CN110400349B - Robot navigation tracking recovery method in small scene based on random forest - Google Patents
Robot navigation tracking recovery method in small scene based on random forest Download PDFInfo
- Publication number
- CN110400349B CN110400349B CN201910593421.2A CN201910593421A CN110400349B CN 110400349 B CN110400349 B CN 110400349B CN 201910593421 A CN201910593421 A CN 201910593421A CN 110400349 B CN110400349 B CN 110400349B
- Authority
- CN
- China
- Prior art keywords
- training
- random
- layer
- images
- nodes
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 92
- 238000007637 random forest analysis Methods 0.000 title claims abstract description 31
- 238000011084 recovery Methods 0.000 title claims abstract description 13
- 238000012549 training Methods 0.000 claims abstract description 88
- 239000011159 matrix material Substances 0.000 claims abstract description 43
- 230000009466 transformation Effects 0.000 claims abstract description 32
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 12
- 238000005070 sampling Methods 0.000 claims abstract description 7
- 238000012163 sequencing technique Methods 0.000 claims description 3
- 238000012360 testing method Methods 0.000 description 6
- 238000004364 calculation method Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 3
- 238000006073 displacement reaction Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 230000004807 localization Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/24323—Tree-organised classifiers
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/70—Determining position or orientation of objects or cameras
- G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a robot navigation tracking recovery method under a small scene based on random forests, which comprises the following steps of (1) selecting a scene to be tracked, and shooting the scene for multiple times by using an RGB-D camera; (2) obtaining a transformation matrix of each group of images by using a three-dimensional reconstruction algorithm to obtain M transformation matrices; (3) obtaining a training label and training characteristics; (4) training a random forest model T by using the training labels and the training characteristics; (5) if the robot fails to track, shooting the scene for 1 time by adopting an RGB-D camera; (6) randomly sampling a plurality of pixel points from an RGB image of the group of images, calculating random characteristics corresponding to each pixel point, sending all the random characteristics into T for training, and outputting 3D world coordinates corresponding to all the random characteristics; and finally, determining the self pose according to the obtained transformation matrix, and completing navigation tracking recovery.
Description
Technical Field
The invention relates to a robot tracking recovery method, in particular to a robot navigation tracking recovery method in a small scene based on a random forest.
Background
With the development of the robot technology, the automatic path finding and mapping functions of the robot play an important role in the robot technology, and in the process, the robot needs to know the position of the robot relative to the real world and the posture of the robot, so that the tracking problem of the robot is always a hotspot problem in robotics.
With the development of vslam (visual Simultaneous Localization and mapping) technology, it becomes very important to restore robot tracking after the robot tracking fails, and for this reason, many tracking restoration methods have been developed. The key point of tracking recovery is to calculate the correct 6D pose of the robot after the tracking fails, and after the correct 6D position is obtained, the robot can know the correct pose of the robot relative to the real world, so that the tracking is recovered.
The method based on local features proposed by S.Se, D.G.Lowe et al has the idea that images with known poses and image feature points with unknown poses are extracted firstly, descriptors are calculated, and pose calculation and optimization are performed by utilizing a P3P algorithm and a RANSAC algorithm.
The keyframe-based methods proposed by g.klein and d.murray infer an assumed pose by querying the similarity of the known image and the keyframe, which is fast, but when the similarity of the unknown image and the known image is very different, the assumed pose may be much different from the true pose.
PoseNet proposed by Alex Kendall, Matthew Grimes and Robert Cipola of Cambridge university utilizes deep learning and utilizes an RGB image to directly predict the image pose, but the method cannot meet the requirement of real-time tracking recovery due to high computational complexity, long time consumption and low precision.
Disclosure of Invention
The invention aims to solve the problems, the training features and the training labels are calculated by using an RGB-D (red, green and blue) -graph, the random forest is trained to predict the coordinates of a 3D scene, and tracking is restored through the coordinate relation, so that the robot navigation tracking restoration method under the small scene based on the random forest is simple in restoration, and high in time period and precision.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows: a robot navigation tracking recovery method in a small scene based on random forests comprises the following steps:
(1) selecting a scene to be tracked, and randomly shooting the scene for M times by using an RGB-D camera to obtain M groups of images, wherein each group of images comprises an RGB image and a depth image;
(2) obtaining a transformation matrix of each group of images by using a three-dimensional reconstruction algorithm to obtain M transformation matrices;
(3) obtaining a training label and training characteristics;
(31) randomly sampling a plurality of pixel points from an RGB image of a first group of images, calculating a 3D world coordinate and random characteristics corresponding to each pixel point, wherein the 3D world coordinate and the random characteristics of the pixel points are in one-to-one correspondence;
(32) obtaining randomly sampled pixel points, 3D world coordinates and random characteristics corresponding to the pixel points by using the remaining M-1 groups of images according to the method in the step (31), using the 3D world coordinates as a training label of the model, and using the random characteristics as training characteristics of the model;
(4) training a random forest model T by using the training labels and the training characteristics;
(41) sequencing all training features, randomly selecting a plurality of training features from the sequenced training features to form a random set, presetting the number of layers adopting a tree balance method as A layers, and presetting the depth threshold of the number as B layers, wherein A is less than B;
(42) acquiring splitting parameters of nodes from a1 st layer to an A th layer by adopting a tree balance method, and acquiring splitting parameters of nodes from an A +1 th layer to a B th layer by adopting a minimum spatial variance method;
(43) forming a random forest model T by the split nodes of each layer;
(5) the robot starts tracking, if the tracking fails, an RGB-D camera is adopted to shoot a scene for 1 time, and a group of images are obtained, wherein the group of images comprise an RGB image and a depth image;
(6) randomly sampling a plurality of pixel points from an RGB image of the group of images, calculating random characteristics corresponding to each pixel point, sending all the random characteristics into T, and outputting 3D world coordinates corresponding to all the random characteristics;
(7) obtaining a prediction transformation matrix according to the 3D world coordinates obtained in the step (6);
(8) and (5) the robot acquires the prediction transformation matrix in the step (7), determines the self pose, and completes navigation tracking recovery.
