be effectively improved by adopting a self-adaptive dynamic threshold value method.
The dynamic threshold method is based on a dynamic threshold constructed by acceleration maximum and minimum values in a window, wherein the size of the window depends on the motion state and the acquisition frequency of the pedestrian, and different window sizes also influence the threshold accuracy. According to the time T of each step under different motion states s And the current sampling frequency F of the mobile phone s The window size setting formula is as follows:
W＝(T s *F s )-1
time per step T in normal walking and fast walking states s 0.507s and 0.469 s, respectively.
Dynamic threshold value TH under current window peak The calculation formula of (a) is as follows:
Figure BDA0003839991940000061
wherein, W max Is the maximum value of the current window, W min For the current window minimum, d refers to the approaching peak extent. Since the peak amplitude per step is approximately the same for both the hand and phone postures, while the peak amplitude per step for both the pocket and arm swing postures is in the form of a large peak and a small peak, d =1.05 for both the hand and phone postures and d =0.9 for both the pocket and arm swing postures are set. Then comparing the peak value after minimum peak distance filtering with the dynamic threshold value in the window, if the current peak value is more than or equal to the dynamic threshold value TH in the window peak The true step is counted.
Compared with the traditional wave crest detection method, the self-adaptive detection method utilizing the minimum peak distance selection and the dynamic threshold value method has the advantages that the step frequency detection accuracy is higher, and meanwhile, the self-adaptive detection method is suitable for the step frequency detection of pedestrians walking in different motion modes.
The foregoing detailed description of the preferred embodiments of the invention has been presented. It should be understood that numerous modifications and variations could be devised by those skilled in the art in light of the present teachings without departing from the inventive concepts. Therefore, the technical solutions available to those skilled in the art through logic analysis, reasoning and limited experiments based on the prior art according to the concept of the present invention should be within the scope of protection defined by the claims.

Claims (10)

1. An adaptive stride frequency detection method based on CNN-LSTM motion pattern recognition is characterized by comprising the following steps: acquiring data acquired by a sensor when a user walks; inputting the collected data into a CNN-LSTM model for motion mode identification and classification; inputting the acquired data and the classification result into a step frequency detection algorithm adopting adaptive threshold peak detection to obtain a step frequency detection result;
the steps of adopting the CNN-LSTM model to carry out motion pattern recognition classification comprise:
making the acquired data into two-dimensional slice data, and inputting the slice data into a CNN (continuous neural network);
extracting motion characteristics of slice data by using the convolutional layer;
performing dimensionality reduction and downsampling through a pooling layer;
expanding the data into a one-dimensional array through the flattening layer and inputting the one-dimensional array into the LSTM layer;
classifying different motion modes of the data by using an LSTM layer;
and outputting a classification result through the full connection layer, wherein the classification result comprises a motion state and a mobile phone posture.
2. The adaptive step frequency detection method based on CNN-LSTM motion pattern recognition, as claimed in claim 1, wherein the sensor data collection comprises three-axis acceleration and three-axis angular velocity.
3. The method of claim 1, wherein the step-frequency generation of the collected data into two-dimensional slice data is performed by sequentially generating 30 pieces of data into two-dimensional slice data with a size of 30 × 6.
4. The CNN-LSTM motion pattern recognition-based adaptive stride frequency detection method of claim 1, wherein the activation function of the full link layer is Soft-max.
5. The CNN-LSTM motion pattern recognition-based adaptive step frequency detection method according to claim 1, wherein the motion states include normal walking and fast walking, and the mobile phone gestures include handheld, talking, arm swinging and pocket.
6. The method of claim 5, wherein the stride frequency detection algorithm comprises the following steps:
filtering the acquired data by using a filter;
screening the filtered data according to the formula below to obtain qualified Peak value Peak t ；
Peak t ＝(a t ≥(a t-k ∶a t-1 )&&a t ≤(a t+1 ∶a t+k ))
Wherein, a t The method comprises the steps that a sum acceleration value is obtained under a t sample, k is a sample quantity threshold value on the left side and the right side of the t, and the k is set according to a motion state classification result of collected data;
the dynamic threshold window W size is calculated according to the following equation,
W＝(T s *F s )-1
wherein, T s For each step of time, F s Sampling frequency for the current mobile phone;
then, the current window is calculated according to the following formulaLower dynamic threshold TH peak ；
Figure FDA0003839991930000021
Wherein, W max Is the maximum value of the current window, W min D is the approaching peak value degree;
and comparing the peak value after the minimum peak distance filtering with the dynamic threshold value in the window, and if the current peak value is greater than or equal to the dynamic threshold value in the window, judging that the current peak value is a real step.
7. The method of claim 6, wherein the filter comprises a low pass filter and a convolution smoothing filter, and the cut-off frequency of the low pass filter is 3Hz.
8. The method of claim 6, wherein the threshold k is set to 20 in a normal walking mode and to 10 in a fast walking mode.
9. The method of claim 6, wherein each step time T is a step frequency of the adaptive detection method based on the CNN-LSTM motion pattern recognition s Set to 0.507s in the normal walking mode and 0.469 seconds in the fast walking mode in the sport mode.
10. The CNN-LSTM motion pattern recognition-based adaptive stride frequency detection method of claim 6, wherein the approach peak degree d is set to 1.05 in handset and phone postures and 0.9 in pocket and arm swing postures.
CN202211102627.9A 2022-09-09 2022-09-09 Self-adaptive step frequency detection method based on CNN-LSTM motion mode identification Pending CN115439934A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211102627.9A CN115439934A (en) 2022-09-09 2022-09-09 Self-adaptive step frequency detection method based on CNN-LSTM motion mode identification

