Method and system for human gait detection
Through the combination of a single three-axis accelerometer and a naive Bayes classifier, the effectiveness of the existing gait detection method under unsupervised conditions is solved, and low-power, real-time gait detection on the watch is achieved, which is suitable for Long-term monitoring of daily life.
Patent Information
- Application Number
- CN202010850567.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2019-08-21
- Filing Date
- 2020-08-21
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2040-12-06
AI Technical Summary
The existing gait detection methods are poorly effective under unsupervised conditions in real world, especially in daily life, where there are interference factors such as stopping in situ movement and unpredictable physical activity, resulting in a significant reduction in detection performance and multi-sensor system function. High consumption and inconvenient for long-term measurement.
Using a single three-axis accelerometer combined with a naive Bayes classifier, the optimization algorithm is optimized to achieve online gait detection with low power without calibration, suitable for internal operation of the watch, and provides real-time feedback .
Accurate gait detection under real-world conditions, reduces power consumption and computing, provides high availability and real-time, is suitable for long-term monitoring without affecting the daily activities of users.
Smart Images

Figure CN112401876B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a gait detection method for detecting a person's gait. The method uses a processing circuit and an accelerometer, which can be worn, for example, on a person's wrist. The present invention also relates to a gait detection sensor configured to implement the method. Background Art
[0002] Physical activity (PA) is one of the fundamental aspects of daily life that is closely related to well-being and is considered a major health indicator for a population. The World Health Organization (WHO) reports that PA is closely linked to fall risk, cognitive function, muscle health, and functional fitness in older adults. PA is particularly important when considering the increasing trend of an aging population. PA is a critical component of healthy aging, multi-therapeutic interventions, and risk reduction for chronic noncommunicable diseases such as diabetes, hypertension, cardiovascular disease, depression, obesity, and certain types of cancer.
[0003] Among the different types of PA, gait (e.g., walking and running) is one of the most important and effective PAs, and an objective assessment of gait can provide useful and valuable information. Thanks to the recent development of wearable technology, various systems based on inertial measurement units (IMUs) have been developed to monitor PA (especially gait bouts), which include accelerometers and gyroscopes attached to the lower limbs, upper body, or both. In some cases, the IMU in the phone is used when the smart phone is fixed to different parts of the body. Although these systems provide gait detection outside the laboratory, they still have some defects. In daily situations, especially when the goal is long-term measurement, wearing multiple sensors on the feet, calves, thighs, buttocks or chest may be cumbersome, uncomfortable and inconvenient. In order to ensure the fixed orientation / position of the sensor and the fixation and alignment of the body part during the entire measurement period, expert intervention and additional testing may be required, which will affect the usability of the system or may easily interfere with the wearer and change his usual daily activities. Furthermore, due to the use of multiple sensors and / or modalities (e.g., gyroscopes), the power consumption of such systems can be high, which limits the duration of the measurements. Existing algorithms also lack the ability to provide real-time data processing and analysis associated with generating real-time feedback due to their high complexity. Therefore, most known systems operate only offline, so that the recorded raw data must be later transferred to a server for further analysis.
[0004] Given the above limitations, alternative solutions for PA (especially gait) monitoring use a single IMU mounted on the wrist. These solutions provide comfort, high usability, and unobtrusive monitoring (e.g., integrated inside a wristwatch), resulting in improved user compliance. Currently known wrist-based methods use abstract modeling, in which several features based on time, frequency, and statistics are extracted and fed into various types of machine learning models (e.g., decision trees or support vector machines (SVM)). Such methods are independent of sensor orientation. Therefore, sensor calibration and alignment are not required. These advantages make the wrist a suitable sensor location for long-term measurement of gait in real-world situations. However, the association of gait activity with wrist movement is more challenging than that of the upper body or lower limbs. The wrist may have "independent" gait movements (e.g., carrying a bag, hands in pockets) and "independent" non-gait movements (e.g., wrists moving while sitting or standing), which can lead to many problems for accurate gait detection under normal conditions.
