Dynamic scene RGB-D SLAM method based on target detection and deep clustering
Through object detection and depth clustering technology, dynamic object areas are extracted and clustered blocks are divided, and feature point errors are calculated, which solves the problem of low positioning accuracy in traditional visual SLAM in dynamic environments, achieving a balance between high precision and real-time, improving the improvement of Robustness.
Patent Information
- Application Number
- CN202410949215.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-16
- Publication Date
- 2024-11-08
AI Technical Summary
Traditional visual SLAM technology has low positioning accuracy in dynamic environments, and existing methods are difficult to ensure both accuracy and real-timeness, especially when dynamic objects occupy most of the field of view, correct feature point matching is affected.
The dynamic object area is extracted by the object detection algorithm, and the dynamic area is divided by K-means clustering, the mean and variance of the reprojection error of the feature points are calculated, the static feature points that meet the conditions are selected, the dynamic feature points are eliminated, the positioning accuracy is improved and real-time performance is ensured. .
It significantly improves the positioning accuracy of the camera in a dynamic environment, reduces the interference of dynamic objects on position estimation, achieves a balance between high accuracy and real-time, and improves robustness.
Smart Images

Figure CN118918184A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of visual navigation technology, in particular to a dynamic scene RGB-D SLAM method based on target detection and depth clustering. Technical Background
[0002] Simultaneous Localization and Mapping (SLAM) technology is a technology that realizes the positioning of the device itself and the construction of the environmental map. Its principle is to use sensors such as cameras, lidar or inertial measurement units (IMU) to collect surrounding environmental information, and then use algorithms to integrate this information to determine the specific position of the device in the location environment and build an environmental map.
[0003] When traditional visual SLAM technology obtains the camera pose, it mostly assumes a static environment and the objects in the scene are stationary. However, in real life, the environment is not always static. When there are moving objects in the environment, the feature points extracted from the moving objects will produce large errors in the pose calculation, resulting in camera tracking failure.
[0004] In order to adapt to dynamic environments, researchers have further optimized the existing visual SLAM framework. In the absence of semantic information as a priori, researchers have used reliable constraints to find the correct feature matching relationship to deal with dynamic SLAM problems. With the development of AI technology, the method of integrating deep learning technology with visual SLAM algorithms has been widely studied in recent years. Deep learning algorithms can obtain prior semantic information, and eliminating dynamic features can enhance the positioning accuracy of the algorithm in dynamic scenes. Through integration with deep learning, visual SLAM technology has achieved good positioning effects in dynamic environments.
[0005] Some existing methods still have some problems. In the absence of semantic information as a priori, dynamic feature points in the scene are identified and eliminated by establishing reliable geometric constraints. The accuracy of their positioning mostly depends on the proportion of dynamic objects in the field of view. When dynamic objects occupy most of the field of view, the correct feature point matches will be treated as dynamic feature points and eliminated, and their positioning accuracy will be greatly affected.
[0006] There are two solutions to improve the performance of visual SLAM in dynamic environments through deep learning. One is to mask dynamic objects through semantic segmentation algorithms. Although this method has high accuracy, it is generally not real-time. The other is to obtain the dynamic area of dynamic objects through target detection algorithms. This method has good real-time performance, but it is easy to mistakenly delete static map points in the obtained dynamic area. Therefore, the accuracy of this method in some scenes is poor. Overall, it is difficult to consider the balance between accuracy and real-time performance when optimizing the performance of visual SLAM in dynamic environments through deep learning methods. Summary of the invention
[0007] In view of the problem that dynamic objects in real environments may affect camera positioning, the present invention provides a dynamic scene RGB-D SLAM method based on target detection and deep clustering.
[0008] A dynamic scene RGB-D SLAM method based on target detection and depth clustering. First, a target detection algorithm is used to obtain a dynamic area of a dynamic object in a scene; secondly, the depth information in the dynamic area is clustered, and the dynamic area is divided into a plurality of cluster blocks; then, a relatively accurate posture is calculated using static feature points outside a detection frame, the reprojection errors of all feature points are calculated, and the mean and variance of the errors are calculated in groups; finally, the dynamic feature points in the cluster blocks that do not meet the conditions are filtered out. The present invention greatly improves the positioning accuracy of a camera in a dynamic environment, while ensuring the real-time and robustness of the system.
