KR20130046741A - Autonomous topological mapping method of medium-large size space using upward single camera - Google Patents

Abstract

PURPOSE: An autonomous phase map generating method in a medium and large sized space using a bottom-up signal camera is provided to reduce map making errors caused by the height of a ceiling by extracting a landmark not influenced by the height of the ceiling. CONSTITUTION: A landmark in a current position of a robot extracted from an input image is matched with a preregistered landmark to obtain landmark difference information(S10). Sub nodes in a predetermined distance from a base node are generated to be registered in a phase map(S20). Landmark difference information is applied to a Kalman filter to correct positions of the registered nodes(S30). [Reference numerals] (AA) Start; (BB) End; (S10) Obtain landmark difference information by matching a landmark in the current position of a robot extracted from an input image with a preregistered landmark; (S20) Generate sub nodes in a predetermined distance from a base node through self-driving of the robot and register in a phase map; (S30) Correct the position of the robot and the positions of the nodes registered in the phase map by applying the landmark difference information to a Kalman filter

Description

AUTOOMOUS TOPOLOGICAL MAPPING METHOD OF MEDIUM-LARGE SIZE SPACE USING UPWARD SINGLE CAMERA}

The present invention relates to a method capable of simultaneously performing autonomous phase mapping and position recognition in a medium-large space using a bottom-up single camera. More specifically, the medium-to-large space has various types of static and dynamic obstacles and various types of environmental structures due to its characteristics, and the present invention uses a camera and a laser scanner together to cope with such environments. In the generation process, we propose a method to create a more intelligent and efficient environment topology map by performing autonomous environment search and obstacle recognition process simultaneously without user setting and manipulation.

Simultaneous Localization and Mapping (SLAM) technology of mobile robots has been studied in recent 20 years. In the early stages of research, mapping methods using extended Kalman Filters (EKF) based on distance sensors such as ultrasonic sensors and laser scanners were mainly used. Recently, mapping methods using vision sensors such as cameras have been actively studied. have.

Unlike distance sensors, vision sensors can acquire a wealth of environmental information, which greatly assists in data association processing, which is very important in SLAM. In particular, the price-performance ratio is excellent, and it can be applied directly to a mobile robot without any special device. Representative researches using single cameras in the SLAM field include Monocular SLAM (MonoSLAM) and Ceiling-View SLAM (CV-SLAM). MonoSLAM is also used for location recognition of humanoid robots, and CV-SLAM is a real-world robot cleaner. As a major example of commercialization technology applied to the product, it played a major role in increasing the efficiency of existing robot cleaning.

Most of the above-mentioned conventional technologies deal with mapping in a narrow area of a general home environment or indoors, and thus have recently played a role of information / promotion, security services, etc. in medium / large spaces such as building lobby, museum, and exhibition hall where the robot is highly utilized. It is difficult to apply existing technology due to environmental variables and functional constraints in applying technology to performing robots. In general, a medium-large space can be defined as a large space that cannot be detected by a 4m laser scanner, a space with a high ceiling (10m or more in height), and a lot of dynamic obstacles. In such a space, the ceiling is very high and has various ceiling patterns and structures. Therefore, if the existing CV-SLAM algorithm is applied as it is, it is difficult to accurately measure the height of the feature points required for location recognition, which causes problems in the mapping.

In the prior art, the method of mapping through the manual operation of the robot in the target environment is mainly used, but since the medium-large space has a large space and many dynamic obstacles, the manual mapping method by the user is difficult to maintain the consistency of work. Various environmental variables that occur frequently, for example, there is a problem that the efficiency of the work is reduced by causing significant time delay and human / physical costs due to obstacles, lighting changes, and the like.

In order to solve these problems, the present invention proposes an automatic phase mapping algorithm for medium and large spaces based on a driving algorithm through autonomous environment exploration / recognition, thereby maximizing the work efficiency and simplifying the overall mapping process. In the application, we propose a method that has excellent mapping ability without any limitation.

In addition, the present invention minimizes the H / W specification required for the mapping for the smooth solution to problems such as system load increase and cost, and at the same time the image processing process that is robust to environmental changes generated by using a vision sensor The software and hardware specifications necessary for the execution of the proposed mapping algorithm are minimized by minimizing the amount of computation required and efficiently managing the mapping process such as feature point extraction and node registration / deletion, which are inevitably frequently occurred during the mapping of large spaces. We propose a method to minimize the problem.

More specifically, the present invention provides a method for generating an autonomous phase map in a medium-to-large space using a bottom-up single camera that enables a robot to autonomously search for an environment and create a phase map of a given environment without requiring any manual operation by a user. Let it be technical problem.

