JP6240595B2 - Self-position estimation apparatus and mobile body equipped with self-position estimation apparatus - Google Patents

Description

  The technology disclosed in this specification relates to a moving object. Specifically, the present invention relates to a technique for estimating the position of a moving object.

  2. Description of the Related Art A moving body that moves within a predetermined moving area and that can estimate its own position within the moving area is known. The self-position estimation of a mobile object may fail for various reasons. At this time, if the mobile body itself cannot determine whether the self-position estimation succeeds or fails, the mobile body continues to move without recognizing the fact even if the self-position estimation fails, There arises a problem that it collides with an object (such as an obstacle) in the moving area.

  Non-Patent Document 1 discloses a moving body having a function capable of self-determining whether self-position estimation is successful. This moving object stores an expression representing the reliability of the self-position estimation result, and when the value obtained by inputting a variable acquired at the time of self-position estimation into the expression exceeds a predetermined value, the mobile body It is determined that the position estimation has succeeded, and it is determined that the position has failed when the position is below a predetermined value. As a result, it can be recognized that the mobile body itself has failed in self-position estimation, so even if the mobile body fails in self-position estimation by incorporating the corresponding means when the self-position estimation has failed into the mobile body, The risk of such a collision can be avoided.

Masashi Yokozuka, Hironori Adachi, Osamu Matsumoto, “Abnormality detection and self-position detection by particle filter using LRF and recovery by GPS”, Automobile Engineering Society Proceedings, Automobile Engineering Society, March 2012, Vol. 43, No. 2, p. 587-592

  In Non-Patent Document 1, an expression representing the reliability of the self-position estimation result is theoretically derived based on an algorithm used for self-position estimation. However, in the formula constructed theoretically, the accuracy of the formula is lowered due to the difference between the theory and the actual, and there is a possibility that the accuracy of determination as to whether the self-position estimation has succeeded or failed has been lowered.

  The present specification discloses a moving body that can determine whether the moving body has succeeded or failed in self-position estimation with higher accuracy than before.

  The present specification discloses a self-position estimation apparatus that estimates the self-position of a moving body that moves within a predetermined movement area. The self-position estimation device includes a self-position estimation unit, an abnormality determination formula storage unit, and an abnormality determination unit. The self position estimating unit estimates the self position. The abnormality determination expression storage unit stores an abnormality determination expression obtained by machine learning of learning data acquired based on self-position estimation during learning movement of the moving body. The abnormality determination unit determines whether the self-position estimation in the self-position estimation unit has succeeded or failed when the mobile body performs the main movement. The self-position estimation unit includes an environment map of the moving area, a moving distance and moving direction of the moving body from the self-position estimated by the immediately preceding self-position estimation, and an object existing in the moving area from the moving body. The self-position is estimated based on the distance and the orientation of the object with respect to the moving object. The learning data includes a plurality of data that associates a plurality of variables when the self-position estimation unit estimates the self-position at the time of learning movement and a classification result when the self-position estimation result at that time is classified as normal or abnormal. Have. The abnormality determination unit determines whether the self-position estimation has succeeded or failed by inputting, into the abnormality determination formula, a plurality of variables acquired by the self-position estimation unit when the moving body performs the main movement. .

  In the above-described mobile body self-position estimation device, the abnormality determination expression storage unit stores an abnormality determination expression serving as a reference for determining whether the self-position estimation has succeeded or failed. This abnormality determination formula is obtained by machine learning of learning data acquired when the moving body learns and moves. For this reason, it is possible to determine whether the self-position estimation has succeeded or failed with higher accuracy than a theoretically constructed abnormality determination formula. In addition, the self-position estimation performance of the moving object is determined for each moving object depending on the accuracy of the observation information acquired (that is, the distance from the moving object to the object existing in the moving region and the orientation of the object with respect to the moving object). There is variation. Since this variation is reflected in the learning data, the abnormality determination formula obtained by machine learning of the learning data is a formula specific to each moving object. In other words, the above moving body can determine whether or not the self-position estimation has succeeded based on an abnormality determination formula adapted to each moving body. In addition, even in a single moving body, the self-position estimation performance changes depending on the running years and the like. In the above moving body, the abnormality determination formula can be updated by periodically making the moving body learn and move. For this reason, the moving body can determine whether or not the self-position estimation has succeeded based on the abnormality determination formula that more reflects the current self-position estimation performance of the mobile body.

  Details of the technology disclosed in this specification and further improvements will be described in detail in the detailed description and examples.

The system configuration of a mobile object is shown. 1 shows a system configuration when a mobile object performs self-position estimation processing. 1 shows a system configuration when a moving body performs learning data generation processing. An example of the movement area | region where a moving body learns and moves is shown. An example of learning data is shown. The system configuration | structure when a moving body implements abnormality determination type | formula production | generation processing is shown. The graph of an abnormality judgment formula is shown. 1 shows a system configuration when a moving body performs an abnormality determination process and an estimation means switching process.