[0009] The method includes the following steps:
[0010] Step 1: Use the target detection algorithm to extract the rectangular detection box of the dynamic object in the image;
[0011] In order to know whether there are dynamic objects in the environment, a deep learning-based target detection algorithm is used to identify the location of dynamic objects in the image. The yolov5 algorithm is used here to detect dynamic objects in the environment. The image is input into the yolov5 network model. After prediction by the yolov5 network model, a rectangular bounding box is output to indicate the location of the dynamic object.
[0012] Step 2: Perform K-means clustering on the depth values corresponding to the pixels in the rectangular bounding box to divide the pixel groups in the detection box into cluster blocks;
[0013] The rectangular detection frame identified by the Yolov5 algorithm can select the approximate location of the dynamic object. The pixels in the frame include not only the pixels of the dynamic object, but also some static background pixels. If all feature points in the rectangular detection frame are matched and eliminated, the influence of the dynamic object can be filtered out, but some static background feature points will also be eliminated. Therefore, it is necessary to separate the dynamic object from the static background in the detection frame;
[0014] After observation, the pixel depth difference between dynamic objects is not large, but the pixel depth difference between dynamic objects and static background is large. Based on this feature, clustering can be used to separate dynamic objects from static background according to pixel depth.
[0015] The K-means clustering algorithm can cluster elements with similar distances together. Suppose the depth dataset in the detection box is S, the number of clustering categories is k, and the Euclidean distance is used as the similarity index. The cost function is the sum of the squares of the distances of each category minimized after clustering. The corresponding cost function is shown in formula (1):
[0016]
[0017] S k,i Indicates the k-th depth element; n j represents the number of depth elements of the jth class; k represents the group class; c j Represents the j-th deep cluster center.
[0018] The process of K-Means algorithm is as follows:
[0019] Step 1: Select the number of cluster centers as k;
[0020] Step 2: Calculate the distance between the depth sample and each cluster center;
[0021] Step 3: Assign the depth samples to the cluster center closest to them;
[0022] Step 4: Calculate the mean of each cluster and use the mean as the new cluster center;
[0023] Step 5: If the cluster centers no longer change or the number of algorithm iterations exceeds the set threshold, exit; otherwise, return to Step 3.
[0024] After the pixels in the detection box are clustered in the above steps, pixels with similar depths will be clustered in the same cluster block.
[0025] Step 3: Calculate the initial pose, use the initial pose to calculate the reprojection error of the feature points, and count the mean and variance of the cluster blocks in different parts;
[0026] The pixels outside the default detection area are static pixels, and the feature points on them are all static feature points. The feature points in the detection area can be divided into dynamic feature points and static feature points. In step 2, the K-means algorithm is used on the depth information in the detection frame to cluster most of the static pixels and dynamic pixels together. To this end, we performed the following steps:
[0027] Step 1: Calculate the initial camera pose using static feature points outside the detection box;
[0028] Step 2: Calculate the reprojection error of all feature points using the initial camera pose;
[0029] The reprojection error is calculated as follows:
[0030] The spatial point corresponding to the dynamic feature point is P c , formula (2) transfers the spatial point from the world coordinate system to the current camera coordinate system:
[0031]
[0032] Then transfer the camera coordinate system to the pixel plane coordinate system:
[0033] P u ' v =KP c =(u',v') (3)
[0034] For a binocular camera, the corresponding pixel coordinates in the right eye image have changes in the horizontal coordinates:
[0035]
[0036] in:
[0037] P w : The position of the map point in the world coordinate system;
[0038] P c : The position of the map point in the current camera coordinate system;
[0039] R cw , t cw : The rotation and translation of the current frame relative to the world coordinate system;
[0040] P u ' v , P u ' vr : The two-dimensional projection of the map point on the current frame, where u', v' are the horizontal and vertical coordinates, u' r is the horizontal coordinate of the right target;
[0041] bf : The baseline between the two eyes of the binocular camera;
[0042] Z c : The depth of the map point relative to the current frame.
[0043] The calculation formula of projection error is as follows:
[0044] e 2 =(P uvr -P u ' vr )·(P uvr -P u ' vr ) T (5)
[0045] Step 3: Classify the reprojection error values of all feature points according to cluster blocks and static feature points, and calculate the mean and variance of each group;
[0046] The mean is calculated as follows:
[0047]
[0048] The formula for calculating variance is as follows:
[0049]
[0050] Step 4: Filter the feature points that meet the conditions in the detection frame according to the mean and variance, and calculate the camera pose.