Preferably, the method comprises the following steps: in the step (2), a Kinectfusion three-dimensional reconstruction method is adopted for each group of images to obtain a transformation matrix.
Preferably, the method comprises the following steps: the step (31) is specifically as follows:
selecting a pixel point, calculating a 3D world coordinate X of the pixel point by adopting the following formula, and calculating a world coordinate W of the pixel point according to the 3D world coordinate X1；
X＝K-1*P*D(P)；
W1＝HX；
Wherein K is the internal reference matrix of the camera, K-1Expressing an inverse matrix of an internal reference matrix, wherein P is a vector of the pixel point, D (P) is a depth value in a depth map corresponding to the pixel point, and H is a transformation matrix corresponding to an RGB map where the pixel point is located;
calculating the random characteristic F of the pixel point by adopting the following formulap,
Wherein, c1And c2Representing two random channels of 3 image channels in an RGB image, I (P, c)1) Representing the channel c at vector P1And δ represents a 2D offset on the pixel coordinate.
Preferably, the method comprises the following steps: in the step (42):
the method for obtaining the splitting parameters of the nodes from the 1 st layer to the A th layer by the tree balancing method specifically comprises the following steps:
A1. constructing nodes of a first layer;
calculating the balance parameter Q of each training feature in the random set by adopting the following formulabSelecting the smallest QbThe corresponding training characteristics are used as splitting parameters of the nodes of the first layer;
in the formula, SLThe number of training features to the left of the current training feature, SRThe number of training features to the right of the current training feature;
A2. respectively forming a data set by using the left and right training features used as the training features of the splitting parameter of the node at the upper layer, and finding out the minimum Q in each data set by using the same method of A1bTaking the corresponding training characteristics as splitting parameters of the next layer of nodes;
A3. repeating the step A2 until a node of layer A is found;
the method for obtaining the splitting parameters of the nodes from the A +1 th layer to the B th layer by the minimum spatial variance method specifically comprises the following steps:
B1. respectively forming training characteristics at two ends of all splitting nodes on the A layer into a data set, and calculating the spatial variance Q of each training characteristic in the data set by adopting the following formulavFind Q in the data setvThe minimum training characteristic is used as a splitting parameter of the node of the A +1 layer;
wherein
Where n denotes the node index of the tree, m denotes the computed training labels, S denotes a randomly selected set of marked pixel points (p, m), S denotes a random selection of the set of marked pixel points (p, m)nRepresenting the complete set at node n, L representing the left sub-tree, R representing the right sub-tree,
representing a set of left or right sub-trees,
represents the mean value of m in S;
B2. according to the method of B1, the splitting parameter from the A +2 layer node to the B layer node is calculated.
Preferably, the method comprises the following steps: in step (7), the specific method for obtaining a prediction transformation matrix is as follows:
and calculating the 3D world coordinates by using a Kabsch method to obtain a pose matrix, and optimizing the pose matrix by using a RANSAC algorithm to finally obtain a prediction transformation matrix.
Compared with the prior art, the invention has the advantages that: the calculation of feature points and descriptors is not needed, the matching of pictures is not needed, and only a random forest model needs to be trained in advance. In the invention, the method for constructing the random forest model is very special, the random forest model is constructed by adopting a tree balance method and a minimum spatial variance method together, and in the tree balance method, Q is obtained according to a formula and a method of the random forest modelbFinally, the method can shorten the time in actual use, and compared with a deep learning method, the method is higher in accuracy and shorter in training and testing time.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a comparison graph of the angle truth value and the evaluation value in example 2;
FIG. 3 is a comparison graph of the shift truth value and the estimation value in example 2;
FIG. 4 is a graph of the angle truth and estimation error of FIG. 2;
FIG. 5 is a graph of the shift values and estimation error of FIG. 4.
Detailed Description
The invention will be further explained with reference to the drawings.
