Laser repositioning method and system based on neural network

Abstract

A laser relocation method based on a neural network comprises the following steps that S1, a robot traverses the whole environment space, environment information is collected by means of a laser radar sensor carried by the robot, and then a complete point cloud map is constructed; s2, the robot continuously works, laser point cloud data and pose information at the current moment are recorded at the same time, then, the laser point cloud data and the pose information at the current moment serve as point cloud key frames to be stored, the operation is continuously carried out till the robot completes traversal of the point cloud map in the step S1, and the point cloud map is obtained; and finally obtaining a point cloud database containing point cloud key frame information. S3, training the robot pose information prediction model according to the point cloud key frame information in the point cloud database, and storing the robot pose information prediction model after training is completed; and S4, repositioning of the robot is realized. The method can effectively improve the repositioning precision and efficiency of the robot, and has a wide application prospect.

Description

Technical Field

[0001] The present invention relates to the technical field of robot relocalization, and specifically relates to a laser relocalization method and system based on a neural network. Background Art

[0002] In many unmanned control systems and mobile robot systems, it is important to estimate the position and attitude of the target for self-relocalization. Traditional attitude estimation methods for relocalization mainly include differential signals (RTK), fixed point position values, lidar template matching, etc. However, at present, the accuracy of relocalization is difficult to meet the requirements, and the stability and redundancy of the robot positioning system need to be improved so that the robot positioning system can output stable positioning results under different conditions. Summary of the Invention

[0003] The purpose of the present invention is to provide a laser relocalization method and system based on a neural network, which can effectively solve the technical problems mentioned in the background art.

[0004] To achieve the above purpose, the present invention provides the following technical solutions: A laser relocalization method based on a neural network, which includes the following steps: S1. The robot traverses the entire environmental space, collects environmental information by means of its own mounted lidar sensor, and then constructs a complete point cloud map. S2. The robot keeps working, and at the same time records the laser point cloud data and pose information at the current moment. Subsequently, the laser point cloud data and its pose information at the current moment are saved as point cloud key frames. Such operations are continuously carried out until the robot completes the traversal of the point cloud map in the above step S1, and finally a point cloud database containing point cloud key frame information is obtained. S3. Using the data in the point cloud database as training data, train the neural network to obtain a robot pose information prediction model. Train the robot pose information prediction model according to the point cloud key frame information in the point cloud database, and save the robot pose information prediction model after training. The steps of training the robot pose information prediction model include: extracting the point cloud key frames and the pose information of the point cloud key frames in the point cloud database, projecting the point cloud key frames into two-dimensional pictures, and using the two-dimensional pictures projected from the point cloud key frames as the input of the neural network. At the same time, using the point cloud key frame pose information as supervised data, train the robot pose information prediction model, and save the trained robot pose information prediction model. S4. Realize robot relocalization, and the steps of relocalization include: S41. When the robot loses its positioning, acquire the point cloud data at the current moment, convert the point cloud data into a 2D image through 2D projection, and then input it into the robot pose information prediction model to obtain the global pose information of the robot. Use the relative rotation variable as the initial value of the iterative closest point algorithm to calculate the relative pose transformation between the actual laser point cloud and the key frame point cloud, and multiply the relative pose transformation by the key frame point cloud to obtain the global pose. S42. Input the global pose into the positioning node, calculate the matching error of the global pose through the positioning node, compare the matching error with the preset error threshold. If the matching error is less than the preset error threshold, the relocalization is successful; otherwise, the relocalization fails.

[0005] A laser relocalization system based on a neural network, comprising: A mapping module, which is used to drive the robot to traverse the entire environmental space, collect environmental information through the lidar sensor on the robot, and construct a complete point cloud map. A point cloud database construction module, which records the laser point cloud data and pose information at the current moment during the operation of the robot, saves the laser point cloud data and its pose information at the current moment as a point cloud key frame. During the movement of the robot, the point cloud database construction module continuously updates the key frame information to ensure that the data in the point cloud database can accurately reflect the changes in the environment where the robot is located until the robot traverses the point cloud map obtained by the mapping module to obtain a point cloud database containing point cloud key frame information. A robot pose prediction module, which acquires the point cloud data of the robot at the current moment, projects the point cloud data of the robot at the current moment into a 2D image and inputs it into the robot pose information prediction model to obtain the pose information of the robot. A relocalization module, which calculates the global pose information of the robot through the pose information of the robot obtained by the robot pose information prediction model to achieve relocalization.