[0051] The size of the reprojection error can reflect the difference between the theoretical value and the actual value of a single feature point in this pose. The error mean and variance of the feature points in each cluster block can reflect the dynamic degree of the feature points in the area. If the error mean and variance of the cluster block feature point set and the error mean and variance of the static feature point set can be kept in a reasonable range, it means that the dynamic degree of the feature points in the cluster block is relatively low. Therefore, in order to eliminate dynamic feature points and obtain accurate camera pose, the following steps are performed:
[0052] Step 1: Calculate the static feature points to get the accurate pose R cw , t cw , use the pose to obtain the reprojection error of the feature points, and divide them into static feature point groups and cluster block groups (the number of cluster block groups is related to K in K-means clustering), and calculate the mean and variance of each group of errors;
[0053] Step 2: Compare the mean and variance of the cluster block group with the static feature point group to filter out the cluster block group with low dynamic degree;
[0054] Step 3: Integrate the feature points in the cluster block groups that meet the conditions into the static feature point group, and repeat Step 1 until no cluster block groups meet the conditions.
[0055] The beneficial effect of the present invention is to improve the traditional ORBSLAM2 algorithm process, solve the problem of low positioning accuracy of the traditional visual SLAM algorithm in a dynamic environment, reduce the interference of dynamic objects on camera pose estimation, and improve the positioning accuracy of the camera. For dynamic objects in the environment, a target detection algorithm is used to detect the pixel area where the dynamic object is located, and the dynamic area is divided into small dynamic clusters by performing a K-Means clustering operation on the depth of the pixels in the dynamic area. The mean and variance of the reprojection error of the feature points in the cluster cluster are compared with the static feature points, and abnormally large cluster blocks are eliminated. While eliminating the dynamic feature points on the dynamic object, better static feature points are retained, and the algorithm is accurately positioned in a dynamic environment, and has better real-time and robustness. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Figure 1 It is a framework for dynamic scene RGB-D SLAM method based on target detection and deep clustering;
[0057] Figure 2 a is the video sequence to be detected;
[0058] Figure 2 b is the result of the target detection algorithm in the video sequence;
[0059] Figure 2 c is the clustering block after K-means clustering of the depth information in the detection frame;
[0060] Figure 2 d is the dynamic clustering block obtained through neural network identification;
[0061] Figure 2 e is the recognition result of dynamic feature points;
[0062] Figure 2 f is the extraction effect of dynamic feature points;
[0063] Figure 3 This is the absolute trajectory error map comparison under the fr3_Walking_static dataset;
[0064] Figure 4 This is the absolute trajectory error map comparison under the fr3_Walking_xyz dataset;
[0065] Figure 5 This is the absolute trajectory error map comparison under the fr3_Walking_halfsphere dataset;
[0066] Figure 6 This is the absolute trajectory error map comparison under the fr3_Walking_rpy dataset. DETAILED DESCRIPTION
[0067] The present invention will be further described below in conjunction with the accompanying drawings.
[0068] like Figure 1 As shown, a dynamic scene RGB-D SLAM method based on target detection and deep clustering includes the following steps:
[0069] Step 1: Use the target detection algorithm to extract the rectangular detection box of the dynamic object in the image;
[0070] In order to know whether there are dynamic objects in the environment, a deep learning-based target detection algorithm is used to identify the location of dynamic objects in the image. The yolov5 algorithm is used here to detect dynamic objects in the environment. The image is input into the yolov5 network model. After prediction by the yolov5 network model, a rectangular bounding box is output to indicate the location of the dynamic object.
[0071] like Figure 2 The image to be detected is shown in a. The image is input into the yolov5 network model. After the model prediction, a rectangular bounding box is output to indicate the location of the dynamic object. The output result is as follows Figure 2 As shown in b;
[0072] Step 2: Perform K-means clustering on the depth values corresponding to the pixels in the rectangular bounding box to divide the pixel groups in the detection box into cluster blocks;
[0073] The rectangular detection frame identified by the Yolov5 algorithm can select the approximate location of the dynamic object. The pixels in the frame include not only the pixels of the dynamic object, but also some static background pixels. If all feature points in the rectangular detection frame are matched and eliminated, the influence of the dynamic object can be filtered out, but some static background feature points will also be eliminated. Therefore, it is necessary to separate the dynamic object from the static background in the detection frame;
[0074] After observation, the pixel depth difference between dynamic objects is not large, but the pixel depth difference between dynamic objects and static background is large. Based on this feature, clustering can be used to separate dynamic objects from static background according to pixel depth.