  The main features of the embodiments described below are listed. The technical elements described below are independent technical elements and exhibit technical usefulness alone or in various combinations, and are not limited to the combinations described in the claims at the time of filing. Absent.

(Feature 1) In the mobile body self-position estimation device disclosed in this specification, the self-position estimation result of the learning data includes the self-position estimated by the self-position estimation unit during the learning movement, and the actual position of the mobile body. If the position error is less than or equal to a predetermined value, it is classified as normal, and if the position error exceeds a predetermined value, it may be classified as abnormal. In this configuration, the self position estimation result is classified as normal or abnormal based on the actual position of the moving object. For this reason, a highly reliable classification result can be acquired, the reliability of learning data is improved, and as a result, an abnormality determination expression with high determination accuracy can be acquired.

(Characteristic 2) The mobile body self-position estimation apparatus disclosed in the present specification may further include a learning data storage unit that stores learning data. According to this configuration, since the learning data is stored in the estimation device itself, if a failure occurs in the abnormality determination accuracy of the moving body, the learning data log is analyzed to verify whether there is any error. This improves the reliability of the moving object.

(Feature 3) A mobile body self-position estimation device disclosed in the present specification generates a learning data generation unit that generates learning data, and machine learning the learning data generated by the learning data generation unit to generate an abnormality determination formula And an abnormality determination formula generation unit that performs the above-described operation. According to this configuration, the self-position estimation apparatus itself can generate learning data, and an abnormality determination formula can be generated based on the learning data. For this reason, the self-position estimation apparatus itself can update the abnormality determination formula, and can appropriately cope with an environmental change (for example, a secular change of the moving body).

(Characteristic 4) In the mobile body self-position estimation apparatus disclosed in this specification, the self-position estimation unit may estimate the self-position by updating the probability distribution of the state of the mobile body. The variable input to the abnormality determination formula may be a variable when the self-position is estimated based on the probability distribution of the latest state of the moving object. According to this configuration, the determination accuracy of the abnormality determination formula is improved as compared with the configuration in which the variable before the probability distribution of the state of the moving body is updated to the latest one is input to the abnormality determination formula.

(Feature 5) In the mobile body self-position estimation apparatus disclosed in the present specification, machine learning may be performed using a multivariate analysis technique.

(Characteristic 6) In the mobile location estimation apparatus disclosed in the present specification, the multivariate analysis may be logistic regression analysis or discriminant analysis.

(Characteristic 7) In the mobile device self-position estimation apparatus disclosed in this specification, machine learning may be performed using a support vector machine technique.

(Characteristic 8) In the mobile body self-position estimation apparatus disclosed in this specification, machine learning may be performed using a neural network technique.

(Feature 9) The self-position estimation apparatus disclosed in the present specification can be mounted on a moving body that moves within a predetermined movement region. The mobile body may include a movement information acquisition unit, an observation information acquisition unit, an environment map storage unit, a self-position estimation device, and an abnormal time process execution unit. The movement information acquisition unit may acquire the movement distance and movement direction of the moving body from the position where the previous self-position estimation is performed. The observation information acquisition unit may observe an object existing in the moving region, and acquire a distance from the moving object to the object and an orientation of the object with respect to the moving object. The environment map storage unit may store an environment map of the moving area. The abnormal time process execution unit, when it is determined by the abnormality determination unit of the self-position estimation device that the self-position estimation has failed, at least the following operation, that is, a warning is sounded, and the self-position estimation unit is temporarily stopped One of recalculating the self-position or switching to another self-position estimation means. According to this configuration, it is possible to suppress the traveling of the moving body from continuing using the same self-position estimation device even though the self-position estimation has failed.

  In the first embodiment, the moving body 10 will be described with reference to FIGS. The moving body 10 is a wheel drive type moving body having wheels on the left and right, and autonomously travels within a predetermined moving area. The left and right wheels of the moving body 10 are driven independently, and the moving body 10 can change the traveling direction by changing the rotational speed of the left and right wheels. When the mobile body 10 travels within the travel area, the mobile body 10 estimates its own position within the travel area, and travels to the destination based on the self-position estimation result. Further, the moving body 10 determines whether the estimated self-position estimation result is normal (that is, self-position estimation is successful) or abnormal (that is, self-position estimation is unsuccessful) based on an abnormality determination formula (described later). And judge yourself. If the self-position estimation result is abnormal, the self-position estimation program is switched to another program. Hereinafter, the traveling in which the moving body 10 self-determines whether the self-position estimation result is normal or abnormal based on the abnormality determination formula is referred to as “main traveling”. Strictly speaking, the mobile object 10 estimates its own position and orientation, but hereinafter, the position and orientation of the mobile object 10 are collectively referred to as “self-position”.

