Laser SLAM inter-frame matching adaptive switching method based on support vector machine

Abstract

The application discloses a laser SLAM interframe matching adaptive switching method based on a support vector machine. First, the AGV is manually controlled to move in a working environment, and a two-dimensional grid map of the environment is constructed by using a laser SLAM algorithm; second, the pose information of the target point of the AGV is input, and an optimal running route is planned by using a path planning algorithm in a navigation algorithm; finally, the data of an inertial measurement unit and a laser radar are synchronously collected, the number of collected environmental feature points and the distance between the surrounding environment and the AGV are substituted into a support vector machine model for judgment, and the environmental state of the AGV at the moment is judged according to the output value, so that the ICP algorithm or the CSM algorithm is adaptively switched. The application overcomes the difficulty of nonlinearity and low dimensionality of warehouse environment representation data, can greatly retain the information of features, is easy to operate, and can meet the needs of most application environments.

Description

Technical Field

[0001] The present invention relates to the technical field of robot positioning, and in particular to a laser SLAM frame matching adaptive switching method based on a support vector machine. Background Art

[0002] In warehouse environments, traditional positioning methods such as GPS, QR codes, magnetic tapes, and ultra-wideband offer high accuracy, but these methods are easily affected by factors such as the application environment and lack flexibility. SLAM (Simultaneous Localization and Mapping) technology primarily addresses the problem of positioning and mapping in unknown environments using only onboard sensors. Due to its high positioning accuracy and flexible positioning methods, it is widely used in the AGV positioning field.

[0003] Among SLAM-based positioning technologies, laser SLAM is a relatively mature one. This technology first calculates the measured data from the odometry and IMU to obtain an estimated pose. Then, using inter-frame matching, it calculates the point cloud data from adjacent LiDAR frames to correct the estimated pose and obtain accurate pose information. Therefore, inter-frame matching is one of the key factors affecting positioning accuracy.

[0004] Existing 2D laser SLAM inter-frame matching methods have certain defects and cannot provide accurate and stable matching information in complex environments. In recent years, adaptive switching positioning technology has been widely used in the field of positioning and navigation. This technology switches between different positioning methods through a switching method, thereby improving positioning accuracy in complex environments. Therefore, using the same technology to switch the laser SLAM inter-frame matching method in a warehouse environment can solve the problem of poor matching effect of the laser SLAM algorithm in complex environments. However, most existing switching positioning technologies switch between GPS and WiFi positioning methods. In warehouse environments, there are problems such as poor signal and difficult equipment installation, which cannot meet the requirements for adaptive switching of laser SLAM inter-frame matching methods. Summary of the Invention

[0005] In response to the above problems, the purpose of the present invention is to provide a laser SLAM inter-frame matching adaptive switching method based on support vector machine, that is, first using 2D laser radar to establish a grid map of the working environment, and then using an adaptive switching method on the established map to switch between the ICP algorithm and the CSM algorithm.

[0006] The present invention comprises the following steps:

[0007] Step 1: On the remote monitoring terminal, manually control the AGV to move in the working environment and use the laser SLAM algorithm to build a two-dimensional grid map of the environment;

[0008] Step 2, input the pose information of the AGV target point, and plan an optimal running route by using the path planning algorithm in the navigation algorithm;

[0009] Step 3, when the AGV starts running, the data of the inertial measurement unit and the laser radar are synchronously collected, wherein the number of environmental feature points η and the distance λ between the surrounding environment and the AGV are obtained by scanning the surrounding environment by the laser radar and after processing;

[0010] The values of η and λ are substituted into the support vector machine model for judgment, and the environmental state of the AGV at the moment is judged according to the output value, so as to adaptively switch to the ICP algorithm or the CSM algorithm.

[0011] The self-adaptive switching method overcomes the nonlinearity and low dimensionality of warehouse environment representation data, can retain feature information to a large extent, is easy to operate, and can meet the needs of most application environments. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 The switching method diagram of the present application. DETAILED DESCRIPTION

[0013] The technical solutions of the present application will be further described in combination with the drawings and through specific embodiments.

[0014] As shown in the drawings, the present application provides a laser SLAM inter-frame matching adaptive switching method based on a support vector machine, and the steps are as follows: Figure 1 Step 1, on the remote monitoring terminal, manually control the AGV to move in the working environment, and construct a two-dimensional grid map of the environment by using the laser SLAM algorithm.