In addition, the present invention provides a method for generating autonomous phase maps in a medium-to-large space using a bottom-up single camera capable of extracting and matching landmarks that are not affected by the height of the ceiling in the input image, thereby reducing the mapping error caused by the ceiling height. To provide a technical problem.

In addition, the present invention provides a method for generating an autonomous phase map in a medium-to-large space using a bottom-up single camera to generate a avoidance path autonomously when the obstacle is detected while driving the robot to safely travel without hitting the surrounding people or structures. It is technical problem to do.

In addition, the present invention allows the robot to follow the nodes registered at a certain distance on the map to correct the position of the robot each time it reaches the node so that the robot does not lose its own position while driving. It is a technical problem to provide a method for generating an autonomous phase map in a medium-to-large space using a bottom-up single camera which is particularly suitable for a service.

In addition, the present invention is to map the map created through the autonomous driving of the robot to have the form of a phase map, it is possible to reduce the amount of data stored in the map and the amount of calculation required to update the map, thereby making it possible to reduce the weight of the algorithm, low cost It is a technical task to provide a method for generating an autonomous phase map in a medium-to-large space using a bottom-up single camera that can be mounted on an embedded board.

In order to solve this problem, the method for generating an autonomous phase map in a medium-large space using a bottom-up single camera according to the present invention is an input image that is a ceiling image of the medium-large space obtained by a bottom-up single camera installed in a robot that autonomously runs the medium-large space. A land that obtains landmark difference information which is a difference between a landmark at the current location of the robot and a registered landmark by matching a landmark at a current location of the robot extracted from the landmark with a previously registered landmark Mark extraction / matching step, node for acquiring surrounding environment information using front and rear laser scanner attached to the robot, and generating auxiliary nodes away from the base node which is the starting point of the robot through autonomous driving and registering them in the phase map Creation / Registration Step and the Landmark Difference The information characterized in that it comprises a position correcting step of applying to correct the position of the node registered with the positions and the phase map of the robot in extended Kalman filter (Extended Kalman Filter).

The landmark is a set of edge points obtained by calculating gradients in the x and y-axis directions and their sizes from the input image having a gray level and sampling a pixel having a value greater than or equal to a predetermined size. It is composed of an edge point list (Edge point list), the landmark is defined using the two-dimensional information of the input image in association with the current position of the robot, and registered with the landmark at the current position of the robot The matching method between landmarks is characterized by obtaining optimal matching information using an iterative closest point (ICP) algorithm.

The basic element of the phase map is composed of an edge which is a connection element between nodes, and among the basic elements of the phase map, the node is an auxiliary node used for the position recognition of the robot and the shortest distance path planning, and the phase map. A boundary node forming a boundary of the, characterized in that it comprises a base node used as a starting point when the robot is running for the environment search.

The newly created node is selected as the location having the best suitability in consideration of the local map generated based on the current location of the robot and the location information of the node registered to date.

The node creation / registration step includes a search step and a use step. In the search step, the robot registers N auxiliary nodes starting from the base node, and in the use step, the robot returns to the base node, which is an initial starting point. The position of the robot and the position of the nodes are corrected by using the landmark difference information, and the shape of the generated phase map has a form of repetition and gradual expansion of polygonal structures including N + 1 nodes.

When the robot can no longer generate a new auxiliary node based on the base node currently running, the node having the lowest uncertainty among the boundary nodes forming the boundary of the phase map is selected as the next base node, and all boundary nodes are selected. When assuming a base node, when the new secondary node can no longer be generated, the creation of the phase map is finished.

When N, the secondary node structure generation condition, is 2, the nodes generated in the topology map are connected to each other using a Delaunay triangulation algorithm to form an edge, and each time a new node is additionally registered. The edge connection process is performed, and the connection strength between nodes has a value between 0 and 1 according to the characteristics of the surrounding environment, and the optimal path planning and autonomous environment search of the robot are performed by using the connection information between the nodes, and the node position is determined. The robot performs an exploration process on the entire space while radially expanding the area around the reference base node for the entire medium and large space exploration, and autonomously avoids obstacles when an obstacle is detected during the movement between nodes. It characterized in that to perform.

The robot, which performs optimal path selection and driving based on edge connection information between nodes and its strength, detects the surrounding environment using a laser scanner mounted on the robot, and if there is an obstacle around the edge, The edge update process to gradually reduce the strength of the connection and increase the connection strength when there are no obstacles around the edge, and when the connection strength of the edge is less than the threshold, the edge connecting the two nodes. Characterized in that the path management process for processing to be disconnected.

