CN108665540A - Robot localization based on binocular vision feature and IMU information and map structuring system - Google Patents

Abstract

A kind of robot localization based on binocular vision feature and IMU information and map structuring system, binocular information collection, feature extracting and matching module include binocular ORB feature extractions submodule, binocular characteristic matching submodule and IMU information collection submodules；Improved IMU initialization and its motion module include IMU angular speed estimation of deviation submodule, acceleration of gravity estimates submodule, IMU acceleration bias estimates submodule and IMU pre-integration submodules；Vision SLAM algorithm initializations and tracking module include that tracking interframe movement submodule and key frame generate submodule；It includes that new key frame is inserted into submodule, part BA optimization submodules and rejects redundancy key frames module that module is built in part；Winding detects and optimization module includes winding detection sub-module and global optimization submodule.The present invention provides that a kind of robustness is good, accuracy is high, adaptable robot localization and map structuring system based on binocular vision feature and IMU information.

Description

Robot positioning and map construction system based on binocular vision features and IMU information

technical field

The invention belongs to the technical field of robots, and in particular relates to a robot positioning and map construction system.

Background technique

Simultaneous Localization and Mapping (SLAM) technology is an important problem in the field of robot navigation. The SLAM problem can be described as: the robot can build a global map of the explored environment, and at the same time, it can use this map to infer at any time. own position. The robot moves freely in the environment through the sensor equipment, locates its own position through the collected information, and builds a map on the basis of the positioning, so as to realize the simultaneous positioning and mapping of the robot. There are two main factors that affect the solution of the SLAM problem, namely, the data characteristics of the sensor and the correlation of the observation data. If the robustness and accuracy of the data association can be improved and the utilization rate of the sensor data can be improved, the robot will be improved. The accuracy of positioning and map construction.

At present, the most common sensors on mobile devices are cameras and IMU sensors. How to design algorithms to use these common sensor devices to achieve high-precision simultaneous positioning and mapping is a current research hotspot.

The present invention innovatively utilizes IMU information to fuse binocular vision information, makes up for the problems existing in binocular vision SLAM, and improves the robustness of the SLAM algorithm. On this basis, a map with real scale information is constructed.

Simultaneous positioning and map creation (SLAM) technology is a classic problem in the field of robotics. The SLAM problem can be described as: the robot starts to move in an unknown environment from an unknown position, estimates its own pose and establishes an environmental map during the movement process, Realize the autonomous positioning and mapping of the robot. Two factors that affect the SLAM system are the correlation of observation data and environmental noise. The correctness of observation of the surrounding environment depends on good data correlation, which in turn affects the construction of the environmental map.

Contents of the invention

In order to overcome the disadvantages of poor robustness, low accuracy and poor adaptability of existing robot positioning and map construction systems, the present invention provides a binocular-based Robot localization and mapping system based on visual features and IMU information.

The technical solution adopted by the present invention to solve its technical problems is:

A robot positioning and map construction system based on binocular visual features and IMU information, the system includes binocular information collection, feature extraction and matching modules, improved IMU initialization and motion modules, visual SLAM algorithm initialization and tracking modules, Local mapping module, loop detection and optimization module; the binocular information collection, feature extraction and matching modules include binocular ORB feature extraction submodule, binocular feature matching submodule and IMU information acquisition submodule; the improved IMU Initialization and its motion module include IMU angular rate deviation estimation sub-module, gravity acceleration estimation sub-module, IMU acceleration deviation estimation sub-module and IMU pre-integration sub-module; the visual SLAM algorithm initialization and tracking module includes tracking inter-frame motion sub-module and key frame generation submodule; the local mapping module includes a new key frame insertion submodule, a local BA optimization submodule and a redundant keyframe elimination module; the loopback detection and optimization module includes a loopback detection submodule and a global optimization submodule module.

Further, in the binocular information collection, feature extraction and matching module, the binocular ORB feature extraction sub-module accepts the pictures taken by the binocular camera during the movement of the robot, uses the ORB feature extraction algorithm to extract the ORB features in the left and right images, and The extracted results are stored; the binocular feature matching module matches the feature points extracted from the left and right pictures to obtain the depth information of the feature points; the IMU information collection module collects the IMU information inside the robot.