[0006] Preferably, when the robot loses its positioning, the relocalization module acquires the point cloud data at the current pose, projects a frame of point cloud data at the current pose into a 2D picture and inputs it into the robot pose information prediction model to obtain the global pose information of the robot; uses the relative rotation variable as the initial value of the iterative closest point algorithm to calculate the relative pose transformation between the actual laser point cloud and the key frame point cloud, multiplies the relative pose transformation by the key frame point cloud to obtain the global pose; inputs the global pose into the positioning node, and the positioning node calculates the matching error of the global pose, compares the matching error with the preset error threshold. If the matching error is less than the preset error threshold, the relocalization is successful; otherwise, the relocalization fails.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention processes the key frame data of the laser through a neural network to obtain the relative pose information of the carrier, assisting the unmanned system and the map to achieve precise positioning; improving the stability and redundancy of the robot positioning system, enabling the robot positioning system to output stable positioning results under different conditions; therefore, the present invention can effectively improve the repositioning accuracy and efficiency of the robot and has a wide range of application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 is a flowchart of a laser repositioning method based on a neural network in the present invention; Figure 2 is a flowchart of a laser repositioning system based on a neural network in the present invention; Figure 3 is a flowchart of the pose prediction of a laser robot based on a neural network in the present invention; Figure 4 is a flowchart of the global pose of repositioning in the present invention; DETAILED DESCRIPTION OF THE INVENTION

[0009] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0010] With the great development of neural network technology and the improvement of computer performance in recent years, using the neural network method for repositioning has become a future trend; using a neural network for robot repositioning can give full play to the advantages of the neural network and the laser sensor, obtain highly generalized and robust repositioning information, and can work in an environment with high similarity and lack of RTK, serving as a new positioning supplement means for repositioning; Please refer to Figures 1-4 as shown, a laser repositioning method based on a neural network includes the following steps: S1. The robot traverses the entire environmental space, collects environmental information with the lidar sensor carried by itself, and then constructs a complete point cloud map; S2. The robot keeps working while recording the laser point cloud data and pose information at the current moment. Subsequently, the laser point cloud data and its pose information at the current moment are saved as key point cloud frames. For example, at regular time intervals, the robot records the current laser point cloud data and its own position and pose information, and stores these data. Such operations continue until the robot completes the traversal of the point cloud map in step S1 above, and finally obtains a point cloud database containing key point cloud frame information. S3. Use the data in the point cloud database as training data to train the neural network, and then obtain a robot pose information prediction model. Train the robot pose information prediction model based on the key point cloud frame information in the point cloud database, and save the robot pose information prediction model after training. The steps of training the robot pose information prediction model include: extracting the key point cloud frames and the pose information of the key point cloud frames in the point cloud database, projecting the key point cloud frames into two-dimensional images, and using the two-dimensional images projected from the key point cloud frames as the input of the neural network. At the same time, use the key point cloud frame pose information as supervised data. Continuously adjust the parameters of the neural network through a large amount of training data to improve the prediction accuracy of the model. Train the robot pose information prediction model and save the trained robot pose information prediction model for calling when needed. S4. Implement robot relocalization. The steps of relocalization include: S41. When the robot loses its position, obtain its point cloud data at the current moment. After converting the point cloud data into a two-dimensional image through two-dimensional projection, input it into the robot pose information prediction model to obtain the global pose information of the robot and achieve relocalization. For example, in a complex environment, due to interference or other reasons, the robot loses its position. At this time, by obtaining the current point cloud data and using the trained robot pose information prediction model for prediction, the position and pose of the robot can be re-determined. Use the relative rotation variable as the initial value of the iterative closest point algorithm, calculate the relative pose transformation between the actual laser point cloud and the key frame point cloud, and multiply the relative pose transformation by the key frame point cloud to obtain the global pose. S42. Transmit the global pose to the localization node, calculate the matching error of the global pose through the localization node, and compare the matching error with a preset error threshold. If the matching error is less than the preset error threshold, the relocalization is successful; otherwise, the relocalization fails.

[0011] A laser relocalization system based on a neural network, including: Mapping module, which is used to drive the robot to traverse the entire environment space, collect environmental information through the lidar sensor on the robot, and construct a complete point cloud map; For example, in an unknown environment, the mapping module controls the robot to explore, and at the same time uses the lidar to obtain the three-dimensional information of the environment, and gradually constructs a complete point cloud map; Point cloud database construction module, which records the laser point cloud data and pose information at the current moment during the operation of the robot, saves the laser point cloud data and its pose information at the current moment as point cloud key frames, and constructs a point cloud database until the robot traverses the point cloud map obtained by the mapping module, and obtains a point cloud database containing point cloud key frame information; For example, as the robot moves, the point cloud database construction module continuously updates the key frame information to ensure that the data in the database can accurately reflect the changes in the environment where the robot is located; Robot pose prediction module, which obtains the point cloud data of the robot at the current moment, projects the point cloud data of the robot at the current moment into a two-dimensional image and inputs it into the robot pose information prediction model to obtain the pose information of the robot; For example, in the case of lost positioning, the robot pose prediction module responds quickly, obtains the current point cloud data, and predicts through two-dimensional projection and the robot pose information prediction model to provide accurate pose information for the robot; Relocalization module, which calculates the global pose information of the robot through the pose information of the robot obtained by the robot pose information prediction model to achieve relocalization.

