CN113358118A - End-to-end autonomous navigation method for indoor mobile robot in unstructured environment - Google Patents

Abstract

The invention discloses an end-to-end autonomous navigation method for an indoor mobile robot in an unstructured environment. 4) Input the first-view image of the robot into the collision avoidance navigation module to generate collision avoidance waypoints; 5) Use cubic splines to generate collision avoidance paths and track them through model predictive control; 6 ) to determine whether the target point has been reached, if so, end the navigation, otherwise, return to step 3; using this robot autonomous navigation method, in a complex medical environment, only a single monocular RGB camera can be used as visual input to achieve autonomous navigation. The advantages are simplicity, flexibility, economy, practicality, high accuracy and robustness.

Description

End-to-end autonomous navigation method for indoor mobile robots in unstructured environments

technical field

The invention relates to an end-to-end indoor robot autonomous navigation method in an unstructured environment, in particular to an end-to-end autonomous navigation method of a medical service robot in an unstructured environment.

Background technique

With the continuous development of social economy, people's requirements for their own health status and medical service level are increasing year by year. Heavy medical tasks and shortage of medical personnel are urgent problems in the medical field. With the development of modern technology, various mobile robots have emerged as the times require, which can replace or assist medical staff in sample delivery, medical waste recycling and other tasks, reducing the repetitive work of medical staff. This provides a new solution to the problem of heavy work and personnel shortages in the medical field.

The above-mentioned robots operate in a complex and changeable medical environment with a large flow of people. In order to reach the target point safely, accurately and quickly in the "human-machine" shared space, reliable autonomous navigation technology is required. The autonomous navigation of classical mobile robots, such as artificial potential field method and rolling window method, has good performance in fixed environment, but in dynamic environment, its adaptability is weak, reliability is poor, and it is easy to fall into local optimal solution.

For dynamic environments, an end-to-end robot navigation method is currently proposed. The method maps the current visual image of the robot or the detection information of the lidar to the speed and rotation angle of the robot through the neural network. To a certain extent, the problem of poor robot navigation reliability in dynamic environments is solved. But for a complex environment with large flow of people, its real-time collision avoidance effect is unsatisfactory. In view of many of the above factors, medical service robots are still difficult to popularize.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to overcome the defects in the current technology and provide an end-to-end indoor robot autonomous navigation method in a non-structural environment, which is mainly aimed at indoor navigation in large buildings such as hospitals. The main point is to solve the problem of using coefficient map navigation through the processing of the floor plan inside the building; to solve the problem of collision avoidance in a complex and dynamic environment through the end-to-end navigation module.

In order to achieve the above object, the present invention adopts the following technical solutions:

1) Obtain the floor plan and construct the semantic-topology map. With a given target room, an undirected graph is used to search for the shortest node path, and all nodes in this path are set as sub-goals.

2) Obtain the first-view scene image of the robot.

3) Combined with the first-view scene image of the robot obtained in step 2 and the current sub-target, the navigation module outputs the collision avoidance waypoint based on the current scene.

After the step 3, it also includes:

4) Based on the collision avoidance waypoint generated in step 3, the robot generates a collision avoidance path through cubic splines, and uses the model predictive control to track the path to the collision avoidance waypoint.

After the step 4, the following steps are also included:

5) Determine whether the current sub-goal is reached, if yes, go to step 6; if not, go back to step 2.

6) Determine whether the coordinates of the target point are reached, if so, stop the navigation; if not, obtain the current sub-target based on the moment in step 6, and return to step 2.

The creation of the topology map includes the following steps:

A) Obtain the floor plan of the building;

B) Use SegLink to detect the text (room number, corridor number) of the floor plan, and use yolo to detect and identify the logo. The specific results of the detection are: use a rectangular frame to identify the room number, corridor number and identifier of the floor plan (including stairs, safety exits, toilet identifiers), record the position coordinates of the upper left corner of the rectangular frame and the length and width of the rectangular frame; use CRNN to identify the above text;

C) Use semantic segmentation technology to identify the elements of doors, walls, rooms, and corridors on the floor plan. The specific results of the identification are: for the elements that are walls in the floor plan, marked as black, for the elements of the floor plan, the elements of doors are marked as pink, for The elements that are corridors in the floor plan are marked with blue, and the elements that are rooms in the floor plan are marked with a variety of random colors (not the above three colors); the specific implementation methods are as follows:

a) The floor plan is preprocessed into a 512×512 pixel image;

b) Use VGG-encoder to perform feature extraction on the preprocessed image;

c) Use VGG-decoder to decode the features separately, and predict task 1, door, wall, and task 2, room, corridor, and overall network (as shown in Figure 1). The loss function is defined as follows:

w _i =NN _i ;