Further, in the improved IMU initialization and its motion module, the IMU angular rate deviation estimation submodule uses the visual SLAM results to estimate the IMU angular rate deviation using an optimization method; the gravity acceleration and IMU acceleration deviation estimation submodule uses the IMU The results of the angular rate deviation estimation sub-module, for consecutive N key frames, use the relationship between them to obtain a linear equation, and solve to obtain the acceleration of gravity; the estimation of the acceleration deviation is to exclude the influence of the gravity acceleration on the IMU acceleration, and re-calculate the continuous current The optimization of the frame state combines two key frames to solve the equation to obtain the re-estimation of the acceleration of gravity, and at the same time obtain the IMU acceleration deviation; the IMU acceleration deviation estimation sub-module uses the results of the gravity acceleration and the IMU acceleration deviation estimation sub-module to calculate the gravity acceleration Eliminate the influence of IMU acceleration, re-solve the equation for consecutive N key frames to obtain a re-estimation of the acceleration of gravity, and at the same time obtain the IMU acceleration deviation; the IMU pre-integration sub-module pre-integrates all IMU data between two consecutive image frames through IMU The models are ensembled into a single observation, resulting in a motion model of the robot and an estimate of the robot's current pose.

Further, in the visual SLAM algorithm initialization and tracking module, the tracking inter-frame motion sub-module calculates the ORB feature matching between two consecutive images, calculates the intrinsic matrix, and then uses cluster optimization to further estimate the estimated demerit; After successfully returning, the system starts to track in a purely visual way. The key frame generation sub-module monitors the tracking intermediate results. When the current frame meets the conditions of the key frame, the current frame is added to the key frame set, and the current key frame is used as the tracking reference frame .

In the local mapping module, the new key frame insertion submodule) is inserted into the key frame queue according to the key frame generated by the previous submodule; the local BA optimization submodule), after the key frame is inserted, the key to the current partial view Frames and map point clouds are clustered and optimized; redundant key frame sub-modules are eliminated, and redundant key frames in the window are deleted using screening strategies according to the inserted key frames to ensure the size of the map.

In the described loopback detection and optimization module, the loopback detection submodule queries the key frame set of the current frame in the database to detect the loopback; after returning successfully, it estimates the poses of all keyframes on the loopback, and the global optimization submodule All key frames in the loop are optimized globally, and the error factors include visual projection factors and IMU factors.

The technical idea of the present invention is: the robot SLAM system proposed by the present invention is a solution based on binocular vision and inertial navigation information fusion SLAM system. In the present invention, the pure visual pose information is used to estimate the IMU deviation model and the direction of the acceleration of gravity; in the visual pose, a high-efficiency feature extraction algorithm is used to extract rich descriptors from the image, and then the extracted feature points are used to perform Matching to obtain its depth information; use the IMU kinematics model based on pre-integration to establish the camera motion model, and use the motion model to make a real-time preliminary estimate of the camera position; based on the preliminary estimate, the pose between the two image frames before and after Estimation, using visual geometric knowledge to estimate the current position of the robot more accurately; using the key frame selection strategy to realize the sparse construction of local map points; on the basis of visual information matching with IMU information, using graph optimization The back-end optimization algorithm uses the spatial map point projection factor and the IMU factor as factors in the factor graph to estimate the map position accurately and in real time. The invention realizes the robot synchronous positioning and map creation algorithm based on the binocular vision and inertial navigation information, and can accurately estimate the movement of the robot and the surrounding environment information.

The beneficial effects of the present invention are mainly manifested in: the method of fusion of visual information and IMU information is adopted, which can well solve problems in visual SLAM and improve the robustness of the SLAM system; the accuracy is high and the adaptability is strong.

Description of drawings

Fig. 1 is a schematic diagram of the system structure of the present invention.

Fig. 2 is a flow diagram of the system of the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings.

