JP2011245285A - Device for estimating condition of moving body - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a device for estimating condition of a moving body capable of measuring the position and direction (posture) of the moving body and estimating the condition of the same based on a self-contained type measuring means of the moving body.SOLUTION: An acceleration vector-measuring means 1002 and an angular speed vector-measuring means 1001 are measuring means installed on the moving body and measure the acceleration vector and the angular speed vector of the moving body. A posture angle estimating device 1004 inputs the acceleration vector (sensor coordinate system) and the angular speed vector (sensor coordinate system) and estimates the posture angle of the same. An acceleration vector/angular speed component disassembling device 1005 outputs output data of total 6 components of an acceleration component in a gravity direction, an angular acceleration component around a gravity axis, an acceleration component in a travelling direction, an angular speed component around the travelling axis, an acceleration component in a lateral direction, an angular speed component around a lateral axis to the moving body from the above three kinds of outputs, and the movement of the moving body is discriminated by the mighty discriminators 1011 to 101x.

Description

  The present invention relates to a state estimation apparatus for a moving body capable of measuring the position of the moving body and estimating a state where the moving body is placed.

  A position / orientation acquisition method based on pedestrian dead reckoning (PDR) that limits the target of positioning to pedestrians has been studied extensively in recent years, and many reports have been made. It is becoming clear that it can be realized simply by wearing an IMU (Inertial Measurement Unit) by using the constraints that pedestrians have, and this is a field where many results have been obtained.

  The IMU used here is commercially available as a packaged sensor set, and various products are available from the low end to the high end from the viewpoint of performance and cost (weight). The use of low-cost IMUs is spreading.

As shown in Non-Patent Document 1 to Non-Patent Document 4, the main methods of PDR can be broadly classified into the following two depending on the IMU mounting location.
(1) Foot-mounted PDR (for example, Non-Patent Document 1, Non-Patent Document 2)
(2) Waist-mounted PDR (for example, Non-Patent Document 3 and Non-Patent Document 4)

  The toe-mountable PDR effectively attaches the IMU to the toes such as shoes and boots, and effectively uses the constraint condition that the foot stays until the foot leaves the ground. Achieve zero-speed update in the reckoning process. As a result, the accumulation of the estimation error of the moving speed can be suppressed, and this method is used in many recent research reports. However, the IMU attached to the toes of the pedestrian creates movements that move the center of gravity other than walking movements with little movement of the center of gravity (for example, sitting / standing up on a chair, passing through) In principle, it is difficult to detect with sufficient accuracy.

  On the other hand, the waist-mounted PDR is an approach that attempts to measure the speed and stride size by identifying the pedestrian's movement as a pattern by attaching an IMU to the waist near the center of gravity of the human body. is there. The IMU attached to the waist cannot perform zero-speed update unlike the foot-mounted PDR method, but has an advantage that it can effectively measure the movement accompanied by the movement of the center of gravity of the person. For this reason, the waist-worn PDR technique has an advantage that it can realize motion recognition in addition to position and orientation measurement using a single IMU. On the other hand, since it is difficult to detect the timing at which the zero speed update should be performed, the measurement of the posture and speed by integrating the acceleration and the angular velocity cannot be realized by the sensor data by the low cost IMU.

  For this reason, the method of the waist-mounted PDR, including our previous research, detects the walking motion by applying pattern recognition technology to the acceleration and angular velocity data, and the magnitude (ie, walking speed or Step length). (Patent Document 1, Patent Document 2)

JP 2004-264028 A JP 2005-114537 A

E. Foxlin, “Pedestrian Tracking with Shoe-Mounted Inertial Sensors”, In IEEE Computer Graphics and Applications, Volume 25, Issue 6 (November 2005), pp. 38-46, 2005. O. Woodman and R. Harle, “Pedestrian localization for indoor environments”, In Proc. Of the 10th Int. Conf. On Ubiquitous computing (UbiComp), 2008. Tom Judd and Toan V, “Use of a New Pedometric Dead Reckoning Module in GPS Denied Environments”, In Proc. IEEE / ION Poisition, Location and Navigation (PLANS2008), pp. 120-128, 2008. M. Kourogi, and T. Kurata, “Personal Positioning Based on Walking Locomotion Analysis with Self-Contained Sensors and a Wearable Camera”, In Proc. 2nd IEEE and ACM Int. Symp. On Mixed and Augmented Reality (ISMAR2003), pp. 103-112, 2003.