[0005] Nowadays, technological advances and the increasing need and interest in long-term monitoring of physical activity mean that a large number of activity trackers, especially smart watches, have emerged on the market for consumers. They provide a non-invasive and easy way to track physical activity under free-living conditions. However, their reliability and validity are not fully understood. Some studies have shown a significant drop in the performance of such commercial products under different conditions (errors of up to 50%).
[0006] A major and challenging issue regarding gait detection is the effectiveness of existing methods in real-world or unsupervised conditions. In fact, most methods have been validated only under supervision (e.g., in a controlled laboratory environment, involving short-term measurements in a confined space) or semi-supervised conditions (e.g., a series of activities lasting longer, under the supervision of an observer to simulate real-life situations). Such conditions may not match real-world situations, where gait activities are context-dependent, spontaneous, and purposeful. In addition, in daily life, gait may be disturbed by "stationary" movements (e.g., shuffling around, swaying, stepping in place, turning, moving in place, jumping), unpredictable PA sequences, large imbalances in the data (gait periods are significantly less frequent than non-gait periods), and the presence of other types of movement (e.g., cycling, fitness, playing sports). It has been shown that the performance of PA classification is significantly degraded when validation is performed in real-life situations. Only a few products have evaluated their methods in fully real-world conditions without any supervision or predefined PA sequences. Those products used a camera carried by the subject as a reference to label the PA in free-living conditions. Summary of the invention
[0007] An object of the present invention is to overcome at least some of the above-mentioned disadvantages of existing PA detection sensors. More specifically, the present invention aims to provide an improved gait detection method and sensor that can be worn on a user's wrist.
[0008] According to a first aspect of the present invention, a method of detecting a gait of an individual is provided.
[0009] For example, the present method uses an accurate and precise algorithm to detect gait bouts in completely free-living conditions using a single triaxial accelerometer on the wrist. Biomechanically derived features are extracted to characterize real-life gait, using a probabilistic estimator such as a Naive Bayes classifier, followed by one or optionally two physically meaningful post-classification procedures to optimize performance. The solution provides a low-power, calibration-free algorithm that requires low computational effort to implement, for example, inside a wristwatch suitable for providing online feedback to the user in everyday situations. The algorithm has been validated in real-world conditions for healthy young and elderly people against an accurate and pre-validated wearable system.
[0010] According to a second aspect of the present invention, a gait detection sensor is provided.
[0011] According to a third aspect of the present invention, there is provided a wristwatch comprising a gait detection sensor according to the second aspect of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] Other features and advantages of the invention will become apparent from the following description of non-limiting example embodiments made with reference to the accompanying drawings, in which:
[0013] Figure 1 is a simplified block diagram illustrating an example gait detection sensor system according to an embodiment of the present invention;
[0014] Figure 2 schematically illustrates how a wrist posture may be defined in the context of the present invention;
[0015] Figure 3 is a diagram illustrating a definition for time-based probability modification, which is used in a gait detection process according to the present invention;
[0016] Figure 4 is a diagram illustrating a definition for intelligent decision making, which may be used in a gait detection process according to the present invention; and
[0017] Figure 5 is a flow chart outlining the method steps of an example process of gait detection according to the present invention. DETAILED DESCRIPTION
[0018] Embodiments of the present invention will now be described in detail with reference to the accompanying drawings. The present invention will be described in the context of a gait detection system using a wrist accelerometer. However, the teachings of the present invention are not limited to this environment or application. For example, the accelerometer need not be worn on the wrist. The same reference numerals are assigned to the same or corresponding functional and structural elements that appear in different figures. As used herein, "and / or" means any one or more of the items connected by "and / or" in a list. For example, "x and / or y" means any element in a ternary set {(x), (y), (x,y)}. In other words, "x and / or y" means "one or both of x and y". As another example, "x, y, and / or z" means any element in a septet set {(x), (y), (z), (x,y), (x,z), (y,z), (x,y,z)}. In other words, "x, y, and / or z" means "one or more of x, y, and z". In addition, the term "include" is used as an open term in this article. This means that the subject matter covers all the listed elements, but may also include additional unmentioned elements. Therefore, the word "comprising" is interpreted in a broader sense as "including", "containing" or "comprising".