Referring to Figure 1 and Figure 2, a robot positioning and map construction system based on binocular visual features and IMU information, including: binocular information collection, feature extraction and matching module (1), improved IMU initialization and its motion module ( 2), visual SLAM algorithm initialization and tracking module (3), local mapping module (4), loop detection and optimization module (5); described binocular information collection, feature extraction and matching module (1) include: binocular ORB feature extraction submodule (1.1), binocular feature matching submodule (1.2), IMU information collection submodule (1.3); the improved IMU initialization and its motion module (2) include: IMU angular rate deviation estimation submodule (2.1), acceleration of gravity estimation submodule (2.2), IMU acceleration deviation estimation submodule (2.3), IMU pre-integration submodule (2.4); described visual SLAM algorithm initialization and tracking module (3) include: tracking between frames Motion sub-module (3.1), key frame generation sub-module (3.2); described local mapping module (4) includes: new key frame insertion sub-module (4.1), local BA optimization sub-module (4.2), elimination of redundant key A frame module (4.3); the loopback detection and optimization module (5) includes: a loopback detection submodule (5.1) and a global optimization submodule (5.2).

In the binocular information collection, feature extraction and matching module (1), the binocular ORB feature extraction submodule (1.1) uses the ORB feature extraction algorithm to extract the left and right images by accepting the pictures taken by the binocular camera during the movement of the robot. The ORB feature stores the extracted results; the binocular feature matching module (1.2) matches the feature points extracted from the left and right pictures to obtain the depth information of the feature points; the IMU information collection module (1.3) collects the IMU information inside the robot;

In the described improved IMU initialization and its motion module (2), the IMU angular rate deviation estimation submodule (2.1) utilizes the visual SLAM result, and adopts an optimization method to estimate the IMU angular rate deviation; gravity acceleration and IMU acceleration deviation estimation submodule (2.2), using the results of the sub-module (2.1), for consecutive N key frames, using the relationship between them to obtain a linear equation, and solving the acceleration of gravity; the estimation of the acceleration deviation is to exclude the influence of the acceleration of gravity on the acceleration of the IMU, and then re- Combining the optimized key frames of the continuous current frame state in pairs, solving the equation to obtain the re-estimation of the acceleration of gravity, and obtaining the IMU acceleration deviation at the same time; Eliminate the influence of acceleration on IMU acceleration, re-solve the equation for consecutive N key frames to obtain a re-estimation of the acceleration of gravity, and at the same time obtain the IMU acceleration deviation; the IMU pre-integration sub-module (2.4) converts all IMU data between two consecutive image frames The IMU pre-integration model is combined into a single observation value to obtain the motion model of the robot and the current pose estimation of the robot;

In the visual SLAM algorithm initialization and tracking module (3), the visual SLAM algorithm initialization and tracking module (3.1) calculates the ORB feature matching between two consecutive images, calculates the intrinsic matrix, and then utilizes cluster optimization to estimate and demerit Further estimation; after the sub-module (3.1) returns successfully, the system starts to track in a purely visual way, the key frame generation sub-module (3.2) monitors the intermediate results of the tracking, and when the current frame meets the conditions of the key frame, the current frame is added to the key frame set Among them, and use the current key frame as the tracking reference frame;

In the local mapping module (4), the new key frame insertion submodule (4.1) is inserted into the key frame queue according to the key frame generated by the previous submodule; the local BA optimization submodule (4.2), after the key frame is inserted , perform bundle optimization on the keyframes of the current partial view and the map point cloud; eliminate redundant keyframe submodules (4.3), and use the screening strategy to delete redundant keyframes in the window according to the inserted keyframes to ensure the size of the map.

In described loopback detection and optimization module (5), loopback detection submodule (5.1) queries current frame key frame set in database, detects loopback; After submodule (5.2) returns successfully, to all keyframes on loopback The pose is estimated, the global optimization sub-module (5.3) performs global graph optimization on all key frames in the loop, and the error factors include visual projection factors and IMU factors.

The implementation of the module is further described in detail below.