  In many manufacturing and service industries (restaurants, nursing care, lodging, building maintenance, etc.), in order to improve the productivity and efficiency of workers, acquire their movements in addition to their positions and orientations. Is strongly demanded. In a site where workers move frequently, in addition to analyzing the flow of workers, the recognition of movements related to the work of the workers is realized by linking them to the flow lines, thereby improving work efficiency and productivity. Can be effectively visualized and used for analysis for business improvement. In particular, in labor-intensive industries, improvements in worker efficiency and productivity have a significant impact because they are directly linked to labor cost savings.

  An approach that uses a self-contained sensor (acceleration, gyroscope, magnetic sensor) as a measuring means is promising as a means for capturing the movement of the worker. This is because most of human movements including walking movements mainly show characteristic patterns in their angular velocities and acceleration vectors.

  From the viewpoint of performance and cost (weight), IMU, which is a product packaged with a set of sensors, is available in a variety of products from low end to high end. However, high cost and high performance (high end) IMUs are available. Wearing by each worker engaged in such an industry may hinder the work contents from the viewpoint of weight and size, and is not realistic from the viewpoint of cost.

  Therefore, the object of the present invention is to measure the position and orientation (attitude) based on the self-contained measuring means including the acceleration / angular velocity vector measuring means attached to the pedestrian among the moving objects, and the state. It is an object of the present invention to provide a state estimation device for a moving body that can estimate the above.

  In order to achieve the above-described object, the state estimation device for a moving body according to the present invention has a basic idea that a measurement target is limited to a pedestrian's motion and a pedestrian dead reckoning that uses a constraint condition of the motion. (PDR) -based approach enables realization with a small and low-cost IMU, and in addition to measuring the position and orientation of pedestrians, the state of a moving body is estimated by acquiring the type of motion .

  Specifically, the state estimation device for a moving body according to the present invention includes an acceleration vector measuring unit that is mounted on the moving body and measures an acceleration vector of the moving body, and an angular velocity that is mounted on the moving body and measures an angular velocity vector of the moving body. A vector measuring unit, a posture angle estimating device for estimating a posture angle of a moving body based on an output of the acceleration vector measuring unit and an output of the angular velocity vector measuring unit, an output of the acceleration vector measuring unit, and the angular velocity vector measuring unit And the output of the posture angle estimator, the acceleration component in the gravity direction and the angular velocity component around the gravity axis, the acceleration component in the traveling direction and the angular velocity component around the traveling axis, the lateral acceleration component and around the horizontal axis An acceleration / angular velocity component decomposing apparatus that decomposes the angular velocity component into a total of six components, and a total of N (N is an integer) states that are to be identified as true or false. A strong classifier that has been learned, and the strong classifier is configured to input a total of six components that are outputs of the acceleration / angular velocity component decomposition apparatus and present a true or false output, respectively. To do.

  Further, in this case, in the mobile state estimation apparatus, the strong classifiers are N strong classifiers each having a threshold value and identifying true or false for each of a total of N states. It is a feature.

  According to the mobile body state estimation apparatus of the present invention, a framework for realizing (1) pedestrian position / orientation estimation and motion recognition based on the same data by an approach based on pedestrian dead reckoning is effective (2 ) Positioning performance is improved by using position correction using action recognition using machine learning, and (3) the recognition rate of action recognition can be improved using position / orientation measurement results. An apparatus for estimating a state of a moving body configured as described above is provided.