According to the present invention, there is an effect that the robot can autonomously search for an environment and create a phase map of a given environment without any manual operation of a user.

In addition, it is possible to extract and match landmarks that are not affected by the height of the ceiling in the input image, thereby reducing the mapping error due to the height of the ceiling.

In addition, when obstacles are detected while the robot is running, an autonomous path is generated autonomously to safely drive without hitting the surrounding people or structures.

In addition, since the robot travels along nodes registered at a certain distance on the map and corrects the position of the robot each time it reaches the node, the robot does not lose its position while driving, so the robot mainly serves in the medium-to-large space. There is an effect that can be applied directly to the position recognition method of.

In addition, since maps created through autonomous driving of robots have the form of a phase map, the amount of data stored in the map and the amount of calculation required for map updating can be reduced, and accordingly, the algorithm can be reduced in weight, thereby providing a low-cost embedded board. There is an effect that can be mounted on.

1 is a diagram schematically illustrating a robot system in which an autonomous phase map generation method is performed in a medium-large space using a bottom-up single camera according to an embodiment of the present invention.
FIG. 2 is a diagram illustrating data input / output in a robot system in which an autonomous phase map generation method is performed in a medium-large space using a bottom-up single camera according to an embodiment of the present invention.
FIG. 3 is a diagram illustrating exploration and exploitation applied in a process of creating and registering a node in an embodiment of the present invention.
4 is a diagram illustrating a shape of a phase map according to the number of auxiliary nodes according to an embodiment of the present invention.
5 is a diagram illustrating a method of generating an autonomous phase map in a medium-large space using a bottom-up single camera according to an embodiment of the present invention.
FIG. 6 is a diagram illustrating a process of obtaining a gradient image from an input image and defining one landmark by sampling an edge point according to an embodiment of the present invention.
FIG. 7 illustrates a process of extracting one landmark from each of two images and then matching them using an iterative closed point (ICP) algorithm to obtain landmark difference information as a result. It is for the drawing.
8 is a diagram illustrating an example of a node creation and registration process in an embodiment of the present invention.
FIG. 9 is a diagram illustrating the types and features of each node constituting a phase map in an embodiment of the present invention.
FIG. 10 is a view showing an image of an environment around a robot detected by the front and rear laser scanners, and calculating a fitness value for a candidate position of a new node according to an embodiment of the present invention.
FIG. 11 is a diagram illustrating a local map and a generated obstacle avoidance path when an obstacle is detected by a laser scanner while the robot is driving according to an embodiment of the present disclosure.
12 is a diagram showing the results of experiments in Experimental Environment 1. FIG.
FIG. 13 is a diagram showing the results of experiments in Experimental Environment 2. FIG.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a diagram schematically illustrating a robot system in which an autonomous phase map generation method is performed in a medium-large space using a bottom-up single camera according to an embodiment of the present invention, and FIG. 2 is a bottom-up single camera according to an embodiment of the present invention. Is a diagram illustrating data input / output in a robot system in which an autonomous phase map generation method in a medium-to-large space is performed.

1 and 2, the robot system in which the present invention is performed is largely composed of a bottom-up single camera, a front and rear laser scanner, a robot platform, and a SLAM (Simultaneous Localization and Mapping (SLAM) module. The bottom-up single camera is installed to face upward of the robot and is used to extract the landmark image from the large and large space ceiling in real time while the robot is driving. Used for mapping and obstacle avoidance, the robot platform is composed of a robot body and a mobile platform, and controls the movement of the robot and provides data for predicting the moving distance. The moving distance of the robot can be measured by counting the number of revolutions of the wheel by mounting an encoder on the wheel of the robot.

Before describing a method of generating an autonomous phase map in a medium-large space using a bottom-up single camera according to an embodiment of the present invention, the principle underlying the present invention will be described.

Algorithms applied to the present invention are largely divided into two parts: phase map generation and exploration through exploration and exploitation strategies.

The phase map generation algorithm through exploration and exploitation strategy is a detailed algorithm for location correction through landmark extraction and matching, node registration, and Extended Kalman Filter (EKF) SLAM. The autonomous driving is divided into the normal driving mode and the obstacle avoidance mode.

<Phase Map Generation through Exploration and Exploitation Strategy>

1. Explore and Use Strategies

Acquiring and registering environmental information through autonomous driving is an essential function for a robot operating in an unknown environment. The present invention relates to an algorithm for progressively searching the unexplored area through the strategy of searching and using, and at the same time accurately predicting the magnetic position of the robot.

