Walking speed estimation device

Abstract

To provide a device which estimates a more accurate walking speed by measuring a walking state of a subject person in a multifaceted manner and analyzing a measurement result thereof.SOLUTION: A walking speed estimation device 100 includes: measurement means 110 which measures an acceleration speed and a walking speed at a plurality of locations of a walking body to be used as a data set; deep layer learning means 140 which generates predicted walking speeds at a plurality of locations by performing deep learning of the data set; error calculation means 150 which compares the measured walking speed and the predicted walking speed to calculate errors at the plurality of locations; determination / selection means 160 which ranks the predicted walking speeds at the plurality of locations on the basis of magnitude of the errors and selects at least two predicted walking speeds having smallest errors; and estimation means 170 which estimates the walking speed by an inference engine for calculating the selected predicted walking speed.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a walking speed estimation device. [Background technology]

[0002] Evaluation of a person's walking state is important as an index of healthy life expectancy. For this reason, various technologies have been disclosed so far as technologies for estimating walking speed, which is one of the evaluations of walking state, such as a technology that calculates walking speed using an angular velocity sensor attached to the person's thigh (Patent Document 1, Non-Patent Document 1), a technology that observes a person's walking cycle using a mobile terminal (Patent Document 2), a technology that acquires data using a smartphone and further performs deep learning on this data to predict displacement (Patent Document 3, Non-Patent Document 2), and a technology that obtains a pedestrian's behavioral intention and behavioral pattern based on map information and positioning data (Patent Document 4). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2016-214377 [Patent Document 2] Patent Publication No. 2021-3605 [Patent Document 3] Patent No. 2022-78967 [Patent Document 4] Patent Publication 2022-2030 [Non-patent literature]

[0004] [Non-Patent Document 1] Akihiro Suzuki, Kenji Ishii, Ryoichi Nagatomi, "Estimation of walking speed and stride length using lumbar body acceleration - walking and Nordic walking -", Biomedical Engineering, Vol. 49, No. 6, 2011, pp. 957-961https: / / www.jstage.jst.go.jp / article / jsmbe / 49 / 6 / 49_6_957 / _article / -char / ja / [Non-Patent Document 2] Takuto Yoshida, Atsushi Nozaki, Satoshi Hiroi, Katsuhiko Kaji, Takuro Yonezawa, Nobuo Kawaguchi, "Walking Speed ​​Estimation Method Using Deep Learning for Gait-Robust PDR", IPSJ SIG Technical Report, Vol.2019-UBI-63 No.3, Vol.2019-ASD-15 No.3, 2019 / 8 / 26 Summary of the Invention [Problem to be solved by the invention]

[0005] A person's walking state varies depending on not only people with disabilities, but also on habits, muscles, bones, and even the use of a walking stick to assist walking, and it cannot be assumed that everyone has a uniform walking state. Even if the acceleration of parts of the body while walking is obtained (Patent Documents 1, 2, 3, Non-Patent Documents 1, 2, etc.), it is difficult to estimate the subsequent walking speed.

[0006] Furthermore, even if the behavioral intention of a pedestrian is analyzed based on positioning data (Patent Document 4), it is difficult to estimate the walking speed in accordance with the situation that the pedestrian encounters at all times.

[0007] The present invention has been made in consideration of the above-mentioned background and provides a device that measures the walking state of a target person from multiple angles and analyzes the measurement results to estimate the walking speed with higher accuracy. [Means for solving the problem]

