A SLAM autonomous navigation recognition method in a closed scenario

Abstract

The present invention belongs to the technical field of image processing, and particularly relates to a SLAM autonomous navigation recognition method in a closed scenario. In the data association of the method, the SLAM data association module tracks the common features of images in different frames, and realizes the matching of the same features through the correlation clustering between frames, so as to judge the movement of the autonomous robot. Compared with the prior art, (1) the clustering matching has stronger practicability than the data association method in the existing SLAM, and the effect of the clustering matching will not be affected by various complex scenarios; (2) the K-means clustering algorithm is simple to calculate and can be embedded into various systems for feature matching, which is very suitable for visual SLAM in a dynamic environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to a SLAM autonomous navigation recognition method in a closed scenario. Background Art

[0002] Visual Simultaneous Localization and Mapping (SLAM) refers to the technology of using a visual sensor mounted on a robot to perceive the surrounding environment, build an environmental model during movement, and estimate its own position without prior information.

[0003] Current visual SLAM methods are usually based on the static assumption that, during the entire visual SLAM process, it is defaulted that the objects in the environment are stationary. However, most actual application environments are dynamic environments with dynamic objects, such as walking people, moving cars, etc., and these dynamic objects affect the accuracy of mapping. Summary of the Invention

[0004] Aiming at the problem that the current visual SLAM methods are usually based on the static assumption, the present invention provides a SLAM autonomous navigation recognition method in a closed scenario.

[0005] To achieve the above object, the technical solution adopted by the present invention is: A SLAM autonomous navigation recognition method in a closed scenario, including the following steps:

[0006] Step 1: External environment data acquisition. The autonomous robot acquires external environment data through its own camera.

[0007] Step 2: Feature detection. The acquired external environment data is input into the SLAM feature extraction module, and the Vision transformer model is used in the SLAM feature extraction module to implement semantic detection of each object in the external environment and extract the features of each object at the same time.

[0008] Step 3: Data association. The SLAM data association module tracks the common features of images in different frames, and realizes the matching of the same features through correlation clustering between frames, and then judges the movement situation of the object.

[0009] Step 4: Loop detection: The SLAM loop detection module judges whether the movement trajectory of the autonomous robot forms a loop; if a loop is detected, the loop information is provided to the backend optimization module for processing.

[0010] Step 5: Back-end optimization and mapping: The back-end optimization module continuously receives the images captured by the camera, calculates the camera poses of adjacent frames, and receives loop closure information to optimize the cumulative error generated in data association. Meanwhile, it generates a global motion trajectory and a perception map of the surrounding environment.

[0011] Preferably, in a closed scenario where the number of objects is fixed, the number of feature vectors in each frame captured by the camera in the feature detection of Step 2 is fixed and denoted as ones, and the feature vectors in each frame are denoted as .

[0012] Preferably, in the data association of Step 3, the features in the frame, a total of feature vectors are input into the K-means clustering algorithm for iterative calculation. The calculation result classifies the same feature vectors into the same cluster and the different feature vectors into the corresponding clusters to achieve the matching of the same features. Thus, the autonomous robot can accurately identify each object in the closed environment in the moving state and then judge its own movement situation.

[0013] Preferably, the iterative calculation process of the K-means is as follows:

[0014] (1) Set the cluster value Select cluster centers and initialize them, denoted as ;

[0015] (2) Define its loss function as the sum of the squared errors of the distances of each feature vector from the center point of its belonging cluster:

[0016]

[0017] where is the number of objects in the closed environment; represents the -th feature vector among the feature vectors, represents the cluster to which belongs,

[0018] (3) For each feature vector , assign it to the nearest cluster;

[0019]

[0020] represents the variable value when the objective function takes the minimum value;

[0021] (4) For each cluster , recalculate the center of the cluster

[0022] ;

[0023] (5) Set the number of iteration steps, and repeat (3) and (4) until convergence; output the final cluster centers and cluster partitions.

[0024] Preferably, the camera described in step one includes one or a combination of a monocular camera, a binocular camera, and an RGB-D camera; the autonomous robot also comes with a laser ranging unit.

[0025] Preferably, in the feature detection of step two, the position information of each object in each frame is also located.