L _Total =ω ₁ L ₁ +ω ₂ L ₂ ;

Among them, L represents the cross-entropy loss of tasks 1 and 2; C represents the type of prediction type; y _i represents the real type of the pixel point; _pi represents the prediction type of the pixel point; N represents the pixel of the prediction type in this task The total number of points; Ni represents the total number of pixels of the i-th prediction type in the task; L _Total represents the overall cross-entropy loss of task 1 and task 2; N _{Total_i} represents the prediction type of task 1 and task ₂ in a prediction image. total number of pixels.

D) Construction of topology structure, through the following steps:

a) In the image processed by the above semantic segmentation, find all the pixels whose color is pink (door element), store them in the door_list list, and store the pixels in the list into different lists door after classification. The following algorithm:

Step 1: Take out the first pixel in door_list and store it in the list door, and use it as a seed point;

Step 2: Determine whether there are 8 adjacent points of the current seed point in the door_list or whether there are other points except the seed point in the door, if so, go to step 3; if not, go to step 4;

Step 3: Take out the 8 neighboring points of the current seed point in the door_list, store them in the list door, and use the next point of the original seed point in the door as the current seed point, and return to step 2;

Step 4: Determine whether the door_list is empty, if it is empty, end the loop; if it is not empty, create a new list door and return to step 1;

Step 5: Obtain the mean value of the coordinates of all elements in the list door, and use the mean value to represent the coordinates of the door;

b1) Associate the color of the house number and identifier obtained in B, push all the list door as independent elements into the stack, and execute the following algorithm:

Step 1: If the stack is empty, the algorithm ends; otherwise, take the center coordinates of the top element of the stack as the seed point;

Step 2: Obtain the adjacent points in the neighborhood of the seed point 4. If there are non-pink, non-black points and the current color does not exist in the adjacency list, the coordinates are stored in the adjacency list, otherwise the current 4 points are used as new seed points;

Step 3: If the current element is traversed or the length of the adjacency list is equal to 2, create a new adjacency list and return to step 1, otherwise return to step 2;

c1) By looking up the elements contained in each adjacency list, determine the node to which the same door belongs; the nodes that both share the same door are determined to be connected areas, and a topology diagram 1 is constructed.

通过给定置信度α＝0.05，求取均值置信水平为0.95的置信区间：b2) Find the mean value μ ₀ of the number of elements in the list door in a, delete the list whose number of elements is less than μ ₀ /2, and find the mean value μ ₁ and the variance of the number of elements in the remaining list door

By giving a confidence level α=0.05, find the confidence interval with a confidence level of 0.95 for the mean:

Associate the house number and identifier obtained in B with the room color, put all the list doors as independent elements on the stack, and execute the following algorithm:

Step 1: If the stack is empty, the algorithm ends; otherwise, take the center coordinates of the top element of the stack as the seed point;

Step 2: Obtain the neighboring points of the 4th neighborhood of the seed point. If there are non-pink, non-black points and the new list does not have the current color, the coordinates are stored in the new list, otherwise the current 4 points are used as new seed points;

Step 3: If the current element is traversed or the length of the adjacency list is equal to 2, create a new adjacency list and return to step 1, otherwise return to step 2.

c2) by looking up the elements contained in each adjacency list, determine the node to which the same door belongs; the nodes that both share the same door are determined to be connected areas, and a topology diagram 2 is constructed;

The creation of the undirected graph includes the following steps:

A) Obtain real room and door location information;

B) Associate the position information obtained in A with the topological graphs 1 and 2 to construct an undirected graph;

C) Convert the undirected graph 2 into an adjacency matrix A (the horizontal and vertical coordinates represent nodes, and the matrix elements 1 and 0 represent whether they are connected);

D) Construct the Laplacian matrix L from the adjacency matrix A:

L=D-A;

E) If the eigenvalue of the Laplacian matrix L has one and only one 0, use the undirected graph 2, otherwise use the undirected graph 1.