[0008] The present invention is a device for estimating a walking speed. One aspect of the device can be configured as follows. That is, the device includes a measurement means for setting a measurement time point, setting a pre-measurement period and a post-measurement period of a predetermined time interval before and after the measurement time point, measuring the acceleration and walking speed of a plurality of points of a walking body during the pre-measurement period and the post-measurement period, and setting both or one of a calculated walking speed and the walking speed calculated from the acceleration during the pre-measurement period as a pre-data set, and both or one of a calculated walking speed and the walking speed calculated from the acceleration during the post-measurement period as a post-data set, a deep learning means for generating a predicted walking speed during the post-measurement period at the plurality of points by deep learning the pre-data set, an error calculation means for comparing the post-data set with the predicted walking speed and calculating an error at the plurality of points, a determination / selection means for ranking the predicted walking speeds at the plurality of points based on the magnitude of the error, and selecting at least two predicted walking speeds with the smallest error, and an estimation means for estimating a walking speed after the post-measurement period by an inference engine that calculates the selected predicted walking speed.

[0009] In this configuration, the measuring means first sets a measurement time point and measures the walking speed of the subject pedestrian before and after the measurement time point. For the measurement, the acceleration of multiple parts of the pedestrian, such as the neck, waist, thighs, heels, and other parts that move a lot when walking, is measured by a sensor or the like worn by the subject, and the walking speed is obtained by integrating the component of this acceleration in the walking direction. In addition, the walking of the pedestrian can be captured by a video camera or the like, and the video can be analyzed (motion capture) to obtain the walking speed. The measuring means also creates a time-series data set of the walking speed obtained here.

[0010] Next, the deep learning means performs deep learning using the previous data set before the measurement time point as training data to obtain a trained model for estimating the walking speed. Then, the error calculation means calculates the error between the actual walking speed, which is the subsequent data set after the measurement time point, and the walking speed estimated by the trained model. The judgment and selection means ranks the predicted walking speeds at the multiple locations based on the magnitude of the error, and selects at least two of the predicted walking speeds with the least error. As a result, the trained model with the least error, in other words, with the highest estimation accuracy, is selected as the inference engine. Such an inference engine becomes the estimation means, and becomes a device for estimating the walking speed of a pedestrian.

[0011] According to this configuration, the measurement means measures the acceleration and speed of multiple points on the body of a pedestrian, and is therefore able to select a more reliable data set of walking speed, independent of the walking characteristics of each pedestrian. Next, the deep learning means learns and trains using these data sets, thereby estimating the walking speed at each point, and is able to compare the trained model with the actual measured value as the same variable of walking speed. Then, the determination and selection means calculates the error between the estimated walking speed and the actually measured walking speed, and based on this error, selects a trained model with higher accuracy, and this trained model can be used as an inference engine.

[0012] In this way, this configuration can provide a device that estimates walking speed with higher accuracy by applying an inference engine based on acceleration and speed measurement points in accordance with the walking characteristics of a pedestrian, compared to Patent Documents 1, 2, 3 and Non-Patent Documents 1 and 2, which acquire the acceleration of a part of the body while walking. Also, it is possible to estimate a walking speed that always corresponds to the situation that the pedestrian encounters, without analyzing the behavioral intention of the pedestrian based on positioning data (Patent Document 4).

[0013] In the above configuration, the deep learning can be configured to use the front data set as a training data set and the rear data set as a test data set.

[0014] In deep learning, learning is performed using a learning / training dataset (Train), and the learned model is confirmed using a test dataset (Test). In this configuration, the test dataset is selected from the subsequent dataset. With this configuration, if the pedestrian's walking has changed significantly since the time of measurement, the judgment / selection means can select a more accurate inference engine. This configuration is also effective when it is difficult to obtain a large number of datasets, and is also effective when estimating walking speed in real time.

[0015] In the above configuration, the deep learning may be configured to include a training data set and a test data set in the previous data set. This configuration can, for example, periodically observe walking states, store several trained models and inference engines selected by a judgment and selection means, and appropriately apply these inference engines according to the situation and state of a pedestrian.

[0016] In the above configuration, the judging / selecting means can calculate a previous period from the change in acceleration during the previous measurement period at multiple points on the body, calculate a later period from the change in acceleration during the later measurement period, calculate a period difference between the previous period and the later period, and select the predicted walking speed by the inference engine based on a data set at at least one or more points where the period difference is small.

