Hard floor and carpet recognition method for sweeping robot based on binary classification algorithm

Abstract

The present invention proposes a method for identifying hard floors and carpets of a sweeping robot based on a binary classification algorithm. The detection method can utilize the driving current values ​​of multiple specified components and the strength values ​​of ground detection sensors to simultaneously obtain ground detection data, and utilize real-time ground detection data to drive the SVM classification model to realize the classification of the sweeping robot in the cleaning state, thereby improving the accuracy and speed of the judgment results.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent devices, and in particular to a method for identifying hard floors and carpets of a sweeping robot based on a binary classification algorithm. Background Art

[0002] With the development of science and technology, sweeping robots with autonomous movement functions are becoming more and more popular. Sweeping robots can save users a lot of cleaning time and help users make better use of time to enjoy life. Existing sweeping robots still have major problems in autonomous cleaning. When cleaning hard floors and carpets, the power requirements for moving parts and cleaning parts are different, and they need to be switched appropriately. When the sweeping robot is cleaning with the power suitable for hard floors, when it encounters carpets, the hair in the carpet will greatly increase the resistance of the sweeper's movement, and the airtightness between the robot and the carpet will be reduced, making the carpet's cleaning ability very poor. Therefore, how to quickly switch to the appropriate working mode according to the different conditions of the ground to be cleaned to ensure the cleaning effect has become a problem to be solved.

[0003] If only the ground detection sensor is used, for example, the black frosted hard floor has a rough surface, strong diffuse reflection, and weak return intensity; the brushed carpet hard floor has a rough surface, strong diffuse reflection, and weak return intensity, so the effect is the same and it is not easy to distinguish. However, the sensor strength of the white carpet and the hard floor is relatively the same.

[0004] The existing patent publication number CN 107943039A discloses a cleaning method of a sweeping robot and a sweeping robot, which judges whether to adopt a carpet mode or a hard floor mode by successively judging whether the driving current values ​​of the walking wheels, the side sweeps and the middle sweeps increase. This judgment method takes too much time and does not allow the sweeping robot to make a quick judgment. At the same time, if the carpet is smooth, misjudgment will occur. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a method for identifying hard floors and carpets of a sweeping robot based on a binary classification algorithm.

[0006] The hard floor and carpet recognition method of the sweeping robot based on the binary classification algorithm includes:

[0007] Step S1: Performing data training on known ground detection data to obtain an SVM classification model, wherein the ground detection data includes a driving current value of a specified component and a ground detection sensor strength value, wherein the specified component includes at least one of left and right driving wheels, a middle sweep motor, and a side sweep motor;

[0008] Step S2: obtaining real-time ground detection data in the cleaning state, using the real-time ground detection data to drive the SVM classification model to achieve classification of the sweeping robot in the cleaning state, specifically: inputting the obtained real-time ground detection data into the pre-established SVM classification model to obtain a credibility value, if the credibility value is greater than or equal to 1, it is judged to be a carpet environment, if the credibility value is less than or equal to -1, it is judged to be a hard ground environment;

[0009] Step S3: Control the working mode of the sweeping robot according to the judgment result.

[0010] Furthermore, the training process of the SVM classification model includes the following steps: selecting a polynomial kernel function of multiple powers, mapping the training samples to an n-dimensional vector space, using the SVM classification model to find the optimal classification hyperplane of the carpet and hard ground feature samples in the sample feature space, obtaining the support vector set representing the carpet and hard ground feature samples and their corresponding VC credibility, forming a discriminant function for judging carpets and hard grounds, and performing classification.