The navigation module includes the following aspects:

A) The main body of the navigation module is composed of the ResNet-50 network and 3 fully connected layers (as shown in Figure 2);

B) Input the ResNet-50 network with the current robot's first-view image, and input the output features together with the target point coordinates and the current speed and angular velocity information into the fully connected layer, and finally output the waypoint position;

C) The above network is trained in a virtual manner as follows:

a) In the virtual environment, the human walking trajectory is simulated by the cubic spline function, and the trajectory is tracked by the model prediction control. The optimization process needs to minimize the following equation:

表示惩罚权重；

表示i时刻人的状态，包括所在位置及转角；

表示人与人的目标点的距离；

表示人与障碍物的惩罚距离；

表示人与障碍物的带符号距离；

表示i时刻人的速度及角速度；in,

represents the penalty weight;

Indicates the state of the person at time i, including the location and corner;

Represents the distance between people and the target point;

Represents the penalty distance between the person and the obstacle;

Indicates the signed distance between the person and the obstacle;

Represents the speed and angular velocity of the person at time i;

b) In the virtual environment, the motion trajectory of the robot is simulated by the cubic spline function, and the trajectory is tracked by the model prediction control. The optimization process needs to minimize the following equation:

表示惩罚权重；

表示i时刻移动机器人的状态，包括所在位置及转角；

表示人与人的目标点的距离；

表示移动机器人与障碍物的惩罚距离；

表示移动机器人与障碍物的带符号距离；

表示移动机器人与人的惩罚距离；

表示i时刻移动机器人距人的带符号距离；in,

represents the penalty weight;

Indicates the state of the mobile robot at time i, including its location and corner;

Represents the distance between people and the target point;

Represents the penalty distance between the mobile robot and the obstacle;

Indicates the signed distance between the mobile robot and the obstacle;

Indicates the penalty distance between the mobile robot and the human;

Represents the signed distance between the mobile robot and the person at time i;

c) Record the robot's current first-view image and current state information in the virtual environment, intercept the path points generated by the model predictive control, take the first-view image and current state information as input, and use the waypoint information as output to conduct supervised learning training ;

The cubic spline function and model predictive control include the following steps:

A) According to the current speed and turning angle, combined with the next way point information, use the cubic spline function to plan the collision avoidance path;

B) The above path is tracked by model predictive control.

The building scene includes a hospital building, a shopping mall building, and a school building.

The invention adopts the deep learning-based autonomous navigation method for a medical service robot in a dynamic environment, and directly uses a monocular RGB camera as a visual input device, which is economical in price, has strong reliability and high stability. The shortest node path based on graph-structure solves the local optimal problem to a large extent; the learning-based navigation module can effectively and accurately predict the trajectory of moving obstacles, and solve the problem of collision avoidance in complex and changeable environments. Through the predictive control drive to ensure the smooth operation of the robot, the cubic spline function ensures the continuity and smoothness of the speed and acceleration before and after the waypoint, and has strong robustness.

Description of drawings

Fig. 1 is the semantic segmentation network structure of the present invention;

Fig. 2 is the network structure of the navigation module of the present invention;

Fig. 3 is the semantic-topological map construction flow chart of the present invention;

Fig. 4 is the indoor undirected graph making process of the present invention;

5 is a schematic diagram of a floor plan in the present invention;

6 is a floor plan topology diagram without adding a junction point in the present invention;

Fig. 7 is the floor plan topology diagram of adding junction point in the present invention;

8 is a schematic diagram of generating the shortest node path in the present invention;

FIG. 9 is a schematic diagram of partial path planning in the present invention.

FIG. 10 is an overall flow chart of the implementation of the method of the present invention.

Detailed ways

In order to enable practitioners in the technical field to understand the present invention more intuitively, the present invention is further described below in conjunction with the embodiments of specific cases, but in practical applications, the present invention is not limited to specific cases.

A deep learning-based autonomous navigation method for a medical service robot in a dynamic environment of the present invention needs to be used in large indoor places such as hospital buildings, and the realized function is to navigate the mobile robot safely, accurately and quickly in this dynamic environment reach the target point.