As shown in FIG. 1, the moving body 10 includes an encoder 12, a laser range finder 14 (hereinafter referred to as LRF 14), a triangulation sensor 16 (described later), and a computer 18.
The encoder 12 detects the rotation angles of the left and right wheels and outputs them to a self-position estimation unit 28 (described later) of the computer 18. That is, the left and right wheels are each driven by a motor (not shown). The motors that drive the left and right wheels are each provided with an encoder 12, and each encoder 12 detects the rotation angle of the left and right wheels. The rotation angles of the left and right wheels detected by the encoder 12 are input to the computer 18.
The LRF 14 emits laser light and measures the time until the emitted laser light is reflected by an object and returned. From the time measured by the LRF 14, the distance from the LRF 14 (that is, the moving body 10) to the object is measured. Further, since the direction in which the laser beam is emitted from the LRF 14 (that is, the incident angle of the laser beam reflected from the object) is known, the orientation of the object with respect to the LRF 14 can be determined. Hereinafter, the distance to the object and the direction of the object are also referred to as “observation information”. The observation information acquired by the LRF 14 is output to the self-position estimation unit 28 of the computer 18. The LRF 14 emits laser light in a predetermined angle range with respect to the traveling direction of the moving body 10 (for example, 60 ° in the left-right direction with respect to the traveling direction).
The computer 18 includes a CPU 20 that performs arithmetic processing, a RAM 22 that temporarily stores arithmetic processing data, and a ROM 24 that stores arithmetic programs executed by the CPU. The computer 18 executes various processes described later. When the CPU 20 executes the calculation program stored in the ROM 24, the CPU 20 functions as the self-position estimation unit 28, the true value calculation unit 32, and the like. The computer 18 calculates a motor drive amount from the self-position estimated by the self-position estimation unit 28, and drives the motor (wheel) with the calculated drive amount. Thereby, the moving body 10 travels to the target position. Since the traveling control of the moving body 10 by the computer 18 can be performed by a known method, processing such as self-position estimation by the computer 18 will be described in detail here.

(Self-position estimation process)
As shown in FIG. 2, the computer 18 includes an environment map storage unit 26 and a self-position estimation unit 28.
The environment map storage unit 26 is provided in the ROM 24 and stores an environment map in which the position and height of an object in the moving area are recorded on the xy coordinate plane. Examples of the objects stored in the environment map include obstacles (for example, walls, floor surfaces, etc.) in the moving area.
The self-position estimation unit 28 estimates the self-position of the moving body 10 based on a known motion model stored in advance in the ROM 24 by Monte Carlo localization identification (hereinafter referred to as MCL). This process is performed by the CPU 20. MCL is a known technique in which a moving body estimates its own position using a particle filter. Specifically, the self-position estimating unit 28 (1) first generates an initial particle set. Each particle is a vector that represents a position candidate of the moving body 10, and includes, as vector elements, the x coordinate, y coordinate, and yaw angle of the moving body 10 in the xy coordinate system. (2) Next, the self-position estimation unit 28, based on the rotation angle of each wheel acquired from the encoder 12, the moving distance and moving direction (hereinafter referred to as moving distance and moving) of the moving body 10 from a reference point (described later). The direction is also referred to as “movement information”). And these movement information is input into a movement model, and each particle is moved according to a movement model. (3) Subsequently, the self-position estimation unit 28 matches the observation information input from the LRF 14 with the environment map acquired from the environment map storage unit 26 for each particle, and sets the likelihood (probability) of each particle. calculate. (4) Then, particles are selected so that the number of particles is large in the high likelihood region and small in the low likelihood region, and noise is added to generate a new set. The self-position estimation unit 28 increases the likelihood of the particles by repeating the processes (2) to (4) above, finds a particle that most closely matches the observation information, and sets the matching particle as the self-position of the mobile object 10. presume. In other words, the self-position estimation unit 28 includes the actual position of the moving body 10 (hereinafter also referred to as a true value) in the particle set (that is, the true value matches one of the particles constituting the particle set). ) Or until any particle in the particle set is very close to the true value, the above processes (2) to (4) are repeated. That is, in one self-position estimation process, the particle set distribution is updated a plurality of times. The above “reference point” refers to the self-position estimated by the immediately preceding self-position estimation process.

(Learning data generation process)
The mobile body 10 performs a travel called a learning travel before performing the main travel. Below, the process which produces | generates learning data based on the learning driving | running | working of the moving body 10 and the data obtained at the time of learning driving | running | working is demonstrated.

  As shown in FIG. 3, the computer 18 includes a true value calculation map storage unit 30, a true value calculation unit 32, a position in addition to the environmental map storage unit 26 and the self position estimation unit 28 used in the self position estimation process. An error calculation unit 34, a learning data generation unit 36, and a learning data storage unit 38 are provided. The environment map storage unit 26 stores an environment map of a moving area in which the moving body 10 performs learning travel. The moving area in which the moving body 10 performs the learning travel may be different from the moving area in which the main traveling is performed. For this reason, when a movement area differs at the time of learning driving | running | working and the time of this driving | running | working, the environment map memory | storage part 26 has memorize | stored the environment map of both movement area | regions. In the following, in order to distinguish the movement area during the learning run from the movement area during the main run, the former is referred to as a “first movement area” and the latter is referred to as a “second movement area”. The environment map of the first movement area is referred to as “first environment map”, and the environment map of the second movement area is referred to as “second environment map”.