[0019] Figure 1 A block diagram of the proposed sensor system 1 is shown, which is also referred to as sensor for short. In this example, the sensor system is designed to be worn on the wrist. Figure 1The sensor 1 shows six functional blocks or units, as explained in more detail later, namely an accelerometer 3, a segmentation unit 5, a feature extraction unit 7, a probability estimator 9 (a Bayesian estimator in this example), a time-based probability modification unit 11, and an intelligent decision making unit 13. Instead of using a Bayesian estimator as a probability estimator, any suitable probability estimator may be used instead, such as a decision tree, an artificial neural network, a support vector machine, etc. It is also noted that the operations of these functional units may be integrated into fewer than six units. For example, the different functionalities of these blocks may all be integrated into the accelerometer 3, which in this example is configured to measure acceleration in a three-dimensional (3D) space. The 3D accelerometer signal ( A x ,A y ,A z ) is fed into a segmentation unit 5, which is configured to sample and segment the accelerometer signal. The segmented signal is then fed into a feature extraction unit 7, which is arranged to extract relevant features from the segmented accelerometer signal. The feature vector H [ n ] (where n represents the window number resulting from the segmentation) is fed into a Bayesian estimator 9 which is arranged to estimate the probability of a gait occurring P Bayes [ n ], and the Bayesian estimator 9 is trained by extracting features from the training data set and at least some of their corresponding labels (gait occurrence or non-gait occurrence). In estimation and decision theory, a Bayesian estimator or Bayesian action is an estimator or decision rule that minimizes the posterior expectation of a loss function (i.e., the posterior expected loss). Equivalently, one can consider maximizing the posterior expectation of a utility function. An alternative way to formulate an estimator within Bayesian statistics would be maximum a posteriori estimation. An estimator can be understood as a rule for calculating an estimate of a given quantity based on observed data. Therefore, one can distinguish between a rule (estimator), a quantity of interest (the quantity being estimated), and its result (the estimator).
[0020] Before applying the algorithm proposed below, the Bayesian estimator is advantageously trained. For training, other algorithms and sensors than those proposed in the present invention are advantageously used. Training data are collected from a group of individuals so that the Bayesian estimator can then be trained by following, for example, the principle of cross-validation.
[0021] Then the probability of gait occurrence is P Bayes [ n] is fed into a time-based probability modification unit 11, which is configured to utilize the temporal information of the activities detected in the past (in this example, it comes from the training data set) to modify the gait occurrence probability based on the histogram of gait duration in real life situations. In this context, the histogram can be understood as an estimate of the probability distribution for gait duration or non-gait duration. The obtained probability value is then P T [ n ] is fed into an intelligent decision making unit 13, which is configured to use intelligent rules to give a classification or label "gait" or "non-gait" according to the probability obtained from the previous step. In the following, a more detailed description of the data processing steps is provided.
[0022] segmentation
[0023] The segmentation unit 5 first samples the received continuous time domain acceleration signal. It should be noted that the various acceleration signals are advantageously sampled by using the same sampling frequency, for example between 100 Hz and 1000 Hz. In this particular example, a sampling frequency of 200 Hz is used. Then, the segmentation unit 5 uses a moving window with a given duration (6 seconds in this particular example) and a given time overlap (5 seconds in this particular example) to generate a segmented wrist acceleration signal ( A x [ n ] ,A y [ n ] ,A z [ n ]),in n Refers to the window number. It was found through experiments that a window length of 6 seconds optimizes performance. Therefore, the window length is advantageously between 1 second and 10 seconds, or more specifically between 4 seconds and 8 seconds. This amount of data is short enough to have the required time resolution, but long enough to have sufficient data for frequency analysis.
[0024] Feature extraction
[0025] Several features such as strength, periodicity, posture, and noisiness were defined or identified based on the biomechanics of wrist movement in order to highlight the essential differences between gait and non-gait bouts from the perspective of wrist movement. The least absolute shrinkage and selection operator (LASSO) feature selection method according to "Feature Selection for Classification: A Review by J. Tang, S. Alelyani, and H. Liu" (Data Classif. Algorithms Appl., page 37, 2014) was used to specify the best possible set of features to optimize the performance of the training dataset. Interestingly, LASSO selected a set of features that covered all the biomechanical criteria used to define the features (i.e., strength, periodicity, posture, noisiness). In total, 13 features were selected in four categories as follows:
[0026] Strength-based features : A key difference between gait and non-gait cycles is the strength of the wrist acceleration signal. To extract this information, the following features are calculated:
[0027] NI [ n ]: This is the strength of the acceleration norm calculated according to formula 1.