The above-mentioned users who use this navigation method need to complete the following indoor undirected graph creation steps (as shown in Figure 3):

Obtain the floor plan of a building (such as a hospital) and build a preliminary indoor map (as shown in Figure 4);

Obtain the location information (coordinates) of the room and the room-corridor junction (door);

Topologicalize the floor plan (the steps are shown in Figure 5), and the topological standards are as follows: the room, hall and other structures are regarded as 1 node; the connected corridor is regarded as 1 node; corridor a is truncated by corridor b on one side (T type intersection), corridors a and b are regarded as 1 node; corridor a and corridor b are cut off from each other (crossroads), corridors a and b are regarded as 2 nodes; corridors cut off by rooms and halls are regarded as 2 nodes ; The room-corridor junction is regarded as an edge; the production results are shown in the figure (as shown in Figures 6 and 7),

Associate each node in the topology map with the obtained location information, and represent it in the form of an undirected graph to build a complete indoor undirected map (undirected graph 1, undirected graph 2). Use the Laplace matrix to judge the connectivity of the entire graph; if the undirected graph (undirected graph 1) that uses the threshold method to eliminate the door error interference is disconnected (there are two or more Laplace matrices) 0 characteristic root), you need to use an undirected graph (undirected graph 2) without interference removal.

After storing in the indoor undirected graph, users can also add or delete data nodes as needed.

After completing the above work, if the user wants to use the navigation function in the hospital, he needs to complete the following steps:

1) Given a target room, use an undirected graph to search for the shortest node path, and mark the nodes on the path as sub-targets;

In this step, the user gives the robot's target room (such as A12), combined with the current position (such as A8), uses the undirected graph described in D to search for the shortest node path of A8-A12, and obtains the result (a): A8-C3- H5-C2-A2. Adding the edges gives the following result (b): A8-(A8-C3)-(C3-H5)-(H5-C2)-(C2-A12)-A12. (Note: The edges C3-A8 and A8-C3 in the figure are the same side.) Set the node in (b) as the sub-target (as shown in Figure 8), and the cross symbol in the figure is the location of the sub-target, which is used as the The robot's next navigation basis;

2) Obtain the current first-view scene image of the robot;

The scene image is acquired by the monocular RGB camera set on the front of the robot, without any depth information and map information. The scene image will be used for the third step path planning; (Note: The fixed position of the camera must meet the following conditions: a. The captured scene image can include most of the obstacles in front of the robot; b. The fixed position should be in the robot Relatively high above, such as the head.)

3) The robot outputs the collision avoidance waypoint to the current sub-target through the current first-view image;

In this step, the robot uses the navigation module to output the collision avoidance waypoint through the scene image captured in step 2, through the established sub-target, that is, the position that the robot is about to reach in step 4;

Since the data set of the neural network-based navigation module contains moving obstacles during the training process, the navigation process can smoothly avoid moving obstacles (as shown in Figure 9);

Since the data set of the neural network-based navigation module contains samples disturbed by noise during the training process, the navigation module can output waypoint information stably and effectively even if the scene image is blurred due to poor lighting;

4) Generate a collision avoidance path from the current position to the waypoint position through a third-order spline function according to the current speed and steering angle, and use model predictive control to track the path, so that the robot arrives at the waypoint;

The path generated by the third-order spline function can ensure that the path, speed and acceleration before and after the waypoint are continuous and smooth (as shown in Figure 6), thus ensuring the stable operation of the robot;

5) judge whether to reach the current sub-target, if so, execute the next step; if not, return to step 2;

Since it is usually difficult for the robot to directly navigate to the sub-target through one waypoint output (as shown in Figure 9), it needs multi-step waypoint output until it navigates to the sub-target, and the sub-target point in the navigation map remains unchanged (Note : In Figure 9, ① represents the original position of the robot, ② represents the historical output waypoint, ③ represents the current position of the robot, ④ represents the collision avoidance waypoint output based on the current visual image, ⑤ represents the pedestrian position, ⑥ represents the field of view of the robot, ⑦ represents the A8-C3 sub-target points);

6) Determine whether the coordinates of the target point are reached, if so, stop the navigation; if not, obtain the current sub-target based on the current moment, and return to step 2.

After navigating to the sub-target point in the fourth step, you need to decide whether to stop the navigation by judging whether the sub-target point is the end point (the target room given in step 1); if the sub-target point is not the end point, you need to get the next The sub-target point continues to navigate until it reaches the end point; for example, when the robot reaches the sub-target point A8-C3 in the process of performing the navigation tasks from A8 to A12, and determines that this point is not the end point, it needs to obtain the sub-target point at the current moment (relative to A8 - the next point of C3), perform subtasks A8-C3 to C3-H5;

The above descriptions are only the basic embodiments of the present invention, and for practitioners in the technical field, improvements and developments made without departing from the principles of the present invention should also be regarded as the protection scope of the present invention.