  FIG. 4 shows a first movement area 40 in which the moving body 10 performs a learning run. As shown in FIG. 4, a ridge 42 as an obstacle is arranged in a part of the first movement region 40. The collars 42 are arranged so as to face each other, and reflectors 44 are fixed to the opposing surface (inner surface) of the collar 42 at equal intervals. The travel of the moving body 10 during the learning travel is controlled by a remote controller (remote controller). A broken line indicates a travel route of the moving body 10 during the learning travel. As indicated by a broken line, the moving body 10 travels both in a region where the rod 42 is not disposed and in a region between the opposite rods 42.

  The true value calculation map storage unit 30 is provided in the ROM 24 and stores a map in which the positions of the ridges 42 and the reflectors 44 are recorded in the first environment map stored in the environment map storage unit 26. . In other words, the first environment map does not store information on the artificially arranged basket 42 and the reflector 44.

  The triangulation sensor 16 emits laser light in all directions (360 °), detects the laser light that is reflected back by the reflector 44, and detects the distance to the reflector 44 and the orientation of the reflector 44 relative to the moving body 10. And outputs this information to the true value calculation unit 32 of the computer 18. The triangulation sensor 16 is mounted on the moving body 10.

  The true value calculation unit 32 specifies the relative position of the moving body 10 with respect to the reflector 44 (strictly, including the relative azimuth) from the distance and azimuth information acquired from the triangulation sensor 16. Since the position of the reflector 44 is recorded on the map acquired from the map storage unit 30 for true value calculation, the position of the moving body 10 on the map is determined by mapping the relative position of the moving body 10 on the map. Can be identified. In this way, the true value calculation unit 32 calculates the actual position (true value) of the moving body 10. This process is performed by the CPU 20. The true value of the moving body 10 is output to the position error calculation unit 34.

On the other hand, the self-position estimating unit 28 estimates the self-position of the moving body 10 in the first moving region 40 using the first environment map, and outputs the result to the position error calculating unit 34. The self-position estimation process and the true value calculation process are performed at the same timing. Further, the self-position estimation unit 28 determines the standard deviation σ _{x in} the x direction, the standard deviation σ _{y in} the y direction, and the likelihood of particles from the distribution of the latest particle set (that is, the particle set when the self position is estimated). The maximum degree w _max is calculated, and these _pieces of information (standard deviation σ _x , standard deviation σ _y , maximum likelihood w _max ) are output to the learning data generation unit 36. The standard deviation σ _x , the standard deviation σ _y , and the maximum likelihood w _max correspond to an example of “variable”.

  The position error calculation unit 34 calculates a difference (position error) between the true value acquired from the true value calculation unit 32 and the self position acquired from the self position estimation unit 28 and outputs the difference to the learning data generation unit 36. This process is performed by the CPU 20.

  If the position error acquired from the position error calculation unit 34 is 15 cm or less, the learning data generation unit 36 considers that the self-position estimation unit 28 has succeeded in self-position estimation, and classifies the self-position estimation result as “normal”. . On the other hand, if the position error exceeds 15 cm, the self-position estimation unit 28 regards that the self-position estimation has failed and classifies the self-position estimation result as “abnormal”. In the learning run, the situation where the position error is classified as “abnormal” is intentionally created. That is, in the first movement area 40, the ridge 42 that is not recorded in the first environment map is arranged. For this reason, when the moving body 10 travels between the eaves 42, the degree of matching between the observation information acquired by the LRF 14 and the first environment map becomes low, and it becomes difficult to find suitable particles. This means that even if a true value is included in the particle set, the moving body 10 cannot find a particle that matches the true value as a matching particle. For this reason, when the moving body 10 travels between the eaves 42, a particle different from the true value is regarded as a compatible particle, and the possibility that the particle is estimated as the self-position becomes extremely high. As a result, the position error between the estimated self-position and the true value increases, and when the position error exceeds 15 cm, it is classified as “abnormal”.

  As described above, in the learning data generation unit 36, the self-position estimation result is normal or abnormal based on the position error, that is, based on the deviation amount of the estimated self-position of the moving body 10 with respect to the “true value”. Is classified. For this reason, a highly reliable classification result can be acquired. In this embodiment, the normal / abnormal classification reference value of the self-position estimation result is set to 15 cm, but the classification reference value is not limited to this. The lower the classification reference value, the higher the determination accuracy of the abnormality determination formula.