[0012] When the robot has a positioning loss, the relocalization module obtains the point cloud data at the current pose, projects a frame of point cloud data at the current pose into a two-dimensional picture and inputs it into the robot pose information prediction model to obtain the global pose information of the robot; uses the relative rotation variable as the initial value of the iterative closest point algorithm, calculates the relative pose transformation between the actual laser point cloud and the key frame point cloud, multiplies the relative pose transformation by the key frame point cloud to obtain the global pose; transmits the global pose to the positioning node, the positioning node calculates the matching error of the global pose, compares the matching error with the preset error threshold, if the matching error is less than the preset error threshold, the relocalization is successful, otherwise, the relocalization fails.

[0013] The present invention processes the laser key frame data through a neural network to obtain the relative pose information of the carrier, assisting the unmanned system and the map to achieve precise positioning; improves the stability and redundancy of the robot positioning system, enabling the robot positioning system to output stable positioning results under different conditions; therefore, the present invention can effectively improve the relocalization accuracy and efficiency of the robot and has a wide range of application prospects.

[0014] The above content is only an example and illustration of the structure of the present invention. Those skilled in the art of this technology can make various modifications, supplements, or use similar methods for substitution to the specific embodiments described. As long as it does not deviate from the structure of the present invention or exceed the scope defined by this claims, it shall fall within the protection scope of the present invention.

Claims

1. A laser relocation method based on a neural network, characterized in that: The following steps are involved: S1. The robot traverses the entire environment space, collects environmental information with the help of its own laser radar sensor, and then builds a complete point cloud map; S2, the robot continues to work, and records the laser point cloud data and posture information at the current moment, and then saves the laser point cloud data and its posture information at the current moment as point cloud key frames, and this operation continues until the robot completes the traversal of the point cloud map in the above step S1, and finally obtains a point cloud database containing point cloud key frame information; S3, using the data in the point cloud database as training data to train the neural network, and then obtain a robot posture information prediction model; According to the point cloud key frame information in the point cloud database, the robot posture information prediction model is trained, and after the training is completed, the robot posture information prediction model is saved; The steps of training the robot posture information prediction model include: extracting point cloud key frames and posture information of the point cloud key frames in the point cloud database, projecting the point cloud key frames into two-dimensional images, and using the two-dimensional images projected from the point cloud key frames as input of the neural network, and using the point cloud key frame posture information as supervision data to train the robot posture information prediction model, and saving the trained robot posture information prediction model; S4. Realize robot repositioning. The steps of repositioning include: S41, when the robot loses its positioning, obtain its point cloud data at the current moment, and convert the point cloud data into a two-dimensional image through two-dimensional projection, and input the image into the robot posture information prediction model to obtain the robot's global posture information; The relative rotation variable is used as the initial value of the iterative closest point algorithm to calculate the relative pose transformation of the actual laser point cloud and the key frame point cloud, and the relative pose transformation is multiplied by the key frame point cloud to obtain the global pose; S42, the global posture is transmitted to the positioning node, the matching error of the global posture is calculated by the positioning node, and the matching error is compared with a preset error threshold. If the matching error is less than the preset error threshold, the repositioning is successful, otherwise, the repositioning fails.

2. A laser relocation system based on neural network, characterized in that ,include: A mapping module, which is used to drive the robot to traverse the entire environment space, collect environmental information through the laser radar sensor on the robot, and build a complete point cloud map; A point cloud database construction module, which records the laser point cloud data and posture information at the current moment during the robot's operation, and saves the laser point cloud data and its posture information at the current moment as a point cloud key frame. During the movement of the robot, the point cloud database construction module continuously updates the key frame information to ensure that the data in the point cloud database can accurately reflect the changes in the robot's environment, until the robot traverses the point cloud map obtained by the mapping module to obtain a point cloud database containing point cloud key frame information; A robot posture prediction module, wherein the robot posture prediction module obtains point cloud data of the robot at the current moment, projects the point cloud data of the robot at the current moment into a two-dimensional image, and inputs the image into a robot posture information prediction model to obtain the robot's posture information; A repositioning module calculates the robot's global posture information through the robot's posture information obtained by the robot's posture information prediction model to achieve repositioning.

3. The laser relocation system based on neural network according to claim 2, characterized in that ,When the robot loses its positioning, the repositioning module obtains the point cloud data under the current ,pose, and then projects a frame of point cloud data under the current ,pose into a two-dimensional picture and inputs it into the robot ,pose information prediction model to obtain the robot’s ,global pose information; The relative rotation variable is used as the initial value of the iterative closest point algorithm, and the relative pose transformation between the actual laser point cloud and the key frame point cloud is calculated. The relative pose transformation is multiplied by the key frame point cloud to obtain the global pose; the global pose is passed to the positioning node, and the positioning node calculates the matching error of the global pose and compares the matching error with the preset error threshold. If the matching error is less than the preset error threshold, the repositioning is successful, otherwise, the repositioning fails.

Images