[0026] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0027] (1) The SLAM autonomous navigation recognition method in a closed scenario of the present invention realizes the same feature matching through the correlation clustering between frames, and then judges the movement of objects; the clustering matching has stronger practicability compared with the data association method in the existing SLAM, and the effect of clustering matching will not be affected by various complex scenarios;

[0028] (2) In the data association of step three, features in the frames, a total of feature vectors are input into the K-means clustering algorithm for iterative calculation. The calculation result classifies the same feature vectors into the same cluster and the different feature vectors into the corresponding clusters to realize the same feature matching, and then judges the movement of objects; the calculation is simple, and it can be embedded into various systems for feature matching, and is very suitable for visual SLAM in a dynamic environment. Description of the Drawings

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Figure 1 It is a schematic diagram of the SLAM autonomous navigation recognition method in a closed scenario provided in Embodiment 1. Detailed Embodiment

[0030] In order to be able to more clearly understand the above objects, features, and advantages of the present invention, the present invention will be further described below with reference to the drawings and embodiments.

[0031] In the following description, many specific details are set forth in order to provide a thorough understanding of the present invention. However, the present invention may be practiced in other ways different from those described herein. Therefore, the present invention is not limited by the limitations of the specific embodiments disclosed in the following specification.

[0032] Embodiment 1

[0033] The following will further describe the present invention in conjunction with Figure 1 An SLAM autonomous navigation recognition method in a closed scenario includes the following steps:

[0034] Step 1: External environment data acquisition. The autonomous robot acquires external environment data through its own camera.

[0035] Step 2: Feature detection. The acquired external environment data is input into the SLAM feature extraction module. In the SLAM feature extraction module, the Vision transformer model is used to implement semantic detection of each object in the external environment, and at the same time, the features of each object are extracted.

[0036] Step 3: Data association. The SLAM data association module tracks the common features of images in different frames, realizes the matching of the same features through the correlation clustering between frames, and then judges the movement of the object. For the successfully matched features, the controller of the autonomous robot corrects the current position through the distance between the autonomous robot and the feature. At the same time, for the same feature in different frames (in the surrounding environment), the controller of the autonomous robot calculates the difference in the distance before and after to judge the real-time motion state of the autonomous robot.

[0037] Step 4: Loop detection. The SLAM loop detection module judges whether the motion trajectory of the autonomous robot forms a loop (when the distance from the autonomous robot to a certain object is the same as the previous time); if a loop is detected, the loop information is provided to the backend optimization module for processing.

[0038] Step 5: Backend optimization and mapping. The backend optimization module continuously receives the images captured by the camera and calculates the camera poses of adjacent frames (using the visual odometry module in SLAM) and receives the loop information, optimizes the cumulative error generated in the data association, and at the same time generates a global motion trajectory and a perception map of the surrounding environment.

[0039] An autonomous robot is a robot that comes with various necessary sensors, controllers, and ranging units (a laser ranging unit in this embodiment) on its body and can independently complete certain tasks under the condition of no external human information input and control during operation.

[0040] The semantic detection of each object is to detect the specific content of each object in the image. For example, given a photo of a person riding a motorcycle, after semantic detection, the person and the motorcycle should be distinguished.

[0041] Use the Vision transformer model to achieve semantic detection of each object in the external environment, and at the same time extract the features of each object: Vision Transformer (ViT) directly applies the pure Transformer architecture to a series of image patches for classification tasks, and can achieve excellent results. It also outperforms state-of-the-art convolutional networks in many image classification tasks, while the required pre-training computing resources are greatly reduced (at least reduced by 4 times).

[0042] In a closed scenario, the number of objects is fixed (10 in this embodiment), and the number of feature vectors in each frame captured by the camera in the feature detection of step two is fixed and set to each, and the feature vectors in each frame are denoted as .

[0043] In the data association of step three, the features in the frame, a total of feature vectors are input into the K-means clustering algorithm for iterative calculation. The calculation result classifies the same feature vectors into the same cluster and the different feature vectors into the corresponding clusters to achieve the matching of the same features, and then judges the movement of the object.