FIG. 5 shows an example of learning data generated by the learning data generation unit 36. As shown in FIG. 5, the learning data generation unit 36 acquires the time at the time of self-position estimation, the position error acquired from the position error calculation unit 34, the classification result of the self-position estimation result, and the self-position estimation unit 28. A data set group (learning data) in which data sets in which the three variables (standard deviation σ _x , standard deviation σ _y , maximum likelihood w _max ) are associated in time series is generated (σ _x , σ _y , W _max alphabet is actually a number). By referring to the learning data, it is possible to know the position error at each time, normality / abnormality of the self-position estimation result, and variables when the self-position is estimated. For example, at time t = 2s, since the position error between the estimated self position and the true value is 7 cm, the self position estimation result is classified as “normal”, and the variables σ _x , σ _y , w _{max at} that time are classified. Are b, g, and l, respectively. At time t = 10 s, since the position error between the estimated self position and the true value is 20 cm, the self position estimation result is classified as “abnormal”, and the variables σ _x , σ _y , and w _{max at} that time are classified. Are c, h, and m, respectively. The learning data generation process is performed by the CPU 20. As described above, the mobile body 10 itself generates the learning data, so that the system configuration can be simplified as compared with the configuration in which the learning data is generated outside the mobile body 10.

  The learning data storage unit 38 is provided in the ROM 24 and stores the learning data generated by the learning data generation unit 36. The learning data stored in the learning data storage unit 38 is output to an abnormality determination formula generation unit 46 described later. In this way, by storing the learning data by the mobile body 10 itself, it is possible to verify whether there is an error by analyzing the log of the learning data when a malfunction occurs in the mobile body 10 or the like. Become. For this reason, it becomes easy to discover the cause when the malfunction occurs in the moving body 10, and the reliability of the moving body 10 can be improved. In this embodiment, the data set includes the time at the time of self-position estimation. However, the present invention is not limited to this configuration. For example, the data set may not include time, and the learning data may be configured by a data group in which the data set is arranged regardless of the self-position estimation time. Further, the learning data may not include a position error.

（Ｚ：目的変数、Ｘ_１〜Ｘ_３：説明変数、ｂ_０：定数、ｂ_１〜ｂ_３：回帰係数）に学習データを適用して、定数ｂ_０及び回帰係数ｂ_１〜ｂ_３の値を求める。学習データを適用する際は、目的変数Ｚに分類結果（正常：０、異常：１）を用い、説明変数Ｘ_１、Ｘ_２、Ｘ_３に学習データの標準偏差σ_ｘ、標準偏差σ_ｙ、最大尤度ｗ_ｍａｘをそれぞれ用いる。これにより、次の数式で表される異常判定式；

が生成される（図７参照）。ｐは、移動体１０の本走行時における自己位置推定結果が異常と判定される確率（以下、異常判定確率とも称する）である。説明変数Ｘ_１〜Ｘ_３に用いる変数（σ_ｘ、σ_ｙ、ｗ_ｍａｘ）は互いに独立しているため、回帰係数ｂ_１〜ｂ_３は、項目間の影響が排除された係数となる。従って、異常判定確率を算出する際は、より正確に項目ごとの影響度を量ることができる。異常判定式生成処理は、ＣＰＵ２０で行われる。
異常判定式記憶部４８は、ＲＯＭ２４に設けられており、異常判定式生成部４６で生成された異常判定式を記憶する。 (Abnormality judgment expression generation processing)
A process in which the moving body 10 generates an abnormality determination formula will be described. As illustrated in FIG. 6, the computer 18 includes an abnormality determination formula generation unit 46 and an abnormality determination formula storage unit 48 in addition to the units used in the self-position estimation process and the learning data generation process.
The abnormality determination formula generation unit 46 acquires the learning data stored in the learning data storage unit 38, performs logistic regression analysis on the learning data, and generates an abnormality determination formula. Logistic regression analysis is a method of multivariate analysis, and when the result is binary, it is a method of generating a statistical regression model that can predict the probability that the result will occur. Specifically, a logistic model represented by the following formula:

(Z: objective _variable, X 1 to X _3: explanatory variable, _{b 0: constant,} b 1 ~b _3: regression coefficient) by applying the learning data, the value of the constant _{b 0} and regression coefficient _b 1 ~b ₃ Ask for. When applying the learning data, the classification result (normal: 0, abnormal: 1) is used for the objective variable Z, and the standard deviation σ _x , standard deviation σ _y of the learning data is used for the explanatory variables X ₁ , X ₂ , X ₃ , Each maximum likelihood w _max is used. As a result, the abnormality determination formula represented by the following formula:

Is generated (see FIG. 7). p is a probability (hereinafter, also referred to as an abnormality determination probability) that the self-position estimation result during the main traveling of the moving body 10 is determined to be abnormal. Since the variables (σ _x , σ _y , w _max ) used for the explanatory variables X _{1 to} X ₃ are independent from each other, the regression coefficients b _{1 to} b ₃ are coefficients in which the influence between items is eliminated. Therefore, when calculating the abnormality determination probability, the influence degree for each item can be measured more accurately. The abnormality determination formula generation process is performed by the CPU 20.
The abnormality determination formula storage unit 48 is provided in the ROM 24 and stores the abnormality determination formula generated by the abnormality determination formula generation unit 46.