It is a block diagram explaining the basic composition of the state estimating device of the mobile object by the present invention. It is a flowchart which shows the process of one strong discriminator in the state estimation apparatus of the moving body by this invention. It is a 1st figure explaining the short-term fluctuation | variation (pattern of a positive value peak to a negative value peak) of the input data in a weak discriminator. It is a 2nd figure explaining the short-term fluctuation | variation (pattern of a negative value peak to a positive value peak) of the input data in a weak discriminator. It is a 3rd figure explaining the short-term fluctuation | variation (positive value peak pattern) of the input data in a weak discriminator. It is a 4th figure explaining the short-term fluctuation | variation (negative value peak pattern) of the input data in a weak discriminator. It is a 5th figure explaining the short-term fluctuation | variation (pattern from a positive value to a negative value) of the input data in a weak discriminator. It is a 6th figure explaining the short-term fluctuation | variation (pattern from a negative value to a positive value) of the input data in a weak discriminator. It is a figure showing the flowchart of the process which learns a strong discriminator. It is a figure showing an example with which a strong discriminator is comprised with five weak discriminators. It is a block diagram explaining another Example in the state estimation apparatus of the moving body by this invention. It is a flowchart of the process of the Example which comprises the strong discriminator which uses position and azimuth | direction information as an input in addition to an acceleration component and an angular velocity component.

  Hereinafter, an embodiment for carrying out the present invention will be described with reference to the drawings. FIG. 1 is a block diagram illustrating a basic configuration of a moving object state estimation apparatus according to the present invention. In FIG. 1, 1001 is an angular velocity vector measuring means, 1002 is an acceleration vector measuring means, 1003 is an inertial measurement unit (IMU), 1004 is an attitude angle estimating device, 1005 is an acceleration / angular velocity component decomposing device, and 1011 is a first strong identification. 1012 is a second strong classifier (weighted weak classifier), and 101x is an xth strong classifier (weighted weak classifier). Here, an example is shown in which a device for identifying x operations / states is configured.

  The acceleration vector measuring means 1002 and the angular velocity vector measuring means 1001 here are measuring means attached to the moving body, and are means for measuring the acceleration vector and each velocity vector of the moving body.

  The angular velocity vector measuring unit 1002 may be implemented not only by the gyro sensor but also by a measuring unit that estimates the angular velocity based on the differential output of the magnetic sensor, and the acceleration vector measuring unit 1002 and the angular velocity vector measuring unit 1001 are one. It is realized by a packaged inertial measurement unit (IMU) 1003.

  The posture angle estimation apparatus 1004 receives the acceleration vector (sensor coordinate system) that is the output of the acceleration vector measuring unit 1002 and the angular velocity vector (sensor coordinate system) that is the output of the angular velocity vector measuring unit 1001 and estimates the posture angle. Device. For example, as described in Japanese Patent Application Laid-Open No. 2004-264028, a posture angle estimation device developed by the present inventor can be used.

  Here, the attitude angle uniquely determines the attitude of the sensor coordinate system in the world coordinate system (NED, North, East Down coordinate system), and is a rotation matrix or quaternion that performs conversion between the two coordinate systems. It can be expressed by numbers.

  The acceleration / angular velocity component decomposition apparatus 1005 includes an acceleration vector (sensor coordinate system) that is an output of the angular velocity vector measuring unit 1002 that is an acceleration measuring unit, and an angular velocity vector (sensor that is an output of the angular velocity vector measuring unit 1001 that is an angular velocity measuring unit. Coordinate system) and the posture angle which is the output of the posture angle estimation device 1004 as input, the acceleration component in the gravity direction with respect to the moving body, the angular velocity component around the gravity axis, the acceleration component in the traveling direction, and the angular velocity around the traveling axis Output data of a total of six components, that is, a component, an acceleration component in the lateral direction, and an angular velocity component around the side axis are output.

  The output data includes a first strong classifier (weighted weak classifier) 1011, a second strong classifier (weighted weak classifier) 1012,..., An xth strong classifier (weighted weak classifier). The discriminator) is input to the discriminator 101x, and each discriminator discriminates the operation of the moving object wearing the IMU.