[0015] Step 2, input the pose information of the AGV target point on the remote monitoring terminal, and plan an optimal running route by using the path planning algorithm in the navigation algorithm.

[0016] Step 3, a nonlinear model capable of converting data samples that cannot be linearly classified in a low-dimensional space into linearly separable data samples in a high-dimensional space is established, and the model is represented as follows:

[0017] f(x)=W T φ(x)+b Formula (1)

[0018] In formula (1), φ(x) represents the feature vector after x is mapped, and W and b are parameters to be solved.

[0019] The minimum function is obtained as:

[0020] ​

[0021]

[0022] sty i (W T φ(x)+b)-1≥0

[0023] The dual problem of minimizing the function is:

[0024]

[0025]

[0026] α i ≥0

[0027] In formula (3), sample x i and x j When mapping to the feature space, if there are many feature variables at this time, then the inner product φ(x i ) T φ(x i ) is very computationally intensive, so we use the function κ(x i ,x j ) to solve this problem and transform the inner product value into the function κ(x i ,x j ) is obtained, so the dual problem can be transformed into

[0028]

[0029]

[0030] α i ≥0

[0031] Finally, the solution can be obtained:

[0032]

[0033] In formula (5), the function κ(x i ,x j ) is the kernel function. Kernel function definition: Assume that the input space is χ and the feature space is H. If there is a mapping φ(x):χ→H from χ to H, such that for all x i ,x j ∈χ, function κ(x i ,x j ) satisfies the condition κ(x i ,x j )=φ(x i )φ(x j ), then we can get the inner product as φ(x i )*φ(xj ),κ(x i ,x j ) is the kernel function, φ(x) is the mapping function. The kernel function used in the present invention is the Gaussian kernel function.

[0034] In the dual problem of linear support vector machine learning, the kernel function κ(x i ,x j ) instead of the inner product, the solution is the nonlinear support vector machine.

[0035] In the open and confined environments of a warehouse, selecting different matching algorithms based on the environment can improve both the algorithm's matching accuracy and the AGV's positioning precision. This algorithm switching problem is essentially a binary classification problem. Currently, mainstream classification methods include neural networks and support vector machines (SVMs). However, neural networks require a large number of training samples, while SVMs achieve the same binary classification results without requiring a large number of data samples. This suggests that SVMs have stronger generalization capabilities. Therefore, the present invention adopts the SVM method.

[0036] Step 4: When the AGV starts running, it synchronously collects data from the inertial measurement unit and the lidar. The number of environmental feature points η and the distance λ between the surrounding environment and the AGV are obtained by processing the lidar scanning the surrounding environment. The values ​​of η and λ are substituted into the model, that is, formula (1), and the output values ​​are used to determine the environmental state of the AGV at that moment.

[0037] In step 5, the AGV's current environment is determined based on the input values ​​of η and λ, and the algorithm to be used is determined based on the output identification value. For example, if the input value of η is small and the input value of λ is large, the output identification value is 0, indicating that the AGV is in a wide environment with few obstacles. The CSM algorithm is used. This method is suitable for large-scale environments, such as the loading and unloading areas of AGVs in industrial production and the cargo transfer areas in logistics warehouses.

[0038] Furthermore, in the above step 4,

[0039] The number of environmental feature points, η, is determined by a feature point extraction algorithm. In the warehouse racking environment where AGVs operate, goods are neatly and regularly arranged. After a LiDAR scan, these items appear as a series of breakpoints formed by object edges. Corners are also formed where the shelves meet the ground. Based on the environmental characteristics of the warehouse racking area, a feature extraction algorithm is used to extract these breakpoints and corners.