(Abnormality judgment processing, estimation means switching processing)
The mobile body 10 performs the above-described learning data generation processing and abnormality determination formula generation processing before performing the main traveling in the second movement region. When the mobile body 10 performs the main travel, it determines whether the self-position estimation has succeeded or failed based on the abnormality determination formula. If the self-position estimation fails, the self-position estimation means (that is, the self-position estimation means) Switch the estimation program. Below, the abnormality determination process and estimation means switching process at the time of the main travel of the moving body 10 will be described. As shown in FIG. 8, the computer 18 includes an abnormality determination unit 50 and an estimation unit switching unit 52 in addition to the units used in the self-position estimation process, the learning data generation process, and the abnormality determination formula generation process.

When the moving body 10 performs the main traveling in the second movement area, the self-position estimating unit 28 estimates the self-position using the second environment map. The three variables σ _x , σ _y , and w _max when the self position is estimated by the self position estimation unit 28 are output to the abnormality determination unit 50. Note that “when the self position is estimated by the self position estimating unit 28” means that the distribution of the particle set in one self position estimation process is the latest state.

The abnormality determination unit 50 converts the three variables σ _x , σ _y , and w _max acquired from the self-position estimation unit 28 into X ₁ , X ₂ , and X ₃ of the abnormality determination formula acquired from the abnormality determination formula storage unit 48, respectively. When p ≦ 0.5, it is determined that the self-position estimation result is normal (ie, self-position estimation is successful), and when p> 0.5, the self-position estimation result is abnormal (ie, self-position estimation). )). The determination result is output to the estimation means switching unit 52. As described above, variables obtained from the distribution of the latest particle set are input to the abnormality determination formulas X _{1 to} X ₃ . For this reason, compared with the structure which inputs the variable of the particle set before a particle set is updated to the newest state, the precision of a determination result becomes higher.

  When the determination result acquired from the abnormality determination unit 50 is “abnormal”, the estimation unit switching unit 52 switches the self-position estimation unit (self-position estimation program) from MCL to odometry. In self-position estimation using MCL, the mobile 10 estimates its own position based on the estimated self-position at the time of the previous self-position estimation. For this reason, once self-position estimation fails (especially once the true value is not included in the particle set), the estimated self-position after that may move away from the true value. The risk of collision with an object (obstacle) increases. However, when the mobile unit 10 fails in the self-position estimation, the mobile unit 10 immediately switches the self-position estimation means (that is, the self-position estimation program using MCL) to another means (self-position estimation program using odometry). For this reason, it can suppress that the estimated self-position of the mobile body 10 leaves | separates from a true value, and can avoid danger, such as the mobile body 10 colliding with an object. In this embodiment, the determination reference value of the self-position estimation result is set to 0.5, but is not limited to this. As the determination reference value is set lower, the abnormality determination accuracy is improved. Further, the estimation means switching unit 52 may switch to a self-position estimation program other than odometry (for example, a self-position estimation program using GPS).

  The effect of the moving body 10 of Example 1 is demonstrated. The moving body 10 stores an abnormality determination expression serving as a reference for determining whether the self-position estimation has succeeded or failed in the abnormality determination expression storage unit 48. This abnormality determination formula is obtained by analyzing learning data acquired when the mobile body 10 learns and travels using a logistic regression analysis technique which is one of machine learning techniques. For this reason, it is possible to determine whether the self-position estimation has succeeded or failed with higher accuracy than a theoretically constructed abnormality determination formula. Further, the self-position estimation performance of the mobile body 10 varies from mobile body to mobile body depending on the number of LRFs 14 mounted on the mobile body 10 and the like. Since this variation is reflected in the learning data, the abnormality determination formula obtained by analyzing the learning data is a formula specific to each moving object. That is, the mobile body 10 can determine whether or not the self-position estimation has succeeded based on the abnormality determination formula corresponding to the performance of the mobile body 10 itself. In addition, the self-position estimation performance of the moving body 10 varies depending on the number of years of travel due to wear of the driving portion (for example, wheels). In the present embodiment, the abnormality determination formula can be updated by causing the mobile body 10 to regularly learn and travel. For this reason, the mobile body 10 can determine whether or not the self-position estimation has succeeded based on the abnormality determination formula that more reflects the current self-position estimation performance of the mobile body 10. For this reason, compared with the structure which sets an abnormality determination formula uniformly irrespective of the performance for every moving body, abnormality determination accuracy can be improved.