Claims (8)

1. an end-to-end indoor mobile robot autonomous navigation method in an unstructured environment, characterized in that, comprising the following steps: 1) Obtain a floor plan and construct a semantic-topology map; given a target room, use an undirected graph to search for the shortest path, and set the path node as a sub-goal; 2) Obtain the first-view scene image of the robot; 3) Combined with the scene image and the current sub-target in step 2, the navigation module outputs the collision avoidance waypoint. 2. The end-to-end indoor mobile robot autonomous navigation method in the unstructured environment as claimed in claim 1, characterized in that, after the step 3, further comprising: 4) Based on the collision avoidance waypoint in step 3, the robot generates a collision avoidance path through cubic splines, and uses the model predictive control to track the path to the collision avoidance waypoint. 3. The end-to-end indoor mobile robot autonomous navigation method in the unstructured environment as claimed in claim 2, characterized in that, after the step 4, further comprising: 5) judge whether to reach the current sub-target, if so, execute step 6; if not, return to step 2; 6) Determine whether the coordinates of the target point are reached, if so, stop the navigation; if not, obtain the current sub-target based on the moment in step 6, and return to step 2. 4. the end-to-end indoor mobile robot autonomous navigation method in the unstructured environment as claimed in claim 1, is characterized in that, the creation of the semantic-topology map, comprises the following steps: A) Obtain the floor plan of the building; B) Use SegLink to perform text detection on the floor plan, and use yolo to detect and identify the logo. The specific results of the detection are: mark the room number, corridor number and identifier of the floor plan with a rectangular frame, and record the position coordinates of the upper left corner of the rectangular frame and the length and width of the rectangular box; use CRNN to recognize the above text; C) Use semantic segmentation technology to identify the elements of doors, walls, rooms, and corridors on the floor plan. The specific results of the identification are: for the elements that are walls in the floor plan, marked as black, for the elements of the floor plan, the elements of doors are marked as pink, for The elements that are corridors in the floor plan are marked with blue, and the elements that are rooms in the floor plan are marked with various random colors; the specific implementation method is as follows: a) The floor plan is preprocessed into a 512×512 pixel image; b) Use VGG-encoder to perform feature extraction on the preprocessed image; c) Use VGG-decoder to decode the features separately, and predict task 1, door, wall, and task 2, room, corridor, and overall network, respectively. The loss function is defined as follows:

w _i =NN _i ; L _Total =ω ₁ L ₁ +ω ₂ L ₂ ;

Among them, L represents the cross-entropy loss of tasks 1 and 2; C represents the type of prediction type; y _i represents the real type of the pixel point; _pi represents the prediction type of the pixel point; N represents the pixel of the prediction type in this task The total number of points; Ni represents the total number of pixels of the i-th prediction type in the task; L _Total represents the overall cross-entropy loss of task 1 and task 2; N _{Total_i} represents the prediction type of task 1 and task ₂ in a prediction image. total number of pixels. D) Construction of topology structure, through the following steps: a) In the image processed by the above semantic segmentation, find all the pixels whose color is pink, store them in the door_list list, and store the pixels in the list into different lists of doors after classification, and execute the following algorithm: Step 1: Take out the first pixel in door_list and store it in the list door, and use it as a seed point; Step 2: Determine whether there are 8 adjacent points of the current seed point in the door_list or whether there are other points except the seed point in the door, if so, go to step 3; if not, go to step 4; Step 3: Take out the 8 neighboring points of the current seed point in the door_list, store them in the list door, and use the next point of the original seed point in the door as the current seed point, and return to step 2; Step 4: Determine whether door_list is empty, if it is empty, end the loop; if not, create a new list door and return to step 1; Step 5: Obtain the mean value of the coordinates of all elements in the list door, and use the mean value to represent the coordinates of the door; b1) Associate the house number and identifier obtained in B with the room color, put all the list doors as independent elements on the stack, and execute the following algorithm: Step 1: If the stack is empty, the algorithm ends; otherwise, take the center coordinate of the top element of the stack as the seed point; Step 2: Obtain the adjacent points in the neighborhood of the seed point 4. If there are non-pink, non-black points and the current color does not exist in the adjacency list, the coordinates are stored in the adjacency list, otherwise the current 4 points are used as new seed points; Step 3: If the current element is traversed or the length of the adjacency list is equal to 2, create a new adjacency list and return to step 1, otherwise return to step 2; c1) By looking up the elements contained in each adjacency list, determine the node to which the same door belongs; the nodes that both share the same door are determined as a connected area, and a topology diagram 1 is constructed; b2) Find the mean value μ ₀ of the number of elements in the list door in a, delete the list whose number of elements is less than μ ₀ /2, and find the mean value μ ₁ and the variance of the number of elements in the remaining list door