[0017] Pedestrians do not always walk steadily, and may lose their walking rhythm due to changes in physical condition, tripping over obstacles, avoiding obstacles, losing balance, etc. People walk on two legs, and in this case, the acceleration and velocity at multiple measurement points on the body that change periodically during normal walking disrupts the periodicity, and the disruption is particularly noticeable at the heels and thighs. This configuration can assist in determining whether it is more preferable to use one of the selected inference engines for the walking state by calculating the periodic difference. This configuration is particularly suitable for estimating walking speed in real time.

[0018] In the above configuration, the deep learning may be configured to be a goal recurrent neural network (RNN). Walking speed is a variable with respect to time, and it is preferable to apply an RNN. That is, an RNN (Recurrent Neural Network) is a recurrent neural network with a recurrent connection in which the output of a certain layer is input retroactively, and the state at each time has information on the input at the previous time ("has memory"). Therefore, this structure makes it possible to express a long-period time series without stacking a large number of layers.

[0019] In the above configuration, the goal regression neural network may be configured to be a long short-term memory neural network (LSTM). LSTM (Long Short-Term Memory) is an RNN with a node having a specific structure, an LSTM block. By introducing this block, LSTM can efficiently learn the long-term dependency of a time series. The LSTM block has the following two characteristic structures. The first is called an input gate / output gate, which helps to efficiently learn the pattern by removing values ​​that deviate from the characteristic pattern of the target time series from the learning target. The second is called a forget gate, which helps the model to follow the state transition of the time series by making the state that has been memorized up to that point forgotten to some extent. In this way, with this configuration, time series data can be learned more reliably.

[0020] In the above configuration, the error calculation means may be configured to use either or both of a Root Mean Squared Error (RMSE) method and a Mean Absolute Percentage Error (MAPE) method.

[0021] Evaluation indices for deep learning regression models include mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAPE). Since the walking training machine needs to follow without delay, it is important to prevent large errors from occurring. Therefore, in this configuration, RMSE, which can emphasize large errors, is used as the evaluation index, and MAPE, which can calculate relative errors instead of absolute errors and is expressed as a percentage, is also applied, allowing for more accurate judgment and selection, and improving the reliability of the estimation engine. Effect of the Invention

[0022] The present invention can provide a device that measures the walking state of a target person from various angles and analyzes the measurement results to estimate the walking speed with higher accuracy. [Brief description of the drawings]

[0023] [Figure 1] 1 is an explanatory diagram showing a configuration of a walking speed estimation device according to an embodiment of the present invention. [Diagram 2] This is an example of a measuring means for measuring the walking speed of a pedestrian. [Diagram 3] FIG. 1 is an explanatory diagram of a goal regression neural network (RNN). [Figure 4] FIG. 1 is an explanatory diagram of a long short-term memory neural network (LSTM). [Diagram 5] FIG. 1 is an explanatory diagram of a time series data set. [Figure 6] 1 is a diagram showing an example of periodicity of acceleration at multiple points on a pedestrian's body measured by a measurement means; FIG. [Figure 7] FIG. 4 is a diagram showing an example of a root mean square error (RMSE) calculated by an error calculation means. [Figure 8] FIG. 13 is a diagram showing an example of a mean absolute error (MAPE) calculated by an error calculation means. [Figure 9] FIG. 13 is a diagram showing an example of an average error in the distance traveled between the measured values ​​and the estimated speed for each step of a pedestrian. [Figure 10] 1 is a table summarizing the results of the root mean square error (RMSE) and the mean absolute rate error (MAPE) at multiple points on a pedestrian's body in a validation test. [Figure 11] 1 is an example of an application of an embodiment of the present invention to a cane walking training machine. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0024] <Description of the embodiment> A first embodiment of the present invention will be described below with reference to Figs. 1 to 5. Fig. 1 is an explanatory diagram showing a configuration of a walking speed estimation device according to an embodiment of the present invention. Fig. 2 is an example of a measuring means for measuring a walking speed of a pedestrian. Fig. 3 is an explanatory diagram of a goal regression neural network (RNN). Fig. 4 is an explanatory diagram of a long short-term memory neural network (LSTM). Fig. 5 is an explanatory diagram of a time-series data set.