In particular, in the first embodiment, MCL is used as the self-position estimating means. In this case, for example, if an object that is not recorded in the second environment map is arranged in the second movement area, the degree of matching between the observation information acquired by the LRF 14 and the environment map becomes low, and the moving object 10 Is likely to fail in self-position estimation. That is, even if a true value is included in the particle set, the moving object 10 cannot regard a particle corresponding to the true value as a conforming particle, and it is highly likely that another particle is regarded as a conforming particle. Become. Even when the moving object 10 of the first embodiment includes a true value in the particle set as described above (that is, when the estimated self-position is not so far from the true value), the abnormality determination formula When the derived abnormality determination probability exceeds 0.5, it can be determined that the self-position estimation result is abnormal. For this reason, the moving body 10 can exhibit high abnormality determination accuracy.
On the other hand, Non-Patent Document 1 does not consider the possibility that the moving body may fail in self-position estimation when a true value is included in the particle set. The technique of Non-Patent Document 1 is a technique premised on a case where a true value is not included in the particle set (that is, a case where the estimated self-position is relatively far from the true value). For this reason, the mobile body of Non-Patent Document 1 can determine that its own position estimation result is abnormal only when the position error is relatively large, and as a result, the abnormality determination accuracy decreases.

In addition, information (that is, variables such as a self-position estimation result and a standard deviation σ _x ) that is the basis of the learning data is acquired by causing the moving body 10 to actually travel in the first movement area 40. For this reason, the self-position estimation performance of the moving body 10 is directly reflected in the learning data. Thus, by generating learning data based on information obtained by actually traveling in a specific area, a more accurate abnormality determination expression can be generated.

  As mentioned above, although the Example of the technique which this specification discloses was described in detail, these are only illustrations, and the mobile body which this specification discloses contains what changed and changed the said Example variously. It is.

  For example, the self-position estimation unit 28 may perform self-position estimation processing using an extended Kalman filter instead of MCL. In this case, an abnormality determination formula is generated using the values obtained from the estimated error covariance matrix (for example, the variance value, the value of each element of the error covariance matrix, the major axis and minor axis of the error ellipse) as explanatory variables. be able to. Moreover, normal / abnormality of the self-position estimation result can be accurately determined by inputting these variables obtained at the time of the actual driving into the abnormality determination formula.

  In addition, the moving body 10 may not have the learning data generation unit 36, and the learning data may be generated outside the moving body 10. Further, the moving body 10 may not have the learning data storage unit 38. That is, the learning data may be deleted when the abnormality determination formula is generated based on the learning data. Further, the moving body 10 may not have the abnormality determination formula generation unit 46, and the abnormality determination formula may be generated outside the moving body 10.

Further, the abnormality determination formula generation unit 46 may use a support vector machine instead of the multivariate analysis. The support vector machine is a well-known technique in pattern recognition modeling, and is a technique for constructing an identification function that outputs a class to which a feature vector belongs in binary when a feature vector with an unknown class assignment is input. The discriminant function is composed of a training sample set with known class membership. For this reason, a discriminant function (abnormality determination formula) can be configured by using the learning data of the first embodiment as a training sample set and learning the learning data with a support vector machine. Then, by inputting the three variables σ _x , σ _y , and w _max acquired from the self-position estimation unit 28 to the discrimination function, it is possible to determine normality / abnormality of the self-position estimation result. Also by this method, the same effects as those of the first embodiment can be obtained.

Further, the abnormality determination formula generation unit 46 may use a neural network instead of the multivariate analysis. The neural network is a well-known technique in pattern recognition modeling, and is a technique for constructing a discriminant function that outputs a class to which a variable belongs in binary when a variable with unknown class assignment is input. In this method, the discrimination function can be configured in two ways, supervised learning and unsupervised learning. In supervised learning, a learning example and a target output for the learning example are given, and the weight of the function in the initial state is adjusted so that the output obtained by inputting the learning example to the function in the initial state matches the target output Thus, an identification function is configured. For this reason, learning variables are learned by a neural network with the three variables σ _x , σ _y , and w _max of the learning data of the first embodiment as learning examples, and the classification result (normal / abnormal) at that time as a target output. Thus, an identification function (abnormality determination formula) can be configured. Then, by inputting the three variables σ _x , σ _y , and w _max acquired from the self-position estimation unit 28 to the discrimination function, it is possible to determine normality / abnormality of the self-position estimation result. Also by this method, the same effects as those of the first embodiment can be obtained. Note that the method of machine learning learning data is not limited to the methods described in the first to third embodiments.

  When analyzing learning data, discriminant analysis may be used instead of logistic regression analysis.

  Moreover, the variable input to the abnormality determination formula when performing the abnormality determination process is not limited to the variable obtained from the latest particle set distribution. For example, a variable obtained from the distribution of the particle set immediately before being updated to the latest particle set may be input to the abnormality determination formula. The type of variable is not limited to standard deviation or maximum likelihood. Further, the number of variables is not limited to three, and may be increased or decreased as appropriate so that the regression model can be accurately regressed.