[0028] , (1)
[0029] in, SA [ f i ] is the magnitude of the spectral acceleration norm calculated according to equations 2 and 3. To estimate the spectrum, a Blackman windowing is used N point fast Fourier transform (FFT), where N is the number of samples in a time window (i.e., in this case N =1200). In addition, f i refers to the frequency resolution of the method and is shown in Equation 4. A logarithmic function is used to reduce the range of this feature, as well as the thick tail of its histogram, which is suitable for further Bayesian modeling.
[0030] (2)
[0031] (3)
[0032] (4).
[0033] MeanA [ n ]: This is the average value of the acceleration norm within a time window.
[0034] Features based on periodicity : Considering the cyclic nature of gait, five features related to the periodicity of acceleration signals are included, as follows:
[0035] NACFmax [ n ]: The autocorrelation function of the acceleration norm is calculated and normalized to the first sample (i.e., the zero-lag sample). Its maximum peak is then reported for each window NACFmax , excluding zero lag samples. Therefore, this feature can be called the maximum peak of the normalized autocorrelation function of the acceleration norm.
[0036] NACFp2p [ n ]: This feature is the peak-to-peak value of the largest peak and valley of the normalized autocorrelation function, excluding the zero lag sample. In other words, this feature is the peak-to-peak value of the normalized autocorrelation function of the acceleration norm.
[0037] SAmax [ n ]: According to formula 5, use SA [ f i ] to estimate the normalized spectrum of the acceleration norm (NSA), and its maximum peak amplitude is calculated as SAmax [ n ]. Therefore, this feature is the maximum peak of the acceleration norm spectrum.
[0038] (5).
[0039] DomSAmax [ n ]: Design a score that shows how clear the maximum peak of the NSA is compared to its neighboring samples. This score can be called the clarity of the maximum peak of the NSA. This feature is calculated according to Formula 6, where f max , f max-1 and f max+1 They refer to the frequencies of the maximum peak of NSA, the previous sample, and the next sample, respectively. Therefore, this feature is the clarity of the maximum peak of the frequency spectrum of the acceleration norm.
[0040] (6).
[0041] Cad [ n]: Not only the periodicity of the signal, but also the value of the period itself is an important piece of information to distinguish between gait cycles and non-gait cycles. Usually, the stride frequency (i.e., the fundamental frequency of the gait signal) is limited to a short range of around 1 Hz (i.e., between 0.3 Hz and 2.5 Hz). The fundamental frequency of the acceleration signal is calculated using the algorithm described in the following document: "A wrist sensor and algorithm to determine instantaneouswalking cadence and speed in daily life walking" by B. Fasel et al. (Med. Biol. Eng. Comput., Vol. 55, pp. 1773–1785, 2017). Therefore, this feature is the cadence or step frequency (steps / min) of the gait.
[0042] Pose-based features : During gait (i.e., running or walking), the wrist usually has a specific predictable posture, while during non-gait cycles, the posture is mostly unpredictable and irregular. Therefore, extracting information about wrist posture should be useful for gait bout detection. θ Defined as the accelerometer on the wrist y Axis and Earth's horizontal plane x Global , y Global > (by the Earth coordinate system x Axis and y The plane formed by the axes and perpendicular to the gravity vector, see Figure 2 ). Assuming the sensor can only rotate around the wrist, the sensor y The axis is almost aligned with the longitudinal axis of the wrist. Figure 2 , and if the dynamic acceleration of the wrist movement remains low, the gravity vector is y The axis and the sensor x Axis and z The plane defined by the axis (< x , z >) is:
[0043] (7)
[0044] , (8)
[0045] in, g represents the acceleration due to gravity, and A <x,z> [n ] is for window n On the plane x , z >The magnitude of the acceleration vector obtained above. Therefore, the angle can be estimated by formula 9 θ [ n ].