[0025] 1 and 2, a walking speed estimation device 100 according to one embodiment of the present invention includes a measurement means 110 that measures the acceleration and speed of a pedestrian's body at multiple locations and creates a walking speed dataset, a deep learning means 140 that performs deep learning using this dataset as learning / training data and generates a learned model, an error calculation means 150 that calculates the error between the walking speed estimated by the learned model and the actual walking speed, a determination / selection means 160 that ranks the predicted walking speeds at the multiple locations based on the magnitude of the error and selects at least two predicted walking speeds with the least error, and an estimation means 170 that estimates the pedestrian's walking speed using the selected learned model (inference engine).

[0026] In this embodiment, as an example, the measurement points for acceleration and speed by the measurement means 110 are six points in total, including the left and right heels 112, 112, the left and right thighs 114, 114, the waist 116, and the neck 118. Referring to Fig. 2, the walking speed of a walker is measured by acceleration sensors 120 attached to multiple points on the walker's body in Fig. 2(a) and a motion capture device 130 that captures images of walking in Fig. 2(b).

[0027] The acceleration sensor 120 is composed of left and right heel sensors 122, 122 of the walker 10, left and right thigh sensors 124, 124, a waist sensor 126, and a neck sensor 128. A general three-axis acceleration sensor can be applied as the acceleration sensor 120. The acquired acceleration is integrated and the walking direction component is extracted, so that it can be converted into walking speed.

[0028] Here, the acceleration data measured during bipedal walking has periodicity. If the measurement data does not show periodicity, there may be a problem with either the measurement or the pedestrian. These may be excluded because they are not suitable as a data set for deep learning.

[0029] The motion capture 130 is a technology for digitally recording the walking motion of an actual pedestrian, and in this embodiment, a general optical motion capture 130 is applied. That is, as shown in FIG. 2(b), the state in which the pedestrian 10 walks along the walking path L from the Start to the End is captured by a plurality of cameras 132, and this motion is digitized. The position of the marker attached to the pedestrian 10 is not particularly limited, but it may be related to the mounting position of the acceleration sensor 120, or a position with higher measurement accuracy may be selected based on the results of an experiment or the like. The motion capture 130 is not limited to the optical type, and an inertial sensor type, a mechanical type (potentiometer, encoder), or the like may be appropriately selected.

[0030] 5, the walking speed data acquired and calculated by the acceleration sensor 120 and the motion capture 130 for a total of six locations, namely the left and right heels 112, 112, the left and right thighs 114, 114, the waist 116, and the neck 118, in the measurement means 110, is made into a time-series data set. This data set is created as the speed at times Δt0, Δt1, Δt3... from Δt hours before each of the measurement times t0, t1, t2.... Each data set R is sent to the deep learning means 140.

[0031] The deep learning means 140 performs deep learning using these data sets R as learning and training data. A convolutional neural network (CNN), which is well known as a deep learning means and was invented based on the network structure of the human brain, cannot handle time series data such as stock price fluctuations and precipitation well. Therefore, in this embodiment, a goal regression neural network (RNN) suitable for time series data is applied. As shown in FIG. 3, the RNN is a repetition of neural network modules, in which not only the current time series value but also the output of the previous intermediate layer is used as the current input.

[0032] When expanded in the time domain, RNNs can be seen as static neural networks, and in theory can express long-term dependencies, so intermediate layer states going back far from a certain time can also be used for learning. However, in practice, when the time series becomes long, weights are applied and the gradient disappears or explodes, which can lead to problems with learning.