[0011] More specifically, the pre-established SVM classification model divides the target to be classified into two categories, one of which is carpet and the other is hard ground. When assuming linear separability, the discriminant function f(x) = wx + b = 0 is defined. This discriminant function is a hyperplane in the n-dimensional vector space, referred to as the dividing surface. Given T samples {xi, yi} l i = 1, where xi∈R n ,yi∈{+1,-1},i=1,2,...N,xi is the ith feature vector,yi is the class label,when yi=1 it is a positive example,when yi=-1 it is a negative example;Geometric interval: for a given data set T and hyperplane wx+b=0,the geometric interval of the hyperplane about the sample point (xi,yi) is defined as

[0012]

[0013] The minimum geometric interval of the hyperplane about all sample points is:

[0014]

[0015] This distance is the so-called distance from the support vector to the hyperplane;

[0016] According to the above definition, the problem of solving the maximum segmentation hyperplane of the SVM classification model can be expressed as the following constrained optimization problem:

[0017]

[0018] Divide both sides of the constraint by γ to get

[0019]

[0020] Because ‖w‖ and γ are both scalars, for the sake of simplicity, let

[0021]

[0022]

[0023] get

[0024] y i (w*x i +b)≥1,i=1,2…,N

[0025] And because maximizing γ is equivalent to maximizing That is, equivalence and minimization in, This is to keep the form simple and not affect the results after derivation. Therefore, the maximum segmentation hyperplane problem of the SVM classification model can be expressed as the following constrained optimization problem:

[0026]

[0027] st y i (wx i +b)≥1,i=1,2…,N

[0028] This is a convex quadratic programming problem with inequality constraints. The Lagrange multiplier method can be used to obtain its dual problem.

[0029] First, we transform the original objective function with constraints into the newly constructed Lagrangian objective function without constraints:

[0030]

[0031] where a i is a Lagrange multiplier, and a i ≥0, let When the sample point does not meet the constraints, that is, it is outside the feasible solution region:

[0032] y i (w*x i +b)<1

[0033] At this time, a i If θ(w) is set to infinity, then θ(w) is also infinite.

[0034] When the full point satisfies the constraints, that is, it is within the feasible solution region:

[0035] y i (w*x i +b)≥1

[0036] At this time, θ(w) is the original function itself. Therefore, combining the two cases can give us our new objective function

[0037]

[0038] So the original constraint problem is equivalent to:

[0039]

[0040] Using the Lagrange function duality, swap the positions of the minimum and maximum: when

[0041] In order to obtain the specific form of solving the dual problem, let the partial derivative of L(w,b,a) with respect to w and b be 0, and we can get:

[0042]

[0043]

[0044] It can be seen that in a*, there is at least one (Proof by contradiction can prove that if all are 0, then w = 0, a contradiction). For this j,

[0045] y i (w*.x i + b*)-1=0

[0046] So we can get:

[0047]

[0048]

[0049] For any training sample ( x i .x j) , there is always a i =0 or y j (wx j +b)=1. i = 0, then the sample will not appear in the final equation for solving the model parameters. i >0, then there must be y j (wx j+b)=1, the corresponding sample point is located on the maximum interval boundary, which is a support vector. This shows an important property of the support vector machine: after training, most of the training samples do not need to be retained, and the final model is only related to the support vector.

[0050] So far, the assumption has been made that the training set data is linearly separable. However, in reality, there is almost no completely linearly separable data. To solve this problem, the concept of "soft interval" is introduced, which allows some points not to satisfy the constraint y. j (wx j +b)≥1

[0051] Using hinge loss, the original optimization problem is rewritten as

[0052]

[0053] st y i (wx i +b)≥1-ξ i

[0054] ξ i ≥0,i=1,2…,N

[0055] where ξ i is the “slack variable”, ξ i =max(0,1-y i (w*x i +b)), which is a hinge loss function. Each sample has a corresponding slack variable, which represents the degree to which the sample does not meet the constraints. C>0 is called the penalty parameter. The larger the C value, the greater the penalty for classification. The idea is consistent with the linear separable solution. Here, the Lagrange function is first obtained by the Lagrange multiplier method, and then its dual problem is solved.

[0056] The linear support vector machine learning algorithm is as follows:

[0057] Input: training data set T = {(x1,y1),(x2,y2),…,(x N ,y N )}where xi∈R n ,yi∈{+1,-1},i=1,2,...N,

[0058] Output: Separating hyperplane and classification decision function, selecting penalty parameter C>0, constructing and solving convex quadratic programming problem

[0059]

[0060] 0≤a i ≤C,i=1,2…,N

[0061] Get the optimal solution

[0062]

[0063] (2) Calculation

[0064]

[0065] Select a * A component of Satisfy the conditions calculate

[0066]

[0067] (3) Find the separating hyperplane

[0068] w * .X+b * =0

[0069] Classification decision function:

[0070] f(x)=sign(w * .X+b * )

[0071] Furthermore, in step S1, the known ground detection data is ground detection data of hard ground and carpet of at least 20 materials placed within a range of one square meter.