[0046] (9).
[0047] The proposed pose-based (i.e., wrist pose) feature is defined as:
[0048] (10).
[0049] Noisiness-based features : During gait, the acceleration signal is quasi-cyclical in nature, where the energy is mainly in low frequency bands, whereas during non-gait periods, the signal is rather random and non-stationary, where the energy is distributed in larger frequency bands. Hence, like the signal-to-noise ratio (SNR), the gait / non-gait power ratio is expected to be higher during gait than non-gait periods. Therefore, several features are designed to capture information about the level of “noise” (i.e. non-gait periods) in the signal (i.e. gait periods). These features are as follows:
[0050] HLR [ n ]= High-low frequency ratio : The ratio between the intensity exhibited at high frequency and at low frequency is defined as a feature according to formula 11. The frequency threshold is set to 3.5 Hz by experiment to optimize performance. However, other frequency thresholds can be used instead. The threshold is advantageously between 1 Hz and 10 Hz, or more specifically between 2 Hz and 5 Hz. Therefore, the feature is the high-low frequency ratio of the energy of the acceleration norm.
[0051] (11).
[0052] ZCR [ n ]: Due to the noisy and erratic nature of wrist movements, the rate of zero crossings in the acceleration norm during non-gait periods should be higher. First, remove the mean of the acceleration norm within the time window. Then, discard any linear trend in the resulting signal, for example by using the "detrend" function in MATLAB®. Finally, count the number of zero crossings and use it as a feature ZCR [ n ]. Therefore, this feature is the zero-crossing rate of the acceleration norm after removing the mean.
[0053] SEF [ n]: As shown in Equation 12, the spectrum edge frequency is estimated as the frequency below which the signal energy is observed α (%), “A comparison of median frequency, spectral edge frequency, a frequency bandpower ratio, total power, and dominance shift in the determination of depthof anesthesia by JC Drummond, CA Brann, DE Perkins, and DE Wolfe” (Acta Anaesthesiol. Scand., vol. 35, no. 8, pp. 693–699, 1991). We found that α = 70 (%) provides the best performance in this application. This feature can be called the spectral edge frequency of the acceleration norm.
[0054] (12).
[0055] RandA [ n ]: Assuming that the wrist acceleration signal is less random during gait than during non-gait periods, it is defined according to the autocorrelation-based test introduced in the following literature RandA [ n ] feature to measure the randomness of the signal: "Introduction to time series and forecasting by PJ Brockwell, RA Davis and MV Calder (Volume 2, Springer, 2002)". According to this test, if the time series comes from a stationary random process (which is almost the case for the acceleration norm of non-gait cycles over a short window of 6 seconds), then the samples of the autocorrelation of the time series will be mainly confined to Thresholds between N is the number of samples in a time window (i.e. 1200). RandA [ n ] is defined as The percentage of autocorrelated samples that are outside the range of . RandA [ n ], the higher the value of , the less random the signal is. Therefore, this feature can be defined as the randomness score of the acceleration norm.
[0056] Kurtosis A [ n ]: Kurtosis is a well-known tool to measure the degree to which a data distribution is prone to outliers, "P. H. Westfall, "Kurtosis as peakedness, 1905–2014. RIP" (Am. Stat., vol. 68, no. 3, pp. 191–195, 2014). It is assumed that the acceleration norm of non-gait periods contains more outliers than gait due to higher signal randomness. Therefore, the kurtosis of the acceleration norm within a time window is calculated as a feature.
[0057] Finally, for each time window n , H [ n ] is constructed as a feature vector including all selected features. However, it will be possible to include only some of the above features in the feature vector, or to include some other features in addition or alternatively.