Next, these strong classifiers will be described.
(Action recognition framework)
AdaBoost or adaptive boosting used in strong classifiers is one of machine learning algorithms, and a plurality of different classifiers whose performance is not sufficient (however, the error rate is less than 0.5) (These are called weak classifiers) are combined by weighting to ultimately produce a classifier with high performance. In order to construct a framework for motion recognition using AdaBoost, it is necessary to determine sensor data sets to be analyzed and to determine weak classifiers that perform motion discrimination from these data sets.

  In the mobile object state estimation apparatus according to the invention, AdaBoost is used as a discriminator, and a plurality of weak discriminators are combined as training data using sensor data appropriately labeled as an operation to be discriminated, and the weight is learned. In this way, the operation recognition is realized.

  FIG. 2 is a flowchart showing the processing of one strong classifier in the moving object state estimation apparatus according to the present invention. One strong classifier corresponds to a recognition processing device for one operation or state. In the process of performing this recognition, first, in S201, an acceleration vector (3-axis / sensor coordinate system) is acquired from the IMU. In S202, an angular velocity vector (3-axis / sensor coordinate system) is acquired from the IMU. Here, the offset output of the angular velocity vector may be compensated by an external temperature compensation table or another drift compensation system (for example, a method as shown in International Publication WO2010 / 1968). In S203, the attitude angle of the IMU with respect to the world coordinate system is estimated based on the acceleration / angular velocity components acquired in S201 and S202. The attitude angle is expressed by, for example, a rotation matrix C (3 × 3 matrix) representing conversion between the IMU sensor coordinate system and the world coordinate system.

  In S204, based on this attitude angle, the acceleration vector in the sensor coordinate system acquired in S201 is decomposed into three acceleration components in the vertical direction, the traveling direction, and the lateral direction. In S205, based on the attitude angle estimated in S203, the angular velocity vector in the sensor coordinate system acquired in S202 is decomposed into three angular velocity components around the yaw axis, pitch axis, and roll axis. Next, an initialization process before entering the main body process of the strong classifier is performed in S206. In S206, the iterator variable i and the summation variable S are each initialized to 0. In S207, for the i-th weak classifier (i), a weak classifier data set, that is, a threshold parameter set and a weight (reliability) α (i) are acquired.

  In S208, the weak discriminator (i) applies the processing by the weak discriminator to the angular velocity component or acceleration component to be processed, and the output W (i) is a value of {+1, -1}. Get as. In S209, the output O (i) as the weak classifier (i) is obtained by multiplying the weight α (i) acquired in S207 by the output W (i) of the weak classifier obtained in S208. In S210, the output O (i) is added to the total variable S, and +1 is added to the iterator variable i. In S211, it is determined whether or not the iterator variable matches N (the number of weak classifiers obtained by learning). If they do not match, the process S207 is executed. If they match, the loop process is terminated and the process in S212 is completed. Move on to processing. In S212, whether the value of the sum variable S is positive or negative is determined. If it is positive, it is determined that there is an operation or state X. If the value of the sum variable is negative, it is determined that there is no action or state X.

(Sensor data to be analyzed)
The data of the acceleration / gyro sensor mounted near the center of gravity of the person is output in the sensor's coordinate system, so if it is used as it is as learning data, it contains local information. It may not be appropriate to describe. Therefore, in the human coordinate system, the acceleration vector and the angular velocity vector are decomposed into three axes, ie, the direction of gravity, the direction of travel, and the direction of the outer product vector (ie, the side) orthogonal to the two directions, and the respective components are separated. Used as input data for machine learning. That is, since there are three components for the acceleration vector and three components for the angular velocity, a total of six components are candidates for analysis.