[0033] Therefore, in this embodiment, further consideration has been made and a long short-term memory neural network (LSTM), which is a type of goal regression neural network (RNN), has been applied.

[0034] LSTM is capable of learning both long-term and short-term dependencies by replacing the intermediate layer nodes of an RNN with LSTM blocks as shown in Figure 4. As LSTM is a type of RNN, it has a chain-like structure, but it does not have a single neural network layer but four interacting layers. Walking is considered to be a periodic movement, so it can learn time-series data, and LSTN, which has no long-term dependencies, is suitable for application.

[0035] The deep learning means 140 performs learning and training using data sets for a total of six locations: the left and right heels 112, 112, the left and right thighs 114, 114, the waist 116, and the neck 118, and generates a learned model.

[0036] The deep learning learning and training data set (Train) and test data set (Test) may be a data set (previous data set) before the measurement time t0 in Fig. 5, or the learning and training data set (Train) may learn using the previous data set, and the test data set (Test) may be a data set (later data set) after the measurement time t0. In the latter case (use of the later data set), if the pedestrian's gait changes significantly after the measurement time, the determination and selection means 160 described later can select an inference engine with higher accuracy.

[0037] The error calculation means 150 calculates the error between the actual walking speed, which is the data set R after the measurement time, and the walking estimated speed T by the learned model. Evaluation indices for regression models of deep learning and machine learning include the mean absolute error MAE, the mean square error MSE, the root mean square error RMSE, and the mean absolute error rate MAPE. When applying the estimation of the walking speed to a walking training machine, for example, it is necessary to follow without delay, so it is important to prevent large errors from occurring. Therefore, in this embodiment, the RMSE, which can emphasize large errors, is applied as the evaluation index. Furthermore, MAPE, which can calculate a relative error instead of an absolute error and is expressed as a percentage, is also applied. Equation 1 and Equation 2 show the definition formulas of RMSE and MAPE. The smaller the values ​​of both evaluation indices, the higher the accuracy.

[0038]

number

[0039]

number

[0040] The error calculation means 150 calculates the root mean square error RMSE and the mean absolute error MAPE for the actual walking speed, which is the subsequent data set R, and the walking speed estimated by the trained model T, and sends the results as errors E to the judgment and selection means 160.

[0041] The judging and selecting means 160 ranks the predicted walking speeds at the multiple locations based on the magnitude of the error E, and selects at least two of the predicted walking speeds with small errors E. Through this process, the judging and selecting means 160 can select multiple trained models with higher estimation accuracy as the inference engine C, and can also remove outliers with large errors E.

[0042] The inference engine C selected by the determination and selection means 160 is stored in the estimation means 170, and an estimated walking speed can be generated by the inference engine C with higher estimation accuracy depending on the walking state of the pedestrian.

[0043] <Application Example of the Embodiment> Hereinafter, the results of the verification test will be described as an application example of this embodiment with reference to Figs. 6-10. Fig. 6 is a diagram showing an example of the periodicity of acceleration at multiple points on a pedestrian's body measured by a measurement means. Fig. 7 is a diagram showing an example of the root mean square error (RMSE) calculated by an error calculation means. Fig. 8 is a diagram showing an example of the mean absolute error (MAPE) calculated by an error calculation means. Fig. 9 is a diagram showing an example of the average error related to the travel distance of the measured value and the speed estimate value for each step of a pedestrian. Fig. 10 is a table summarizing the results of the root mean square error (RMSE) and the mean absolute error (MAPE) at multiple points on a pedestrian's body in the verification test.

[0044] The overall configuration of the verification test is to measure the walking state of a pedestrian as shown in Figure 2. The test verifies the fluctuation in prediction accuracy caused by changes in the time from measurement to prediction. The prediction time to be verified is changed within the range of one walking cycle.