[0058] Bayesian Estimator
[0059] The gait occurrence probability for each window is estimated by using the Bayesian estimator according to Formula 13:
[0060] , (13)
[0061] in, is the observed eigenvector H [ n ] is the probability of the gait occurrence condition on . In addition, and They are respectively in the gait (G) classification and the non-gait (NG) classification. H [ n ] is the probability of occurrence. In addition, P G and P NGare the prior probabilities of gait and non-gait occurrence, respectively. The multivariate multinomial distribution of the Bayesian estimator (“mvmn” in MATLAB®) is considered. In addition, in order to manage the inherent imbalance of samples between gait and non-gait periods (in real-world situations, non-gait samples are relatively more frequent than gait samples), the Laplace smoothing parameter is utilized in the calculation of the prior probability, “Smoothing multinomial naïve bayes in the presence of imbalance” by AY Liu and CE Martin (In International Workshop on Machine Learning and Data Mining for Pattern Recognition, 2011, pp. 46–59), as follows:
[0062] (14)
[0063] , (15)
[0064] in, N NG and N G is the total number of samples or windows observed for non-gait and gait periods. l It is set based on experience ( N G + N NG ) / 10.
[0065] Time-based probability modification
[0066] Use information from past detected activities to increase the certainty of decisions made about current activities. Figure 3 As shown in q [ n -1] and d [ n -1] are defined as until the window n -1 The type and duration of the last activity detected (i.e., the last activity was from the window n - d [ n -1] and continues until the window n -1, window n -1 inclusive), and Indicates that the type of the previous activity is known ( q [ n -1]) and duration ( d [ n -1]) in the window nhas the same activity as in the previous window (i.e., q [ n ]= q [ n -1]). To this end, two exponential functions (see Formulas 16 and 17) are fitted to the probability density function of the duration of gait and non-gait bouts in daily life obtained from the training phase of the method. Therefore, the parameters in Formulas 16 and 17 (i.e., β G , γ G , τ G etc.). Then, because the probability given by Bayes ( P Bayes ) is usually better than is more reliable, so by using Equation 18 Mapped to a shorter range of [0.05, 0.20] (adjusted experimentally) to reduce The modification effect of However, other suitable ranges may be used instead. Finally, the time window is calculated by using Equation 19 n The modified gait occurrence probability ( P T ), where the term “ min-max " function limits the probability to the range [0, 1].
[0067] (16)
[0068] (17)
[0069] (18)
[0070] , (19)
[0071] in, The definition is as follows:
[0072] (20).
[0073] Intelligent decision making
[0074] when P T [ n ]When it is far enough from 0.5, it is easy to make a decision window n is gait. PT [ n ] is close to 0.5, making decisions becomes challenging. P T [ n ] close to 0.5 may occur near the transient state between activities because part of the feature window is gait and the other part is non-gait. Therefore, the following algorithm is designed to be based on P T [ n ] to make intelligent decisions. P T [ n ]<0.3 and P T [ n ]>0.7, the decisions are NG and G respectively. Here, the decision refers to the n mark L [ n ]. These windows can be called reliable windows. On the other hand, if (called a fuzzy window or uncertainty region), then examine the time period between the last reliable window and the next reliable window, or more specifically, its duration. For example, for a fuzzy window n ,window m and k are the previous and next reliable windows respectively ( m < n < k , see Figure 4 ).if , then the decision threshold is changed from the conventional 0.5 to Otherwise, if P T [ n ]>0.6 or P T [ n ]<0.4, the decisions are G and NG respectively, and for , the last reliable decision is assigned to the window n (Right now, L [ n ]= L [ m ]). It should be noted that other suitable parameter values can be used to replace the above parameter values. The following algorithm briefly explains the procedure of the proposed intelligent decision making.
[0075]
[0076] (START starts, END ends, IF if, THEN then, ELSE otherwise).
[0077] exist Figure 5 The above process is outlined in the flowchart of . In step 101, the accelerometer 3 measures to obtain one or more acceleration signals, which are continuous signals in the time domain. In step 103, the segmentation unit 5 samples these continuous time domain signals, and in step 105, segments them. In step 107, the feature extraction unit 7 extracts a set of feature vectors from the acceleration signals segmented for the corresponding window. In step 109, the Bayesian estimator 9 uses the extracted feature set and determines the gait probability occurrence value for the corresponding window. In step 111, the time-based probability modification unit 11 modifies the probability value obtained in step 109 to improve the accuracy of the probability value. In step 113, the intelligent decision making unit 13 marks the modified probability value obtained in step 111. The modified probability value is marked as "gait" or "non-gait". It is noted that the intelligent decision making unit 13 is optional and labels may be derived directly from the modified probability values, e.g. such that (modified) probability values greater than 0.5 will be interpreted as “gait”, while other probability values will be labeled as “non-gait”.