(Weak classifier used for motion identification)
Regarding the short-term fluctuations of each component of the input data subject to analysis described above,
(1) A pattern in which a negative peak below a certain threshold appears following a positive peak above a certain threshold (FIG. 3),
(2) A pattern in which a positive peak above a certain threshold appears after a negative peak below a certain threshold (FIG. 4),
(3) A pattern (FIG. 5) that appears only for positive peaks above a certain threshold.
(4) A pattern that appears only for negative peaks below a certain threshold value (FIG. 6),
(5) A pattern (FIG. 7) that changes from a positive value above a certain threshold value to a negative value below a certain threshold value,
(6) A pattern that changes from a negative value below a certain threshold value to a positive value above a certain threshold value (FIG. 8) can be cited as a pattern that can be used as a weak classifier.

Regarding the characteristics of long-term fluctuation of each component, when performing frequency analysis with a Fast Fourier Transform (FFT) etc. and looking at each component in the frequency domain,
(A) There is a pattern in which the peak power in the frequency spectrum of the frequency domain of the component has a ratio equal to or greater than a certain threshold with respect to the total power.
Also, between the two components,
(A) A pattern in which the cross-correlation between components is greater than or equal to a certain threshold value, or less than a certain threshold value, (b) A range in which the difference value between the positive and negative peak times of each component is constant ± threshold value Patterns that are within,
(C) When each component is converted to the frequency domain by FFT or the like, three patterns, in which the phase difference of the peak frequency in the power spectrum falls within a certain central value ± threshold, can be used as weak classifiers. As a pattern.

(Strong classifier learning method)
The above-described strong classifier is learned in a framework called AdaBoost (Adaptive Boosting). FIG. 9 shows an example of a flowchart of the processing. What is necessary as a premise of this processing is learning data (M) and weak classifiers (N). This learning data includes time-series data of angular velocity components, acceleration components, and the position / orientation of the wearer in accordance with the world coordinate system.

  In the processing procedure S901, the number N of weak classifiers used by the strong classifiers is set at a stage before learning. As for the magnitude of N, it is known that the recognition accuracy improves as a sufficiently large value is set, but the accuracy improvement is saturated at a certain value or more. On the other hand, as N is increased, the calculation cost is proportionally increased in the recognition phase. For this reason, it is desirable to set the N value that is considered to be optimal after learning with different N values for the sample data in advance and applying it. In S902, N weak classifiers including the threshold parameter set are determined and set. In S903, the learning data (j) prepared in advance (a total of M pieces) is given a weight β (j) and initialized.

  In S904, a loop initialization process described later is performed. Set the iterator variable i to 0 and initialize it. In S905, the acceleration, angular velocity, position, and orientation of the learning data are applied to the weak classifier (i), and the optimum threshold parameter set for the learning data and its weight α (i) (reliability) Is calculated. Note that the optimum threshold parameter set search range here is effective when the acceleration component is set to a step size of about 0.01G between + 3G and -3G. Concerning the acceleration / angular velocity component, collecting typical data for such an operation / state and measuring the range in advance can contribute to the setting of this value.

  In S906, an output W (i) {+1 or -1} is obtained by applying to the weak classifier (i) using the threshold parameter set that is optimal for the learning data. In S907, the weight β (j) of the learning data is updated based on the output W (i) obtained in S906. In S908, the iterator variable i ← i + 1 is updated by loop processing. In S909, it is determined whether or not the iterator variable has reached N. If it does not match N, the process returns to S905 to continue the process, and if it matches, the process proceeds to S910.

  In S910, a strong classifier is configured based on each weak classifier (i) and its weight α (i) obtained in S905 to S907.

  FIG. 10 is an example of a strong classifier configured by the learning result as described above. FIG. 10 shows a case where the strong classifier is composed of five weak classifiers. Different components, weights, and threshold values are set for each weak classifier. Each strong classifier is composed of one or more weak classifiers having such different parameters.

  According to the state estimation device for a moving body of the present invention, a movement that realizes a method for identifying a motion at the same time as estimating the position and orientation (posture) by using a low-cost IMU mounted on the waist of a pedestrian. A body state estimation device is provided.