[0040] The distance λ between the surrounding environment and the AGV is determined by the laser radar ranging algorithm. When the laser radar is used to scan the environment, a set of distance data l is returned in one scanning cycle. i, take the length value corresponding to the point closest to the lidar as the distance between the environment and the lidar in the current frame of data:

[0041] Furthermore, in step 5 above, the number of feature points η obtained in the laser radar scanning environment and the distance value λ between the AGV body and the surrounding objects are input into the support vector machine model to obtain the output identification including:

[0042] Step 5-1: The number of feature points η obtained by the laser radar scanning and the distance value λ between the AGV body and surrounding objects are input into the model as classification features;

[0043] In step 5-2, the two input values ​​of η and λ are compared with the preset values ​​in the support vector machine model. When the η value is less than the corresponding preset value and the λ value is greater than the corresponding preset value, it means that there are fewer obstacles around the AGV and it is in an open environment. At this time, the identification value output by the model is 0, and the CSM algorithm is used. This method is suitable for large-scale environments, such as the cargo transfer area in a logistics warehouse; when the η value is greater than the corresponding preset value and the λ value is less than the corresponding preset value, it means that there are more obstacles around the AGV and it is in a narrow environment. At this time, the identification value output by the model is 1, and the ICP algorithm is used. The range is identified as the shelf area, and its space aisles are relatively narrow.

[0044] Example:

[0045] Step 1: When the AGV is in motion, the lidar continuously scans the surrounding environment.

[0046] Step 2: Scan the surrounding environment to obtain the number of feature points η and the distance value λ of the surrounding objects and bring them into the established model formula (1).

[0047] In step 3, the numerical values ​​obtained by the AGV are fed into the support vector machine model. The model classifies and identifies the following two inputs: 1. A large η value and a small λ value are assigned an identification value of 1; 2. A small η value and a large λ value are assigned an identification value of 0. Different identification values ​​are then matched to different algorithms. Therefore, the present invention uses an adaptive switching method to improve the positioning accuracy of the AGV. The process of this adaptive switching method is as follows:

[0048] Step 3.1: The sensor detection data carried by the AGV is used to substitute the number of feature points η obtained by the laser radar scanning and the distance value λ of the surrounding objects into the established nonlinear model formula (1) for judgment. If the identification value output by the model is 1, it means that there are many obstacles around the AGV and it is in a narrow environment. The host computer switches the matching algorithm to the ICP algorithm, which can make full use of the environmental structural features for matching and avoid the matching error caused by the AGV's wheels slipping in a curved environment, thereby improving the positioning accuracy and speed of the AGV.

[0049] In step 3.2, if the identification value output by the model is 0, it means that there are few obstacles around the AGV and it is in a wide environment. The host computer will switch the matching algorithm to the CSM algorithm to ensure the positioning accuracy of the AGV in a large scene environment.

Claims

1. Laser SLAM inter-frame matching adaptive switching method based on support vector machine, characterized by The method comprises the following steps: Step 1: On the remote monitoring terminal, manually control the AGV to move in the working environment and use the laser SLAM algorithm to build a two-dimensional grid map of the environment; Step 2: Input the position information of the AGV target point and use the path planning algorithm in the navigation algorithm to plan an optimal operation route; Step 3: When the AGV starts running, it synchronously collects data from the inertial measurement unit and the lidar. The number of environmental feature points η and the distance λ between the surrounding environment and the AGV are obtained by processing the surrounding environment scanned by the lidar. Substitute the values ​​of η and λ into the support vector machine model for judgment, and judge the environmental state of the AGV at that moment based on the output values, so as to adaptively switch to the ICP algorithm or CSM algorithm; specifically: Step 3-1: The number of feature points η obtained by the laser radar scanning and the distance value λ between the AGV body and surrounding objects are input into the support vector machine model as classification features; Step 3-2: Compare the two input values ​​of η and λ with the preset values ​​in the support vector machine model. When the η value is less than the corresponding preset value and the λ value is greater than the corresponding preset value, it means that there are few obstacles around the AGV and it is in an open environment. At this time, the identification value output by the support vector machine model is 0, and the CSM algorithm is used. When the η value is greater than the corresponding preset value and the λ value is less than the corresponding preset value, it means that there are many obstacles around the AGV and it is in a narrow environment. At this time, the identification value output by the support vector machine model is 1, and the ICP algorithm is used. The number of environmental feature points η is determined by a feature point extraction algorithm, which is used to extract breakpoints and corner points; the breakpoints are a series of breakpoints formed by the edges of objects between goods, and the corner points are the locations where the shelves touch the ground.

2. The laser SLAM frame matching adaptive switching method based on support vector machine according to claim 1, wherein: The distance λ between the surrounding environment and the AGV is determined by the lidar ranging algorithm. When the lidar is used to scan the environment, a set of distance data will be returned in one scanning cycle. The length value corresponding to the point closest to the lidar is taken as the distance between the environment and the lidar in the current frame of data.

Images