[0078] The principles of an accurate and precise window-based algorithm are explained above to identify gait bouts and estimate their duration in an unsupervised real-world situation using a single low-power accelerometer mounted on the wrist. The probability density functions of the biomechanically derived features (i.e., in terms of intensity, periodicity, posture, and noisiness) illustrate the high ability of the selected features to distinguish between gait and non-gait bouts. The LASSO score shows that in this application, NACFmax , SAmax , NI and WristPost is the best or most meaningful feature among them. In addition, periodicity is a better criterion for distinguishing gait and non-gait bouts.
[0079] By using only one low-power accelerometer, optimizing feature computations and using some implementation techniques, an optimized implementation of the proposed sensor shows very low power consumption (135.5mAh per year) under real-world conditions. The implemented method provides approximately one year of continuous valid gait measurement with a single common battery cell (250mAh). This is a great advantage, as many medical and sports applications critically require long-term measurements of physical activity in real-life situations. The simplicity of the proposed method and its low computation time (1 millisecond per window) provide the possibility of real-time and on-board analysis of physical activity, which provides the possibility of generating real-time feedback, which can be important in many applications, such as intervention planning.
[0080] The proposed method has been validated in unsupervised everyday situations for young and elderly people, offering great potential for use in clinical settings for monitoring patients with limited mobility due to disease. For example, the system can be used in larger groups of elderly people to analyze the impact of various factors (such as aging, obesity, and frailty) on the quality and quantity of physical activity in everyday life situations. More importantly, the proposed method can be used as a preliminary stage for many physical activity analysis algorithms where accurate detection of gait bouts is required, such as gait cadence estimation and gait speed estimation.
[0081] In summary, in the above examples, the present invention presents an accurate and precise method for detecting gait bouts in free-living situations. Biomechanically derived features are combined with a naive Bayes classifier, followed by two physically meaningful post-classification steps to cope with the difficulties posed by wrist movements in challenging real-world situations. Such a wrist-based low-power and calibration-free system (sensor-to-body alignment does not require a calibration phase) provides a versatile measurement tool with high usability and autonomy, which is ideal for long-term monitoring of physical activity in free-living situations. In addition, the simplicity and real-time nature of the proposed method allow the method to be implemented inside a wristwatch, which protects the user's privacy. This also provides the possibility of giving online meaningful feedback to the user in daily life to promote a more active lifestyle.
[0082] Although the present invention has been illustrated and described in detail in the drawings and the foregoing description, such illustration and description should be considered illustrative or exemplary rather than restrictive, and the present invention is not limited to the disclosed embodiments. Based on a study of the drawings, the disclosure, and the appended claims, those skilled in the art will understand other embodiments and variations and be able to implement them when implementing the claimed invention.
[0083] In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. The mere fact that different features are recited in mutually different dependent claims does not indicate that a combination of these features cannot be used to advantage. Any reference signs in the claims should not be construed as limiting the scope of the invention.
Claims
1. A method for detecting the gait of an individual by using a gait detection sensor worn by the individual, the gait detection sensor comprising an accelerometer and a processing unit, the method comprising: The accelerometer obtains an acceleration signal representing one or more sensor acceleration values; The processing unit samples the acceleration signal to obtain a sampled acceleration signal; The processing unit divides the sampled acceleration signal into a given number of windows with a given duration to obtain a divided acceleration signal; A processing unit extracts a feature set from the segmented acceleration signal, wherein the feature set characterizes the acceleration signal; The processing unit determines a probability value for a corresponding window n using the extracted feature set by using a trained probability estimator, wherein n is a positive integer greater than zero, the probability value giving an estimated probability value of the occurrence of a gait of the individual during the corresponding window; The processing unit modifies the estimated probability value by using a histogram of previously detected gait durations to obtain a modified probability value, the histogram including an estimated value of a probability distribution for a gait duration or a non-gait duration; as well as The processing unit determines whether the corresponding window represents a gait occurrence based on the modified probability value and by using a determination threshold.