1.1 Exploration

First, referring to FIG. 3A, a process of registering N nodes, that is, an auxiliary node, a predetermined distance away from a base node, that is, a base node of a search, is performed. The newly registered auxiliary node and the base node should be separated from each other by a certain distance, and maintained at a certain distance from the surrounding obstacles. The base node and the N secondary nodes form a polygon with (N + 1) sides, where the secondary node is registered such that the minimum angle of the polygon's cabinet is close to 180 * (N-1) / (N + 1). do. This is to make the length of one side of the polygon as the same as possible.The longer the length of the side, i.e., the smaller the minimum value of the cabinet, is the section that the robot must drive using only the odometry data provided by the mobile platform. This becomes long, and in this case, the error about the position of the robot becomes large.

1.2. Exploitation

Referring to (b) of FIG. 3, the robot registers N auxiliary nodes and then returns to the base node again. Since the robot predicts the node and its position using only driving data while registering the auxiliary node, an error occurs from the actual position. In order to reduce this error, it is determined that the robot arrives at the base node correctly using a single bottom-up camera, which is a vision sensor mounted on the robot, and the position and node of the final robot are determined by using the predicted position of the robot and the difference of the position of the base node. Correct the position of.

1.3. Structure of the phase map

Referring to FIG. 4, assuming that the number of auxiliary nodes is N, the prepared phase map has a structure in which (N + 1) square patterns are continuous. Each node has a landmark that describes the node, which can be defined differently depending on the sensor used.

In consideration of the accuracy of the driving data and the efficiency of the mapping time, it is determined that N = 2 is optimal. Therefore, the following detailed description assumes N = 2.

5 is a diagram illustrating a method of generating an autonomous phase map in a medium-large space using a bottom-up single camera according to an embodiment of the present invention.

Referring to FIG. 5, an autonomous phase map generation method in a medium-large space using a bottom-up single camera according to an embodiment of the present invention includes a landmark extraction / matching step S10, a node generation / registration step S20, and a position correction It comprises a step S30.

In the landmark extraction / matching step (S10), a process of acquiring landmark difference information, which is a difference between two landmarks, by matching a landmark at a current position of a robot extracted from an input image with a landmark registered in advance This is done. The input image is a ceiling image of a medium-large space obtained by a single bottom-up camera installed in a robot that autonomously runs a medium-large space.

For example, a landmark calculates a gradient and its magnitude in the x- and y-axis directions from an input image having a gray level, and calculates a gradient of edge points obtained by sampling a pixel having a predetermined value or more. It can be composed of an edge point list, which is a set, and these landmarks can be defined by using two-dimensional information of the input image in association with the current position of the robot, and the landmarks at the current position of the robot Matching of the registered landmarks may be performed through an iterative closest point (ICP) algorithm.

The extraction and matching process of such landmarks will be described in more detail as follows.

Since the present embodiment uses a bottom-up single camera, nodes registered on the phase map are associated with landmarks extracted from the ceiling image obtained at the location. In order to take into account changes in various image conditions that occur when using a bottom-up single camera as a vision sensor, an edge point set that is not sensitive to changes in illumination is used without using raw data of the acquired image. The landmark is defined by extraction, and the matching degree between the two landmarks, that is, the landmark at the current position of the robot and the registered landmark, is calculated using the ICP algorithm.

First, an example of a process of extracting a landmark will be described with reference to FIG. 6.

6, reference numeral (a) is an input image, (b) is a gradient image, and (c) an edge point is sampled and defined.

To extract a landmark from the currently input image, a gradient image is obtained first. The gradient is calculated in each direction of the x and y axes of the image coordinate system, and the magnitude is squared to calculate the magnitude at each pixel. Extracting a pixel having a magnitude value greater than or equal to a predetermined value increases the number of extracted points, and thus extracts only pixels having a local maximum of magnitude. To simplify local maximum calculations, we first create an integral image. The extraction condition is shown in Equation 1 below.

)으로 정의되며, 도 6의 도면부호 (c)는 이 랜드마크를 나타낸다.M (i, j) is the magnitude in pixel i, j and S is the window size (window_size) of the local neighbor (nh). Therefore, if the value of v is greater than or equal to a predetermined threshold, the pixels i and j are extracted and registered as one edge point forming a landmark. The final landmark extracted from the image is a set of edge points with N _p points (

And reference numeral (c) of FIG. 6 denotes this landmark.

Next, an example of a landmark matching process will be described with reference to FIG. 7.