[0075] The K-means clustering algorithm can cluster elements with similar distances together. Suppose the depth dataset in the detection box is S, the number of clustering categories is k, and the Euclidean distance is used as the similarity index. The cost function is the sum of the squares of the distances of each category minimized after clustering. The corresponding cost function is shown in formula (1):
[0076]
[0077] S k,i Indicates the k-th depth element; n j represents the number of depth elements of the jth class; k represents the group class; c j Represents the j-th deep cluster center.
[0078] The process of K-Means algorithm is as follows:
[0079] Step 1: Select the number of cluster centers as k;
[0080] Step 2: Calculate the distance between the depth sample and each cluster center;
[0081] Step 3: Assign the depth samples to the cluster center closest to them;
[0082] Step 4: Calculate the mean of each cluster and use the mean as the new cluster center;
[0083] Step 5: If the cluster centers no longer change or the number of algorithm iterations exceeds the set threshold, exit; otherwise, return to Step 3.
[0084] After the pixels in the detection box are clustered in the above steps, pixels with similar depths will be clustered in the same cluster block.
[0085] Will Figure 2 The detection box result of b is clustered by K-means, and the number of K-means clustering layers is set to 4. The result of segmenting the cluster blocks is as follows: Figure 2 As shown in c;
[0086] Step 3: Calculate the initial pose, use the initial pose to calculate the reprojection error of the feature points, and count the mean and variance of the cluster blocks in different parts;
[0087] The pixels outside the default detection area are static pixels, and the feature points on them are all static feature points. The feature points in the detection area can be divided into dynamic feature points and static feature points. In step 2, the K-means algorithm is used on the depth information in the detection frame to cluster most of the static pixels and dynamic pixels together. To this end, we performed the following steps:
[0088] Step 1: Calculate the initial camera pose using static feature points outside the detection box;
[0089] Step 2: Calculate the reprojection error of all feature points using the initial camera pose;
[0090] The reprojection error is calculated as follows:
[0091] The spatial point corresponding to the dynamic feature point is P c , formula (2) transfers the spatial point from the world coordinate system to the current camera coordinate system:
[0092]
[0093] Then transfer the camera coordinate system to the pixel plane coordinate system:
[0094] P u ' v =KP c =(u',v') (3)
[0095] For a binocular camera, the corresponding pixel coordinates in the right eye image have changes in the horizontal coordinates:
[0096]
[0097] in:
[0098] P w : The position of the map point in the world coordinate system;
[0099] P c : The position of the map point in the current camera coordinate system;
[0100] R cw , t cw : The rotation and translation of the current frame relative to the world coordinate system;
[0101] P u ' v , P u ' vr : The two-dimensional projection of the map point on the current frame, where u', v' are the horizontal and vertical coordinates, u' r is the horizontal coordinate of the right target;
[0102] b f : The baseline between the two eyes of the binocular camera;
[0103] Z c : The depth of the map point relative to the current frame.
[0104] The calculation formula of projection error is as follows:
[0105] e 2 =(Puvr -P u ' vr )·(P uvr -P u ' vr ) T (5)
[0106] Step 3: Classify the reprojection error values of all feature points according to cluster blocks and static feature points, and calculate the mean and variance of each group;
[0107] The mean is calculated as follows:
[0108]
[0109] The formula for calculating variance is as follows:
[0110]
[0111] In this experiment, the number of clusters is 4, so the calculated error mean and variance are 5 groups, namely, 1 group of static feature points, such as Figure 2 As shown in d, there are 4 groups of cluster block feature points, such as Figure 2 As shown in e;
[0112] Step 4: Filter the feature points that meet the conditions in the detection frame according to the mean and variance, and calculate the camera pose.