  Some of the movements that differ from walking movements (for example, bending over and sitting in a chair, etc.) most often occur only in a specific location, and therefore identify the movement by identifying that movement. It becomes possible to narrow down. Conversely, if the position and its attribute are specified by other means, the actions that can occur at the position are limited, and the context information contributes to the improvement of the action identification rate.

  Also, by acquiring the posture state of a person (standing, sitting, standing on knees, etc.), it is possible to determine whether or not walking can occur. It is expected that it can be eliminated.

  FIG. 11 shows another embodiment in which one of the strong classifiers is configured using the position and the direction. In addition to the basic configuration of the present invention shown in FIG. 1, a positioning device 1107 and a traveling direction estimation device 1106 are provided. The positioning device 1107 can be configured by a device that estimates the position of a positioning target (pedestrian) using a so-called PDR (Pedestrian Dead Reckoning) framework. Further, the traveling direction estimation device 1106 can be configured by the above-described PDR framework, and is one of the essential components for configuring the positioning device 1107.

  The data flow is as follows. First, an acceleration / angular velocity vector (3-axis / sensor coordinate system) is measured by the IMU, and this is input to the attitude angle estimation device 1104, the positioning device 1107, the traveling direction estimation device 1106, and the acceleration / angular velocity component decomposition device 1105, respectively. . Note that the attitude angle estimation device 1104 may be provided with a magnetic vector measuring unit separately, and a magnetic vector (three axes / sensor coordinate system) measured thereby may be added as an input to perform azimuth correction based on geomagnetism. . The posture angle estimation device 1104 estimates, for example, a rotation matrix C (3 × 3 matrix) describing the relationship between the sensor coordinate system and the world coordinate system as a posture angle, and this posture angle is an acceleration / angular velocity component decomposition device 1105 and a positioning device 1107. , And given as an input to the traveling direction estimation device 1106.

  The acceleration / angular velocity component decomposition apparatus 1105 decomposes the acceleration vector input by the IMU into three components in the vertical direction, the traveling direction, and the horizontal direction from the sensor coordinate system. The angular velocity vector is decomposed into three components around the yaw axis, pitch axis, and roll axis, and a total of six components are input to the strong discriminator 111k. The positioning device 1107 measures the position of the positioning target using a PDR framework or the like based on the acceleration / angular velocity component obtained from the IMU, the posture angle obtained from the posture angle estimation device 1104, and the traveling direction obtained from the traveling direction estimation device 1106. It is a device to do. The estimation result position and uncertainty (for example, the error variance covariance matrix in the Kalman filter) are given as inputs to the strong classifier 111k. The acceleration / angular velocity vector that is input to the positioning device 1107 may be an acceleration / angular velocity vector converted into the world coordinate system by the acceleration / angular velocity component decomposition device 1105 or the like instead of the sensor coordinate system.

  The traveling direction estimation device 1106 is a device that estimates an orientation that is a traveling direction of a positioning target (for example, a pedestrian) based on the acceleration / angular velocity component obtained from the IMU and the posture angle obtained from the posture angle estimation device 1104. By using the PDR framework, the traveling direction of the positioning target is estimated, and this is given as an input to the positioning device 1107 and the strong classifier 111k, respectively. The strong discriminator 111k receives the component-resolved acceleration / angular velocity component and position / orientation, performs recognition processing for the motion / state k, returns True when the motion / state k is considered to exist, and otherwise Returns False.

  FIG. 12 is a flowchart of the process for realizing the strong classifier configured in FIG. 11 in consideration of the position / orientation information in addition to the acceleration / angular velocity component. First, in S1201, an acceleration vector (3-axis / sensor coordinate system) is acquired from the IMU, and in S1202, an angular velocity vector (3-axis / sensor coordinate system) is acquired from the IMU. In S1203, the posture angle estimation device 1104 estimates the posture angle based on the acceleration vector and the angular velocity vector. In S1204, the traveling direction of the IMU is estimated based on the acceleration vector and angular velocity vector obtained in S1201 and S1202, and the attitude angle obtained in S1203. In S1205, the position and its uncertainty are estimated based on the acceleration vector and angular velocity vector acquired in S1201 and S1202, the attitude angle obtained in S1203, and the traveling direction obtained in S1204.