Referring to FIG. 7, a registered landmark is a set of edge points. Accordingly, it is possible to determine matching between the currently extracted landmark P1 and the registered landmark P2 through point-to-point matching. In this embodiment, two landmarks are matched using an iterative closest point (ICP) matching technique. Table 1 below summarizes the ICP matching process.

Input (Two Point Sets; P1, P2)
e = infinity
iteration = 0
Repeat until e <eth or iteration> Niter
1. Find the point in P2 that is closest to all points in P1.
2. If distance is more than certain, exclude from correspondence.
3. Rotation and translation components are obtained using two matched point sets (P1 ', P2').
4. After conversion, add the Euclidean distance error (e) for each correspondence.
5. Increase iteration
Output (Error, dx, dy, dth)

When the ICP matching error is within a certain value, it is determined that the two landmarks match. In addition, as a result of ICP matching, the rotation degree (dth) of the two landmarks and the movement degree (dx, dy) of the x and y axes can be known, and this information is later used for visual servoing of the robot in the use step. do.

In the node generation / registration step (S20), a process of generating auxiliary nodes away from a base node, which is a starting point of the robot, by a predetermined distance, and registering them on a phase map through autonomous driving of the robot is performed.

For example, referring to FIG. 8, the node generation / registration step S20 is composed of a search step and a use step. In the search step, the robot registers N auxiliary nodes starting from the base node. It may be configured to return to the base node to correct the position of the robot and the position of the nodes by using the landmark difference information, and the shape of the generated phase map may be configured to have a repeating pattern of a polygonal structure having N + 1 sides. Can be. In addition, the newly created node may be configured to be selected as the location having the best suitability in consideration of the local map generated based on the current location of the robot and the location information of the node registered so far.

The generation and registration algorithm of such a node will be described in detail as follows.

On the basis of the current base node, registration of a new auxiliary node is registered after selecting an optimal candidate position (Maximum fitness node) using distance information with previously registered nodes and distance information with obstacles. After registering N nodes (in this embodiment, based on the case where N is 2) instead of registering a new node indefinitely, the reason for returning to the base node again is for periodic loop closing. This driving strategy is to accurately predict the position of the robot and the position of the node.

First, the components of the phase map in the present embodiment will be described.

FIG. 9 is a diagram illustrating the types and features of each node constituting a phase map in an embodiment of the present invention.

Referring to FIG. 9, a phase map created through autonomous driving of a robot is represented as a three-dimensional environment on a personal computer (PC) to aid visual understanding, and the detailed items are the location of the node, its uncertainty, and the connection of the nodes. It consists of information, attributes of a node, and landmarks associated with the node. Description of the components of such a phase map is shown in Table 2 below.

Component Explanation The location of the node 1. Mean position (x _i , y _i , th _i ) of each predicted node in the global coordinate system
2. The direction th _i of the node matches the direction of the associated landmark Node Uncertainty 1.3 expressed as 3 by 3 matrix (P _i )
2. Diagonal term expresses the degree of uncertainty Node's connection information 1. Registered nodes are connected to each other using Delaunay Triangulation algorithm
2. Using node connection information for shortest path planning and automatic environment search
3. Update node position using node-to-node correlation The attributes of the node Each node has one of the following properties: normal, boundary, and base
2. General Node-Used for location recognition and shortest path planning
3. Boundary node-Node that forms the boundary of the map, can be set as the starting point of environment search
4. Base node-the starting node of the current environment search land mark 1. Extracting landmarks using images obtained from the location of the robot at the time of node registration
2. Each node constituting the map has one landmark
3. Landmarks consist of an edge point set
4. The edge point is defined as (x, y) on the image

Next, the generation process of the node will be described in more detail.

As briefly described above, the newly created node is selected as the location having the best suitability in consideration of the local map generated based on the current location of the robot and the location information of the node registered so far. That is, the secondary node to be newly registered in the search phase is selected based on the maximum fitness selection rule.

This will be described in more detail as follows.

(a) Selecting the maximum fitness node

Two factors are considered when selecting a new node. The first is that it should be more than a certain distance from the previously registered nodes, and the second is that the distance to the obstacle closest to the candidate node should be more than a certain distance. This condition is expressed by the following equation.

In Equation 2, fitness is goodness of fit, dist_node is distance from a node, dist_obstacle is distance from an obstacle, and α is a weight factor.

In other words, it minimizes redundancy with previously registered nodes and gives high fitness to positionally secure nodes that are separated from obstacles by a certain distance, thereby securing scalability and stability at the same time. The weight factor α can vary in value depending on where to focus more on scalability and stability.