  Moreover, the mobile body 10 may be provided with a function of sounding an alarm when the self-position estimation result is determined to be abnormal, or may be provided with a function of temporarily stopping, instead of the estimation means switching unit 52. In addition, the self-position estimating unit 28 may have a function of recalculating the self-position. By sounding the alarm, it is possible to inform the manager of the mobile body 10 that the mobile body 10 has failed in the self-position estimation. Moreover, the danger by continuing driving | running | working as it is can be avoided by the mobile body 10 stopping temporarily. In addition, when the moving body 10 fails in self-position estimation due to a moving presence (for example, a pedestrian or a vehicle), the pedestrian or the vehicle is There is a possibility that the self-position estimation result may be "normal" away from the observation range. For this reason, the self-position estimation process using MCL can be continued by recalculating the self-position.

  In addition, the learning travel may be performed by simulation instead of actually traveling a specific area on the moving body.

  Moreover, the moving body 10 may not be a wheel drive type, for example, may be a bipedal walking type. Further, the moving body 10 may have a plurality of LRFs 14.

  Further, in the true value calculation unit 32, the true value of the moving body 10 is calculated using the triangulation sensor 16 and the reflector 44, but the method of calculating the true value is not limited to this. For example, the true value may be calculated using a motion capture or magnetic rail technique.

  Specific examples of the present invention have been described in detail above, but these are merely examples and do not limit the scope of the claims. The technology described in the claims includes various modifications and changes of the specific examples illustrated above. The technical elements described in this specification or the drawings exhibit technical usefulness alone or in various combinations, and are not limited to the combinations described in the claims at the time of filing. In addition, the technology illustrated in the present specification or the drawings achieves a plurality of objects at the same time, and has technical utility by achieving one of the objects.

10: mobile body, 12: encoder, 14: laser range finder (LRF), 28: self-position estimation unit, 36: learning data generation unit, 38: learning data storage unit, 46: abnormality determination formula generation unit, 48: abnormality Determination formula storage unit, 50: Abnormality determination unit

Claims (10)

An apparatus for estimating a self-position of a moving body that moves within a predetermined moving area,
A self-position estimating unit for estimating the self-position;
An abnormality determination expression storage unit that stores an abnormality determination expression obtained by machine learning of learning data acquired based on self-position estimation during learning movement of the mobile body;
An abnormality determination unit that determines whether the self-position estimation in the self-position estimation unit is successful or unsuccessful when the mobile body performs a main movement;
The self-position estimation unit exists in the moving area from the moving body, the environment map of the moving area, the moving distance and moving direction of the moving body from the self-position estimated by the immediately preceding self-position estimation, and Estimating the self-position based on the distance to the moving object and the orientation of the object relative to the moving object,
The learning data is data in which a plurality of variables when the self-position estimation unit estimates the self-position at the time of learning movement and a classification result when the self-position estimation result at that time is classified as normal or abnormal Have multiple
The abnormality determination unit has succeeded or failed in self-position estimation by inputting a plurality of variables acquired by the self-position estimation unit when the moving body performs a main movement in the abnormality determination formula. A mobile body self-position estimation apparatus for determining whether or not   The self-position estimation result of the learning data is classified as normal when the position error between the self-position estimated by the self-position estimation unit during learning movement and the actual position of the moving object is equal to or less than a predetermined value. 2. The mobile body self-position estimation apparatus according to claim 1, wherein the position error is classified as abnormal when the position error exceeds a predetermined value.   The mobile body self-position estimation apparatus according to claim 1, further comprising a learning data storage unit that stores the learning data. A learning data generation unit for generating the learning data;
The movement according to claim 1, further comprising: an abnormality determination expression generation unit that generates an abnormality determination expression by machine learning of the learning data generated by the learning data generation unit. Body self-position estimation device. The self-position estimation unit estimates the self-position by updating the probability distribution of the state of the moving object,
5. The movement according to claim 1, wherein the variable input to the abnormality determination formula is a variable when a self-position is estimated based on a probability distribution of a latest state of the moving body. Body self-position estimation device.   The said machine learning is a self-position estimation apparatus of the moving body as described in any one of Claims 1-5 performed using the method of multivariate analysis.   The mobile body self-position estimation apparatus according to claim 6, wherein the multivariate analysis is logistic regression analysis or discriminant analysis.   The mobile machine self-position estimation apparatus according to any one of claims 1 to 5, wherein the machine learning is performed using a support vector machine technique.   The mobile machine self-position estimation apparatus according to any one of claims 1 to 5, wherein the machine learning is performed using a neural network technique. A moving body that moves within a predetermined movement area,
A movement information acquisition unit that acquires the movement distance and movement direction of the mobile body from the position where the previous self-position estimation was performed;
An observation information acquisition unit for observing an object existing in the moving region and acquiring a distance from the moving object to the object and an orientation of the object with respect to the moving object;
An environmental map storage unit storing an environmental map of the moving area;
The self-position estimation apparatus according to any one of claims 1 to 6,
When it is determined that the self-position estimation has failed in the abnormality determination unit of the self-position estimation device, at least the following operation, that is, an alarm is sounded or temporarily stopped, the self-position estimation unit re-establishes the self-position. A moving object comprising: an abnormality processing execution unit that executes one of calculating or switching to another self-position estimation means.

Images