[0045] The test conditions were as follows: one able-bodied person in his twenties walked in a straight line on an 8m walking path L 30 times in total. In the test, the acceleration sensor 120 was attached to the center of the waist sensor 126 and directly above the heel sensors 122, 122 on the outside of both feet. The marker for the optical motion capture 130 was attached in the same position as the waist acceleration sensor.

[0046] The test procedure is as follows. (1) We confirmed the safety of the test site and the health of the subjects. We explained the test contents and obtained the consent of the pedestrians (subjects). (2) The subjects were asked to change into clothes for measurement that were easy to move in, so that it would be easy to attach the acceleration sensor 120. The subjects were asked to wear shoes that they were usually accustomed to wearing. (3) An acceleration sensor and a motion capture 130 marker were attached to the subject. (4) To ensure that the subjects walked at a constant cycle, the subjects' walking cycle was measured and then a metronome was used to play sounds periodically to stabilize the cycle. (5) An initial rest of at least 2 seconds was provided for each trial. The subject touched the toes to the ground three times to synchronize the acceleration sensors 120. After that, the subject walked along the walking path L. After each measurement, the subject was asked whether he or she felt any discomfort while walking.

[0047] Following the above test procedure, we conducted testing and performed scaling by standardization based on the test results, and then created an LSTM learning model according to the hyperparameters shown in the table below.

[0048] [Table 1]

[0049] This model was implemented using the Windows 10 OS and the programming language Python 3.6 in Jupyter Notebook on Anaconda 3. Tensorflow and Keras were used as machine learning frameworks, and Scikitlearn was used as the machine learning library.

[0050] The input to the learning model needs to be one complex time series data. Therefore, the data was trimmed at the timing of a large increase in the acceleration sensor 120 of the left heel 112 due to the landing of the left foot, and a range in which walking was stable was prepared for each trial. Each trial was joined together to create one complex time series data. A moving average was applied to 10 data in the forward direction so that the waveform of the joined timing would be smooth.

[0051] As an application direction of the walking speed estimation device 100 according to this embodiment, we are considering the operation control of a walking training machine (see FIG. 11). Since this walking training machine needs to follow without delay, it is important to prevent large errors from occurring. Therefore, in the verification test according to this embodiment, RMSE, which can emphasize large errors, was used as an evaluation index. In addition, MAPE, which can calculate relative errors instead of absolute errors and is expressed as a percentage, was also used.

[0052] In addition, K-fold cross validation was used as a method to evaluate the generalization performance of deep learning and machine learning. K-fold cross validation is a method in which data is divided into K parts, and learning is performed K times so that all K pieces of data are used as validation data once, and then accuracy is verified. In this test, K=10 was set, and 10 learning models were constructed. The maximum, minimum, and average values ​​of the accuracy were then calculated.

[0053] When processing the test results, the average, maximum, and minimum walking speed of the pedestrian (subject) were calculated. Note that these maximum and minimum speeds were calculated as the average of the maximum and minimum peaks for each walking cycle, so as not to reference outliers. The average pedestrian speed was 0.997 m / s, the maximum speed was 1.219 m / s, and the minimum speed was 0.847 m / s.

[0054] Next, to check whether the fluctuations in each acceleration were regular, the sum of squares of the triaxial acceleration in the first trial was calculated. The results are shown in Figure 6. Note that the scale of the vertical axis is not taken into consideration because scaling is performed. Referring to Figures 6(a) and (b), the acceleration waveforms of the right and left ankles are almost the same, occurring approximately every 1.2 seconds. The phase is 0.6 seconds on the left and right, and naturally there is a half-cycle shift. Referring to Figure 6(c), the acceleration waveform of the waist is also almost the same as the waveform of the heel, occurring approximately every 1.2 seconds.

[0055] Since the walking cycle was approximately 1.2 seconds, the evaluation indices RMSE and MAPE were calculated as the accuracy of speed prediction when the Predict Time was changed at 1.2 seconds or less. The results are shown in Figures 7 and 8, respectively. These figures show the maximum, minimum and average accuracy evaluation values ​​of 10 trained models (prediction models) created by 10-fold cross-validation for each prediction time.