This will be described in more detail with reference to FIG. 10.

FIG. 10 is a view showing an image of an environment around a robot detected by the front and rear laser scanners, and calculating a fitness value for a candidate position of a new node according to an embodiment of the present invention.

Referring first to (a) of FIG. 10, in order to calculate a distance from an obstacle, first, a 8 × 8 m space around a robot is detected by using a front and rear laser scanner, and a local map in the form of a binary image is created. By applying the distance transform algorithm to this image, you can find the distance value of the closest obstacle to a specific location in a local map.

Next, the fitness at all candidate positions is expressed by gray level through Equation 2, as shown in FIG. In FIG. 10A, the upper left and lower center portions are portions in which obstacles are detected close to the robot, and thus the fitness is low in black, and the region having high fitness is shown in white. Based on this information, as shown in FIG. 10B, the final node candidate group is represented by a white band region.

(b) Base Node Change Condition

If a new secondary node can no longer be generated based on the current base node, the node having the lowest uncertainty among the boundary nodes forming the boundary of the phase map may be selected as the next base node.

That is, when a fitness of a new candidate node is obtained from a current base node, if there is no candidate node having a fitness of a predetermined value or more, the base node is changed to find and register a new candidate node again. The next base node can be selected from only boundary nodes, not all nodes registered to date. The choice of base node is determined using the degree of uncertainty of the candidate base node. If this is expressed as an expression, Equation 3 is obtained.

In Equation 3, Pi of the 3 x 3 matrix represents the degree of uncertainty of each boundary node, and the sum of the diagonal components of the matrix determines the uncertainty of the node. Therefore, the next base node becomes the boundary node with the least uncertainty.

When the next base node is selected, the node repeats the search and use routines again.

(c) Mapping Termination Rules

Assuming all boundary nodes as base nodes, it can be configured to terminate the creation of the phase map when it is no longer possible to create a new secondary node.

This will be described in more detail as follows.

If there are no more nodes to register in the current base node, the procedure of changing the base node and registering a new node again is repeated. In this case, the base node once selected cannot be selected as the next base node. Therefore, when mapping in a closed space, there is a point in time when the base node that can be selected no longer exists. At this point, it is determined that the mapping is completed. That is, the robot visits all boundary nodes at least once and confirms that it cannot be a base node, and then automatically completes the mapping.

In the position correction step (S30), the node registered in the position and phase map of the robot by applying the landmark difference information obtained in the landmark extraction / matching step (S10) described above to the Extended Kalman Filter (EKF) The process of correcting the position of is performed.

This will be described in more detail as follows.

When the auxiliary node is registered and returned to the base node again, the position of the current robot and the node registered so far are corrected through the EKF update. In EKF SLAM, the position of the current robot (x, y, th) and the positions of registered nodes (x _n , y _n , th _n ) are stored in state variables, and covariance of the uncertainty of each state Expressed as a (covariance) matrix. If this is expressed as an expression, Equation 4 is obtained.

EKF SLAM consists of two phases, Prediction and Update. When driving between nodes, only a prediction step is performed, and an update step is performed only when it is determined that the node is reached correctly.

(a) Prediction

In this prediction step, the current robot position is predicted by using odometry data provided by the robot platform. If this is expressed as an expression, Equation 5 is obtained.

In Equation 5, F _x and F _u are Jacobian matrices obtained by differentiating the motion model F with respect to x _t and u _t , and v _t is process noise having a covariance Q _t . Since all registered nodes are static regardless of the movement of the robot, there is no change in the prediction stage, and only state and covariance related to the robot are updated according to the movement amount of the robot.

(b) Update

를 구하게 된다. 이노베이션 벡터(innovation vector)를 이용하여 로봇의 현재 위치와 현재까지 등록된 노드의 위치를 갱신한다. 이를 수식으로 표현하면 다음 수학식 6과 같다.When it is determined that the robot arrives at the node correctly, an innovation vector is obtained by matching the landmark extracted from the current image with the landmark of the registered node.

Will be obtained. Update the current position of the robot and the position of the node registered so far by using an innovation vector. If this is expressed as an expression, Equation 6 is obtained.

In Equation 6, H _t is a Jacobian matrix obtained by differentiating an observation function H with respect to x _t , w _t is observation noise having a covariance R _t , and K _t is an update. Kalman gain to determine the degree.

In the present embodiment, when N, the number of auxiliary nodes, is 2, the nodes generated in the phase map have a structure connected to each other using a Delaunay triangulation algorithm, and the robots are connected using the connection information between the nodes. It may be configured to perform a shortest path planning and autonomous environment search of and update node location.