In the binocular information collection, feature extraction and matching module (1):

The binocular ORB feature extraction sub-module (1.1) is used to extract the ORB feature of the image after the robot collects the left and right images each time. The reason why the ORB feature is selected as the image feature is that the ORB feature has good viewing angle and illumination invariance , and the ORB feature, as a binary visual feature, has the characteristics of fast extraction and matching speed, which is very suitable for SLAM systems that require real-time performance. After the ORB features are extracted, the extracted results are stored.

The binocular feature matching sub-module (1.2), since the left and right images have undergone epipolar correction, only need to match in the same line when searching for matching points. After each robot collects image extraction feature points, use the descriptor matching of feature points, The absolute scale of the feature points can be calculated using the triangulation method of binocular vision, and the calculation formula is as follows:

In the above formula, f is the focal length of the camera, B is the distance between the optical centers of the two cameras; D is the parallax, that is, the distance difference between the same feature point in the left and right pictures, and z is the depth of the feature point;

The IMU information acquisition sub-module (1.3) is used to collect the IMU data inside the robot in real time, and send it to the IMU pre-integration sub-module (4.1) for processing, because the frequency of the IMU data collected by the IMU information acquisition module Much higher than the frequency of visual image acquisition, the use of IMU pre-integration can reduce the computational pressure of the factor graph-based loop closure detection and optimization module (5).

In the improved IMU initialization and its motion module (2), the IMU angular rate deviation estimation submodule (2.1) utilizes the results of binocular vision SLAM to add the IMU angular rate deviation variable to the robot state vector, and to the IMU Angular rate bias estimation. The optimization formula is:

Among them, N is the number of current keyframes, is the camera center C coordinate calculated by the pure visual SLAM algorithm Multiplied with the calibration matrix R _CB between the camera and the IMU center (denoted as B), ΔR _i,i+1 is the rotation between i and i+1 key frames obtained according to the IMU pre-integration sub-module (2.4) matrix; is the first-order approximation of the equation of ΔR _i,i+1 changing with the angular rate deviation obtained from the IMU pre-integration sub-module (2.4); by solving the optimization equation, the IMU angular rate deviation b _g is obtained.

Gravitational acceleration estimation sub-module (3.2), using the result of IMU angular rate deviation estimation sub-module (2.1), for the relationship between two connected key frames, using the pre-integrated measured value Δv, the equation is obtained:

In the above formula, R _WC represents the rotation matrix of camera center C relative to world coordinates W, R _WB represents the rotation matrix of IMU center B relative to world coordinates, p _CB represents the position of IMU center B in camera coordinates, and Δt is the frame time difference between Represents the camera speed between frames, s is the absolute scale of the feature point, and can be obtained through calculation, the gravitational acceleration g _w .

The IMU acceleration deviation estimation sub-module (2.3) uses the results of the gravity acceleration estimation sub-module (2.2) and utilizes the linear relationship between three consecutive key frames:

Use 1, 2, 3 to represent three consecutive key frames of i, i+1, and i+2. The calculation equations for the items in the above formula are as follows:

In the above formula, [ ] _{(; 1:2)} refers to the use of the first two columns of the matrix, R _WC represents the rotation matrix of the camera center C relative to the world coordinate W, R _WB represents the rotation matrix of the IMU center B relative to the world coordinates, R _WI represents the rotation moment of the current coordinates relative to the world coordinates, p _B represents the position of the IMU center B in the world coordinates, I is the identity matrix Δt is the time difference between frames, Δv represents the camera speed between frames, and δθ _xy is to be estimated The direction angle of gravitational acceleration, b _a is the acceleration deviation. Stack all three consecutive keyframe equations into a linear equation, denoted as:

A3(N-2)×5X5×1=B3(N-2)×1

Singular value decomposition is performed through the thread equation above to obtain the gravity acceleration direction δθ _xy and the IMU acceleration deviation b _a .