[0044] The iterative calculation process of the K-means is as follows:

[0045] (1) Set the cluster value Select cluster centers and initialize them, denoted as ;

[0046] (2) Define its loss function as the sum of the squared errors of the distances of each feature vector from the center point of the cluster to which it belongs:

[0047]

[0048] where is the number of objects in the closed environment; represents the th feature vector among feature vectors, represents the cluster to which belongs, and

[0049] (3) For each eigenvector , assign it to the nearest cluster; cluster the eigenvectors according to the principle of the minimum distance (the minimum distance to the center point of a certain cluster);

[0050]

[0051] represents the variable value when the objective function takes the minimum value;

[0052] (4) For each cluster , recalculate the center of this cluster; use the sample mean of K clusters to update iteratively;

[0053] ;

[0054] (5) Set the number of iteration steps, and repeat (3) and (4) until convergence; output the final cluster centers and clustering partitions.

[0055] The camera described in Step 1 includes one or a combination of a monocular camera, a binocular camera, and an RGB-D camera.

[0056] In the feature detection of Step 2, the position information of each object in each frame is also located.

[0057] ,

[0058] and respectively represent the physical positions of the upper left side points of the feature in each frame, and represent the length and width of the feature .

[0059] Embodiment 2

[0060] The difference between this embodiment and Embodiment 1 is: set the cluster value .

[0061] The above are only the preferred embodiments of the present invention, and are not limitations of the present invention in other forms. Any person skilled in the art may use the technical content disclosed above to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification and equivalent change made to the above embodiments according to the technical essence of the present invention still belong to the protection scope of the technical solution of the present invention.

Claims

1. A method for SLAM autonomous navigation recognition in a closed scenario, characterized in that: It includes the following steps: Step 1, external environment data acquisition. The autonomous robot obtains external environment data through its own camera; Step 2, feature detection. The obtained external environment data is input into the SLAM feature extraction module. In the SLAM feature extraction module, the Vision transformer model is used to realize semantic detection of each object in the external environment, and at the same time, the features of each object are extracted. In the feature detection, the position information of each object in each frame is also located, and the formula is expressed as: ， Among them, and respectively represent the physical positions of the upper left corner points in each frame for the feature , and represent the length and width of the feature . Step 3, data association. The SLAM data association module tracks the common features of images in different frames, and realizes the matching of the same features through correlation clustering between frames, and then judges the movement of the object. The data association includes inputting m features in n frames, a total of n*m feature vectors into the K-means clustering algorithm for iterative calculation. The calculation result classifies the same feature vectors into the same cluster and the different feature vectors into the corresponding clusters to realize the matching of the same features; Step 4, loop detection: The SLAM loop detection module judges whether the movement trajectory of the autonomous robot forms a loop; if a loop is detected, the loop information is provided to the backend optimization module for processing; Step 5, backend optimization and mapping: The backend optimization module continuously receives the images taken by the camera and calculates the camera poses of adjacent frames and receives the loop information, optimizes the cumulative error generated in the data association, and at the same time generates the global movement trajectory and the perception map of the surrounding environment.

2. The method for SLAM autonomous navigation recognition in a closed scenario according to claim 1, characterized in that: In a closed scenario, the number of objects is fixed. In the feature detection of step 2, the number of feature vectors in each frame captured by the camera is fixed and denoted as , and the feature vectors in each frame are denoted as .

3. The method for SLAM autonomous navigation recognition in a closed scenario according to claim 2, characterized in that: The iterative calculation process in the K-means clustering algorithm is as follows: (1) Set the cluster value Select cluster centers and initialize them, denoted as ; is the number of objects in the closed environment; (2) Define its loss function as the sum of squared errors of the distances of each feature vector from the center point of its cluster: Among them represents the th eigenvector in a set of eigenvectors, represents the cluster to which it belongs, represents the center point corresponding to the cluster; (3) For each eigenvector , assign it to the nearest cluster; Denote the variable values when the objective function takes the minimum value; (4) For each cluster , recalculate the center of the cluster ； (5) Set the number of iteration steps, and repeat (3) and (4) until convergence; output the final cluster centers and cluster partitions.

4. The method for SLAM autonomous navigation recognition in a closed scenario according to claim 1, characterized in that: The camera in Step 1 includes one or a combination of a monocular camera, a binocular camera and an RGB-D camera; the autonomous robot also has a laser ranging unit.

Images