In addition, the robot performs exploration of the entire space while radially expanding the area around the reference base node for the entire medium-to-large space exploration, and the node information created so far and the laser mounted to face the robot forward and backward The apparatus may be configured to drive a shortest path based on node information set using scanners, and perform an obstacle avoidance operation when an obstacle is detected during movement between nodes.

In addition, each registered node is connected to each other using a Delaunay triangulation algorithm to form an edge, and each time a new node is registered, edge connection is performed, and the connection strength between nodes is a value between 0 and 1. It can be configured to have.

In addition, the driving environment is detected by using a laser scanner mounted on the robot that is running. If there is an obstacle around the edge, the connection strength between nodes gradually decreases. The connecting edges can be configured to be disconnected.

In addition, when continuously receiving an observation that an obstacle is not around a disconnected edge, the connection strength between two nodes constituting the progressively disconnected edge increases, and the two nodes reconnect when the connection strength between the two nodes exceeds the threshold. And to reconstruct the edge.

Hereinafter, specific driving of the robot will be described.

The robot's driving is divided into general driving mode and obstacle avoidance mode.

(a) Normal driving mode

There are two modes of normal driving of the robot: goal following and node following. When a new auxiliary node is selected, it will go straight to that node. At this time, it will be in target following mode, and when it is close to a node to be newly registered, it will be in node following mode. When an obstacle is detected during driving, the obstacle is converted to an obstacle avoidance mode to be described later, and an obstacle path is generated to follow the path.

When following a node, the node is accessed through visual servoing using the ICP matching result. The ICP matching result is the x, y-axis difference (dx, dy) and rotation amount (dth) of two landmarks on the image, and this value is used to generate a temporary goal. If this is expressed as an equation, Equation 7 is obtained.

Every time the result changes through ICP matching, the goal changes temporarily and is determined to be very close to the node.If the dx, dy, and ICP matching errors are smaller than the specified value, the node following mode is changed. Exit and switch back to Goal following mode.

(b) obstacle avoidance mode

If an obstacle appears around the robot during normal driving mode, it is converted to obstacle avoidance mode. The detection of the obstacle is performed by the front and rear laser scanners mounted on the robot, and the result is displayed in black as shown in FIG. The center of the local map is the robot's position, and the last point of the avoidance path is the temporary target. An obstacle avoidance path is generated using a distance transform algorithm, and the avoidance path generated in FIG. 11B is represented by a red line.

(c) Create shortest distance path from node to any node

When the base node is changed, the shortest distance path from the current node to the base node is generated and traveled. At this time, it finds the intermediate nodes to go to the target base node by using the connection information between the nodes registered so far and the Dijkstra shortest path generation algorithm.

For the evaluation of the present embodiment, two exemplary medium-large spaces were selected and actual field experiments were conducted, and the result of generating a stable phase map for the environment was proved.

12 is a diagram showing the results of experiments in Experimental Environment 1. FIG.

In the experimental environment 1 of FIG. 12, the spaces are 15 m, 20 m, and 20 m in width, length, and height, respectively, and have no obstacle in the center. After placing the robot in the center of the space, the node was registered through autonomous driving, and the total driving time is about 15 minutes. The total number of nodes created is 26 and the average distance between nodes is 2.7m.

FIG. 13 is a diagram showing the results of experiments in Experimental Environment 2. FIG.

The space in Experimental Environment 2 of FIG. 13 is 14 m, 16 m, and 7 m in width, length, and height, respectively. There were sections in which people around the robot were recognized as dynamic obstacles while creating a map and made obstacle avoidance paths and traveled. Some sections remained in a temporary waiting state because they could not create escape paths. After placing the robot in the center of the space, the node was registered through autonomous driving and the total running time is about 24 minutes. The total number of nodes created is 19 and the average distance between nodes is 2.8m.

As described in detail above, according to the present invention, a robot can autonomously search for an environment without any manual manipulation of a user, and thus, a phase map of a given environment can be created.

In addition, it is possible to extract and match landmarks that are not affected by the height of the ceiling in the input image, thereby reducing the mapping error due to the height of the ceiling.

In addition, when obstacles are detected while the robot is running, an autonomous path is generated autonomously to safely drive without hitting the surrounding people or structures.

In addition, since the robot travels along nodes registered at a certain distance on the map and corrects the position of the robot each time it reaches the node, the robot does not lose its position while driving, so the robot mainly serves in the medium-to-large space. There is an effect that can be applied directly to the position recognition method of.