[0113] The size of the reprojection error can reflect the difference between the theoretical value and the actual value of a single feature point in this pose. The error mean and variance of the feature points in each cluster block can reflect the dynamic degree of the feature points in the area. If the error mean and variance of the cluster block feature point set and the error mean and variance of the static feature point set can be kept in a reasonable range, it means that the dynamic degree of the feature points in the cluster block is relatively low. Therefore, in order to eliminate dynamic feature points and obtain accurate camera pose, the following steps are performed:
[0114] Step 1: Calculate the static feature points to get the accurate pose R cw , t cw , use the pose to obtain the reprojection error of the feature points, and divide them into static feature point groups and cluster block groups (the number of cluster block groups is related to K in K-means clustering), and calculate the mean and variance of each group of errors;
[0115] Step 2: Compare the mean and variance of the cluster block group with the static feature point group to filter out the cluster block group with low dynamic degree;
[0116] Step 3: Integrate the feature points in the cluster block groups that meet the conditions into the static feature point group, and repeat Step 1 until no cluster block groups meet the conditions.
[0117] During the experiment, we found that when comparing the reprojection error mean and variance of the clustered block feature points with the static feature points, the settings are as follows:
[0118]
[0119] in:
[0120] The mean and variance of the reprojection error of the static feature point group;
[0121] The mean and variance of the reprojection error of the clustered block feature point group, i=1,2…n, n is the number of clusters.
[0122] When formula (8) is satisfied, the camera accuracy is optimal and the dynamic feature point removal effect is as follows: Figure 2 f.
[0123] Example 1
[0124] Experiments were conducted using four dynamic sequences, fr3_Walking_static, fr3_Walking_xyz, fr3_Walking_halfsphere, and fr3_Walking_rpy, from the public TUM dataset. In these datasets, two people walked back and forth in an office scene. Walking indicates a high-dynamic sequence, and the suffixes xyz, rpy, halfsphere, and static represent different camera movement modes. The experimental platform used a laptop with Ubuntu 20.04 system and an i7-7700HQ processor.
[0125] The method of the present invention is quantitatively compared with ORB-SLAM2. The absolute trajectory error (ATE) comparison results of the experiment are shown in Table 1 below. The root mean square error (RMSE), standard deviation error (SD), and the improvement of the two indicators are calculated respectively.
[0126] Table 1
[0127]
[0128] As can be seen from Table 1, the performance of the proposed method in the high dynamic data set sequence is significantly improved compared with the ORBSLAM2 algorithm, and the improvement of RMSE and SD reaches 94.76% and 95.09% on average.
[0129] The absolute trajectory error graphs of the four data sets are Figure 3-6In each figure, the left figure is the absolute trajectory error diagram of the method of the present invention, and the right figure is the absolute trajectory error diagram of the ORBSLAM2 algorithm, wherein the dotted line is the true trajectory of the data set, and the solid line is the predicted trajectory of the visual SLAM algorithm. It can be seen that the method of the present invention has a higher degree of fit with the true trajectory and has a better tracking effect.
[0130] The running speed of the method of the present invention is shown in Table 2, wherein the tracking threads of the yolo detection algorithm and the SLAM algorithm run in parallel, the average time for the algorithm of the present invention to process each frame of data is 43.561ms, and the running speed can reach 23FPS, which can better meet the real-time requirements while improving the accuracy and robustness of the algorithm.