The IMU pre-integration sub-module integrates all IMU data between two consecutive image frames into a single observation value through the B pre-integration model, and obtains the motion model of the robot and the current position estimation of the robot. Mark the IMU as B, and the acceleration and angular rate output by the IMU are expressed as α _B and ω _B , then the motion equation of the robot between the two IMU outputs i and i+1 is:

The direct IMU pre-integration results of the two key frames k and k+1 are recorded as ΔR, Δv, and Δp, and the motion model between the two key frames is obtained:

In the above formula, R _WB represents the rotation of the IMU center B relative to the world coordinates, v _B represents the speed of the IMU, p _B represents the position of the IMU center B in the world coordinates, b _g and b _a represent the deviation between the angular rate and the acceleration . Jacobian matrix and Represents the first-order approximation of the IMU pre-integration observation value variation equation with the deviation.

In the visual SLAM algorithm initialization and tracking module (3), the tracking inter-frame motion submodule (3.1), after utilizing the ORB feature extraction submodule (1.1) to extract features, performs two continuous initialization images Fc, Fr on the feature Feature matching, using the PnP algorithm to solve the pose relationship between the two. Afterwards, the SLAM system can be used to track and continuously estimate the position of the robot. The reason for using cluster optimization to further estimate the estimation is that cluster optimization can modify the pose matrix to make it more consistent with the ORB feature matching results.

The key frame generation sub-module (3.2) decides whether to select the current frame as a key frame according to the rules. Its rules are as follows:

1) At least 20 frames away from the last relocation;

2) Partial mapping is idle or 20 frames have passed since the last key frame insertion;

3) The current frame contains at least 50 feature points.

In the local mapping module (4), the new key frame is inserted into the submodule (4.1), and the key frame generated by the submodule (3.2) is inserted into the queue, and the relevant view of the update content and the point cloud with the same map as the key frame are generated edge while updating the node's links to other keyframes. Then, the unqualified cloud points of the local view are eliminated, and new cloud points are regenerated.

Local BA optimization sub-module (4.2), local BA optimizes the current keyframe, all keyframes connected to it in the interactive view set, and all map cloud points observed by these keyframes. All other keyframes that can observe these cloud points, but are not connected to the current keyframe will be optimized; where the projection error is:

E _proj (i)＝ρ((x-π(X _C )) ^T ∑ _I (x-π(X _C )))

Among them, π(X _C ) is the projection formula, and x represents the projected coordinates of the space point on the current frame. The conversion from the coordinate value X _W of the world coordinate system to the coordinate value X _C of the camera coordinate system is:

In the above formula, R _CB represents the rotation matrix of the IMU center B relative to the coordinates of the current camera center C, and R _BW represents the rotation matrix of the world coordinate W to the IMU center B. p _CB represents the translation vector of IMU center B relative to the current coordinates of the current camera center C, p _WB represents the translation vector of IMU center B relative to the current world coordinate W;

The IMU factor can be expressed as:

The distribution of each error term is expressed as:

In the above formula Represents the first-order approximation of the rotation transformation ΔR between two frames relative to the acceleration of gravity, Represents the first-order approximation of the velocity Δv between two frames relative to the acceleration of gravity, Represents a first-order approximation of the velocity change Δv between two frames with respect to acceleration. R _BW represents the rotation matrix of world coordinate W to IMU center B, and R _WB represents the rotation matrix of IMU coordinate B to world coordinate W. b _g , b _a represent the deviation of angular rate and acceleration, respectively, and e _R , e _p , _ev , e _b represent the error terms caused by rotation, translation, velocity and IMU deviation, respectively. We regard it as a Gaussian distribution, ∑ _I is the pre-integrated information matrix, and ∑ _R is the biased random walk information matrix representing the huber robust loss equation.

Elimination of redundant keyframes module (4.3), due to the short-term nature of the IMU, allows the tracking thread to discard redundant keyframes if the difference between two consecutive keyframes in the local window of the local BA does not exceed 0.5 seconds. But to be able to perform global optimization, refine the map after the loop is closed or at any time, we do not allow any two consecutive keyframes to differ by more than 3 seconds. If we turn off full BA with IMU constraints, we only need to constrain the time offset between keyframes in the local window.