In addition, since maps created through autonomous driving of robots have the form of a phase map, the amount of data stored in the map and the amount of calculation required for map updating can be reduced, and accordingly, the algorithm can be reduced in weight, thereby providing a low-cost embedded board. There is an effect that can be mounted on.

Although the preferred embodiments of the present invention have been disclosed for illustrative purposes, those skilled in the art will appreciate that various modifications, additions and substitutions are possible, without departing from the scope and spirit of the invention as disclosed in the accompanying claims. In addition, it is obvious that any person skilled in the art may make various modifications and imitations without departing from the scope of the technical idea of the present invention.

S10: Landmark Extraction / Matching Step
S20: Node Creation / Registration Step
S30: Position correction step

Claims (8)

An autonomous phase map generation method in a medium-large space using a bottom-up single camera,
By matching a landmark registered in advance with the landmark at the current position of the robot extracted from the input image which is the ceiling image of the medium-large space obtained by a single camera up-down installed in the robot that autonomously runs the medium-large space, A landmark extraction / matching step of acquiring landmark difference information which is a difference between the landmark at the current position of the robot and the registered landmark;
A node generation / registration step of acquiring surrounding environment information using front and rear laser scanners attached to the robot, and generating auxiliary nodes spaced from a base node which is a starting point of the robot by autonomous driving and registering them in a phase map; And
And a position correction step of correcting the position of the robot and the position of the node registered in the phase map by applying the landmark difference information to an extended Kalman filter. Autonomous Phase Map Generation Method in Medium to Large Spaces. The method according to claim 1,
The landmark is a set of edge points obtained by calculating gradients in the x and y-axis directions and their sizes from the input image having a gray level and sampling a pixel having a value greater than or equal to a predetermined size. Consists of an edge point list,
The landmark is defined using two-dimensional information of the input image in association with the current position of the robot,
The matching method between the landmark at the current position of the robot and the registered landmark may be obtained by using an ICP (Iterative Closest Point) algorithm to obtain optimal matching information. Autonomous phase map generation method. The method according to claim 1,
The basic element of the phase map is composed of an edge which is a connection element between nodes, and among the basic elements of the phase map, the node is an auxiliary node used for the position recognition of the robot and the shortest distance path planning, and the phase map. Method for generating an autonomous phase map in a medium-to-large space using a bottom-up single camera, characterized in that it comprises a boundary node forming a boundary of, a base node used as a starting point when the robot for driving the environment. The method of claim 1 or 3,
The newly created node is selected as the location having the best suitability in consideration of the local map generated based on the current position of the robot and the location information of the node registered so far, the medium-large space using a bottom-up single camera. Autonomous Phase Map Generation Method in. The method according to claim 1 or 4,
The node creation / registration step includes a search step and a use step,
In the searching step, the robot registers N auxiliary nodes starting from the base node,
In the use step, the robot returns to the base node, which is an initial starting point, and corrects the position of the robot and the position of the nodes by using the landmark difference information.
The shape of the generated phase map has a form of repetition and incremental expansion of polygonal structures including N + 1 nodes. 6. The method of claim 5,
If the robot can no longer create a new auxiliary node based on the base node currently running, the node having the lowest uncertainty among the boundary nodes forming the boundary of the phase map is selected as the next base node.
Assuming that all boundary nodes are base nodes, when a new auxiliary node can no longer be generated, the generation of the phase map is terminated. The method of claim 6,
When N, the secondary node structure generation condition, is 2, the nodes generated in the topology map are connected to each other using a Delaunay triangulation algorithm to form an edge, and each time a new node is additionally registered. Edge connection process is performed, and the connection strength between nodes has a value between 0 and 1 depending on the environment characteristics.
Perform the optimal path planning and autonomous environment search of the robot using the connection information between the nodes, update the node location,
The robot performs an exploration process on the entire space while expanding the region radially around the reference base node for the entire medium-large space exploration,
When an obstacle is detected during movement between nodes, the obstacle avoidance operation is autonomously characterized in that the autonomous phase map generation method in a medium-to-large space using a bottom-up single camera. The method of claim 7, wherein
The robot, which performs optimal path selection and driving based on edge connection information between nodes and its strength, detects the surrounding environment using a laser scanner mounted on the robot, and if there is an obstacle around the edge, The edge update process to gradually reduce the strength of the connection and increase the connection strength when there are no obstacles around the edge, and when the connection strength of the edge is less than the threshold, the edge connecting the two nodes. The method for generating an autonomous phase map in a medium-to-large space using a bottom-up single camera is characterized by performing a path management process for processing to be disconnected.

Images