[0131] Table 2
[0132]
Claims
1. A dynamic scene RGB-D SLAM method based on target detection and deep clustering, characterized by: The following steps are involved: Step 1: Use the target detection algorithm to extract the rectangular detection box of the dynamic object in the image; In order to know whether there are dynamic objects in the environment, a deep learning-based target detection algorithm is used to identify the location of dynamic objects in the image. The yolov5 algorithm is used here to detect dynamic objects in the environment. The image is input into the yolov5 network model. After prediction by the yolov5 network model, a rectangular bounding box is output to indicate the location of the dynamic object. Step 2: Perform K-means clustering on the depth values corresponding to the pixels in the rectangular bounding box to divide the pixel groups in the detection box into cluster blocks; The rectangular detection frame identified by the Yolov5 algorithm can select the approximate location of the dynamic object. The pixels in the frame include not only the pixels of the dynamic object, but also some static background pixels. If all feature points in the rectangular detection frame are matched and eliminated, the influence of the dynamic object can be filtered out, but some static background feature points will also be eliminated. Therefore, it is necessary to separate the dynamic object from the static background in the detection frame; After observation, the pixel depth difference between dynamic objects is not large, but the pixel depth difference between dynamic objects and static background is large. Based on this feature, clustering can be used to separate dynamic objects from static background according to pixel depth. The K-means clustering algorithm can cluster elements with similar distances together. Suppose the depth dataset in the detection box is S, the number of clustering categories is k, and the Euclidean distance is used as the similarity index. The cost function is the sum of the squares of the distances of each category minimized after clustering. The corresponding cost function is shown in formula (1): S k,i Indicates the k-th depth element; n j represents the number of depth elements of the jth class; k represents the group class; c j Represents the j-th deep cluster center. The process of K-Means algorithm is as follows: Step 1: Select the number of cluster centers as k; Step 2: Calculate the distance between the depth sample and each cluster center; Step 3: Assign the depth samples to the cluster center closest to them; Step 4: Calculate the mean of each cluster and use the mean as the new cluster center; Step 5: If the cluster centers no longer change or the number of algorithm iterations exceeds the set threshold, exit; otherwise, return to Step 3. After the pixels in the detection box are clustered in the above steps, pixels with similar depths will be clustered in the same cluster block. Step 3: Calculate the initial pose, use the initial pose to calculate the reprojection error of the feature points, and count the mean and variance of the cluster blocks in different parts; The pixels outside the default detection area are static pixels, and the feature points on them are all static feature points. The feature points in the detection area can be divided into dynamic feature points and static feature points. In step 2, the K-means algorithm is used on the depth information in the detection frame to cluster most of the static pixels and dynamic pixels together. To this end, we performed the following steps: Step 1: Calculate the initial camera pose using static feature points outside the detection box; Step 2: Use the initial camera pose to calculate the reprojection error of all feature points; The reprojection error is calculated as follows: The spatial point corresponding to the dynamic feature point is P c , formula (2) transfers the spatial point from the world coordinate system to the current camera coordinate system: Then transfer the camera coordinate system to the pixel plane coordinate system: P u ' v =KP c =(u',v') (3) For a binocular camera, the corresponding pixel coordinates in the right eye image have changes in the horizontal coordinates: in: P w : The position of the map point in the world coordinate system; P c : The position of the map point in the current camera coordinate system; R cw , t cw : The rotation and translation of the current frame relative to the world coordinate system; P u ' v , P u ' vr : The two-dimensional projection of the map point on the current frame, where u', v' are the horizontal and vertical coordinates, u' r is the horizontal coordinate of the right target; b f : The baseline between the two eyes of the binocular camera; Z c : The depth of the map point relative to the current frame. The calculation formula of projection error is as follows: e 2 =(P uvr -P u ' vr )·(P uvr -P u ' vr ) T (5) Step 3: Classify the reprojection error values of all feature points according to cluster blocks and static feature points, and calculate the mean and variance of each group; The mean is calculated as follows: The formula for calculating variance is as follows: Step 4: Filter the feature points that meet the conditions in the detection frame according to the mean and variance, and calculate the camera pose. The size of the reprojection error can reflect the difference between the theoretical value and the actual value of a single feature point in this pose. The error mean and variance of the feature points in each cluster block can reflect the dynamic degree of the feature points in the area. If the error mean and variance of the cluster block feature point set and the error mean and variance of the static feature point set can be kept in a reasonable range, it means that the dynamic degree of the feature points in the cluster block is relatively low. Therefore, in order to eliminate dynamic feature points and obtain accurate camera pose, the following steps are performed: Step 1: Calculate the static feature points to get the accurate pose R cw , t cw , use the pose to obtain the reprojection error of the feature points, and divide it into static feature point groups and cluster block groups, and calculate the mean and variance of each group of errors; Step 2: Compare the mean and variance of the cluster block group with the static feature point group to filter out the cluster block group with low dynamic degree; Step 3: Integrate the feature points in the cluster block groups that meet the conditions into the static feature point group, and repeat Step 1 until no cluster block groups meet the conditions.
2. The dynamic scene RGB-DSLAM method based on target detection and deep clustering according to claim 1, characterized in that: The number of cluster block groups is related to K of K-means clustering.
Citation Information
Cited By
Machine vision target positioning method and system based on bounding box adaptive adjustment
CN120279252A