In the loop detection and optimization module (5), the loop detection submodule (5.1) calculates the bag-of-words vector of the current key frame, calculates the similarity of adjacent images related to its view, and retains the lowest score. Then, we search the image recognition database, discard those keyframes whose score is lower than the minimum score, and query the current frame in the keyframe set in the database.

The global optimization sub-module (5.2) estimates all key frames on the loop. At this time, only the visual projection factor is considered, and the projection factor formula is shown in the optimization sub-module (4.2) of the current frame state.

Taking a robot device equipped with a binocular camera and an IMU as an example, the execution process of this embodiment on this device is introduced in detail.

First, in the binocular and IMU information acquisition and depth information acquisition module (1), feature extraction and feature matching are performed on the binocular visual information, and the IMU information is collected. These two parts transfer the acquired sensor information to the subsequent module. Because the robot is moving in space, these two parts will continuously collect and extract the new data obtained by the sensor. Use the binocular camera to collect images, send the images to the feature extraction algorithm, save the position and descriptor of the extracted image features, and use the feature points extracted from the left and right pictures to match and calculate the depth of the feature points and save them. At the same time, the signal of the IMU sensor is collected and sent to the IMU pre-integration algorithm for processing.

When binocular vision SLAM is performed for a period of time, the improved IMU initialization and its motion module (2) start running. The first is the estimation of the IMU angular rate deviation, and the estimation result of the angular rate deviation of the IMU will be obtained. Secondly, the gravitational acceleration pre-estimation makes a preliminary estimate of the gravitational acceleration, and the gravitational acceleration can be obtained by using the acceleration deviation estimation result and scale, which can make a more accurate estimation of the IMU acceleration deviation. IMU acceleration deviation estimation and gravity acceleration re-estimation are to estimate the IMU acceleration deviation and re-estimate the gravity acceleration at the same time. Using the binocular and IMU information acquisition and the IMU information data obtained by the depth information acquisition module (1), a motion model based on IMU pre-score is established.

Using the extracted feature point information, the visual SLAM algorithm initialization and tracking module (3) initializes the visual SLAM part. Initialize the tracking module, key frame generation module and map points in turn. The initialized system saves the initialized map points and the initial location information. After the vision initialization is completed, the system will execute the tracking algorithm part to estimate the pose between frames. At the same time, the key frame selection strategy is used to generate key frames. Using the kinematic model of the IMU and the binocular vision pose information, the speed of each key frame is estimated, and the initialization of the algorithm that fuses the IMU information is completed. This part can make the result of binocular vision SLAM more robust and reduce the failure of tracking.

The tracking algorithm solves the associated data problem, after which local map construction will proceed. According to the key frame generated in the previous module, in the local view, local BA optimization is performed on the map point cloud seen by the current key frame and its related views. The optimization includes the simultaneous optimization of IMU and visual pose. After optimization, keyframes at certain time intervals in the queue can be deleted to avoid increasing the size of the map within the region.

The loop detection and optimization module (5) based on the graph optimization algorithm is used to globally optimize the constructed map to obtain more accurate estimates of the robot's position and map points. This module is divided into two parts. Firstly, the detection loop sub-module (5.1) calculates the bag-of-words vector of the current key frame, calculates the similarity of the adjacent images related to its view, and keeps the lowest score. Then, we retrieve the image recognition database, discard those key frames whose scores are lower than the minimum score, and query the current frame in the key frame set in the database. When a loop is detected, the poses of all key frames on the loop are estimated. . The global pose optimization sub-module (5.2) is performed when each key frame is inserted, in order to further estimate the positions of the key frames and map points, and to further optimize the poses of all current key frames in the current situation, this In one step, the result of loop closure detection is used, which can eliminate the influence of accumulated errors.