  In S1206, based on the acceleration vector and angular velocity vector obtained in S1201 and S1202, and the attitude angle obtained in S1203, the acceleration vector is decomposed into vertical, traveling, and lateral components, and the angular velocity vector is Breaks down into components around the pitch and roll axes. In S1207, using the acceleration component and angular velocity component obtained in S1206, the position obtained from S1205 and its uncertainty, and the direction of travel obtained in S1204 as input, processing by a strong classifier that recognizes the motion or state k is performed. And output the result as True / False.

  From the prototype IMU (acceleration, magnetism, gyroscope, barometric pressure sensor) attached to the waist near the center of gravity of the person, considering the service industry such as restaurant, nursing care, and building maintenance as the main target domain. Based on this data, it can be used as a method for identifying human movements in addition to position, altitude and direction.

  As a function to be provided as a positioning system for the waist-mounted PDR, a function for identifying a state in which a pedestrian moves in a normal mode is necessary. (1) Forward walking + step up / down, (2) sidewalk, (3 ) It is to discriminate three of the backward walking. In addition to these three basic operations, the mobile body state estimation device of the present invention is configured to be able to discriminate a plurality of characteristic operations involving the movement of the center of gravity. By identifying the motion, the estimation result can be corrected to the position where the motion can occur, and the frequency of motion occurrence is used as a prior probability based on the position estimation result, thereby improving the accuracy of motion recognition. Can be increased.

1001 Angular velocity vector measuring means,
1002 Acceleration vector measuring means 1003 Inertial measurement unit (IMU)
1004 Posture angle estimation apparatus 1005 Acceleration / angular velocity component decomposition apparatus 1011 First strong classifier (weighted weak classifier)
1012 Second strong classifier (weighted weak classifier)
101x xth strong classifier (weighted weak classifier)

Claims (3)

An acceleration vector measuring means mounted on the moving body for measuring the acceleration vector of the moving body;
Angular velocity vector measuring means mounted on the moving body for measuring the angular velocity vector of the moving body;
A posture angle estimating device for estimating a posture angle of a moving body based on an output of the acceleration vector measuring unit and an output of the angular velocity vector measuring unit;
Based on the output of the acceleration vector measuring means, the output of the angular velocity vector measuring means, and the output of the attitude angle estimating device, the acceleration component in the gravity direction and the angular velocity component around the gravity axis, the acceleration component in the traveling direction and the rotation around the traveling axis. An acceleration / angular velocity component decomposing device that decomposes into six components, an angular velocity component, a lateral acceleration component, and an angular velocity component around the horizontal axis;
A pre-learned strong classifier that identifies true or false for each of a total of N (N is an integer) states;
With
The strong classifier receives a total of six components, which are outputs of the acceleration / angular velocity component decomposition apparatus, and presents a true or false output, respectively, and presents a state estimation apparatus for a moving body. The apparatus for estimating a state of a moving body according to claim 1,
The strong discriminator has threshold values of components different for each strong discriminator, and is N strong discriminators that discriminate true or false for a total of N states. Mobile body state estimation device. The apparatus for estimating a state of a moving body according to claim 1,
A traveling direction estimation device that estimates an orientation that is a traveling direction of the moving body based on an output of the acceleration vector measurement unit, an output of the angular velocity vector measurement unit, and an output of the posture angle estimation device;
Based on the output of the acceleration vector measuring unit, the output of the angular velocity vector measuring unit, the output of the posture angle estimating device, the output of the posture angle estimating device, and the output of the traveling direction estimating device, the movement A positioning device that measures the position of the body,
The strong discriminator, in addition to the six components, further presents a true or false output, respectively, with an azimuth as an output of the traveling direction estimation device and a position as an output of the positioning device as inputs. A state estimation device for a moving body.