[0056] The maximum and average values ​​of RMSE and MAPE tend to increase in the short prediction time range. After that, a curve is drawn that has a minimum value at 0.40 seconds and a maximum value at 0.65 seconds. The minimum value at 0.65 seconds decreases to a value close to the value at 0.10 seconds. The minimum value shows a similar trend in the range up to 0.75 seconds, but the amount of deviation is small and it remains almost flat. A sharp increase tendency is observed after 0.75 seconds.

[0057] RMSE and MAPE are indices that evaluate absolute errors. Therefore, it is not clear whether there will be an error in the final distance traveled if the measured value and predicted value are used. Therefore, we calculated the average error in the distance traveled between the measured value and the predicted value for each step. The results are shown in Figure 9.

[0058] The error in the distance traveled drops around 0.60 seconds, just like the RMSE and MAPE, but the degree of error does not change significantly at any prediction time, with an average error of 9.2 mm occurring per step.

[0059] The average walking speed of the pedestrians (subjects) was approximately 1 m / s, and the variation in walking speed was from -0.15 m / s to +0.2 m / s. When predicting 0.6 seconds into the future, which is half a walking cycle ahead, the RMSE was 0.037 m / s, the MAPE was 3.0%, and the error in the distance traveled per step was 8.2 mm, verifying that this is a model with good prediction accuracy.

[0060] Using the models verified as described above, we create multiple models that predict speed from acceleration data of each body part, and rank them by accuracy based on RMSE and MAPE. Specifically, as shown in Figure 10, we store information on the accuracy of each model and the body parts that use acceleration.

[0061] FIG. 10 shows the results of processing up to the error calculation means 150 shown in FIG. 1 for the heel 112, thigh 114, waist 116, and neck 118. Models created using only the parts marked with "o" in the column direction of the heel, waist, thigh, and neck are arranged in ascending order of error RMSE and MAPE. "-" indicates a part not used in learning the model. Therefore, we will consider the RMSE and MAPE marked with "o". Also, the frames surrounding the numbers in the column direction of RMSE and MAPE in FIG. 10 are so-called bar graphs, which are made visually easy to see.

[0062] Referring to Figure 10, the estimation accuracy improves when the ankle is included. In the absence of the ankle, the accuracy of the model including the thigh improves, and here the accuracy of the model combining the thigh and neck improves.

[0063] To summarize the above, the following is true. (1) The accuracy of the neck can be improved by combining it with the lower body, which has low correlation. (2) The temporary increase in waist size is thought to be due to a misperception of forward / backward swaying versus up / down / left / right swaying. (3) The thigh has a greater chance of being replaced when the ankle is missing, although not as easily as the ankle. (4) The ankle is the most important measurement point because it has a large gait characteristic and has the greatest impact on predictions. However, sufficient verification is required for abnormal values ​​and outliers.

[0064] When ranking predicted walking speeds at multiple locations and selecting at least two predicted walking speeds with fewer errors, a measurement data set including the heel is effective. However, even if the heel cannot be measured or there is a problem or error in the measurement value of the heel, it was confirmed that a highly accurate estimation means can be obtained by combining other measurement locations.

[0065] As described above, from the results of the verification test according to this embodiment, the measurement means of the present invention measures the acceleration and speed of multiple parts of the body of a pedestrian, and thus it is possible to select a more reliable data set of walking speed, independent of the walking characteristics of each pedestrian. Next, the deep learning means learns and trains using these data sets, thereby estimating the walking speed at each part and enabling comparison as the same variable of walking speed. Then, it was verified that the judgment and selection means calculates the error between the estimated walking speed and the actually measured walking speed, and based on this error, selects a trained model with higher accuracy, and this trained model can be used as an inference engine.