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211102627.9A CN115439934A (en) 2022-09-09 2022-09-09 Self-adaptive step frequency detection method based on CNN-LSTM motion mode identification

Publications (1)

Publication Number Publication Date
CN115439934A true CN115439934A (en) 2022-12-06

Family

ID=84246741

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211102627.9A Pending CN115439934A (en) 2022-09-09 2022-09-09 Self-adaptive step frequency detection method based on CNN-LSTM motion mode identification

Country Status (1)

Country Link
CN (1) CN115439934A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117592003A (en) * 2024-01-18 2024-02-23 之江实验室 Motion mode identification method, device and medium based on multimode bilinear pooling

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117592003A (en) * 2024-01-18 2024-02-23 之江实验室 Motion mode identification method, device and medium based on multimode bilinear pooling

Similar Documents

Publication Publication Date Title
Batool et al. Sensors technologies for human activity analysis based on SVM optimized by PSO algorithm
CN106846729B (en) Tumble detection method and system based on convolutional neural network
US8000500B2 (en) System and method for analyzing of human motion based on silhouettes of real time video stream
CN105678222A (en) Human behavior identification method based on mobile equipment
CN107292242A (en) A kind of iris identification method and terminal
CN107358248B (en) Method for improving falling detection system precision
CN103942577A (en) Identity identification method based on self-established sample library and composite characters in video monitoring
CN108510011B (en) User travel mode analysis method based on mobile phone multi-sensor
CN107339992A (en) A kind of method of the semantic mark of the indoor positioning and terrestrial reference of Behavior-based control
CN111505632A (en) Ultra-wideband radar action attitude identification method based on power spectrum and Doppler characteristics
CN111089604B (en) Body-building exercise identification method based on wearable sensor
US9398867B2 (en) Efficient activity classification from motion inputs
CN107277222A (en) User behavior state judging method based on mobile phone built-in sensors
CN108171278A (en) A kind of recognizing model of movement method and system based on training data
CN115439934A (en) Self-adaptive step frequency detection method based on CNN-LSTM motion mode identification
Hwang et al. Multi-modal human action recognition using deep neural networks fusing image and inertial sensor data
CN107122711A (en) A kind of night vision video gait recognition method based on angle radial transformation and barycenter
CN112464738A (en) Improved naive Bayes algorithm user behavior identification method based on mobile phone sensor
CN114897025A (en) Human body posture recognition model establishing method and human body posture recognition method
CN110163264B (en) Walking pattern recognition method based on machine learning
CN113269076B (en) Violent behavior detection system and detection method based on distributed monitoring
CN112580486B (en) Human behavior classification method based on radar micro-Doppler signal separation
CN108563939B (en) Human body identity recognition based on gait track curve characteristics
CN107688828B (en) Bus crowding degree estimation method based on mobile phone sensor
Saeedi et al. Context aware mobile personal navigation services using multi-level sensor fusion

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