[0072] Furthermore, in step S1, the hard floor at least includes floor tiles, bottom plates, and concave-convex tiles, and the carpet at least includes: long-hair carpet, short-hair carpet, wool carpet, and polyester carpet.

[0073] Furthermore, in step S1, the sweeping robot is provided with at least three ground detection sensors, which are respectively distributed at the left end, the front end and the right end thereof, and the light intensity value of the ground detection sensor is detected by a transmitting end and a receiving end provided by the ground detection sensor.

[0074] Furthermore, in step S1, the ground detection sensor collects data at a speed of 30 cm / s, and collects data once every 5 ms.

[0075] Furthermore, in step S1, the data collected by the left and right driving wheels are linear rotation speeds, each set of linear rotation speeds corresponds to a current value, and the linear rotation speeds of the left and right driving shafts are different, corresponding to different current values.

[0076] Further, in step S1, when the linear speeds and directions of the left and right driving wheels are the same, the vehicle moves in a straight line. When encountering an obstacle, the rotation directions of the left and right driving wheels are opposite but the linear speeds are the same, the vehicle turns in place. When encountering an obstacle, the rotation directions of the left and right driving wheels are opposite but the linear speeds are different, the vehicle turns.

[0077] Furthermore, in step S1, when the left and right driving wheels collect current values, first, they need to be accelerated to 30m / s constantly for 2-3s, of which 10 groups are collected in 1s, and then the constant speed is collected, once every 5ms. When the sweeping robot needs to be stopped, it takes 1-2s to reduce the speed to 0m / s, of which 10 groups are collected in 1s.

[0078] Furthermore, in step S1, when the left and right driving wheels collect current values, if an obstacle is detected, the speed will be reduced to 0 m / s after 2 seconds, and if no obstacle is detected, the speed will be reduced to 0 m / s after 1 second.

[0079] Furthermore, in step S1, the middle sweeping motor is a roller brush motor driven by a DC motor, and the resistance on the hard ground and the carpet is different, so different current values ​​are generated. The rotation speed of the roller brush is constant, and 10 groups are collected in 1 second.

[0080] Furthermore, in step S1, the side sweeping motor is set on the right side of the sweeping robot, with a constant rotation speed of 200 revolutions per minute, and 10 groups are collected in one second.

[0081] Furthermore, in step S1, the ground detection data needs to collect data of the lateral drive, vertical drive, and oblique drive of the sweeping robot, so that the sample size is large and the credibility is higher.

[0082] Beneficial effects of the present invention: The present invention proposes a method for identifying hard floors and carpets of a sweeping robot based on a binary classification algorithm. The detection method can utilize the driving current values ​​of multiple specified components and the strength values ​​of ground detection sensors to simultaneously acquire ground detection data, and utilize real-time ground detection data to drive the SVM classification model to realize the classification of the sweeping robot in the cleaning state, thereby improving the accuracy and speed of the judgment results. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 The figure is a flow chart of the method for identifying hard floors and carpets of a sweeping robot based on a binary classification algorithm of the present invention.

[0084] Figure 2 The figure is a schematic diagram of the SVM classification model discrimination of the hard floor and carpet recognition method of the sweeping robot based on the binary classification algorithm of the present invention. DETAILED DESCRIPTION

[0085] like Figure 1As shown, the hard floor and carpet recognition method of the sweeping robot based on the binary classification algorithm includes:

[0086] Step S1: Performing data training on known ground detection data to obtain an SVM classification model, wherein the ground detection data includes a driving current value of a specified component and a ground detection sensor strength value, wherein the specified component includes at least one of left and right driving wheels, a middle sweep motor, and a side sweep motor;

[0087] Step S2: obtaining real-time ground detection data in the cleaning state, using the real-time ground detection data to drive the SVM classification model to achieve classification of the sweeping robot in the cleaning state, specifically: inputting the obtained real-time ground detection data into the pre-established SVM classification model to obtain a credibility value, if the credibility value is greater than or equal to 1, it is judged to be a carpet environment, if the credibility value is less than or equal to -1, it is judged to be a hard ground environment;

[0088] Step S3: Control the working mode of the sweeping robot according to the judgment result.