2. The method according to claim 1, wherein: The probability estimator comprises a Bayesian estimator.
3. The method according to claim 1, wherein: The extracted feature set includes one or more biomechanical features in at least one of the following feature categories: strength of the acceleration signal, periodicity of the acceleration signal, posture of the individual at a sensor location, and noisiness of the acceleration signal.
4. The method according to claim 3, wherein: The one or more biomechanical features are at least one of the following features: the strength of the acceleration norm, the average of the acceleration norm, the maximum peak of the normalized autocorrelation function of the acceleration norm, the peak-to-peak value of the normalized autocorrelation function of the acceleration norm, the maximum peak of the spectrum of the acceleration norm, the clarity of the maximum peak of the spectrum of the acceleration norm, the step frequency of the gait, the posture of the individual at the sensor location, the ratio of the energy of the high frequency to the low frequency of the acceleration norm, the zero crossing rate of the acceleration norm after removing the average, the spectrum edge frequency of the acceleration norm, the randomness score of the acceleration norm, and the kurtosis of the acceleration norm.
5. The method according to claim 1, wherein: The segmented acceleration signal includes three substantially orthogonal acceleration components (A x [n],A y [n],A z [n]).
6. The method according to claim 1, wherein: The method further comprises training the probability estimator by using a training data set before determining the estimated probability value.
7. The method according to claim 1, wherein: The modified probability value is obtained as: Where P Bayes [n] is the estimated probability value of the individual's gait occurring during the corresponding window, is the probability of having the same gait activity in window n as in the previous window n-1, which ranges from less than 1, d represents the activity duration, q represents the activity type, and 8. The method according to claim 1, wherein: If the modified probability value P T [n] is determined to be in the ambiguity region such that a ≤ P T [n] ≤ (1 - a), then the processing unit checks the duration of the period between the previous non-ambiguous window m and the next non-ambiguous window k such that m < n < k, in the non-ambiguous window P T [n] > 1 - a or P T [n] < a, then if k - m + 1 ≤ b, where b is a positive integer, then the determination threshold is changed from 0.5 to 1 - mean(P T [m < n < k]).
9. The method according to claim 8, wherein: a is between 0.1 and 0.3, and b is between 5 and 15.
10. The method according to claim 8, wherein: a is between 0.15 and 0.25, and b is between 8 and 12.
11. The method according to claim 8, wherein: If the condition k-m+1≤b is not satisfied, the determination threshold is set to 0.5+c for gait occurrence and to 0.5-c for non-gait occurrence, so that if P T [n]>(0.5+c), then gait is determined to occur, and if P T [n] < (0.5-c), then non-gait is determined, and for (0.5-c) ≤ P T [n]≤(0.5+c), assign the last unambiguous decision to window n.
12. The method according to claim 11, wherein: c is between 0.02 and 0.
2.
13. The method according to claim 11, wherein: c is between 0.05 and 0.
15.
14. The method according to claim 1, wherein: The accelerometer is worn on the individual's wrist.
15. The method according to claim 1, wherein: The determination is made based on an acceleration signal from only a single accelerometer.
16. A gait detection sensor for detecting the gait of an individual, the gait detection sensor comprising an accelerometer and a processing unit, the gait detection sensor comprising means for: Obtaining an acceleration signal representing one or more sensor acceleration values; Sampling the acceleration signal to obtain a sampled acceleration signal; · segmenting the sampled acceleration signal into a given number of windows with given durations to obtain segmented acceleration signals; Extracting a feature set from the segmented acceleration signal, wherein the feature set characterizes the acceleration signal; Determine a probability value for a corresponding window n using the extracted feature set by using the trained probability estimator, wherein n is a positive integer greater than zero, the probability value giving an estimated probability value of the occurrence of a gait of the individual during the corresponding window; modifying the estimated probability value by using a histogram of previously detected gait durations to obtain a modified probability value, the histogram comprising an estimated value of a probability distribution for a gait duration or a non-gait duration; and Determining whether the corresponding window represents a gait occurrence based on the modified probability value and by using a determination threshold.
17. A wristwatch comprising the gait detection sensor according to claim 16.