[0066] <Examples of use of the present invention> The walking speed estimation device 100 according to the present invention can be used in a walking training machine 200 that assists walking by mounting a motor on the caster 210 shown in Fig. 11. That is, the acceleration of each part of the walker 10 is measured, and parts with low periodicity are picked out even though the walking movement is periodic (because the movement changes depending on the patient's physical condition on that day, etc.). Then, models that use the picked-up parts are eliminated, and the model with the highest accuracy at the time of model creation is selected from the other models, and it is determined as the speed prediction model to be used on that day.

[0067] The determined speed prediction model is used as an estimation means 170 to predict the near future walking speed of the walker in real time and the casters 210 are controlled so that the walking training machine 200 follows the walker 10 without any delay, and even a walker who does not have the strength to pull can undergo walking training.

[0068] In particular, the walking state of a pedestrian 10 using a walking training machine 200 may differ from that of an ordinary healthy person, and by applying the walking speed estimation device 100 according to the present invention, it is possible to realize prediction of the walking speed according to the state of the pedestrian.

[0069] As described above, the present invention can provide a device that measures the walking state of a target person from various angles and analyzes the measurement results to estimate the walking speed with higher accuracy. Note that the aspects of the present disclosure are not limited to the above-mentioned embodiments, and modifications are possible without departing from the spirit of the present disclosure. [Explanation of symbols]

[0070] 10. Pedestrians 100 Walking speed estimation device 110. Measurement methods 112...Heel 114...thigh 116 Waist 118...Neck 120 Acceleration sensor 122 Heel sensor 124 Thigh sensor 126 Waist sensor 128 Neck sensor 130···Motion Capture 132...Camera 140 Deep Learning Methods 150...Error calculation means 160... Judgment and selection means 170... Estimation means 200 ···Walking training machine 210 Caster R Dataset T Estimated walking speed E...error C···Inference Engine

Claims

1. A device for estimating walking speed, comprising: a measuring means for setting a measurement time point, setting a pre-measurement period and a post-measurement period of a predetermined time interval before and after the measurement time point, measuring the acceleration and walking speed of a plurality of points of a walking body during the pre-measurement period and the post-measurement period, and setting both or one of a calculated walking speed and the walking speed calculated from the acceleration during the pre-measurement period as a pre-data set, and both or one of a calculated walking speed and the walking speed calculated from the acceleration during the post-measurement period as a post-data set; A deep learning means for generating a predicted walking speed for the subsequent measurement period at the plurality of locations by deep learning the previous data set; an error calculation means for comparing the subsequent data set with the predicted walking speed to calculate an error at the plurality of locations; a determining / selecting means for ranking the predicted walking speeds at the plurality of locations based on the magnitude of the error and selecting at least two of the predicted walking speeds having a small error; an estimation means for estimating a walking speed after the post-measurement period by an inference engine that calculates the selected predicted walking speed; A walking speed estimation device comprising:

2. The deep learning The previous data set is a training data set; The walking speed estimation device according to claim 1 , wherein the latter data set is a test data set.

3. The deep learning The walking speed estimation device according to claim 1 , wherein the previous data set includes a training data set and a test data set.

4. The determining and selecting means is Calculating a previous period from a change in acceleration during the previous measurement period; Calculating a later period from a change in acceleration during the later measurement period; Calculating a period difference between the previous period and the subsequent period; The walking speed estimation device according to claim 1 , further comprising: selecting the predicted walking speed by the inference engine based on a data set at at least one or more points where the periodic difference is small.

5. The walking speed estimation device according to claim 1 , wherein the deep learning is a goal regression neural network (RNN).

6. 6. The walking speed estimation device according to claim 5, wherein the goal regression type neural network is a long short-term memory neural network (LSTM).

7. 2. The walking speed estimation device according to claim 1, wherein the error calculation means uses either or both of a Root Mean Squared Error (RMSE) method and a Mean Absolute Percentage Error (MAPE) method.

Images