[0089] Furthermore, the training process of the SVM classification model includes the following steps: selecting a polynomial kernel function of multiple powers, mapping the training samples to an n-dimensional vector space, using the SVM classification model to find the optimal classification hyperplane of the carpet and hard ground feature samples in the sample feature space, obtaining the support vector set representing the carpet and hard ground feature samples and their corresponding VC credibility, forming a discriminant function for judging carpets and hard grounds, and performing classification.

[0090] More specifically, the pre-established SVM classification model divides the target to be classified into two categories, one of which is carpet and the other is hard ground. When assuming linear separability, the discriminant function f(x) = wx + b = 0 is defined. This discriminant function is a hyperplane in the n-dimensional vector space, referred to as the dividing surface. Given T samples {xi, yi} l i = 1, where xi∈R n ,yi∈{+1,-1},i=1,2,...N,xi is the ith feature vector,yi is the class label,when yi=1 it is a positive example,when yi=-1 it is a negative example;Geometric interval: for a given data set T and hyperplane wx+b=0,the geometric interval of the hyperplane about the sample point (xi,yi) is defined as

[0091]

[0092] The minimum geometric interval of the hyperplane about all sample points is:

[0093]

[0094] This distance is the so-called distance from the support vector to the hyperplane;

[0095] According to the above definition, the problem of solving the maximum segmentation hyperplane of the SVM classification model can be expressed as the following constrained optimization problem:

[0096]

[0097] Divide both sides of the constraint by γ to get

[0098]

[0099] Because ‖w‖ and γ are both scalars, for the sake of simplicity, let

[0100]

[0101]

[0102] get

[0103] y i (w*x i +b)≥1,i=1,2…,N

[0104] And because maximizing γ is equivalent to maximizing That is, equivalence and minimization in, This is to keep the form simple and not affect the results after derivation. Therefore, the maximum segmentation hyperplane problem of the SVM classification model can be expressed as the following constrained optimization problem:

[0105]

[0106] st y i (wx i +b)≥1,i=1,2…,N

[0107] This is a convex quadratic programming problem with inequality constraints. The Lagrange multiplier method can be used to obtain its dual problem.

[0108] First, we transform the original objective function with constraints into the newly constructed Lagrangian objective function without constraints:

[0109]

[0110] where a i is a Lagrange multiplier, and a i ≥0, let When the sample point does not meet the constraints, that is, it is outside the feasible solution region:

[0111] y i (w*x i +b)<1

[0112] At this time, a i If θ(w) is set to infinity, then θ(w) is also infinite.

[0113] When the full point satisfies the constraints, that is, it is within the feasible solution region:

[0114] y i (w*x i +b)≥1

[0115] At this time, θ(w) is the original function itself. Therefore, combining the two cases can give us our new objective function

[0116]

[0117] So the original constraint problem is equivalent to:

[0118]

[0119] Using the Lagrange function duality, swap the positions of the minimum and maximum: when In order to obtain the specific form of solving the dual problem, let the partial derivative of L(w,b,a) with respect to w and b be 0, and we can get:

[0120]

[0121]

[0122] It can be seen that in a*, there is at least one (Proof by contradiction can prove that if all are 0, then w = 0, which is a contradiction). For this j,

[0123] y i (w*.x i + b*)-1=0

[0124] So we can get:

[0125]

[0126]

[0127] For any training sample ( x i .x j) , there is always a i =0 or y j (wx j +b)=1. i= 0, then the sample will not appear in the final equation for solving the model parameters. i >0, then there must be y j (wx j +b)=1, the corresponding sample point is located on the maximum interval boundary, which is a support vector. This shows an important property of the support vector machine: after training, most of the training samples do not need to be retained, and the final model is only related to the support vector.