By giving a confidence level α=0.05, find the confidence interval with a confidence level of 0.95 for the mean:

Associate the house number and identifier obtained in B with the room color, put all the list doors as independent elements on the stack, and execute the following algorithm: Step 1: If the stack is empty, the algorithm ends; otherwise, take the center coordinate of the top element of the stack as the seed point; Step 2: Obtain the adjacent points in the neighborhood of the seed point 4. If there are non-pink, non-black points and the current color does not exist in the adjacency list, the coordinates are stored in the adjacency list, otherwise the current 4 points are used as new seed points; Step 3: If the current element is traversed or the length of the adjacency list is equal to 2, create a new adjacency list and return to step 1, otherwise return to step 2; c2) By looking up the elements contained in each adjacency list, determine the node to which the same door belongs; the nodes that both share the same door are determined to be connected areas, and a topology diagram 2 is constructed. 5. The end-to-end indoor mobile robot autonomous navigation method in an unstructured environment as claimed in claim 1, wherein the creation of the undirected graph comprises the following steps: A) Obtain real room and door location information; B) Associate the position information obtained in A with the topological graphs 1 and 2 to construct undirected graphs 1 and 2; C) Convert the undirected graph 2 into an adjacency matrix A (the horizontal and vertical coordinates represent nodes, and the matrix elements 1 and 0 represent whether they are connected); D) Construct the Laplacian matrix L from the adjacency matrix A: L=D-A;

E) If the eigenvalue of the Laplacian matrix L has one and only one 0, use the undirected graph 2, otherwise use the undirected graph 1. 6. The end-to-end indoor mobile robot autonomous navigation method in the unstructured environment as claimed in claim 1, wherein the navigation module comprises the following aspects: A) The main body of the navigation module is composed of the ResNet-50 network and 3 fully connected layers; B) Input the ResNet-50 network with the current robot's first-view image, and input the output features together with the target point coordinates and the current speed and angular velocity information into the fully connected layer, and finally output the waypoint position; C) The above network is trained in a virtual manner as follows: a) In the virtual environment, the human walking trajectory is simulated by the cubic spline function, and the trajectory is tracked by the model prediction control. The optimization process needs to minimize the following equation:

in,

represents the penalty weight;

Indicates the state of the person at time i, including the location and corner;

Represents the distance between people and the target point;

Represents the penalty distance between the person and the obstacle;

Indicates the signed distance between the person and the obstacle;

Represents the speed and angular velocity of the person at time i; b) In the virtual environment, the motion trajectory of the robot is simulated by the cubic spline function, and the trajectory is tracked by the model prediction control. The optimization process needs to minimize the following equation:

in,

represents the penalty weight;

Indicates the state of the mobile robot at time i, including its location and corner;

Represents the distance between people and the target point;

Represents the penalty distance between the mobile robot and the obstacle;

Indicates the signed distance between the mobile robot and the obstacle;

Indicates the penalty distance between the mobile robot and the human;

Represents the signed distance between the mobile robot and the person at time i; c) Record the robot's current first-view image and current state information in the virtual environment, intercept the waypoints generated by model predictive control, and use the first-view image and current state information as input and waypoint information as output to perform supervised learning train. 7. The end-to-end mobile robot autonomous navigation method in the unstructured environment as claimed in claim 2, wherein the cubic spline function and model predictive control comprise the following steps: A) According to the current speed and turning angle, combined with the next way point information, use the cubic spline function to plan the collision avoidance path; B) The above path is tracked by model predictive control. 8. The end-to-end indoor mobile robot autonomous navigation method in an unstructured environment according to any one of claims 1 to 7, wherein the building scene includes a hospital building, a shopping mall building, and a school building.

Images