Claims (6)

1. A robot positioning and map construction system based on binocular vision features and IMU information, characterized in that the system includes binocular information collection, feature extraction and matching modules, improved IMU initialization and motion modules, visual SLAM Algorithm initialization and tracking module, local mapping module, loop detection and optimization module; the binocular information collection, feature extraction and matching module includes binocular ORB feature extraction submodule, binocular feature matching submodule and IMU information collection submodule ; The improved IMU initialization and its motion module include IMU angular rate deviation estimation submodule, gravity acceleration estimation submodule, IMU acceleration deviation estimation submodule and IMU pre-integration submodule; the visual SLAM algorithm initialization and tracking module include Track inter-frame motion sub-module and key frame generation sub-module; the local mapping module includes a new key frame insertion sub-module, a local BA optimization sub-module and a redundant key frame elimination module; the loop detection and optimization module includes a loop detection submodules and globally optimized submodules. 2. the robot localization and map construction system based on binocular vision feature and IMU information as claimed in claim 1, is characterized in that, in described binocular information collection, feature extraction and matching module, binocular ORB feature extraction submodule By accepting the pictures taken by the binocular camera during the movement of the robot, use the ORB feature extraction algorithm to extract the ORB features in the left and right images, and store the extracted results; the binocular feature matching module matches the feature points extracted from the left and right pictures to obtain the feature points The in-depth information; the IMU information collection module collects the internal IMU information of the robot. 3. The robot localization and map construction system based on binocular vision features and IMU information as claimed in claim 1 or 2, is characterized in that, in the described improved IMU initialization and motion module thereof, the IMU angular rate deviation estimation submodule Using the results of visual SLAM, the optimization method is used to estimate the IMU angular rate deviation; the gravity acceleration and IMU acceleration deviation estimation sub-module uses the results of the IMU angular rate deviation estimation sub-module, and uses the relationship between consecutive N key frames to obtain The linear equation is solved to obtain the acceleration of gravity; the estimation of the acceleration deviation is to exclude the influence of the acceleration of gravity on the acceleration of the IMU, and to combine the optimization of the continuous current frame state with two key frames to solve the equation to obtain the re-estimation of the acceleration of gravity, and at the same time obtain IMU acceleration deviation; the IMU acceleration deviation estimation sub-module uses the results of gravity acceleration and IMU acceleration deviation estimation sub-module to exclude the influence of gravity acceleration on IMU acceleration, and re-solve the equation for consecutive N key frames to obtain a re-estimation of gravity acceleration , and get the IMU acceleration deviation at the same time; the IMU pre-integration sub-module integrates all the IMU data between two consecutive image frames into a single observation value through the IMU pre-integration model, and obtains the motion model of the robot and the current pose estimation of the robot. 4. The robot positioning and map construction system based on binocular vision features and IMU information as claimed in claim 1 or 2, wherein, in the initialization of the visual SLAM algorithm and the tracking module, the tracking inter-frame motion submodule calculates continuously Match the ORB features between the two images, calculate the intrinsic matrix, and then use the bundle optimization to further estimate the estimated score; after the successful return, the system starts to track in a purely visual way, and the key frame generation sub-module monitors and tracks the intermediate results. When the current frame meets the conditions of the key frame, the current frame is added to the key frame set, and the current key frame is used as the tracking reference frame. 5. as claimed in claim 1 or 2, based on the robot localization of binocular vision feature and IMU information and map construction system, it is characterized in that, in described local mapping module, new key frame inserts submodule) according to last submodule The key frame generated by the module is inserted into the key frame queue; the local BA optimization sub-module), after the key frame is inserted, the key frame of the current partial view and the map point cloud are bundled and optimized; the redundant key frame sub-module is eliminated, according to For the inserted keyframes, use the screening strategy to delete redundant keyframes in the window to ensure the size of the map. 6. The robot positioning and map construction system based on binocular vision features and IMU information as claimed in claim 1 or 2, wherein, in the loop detection and optimization module, the loop detection submodule puts the current frame in the database The key frame set is queried and the loop is detected; after the successful return, the poses of all key frames on the loop are estimated, and the global optimization sub-module performs global graph optimization on all key frames in the loop, and the error factors include visual projection factors and IMU factor.