[0128] So far, the assumption has been made that the training set data is linearly separable. However, in reality, there is almost no completely linearly separable data. To solve this problem, the concept of "soft interval" is introduced, which allows some points not to satisfy the constraint y. j (wx j +b)≥1

[0129] Using hinge loss, the original optimization problem is rewritten as

[0130]

[0131] st y i (wx i +b)≥1-ξ i

[0132] ξ i ≥0,i=1,2…,N

[0133] where ξ i is the “slack variable”, ξ i =max(0,1-y i (w*x i +b)), which is a hinge loss function. Each sample has a corresponding slack variable, which represents the degree to which the sample does not meet the constraints. C>0 is called the penalty parameter. The larger the C value, the greater the penalty for classification. The idea is consistent with the linear separable solution. Here, the Lagrange function is first obtained by the Lagrange multiplier method, and then its dual problem is solved.

[0134] The linear support vector machine learning algorithm is obtained as follows:

[0135] Input: training data set T = {(x1,y1),(x2,y2),…,(x N ,y N )}where xi∈R n ,yi∈{+1,-1},i=1,2,...N,Output: separation hyperplane and classification decision function, select penalty parameter C>0, construct and solve convex quadratic programming problem

[0136]

[0137] 0≤a i ≤C,i=1,2…,N

[0138] Get the optimal solution

[0139]

[0140] (2) Calculation

[0141]

[0142] Select a * A component of Satisfy the conditions calculate

[0143]

[0144] (3) Find the separating hyperplane

[0145] w * .X+b * =0

[0146] Classification decision function:

[0147] f(x)=sign(w * .X+b * )

[0148] In step S1, the known ground detection data is the ground detection data of hard floors and carpets of at least 20 materials placed within a range of one square meter, the hard floors at least include floor tiles, bottom plates, and concave-convex tiles, the carpets at least include: long-haired carpets, short-haired carpets, wool carpets, and polyester carpets. The sweeping robot is provided with at least three ground detection sensors, which are respectively distributed at the left end, the front end, and the right end thereof. The light intensity value of the ground detection sensor is detected by a transmitting end and a receiving end provided by the ground detection sensor. The speed of the ground detection sensor during collection is 30cm / s, and at the same time, the collection is performed once every 5ms. The data collected by the left and right driving wheels are linear rotation speeds, and each set of linear rotation speeds corresponds to a current value. The linear rotation speeds of the left and right driving shafts are different, corresponding to different current values. When the linear rotation speeds and directions of the left and right driving wheels are the same, they go in a straight line. When encountering an obstacle, the rotation directions of the left and right driving wheels are opposite but the linear rotation speeds are the same, then they turn in place. When encountering an obstacle, the rotation directions of the left and right driving wheels are opposite. It is opposite but the linear speed is different, then it turns. When the left and right driving wheels collect current values, first, it needs to accelerate to 30m / s constantly after 2-3s, of which 10 groups are collected in 1s, and then the constant speed is collected, once every 5ms. When the sweeping robot needs to be stopped, it needs to drop the speed to 0m / s after 1-2s, of which 10 groups are collected in 1s. When the left and right driving wheels collect current values, if an obstacle is detected, the speed will drop to 0m / s after 2s. If no obstacle is detected, the speed will drop to 0m / s after 1s. The middle sweeping motor is a roller brush motor driven by a DC motor. The resistance on hard floors and carpets is different, so different current values ​​are generated. The speed of the roller brush is constant, and 10 groups are collected in 1 second. The side sweeping motor is set on the right side of the sweeping robot. The speed is constant, which is 200 rpm, and 10 groups are collected in 1 second. The ground detection data needs to collect the data of the horizontal drive, vertical drive, and oblique drive of the sweeping robot, so that the sample size is large and the credibility is higher.

[0149] Beneficial effects of the present invention: The present invention proposes a method for identifying hard floors and carpets of a sweeping robot based on a binary classification algorithm. The detection method can utilize the driving current values ​​of multiple specified components and the strength values ​​of ground detection sensors to simultaneously acquire ground detection data, and utilize real-time ground detection data to drive the SVM classification model to realize the classification of the sweeping robot in the cleaning state, thereby improving the accuracy and speed of the judgment results.

[0150] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for identifying hard floors and carpets of a sweeping robot based on a binary classification algorithm. A method for identifying hard floors and carpets of a sweeping robot based on a binary classification algorithm includes: Step S1: Data training is performed on known ground detection data to obtain an SVM classification model, wherein the ground detection data includes a driving current value of a specified component and a ground detection sensor intensity value, wherein the specified component includes at least one of a left and right driving wheel, a middle sweeping motor, and a side sweeping motor, and the sweeping robot is provided with at least three ground detection sensors, which are respectively distributed at the left end, the front end, and the right end thereof, and the light intensity value of the ground detection sensor is detected by a transmitting end and a receiving end provided by the ground detection sensor, the middle sweeping motor is a roller brush motor driven by a DC motor, and the resistance on the hard floor and the carpet is different, so different current values ​​are generated, and the rotation speed of the roller brush is constant, and 10 groups are collected in 1 second, and the side sweeping motor is arranged on the right side of the sweeping robot, and the rotation speed is constant, which is 200 revolutions / min, and 10 groups are collected in 1 second, and the known ground detection data is the ground detection data of hard floors and carpets of at least 20 materials placed within a square meter; Step S2: obtaining real-time ground detection data in the cleaning state, using the real-time ground detection data to drive the SVM classification model to achieve classification of the sweeping robot in the cleaning state, specifically: inputting the obtained real-time ground detection data into the pre-established SVM classification model to obtain a credibility value, if the credibility value is greater than or equal to 1, it is judged to be a carpet environment, if the credibility value is less than or equal to -1, it is judged to be a hard ground environment; Step S3: Control the working mode of the sweeping robot according to the judgment result.

2. The method for identifying hard floors and carpets by a sweeping robot based on a binary classification algorithm according to claim 1, characterized in that: The training process of the SVM classification model includes the following steps: selecting a polynomial kernel function of multiple powers, mapping the training samples to an n-dimensional vector space, using the SVM classification model to find the optimal classification hyperplane of the carpet and hard floor feature samples in the sample feature space, obtaining the support vector set representing the carpet and hard floor feature samples and their corresponding VC credibility, forming a discriminant function for judging carpets and hard floors, and performing classification.

3. The method for identifying hard floors and carpets by a sweeping robot based on a binary classification algorithm according to claim 1, characterized in that: In step S1, the hard floor at least includes floor tiles, bottom plates, and concave-convex tiles, and the carpet at least includes: long-hair carpet, short-hair carpet, wool carpet, and polyester carpet.

4. The method for identifying hard floors and carpets by a sweeping robot based on a binary classification algorithm according to claim 1, characterized in that: In step S1, the ground detection sensor collects data at a speed of 30 cm / s and collects data once every 5 ms.

5. The method for identifying hard floors and carpets by a sweeping robot based on a binary classification algorithm according to claim 1, characterized in that: In step S1, the data collected by the left and right driving wheels are linear rotation speeds, each set of linear rotation speeds corresponds to a current value, and the linear rotation speeds of the left and right driving shafts are different, corresponding to different current values.

6. The method for identifying hard floors and carpets by a sweeping robot based on a binary classification algorithm according to claim 1, characterized in that: In step S1, when the left and right driving wheels collect current values, first, they need to be constantly accelerated to 30m / s for 2-3s, of which 10 groups are collected in 1s, and then the constant speed is collected, once every 5ms. When the sweeping robot needs to be stopped, it takes 1-2s to reduce the speed to 0m / s, of which 10 groups are collected in 1s.

7. The method for identifying hard floors and carpets by a sweeping robot based on a binary classification algorithm according to claim 1, characterized in that: In step S1, the ground detection data needs to collect data of the lateral driving, vertical driving, and oblique driving of the sweeping robot, so that the sample size is large and the credibility is higher.

Images