Example 1: referring to fig. 1, a method for recovering robot navigation tracking in a small scene based on a random forest comprises the following steps:
(1) selecting a scene to be tracked, and randomly shooting the scene for M times by using an RGB-D camera to obtain M groups of images, wherein each group of images comprises an RGB image and a depth image;
(2) obtaining a transformation matrix of each group of images by using a three-dimensional reconstruction algorithm to obtain M transformation matrices;
(3) obtaining a training label and training characteristics;
(31) randomly sampling a plurality of pixel points from an RGB image of a first group of images, calculating a 3D world coordinate and random characteristics corresponding to each pixel point, wherein the 3D world coordinate and the random characteristics of the pixel points are in one-to-one correspondence;
(32) obtaining randomly sampled pixel points, 3D world coordinates and random characteristics corresponding to the pixel points by using the remaining M-1 groups of images according to the method in the step (31), using the 3D world coordinates as a training label of the model, and using the random characteristics as training characteristics of the model;
(4) training a random forest model T by using the training labels and the training characteristics;
(41) sequencing all training features, randomly selecting a plurality of training features from the sequenced training features to form a random set, presetting the number of layers adopting a tree balance method as A layers, and presetting the depth threshold of the number as B layers, wherein A is less than B;
(42) acquiring splitting parameters of nodes from a1 st layer to an A th layer by adopting a tree balance method, and acquiring splitting parameters of nodes from an A +1 th layer to a B th layer by adopting a minimum spatial variance method;
(43) forming a random forest model T by the split nodes of each layer;
(5) the robot starts tracking, if the tracking fails, an RGB-D camera is adopted to shoot a scene for 1 time, and a group of images are obtained, wherein the group of images comprise an RGB image and a depth image;
(6) randomly sampling a plurality of pixel points from an RGB image of the group of images, calculating random characteristics corresponding to each pixel point, sending all the random characteristics into T, and outputting 3D world coordinates corresponding to all the random characteristics;
(7) obtaining a prediction transformation matrix according to the 3D world coordinates obtained in the step (6);
(8) and (5) the robot acquires the prediction transformation matrix in the step (7), determines the self pose, and completes navigation tracking recovery.
In this embodiment: in the step (2), a Kinectfusion three-dimensional reconstruction method is adopted for each group of images to obtain a transformation matrix.
The step (31) is specifically as follows:
selecting a pixel point, calculating a 3D world coordinate X of the pixel point by adopting the following formula, and calculating a world coordinate W of the pixel point according to the 3D world coordinate X1；
X＝K-1*P*D(P)；
W1＝HX；
Wherein K is the internal reference matrix of the camera, K-1Expressing an inverse matrix of an internal reference matrix, wherein P is a vector of the pixel point, D (P) is a depth value in a depth map corresponding to the pixel point, and H is a transformation matrix corresponding to an RGB map where the pixel point is located;
calculating the random characteristic F of the pixel point by adopting the following formulap：
Wherein, c1And c2Representing two random channels of 3 image channels in an RGB image, I (P, c)1) Representing the channel c at vector P1And δ represents a 2D offset on the pixel coordinate.
In the step (42):
the method for obtaining the splitting parameters of the nodes from the 1 st layer to the A th layer by the tree balancing method specifically comprises the following steps:
A1. constructing nodes of a first layer;
calculating the balance parameter Q of each training feature in the random set by adopting the following formulabSelecting the smallest QbThe corresponding training characteristics are used as splitting parameters of the nodes of the first layer;
in the formula, SLThe number of training features to the left of the current training feature, SRThe number of training features to the right of the current training feature;
A2. respectively forming a data set by using the left and right training features used as the training features of the splitting parameter of the node at the upper layer, and finding out the minimum Q in each data set by using the same method of A1bTaking the corresponding training characteristics as splitting parameters of the next layer of nodes;
A3. repeating the step A2 until a node of layer A is found;
the method for obtaining the splitting parameters of the nodes from the A +1 th layer to the B th layer by the minimum spatial variance method specifically comprises the following steps:
B1. respectively forming training characteristics at two ends of all splitting nodes on the A layer into a data set, and calculating the spatial variance Q of each training characteristic in the data set by adopting the following formulavFind Q in the data setvThe minimum training characteristic is used as a splitting parameter of the node of the A +1 layer;
wherein
Where n denotes the node index of the tree, m denotes the computed training labels, S denotes a randomly selected set of marked pixel points (p, m), S denotes a random selection of the set of marked pixel points (p, m)nRepresenting the complete set at node n, L representing the left sub-tree, R representing the right sub-tree,
representing a set of left or right sub-trees,
represents the mean value of m in S;
B2. according to the method of B1, the splitting parameter from the A +2 layer node to the B layer node is calculated.
In step (7), the specific method for obtaining a prediction transformation matrix is as follows:
and calculating the 3D world coordinates by using a Kabsch method to obtain a pose matrix, and optimizing the pose matrix by using a RANSAC algorithm to finally obtain a prediction transformation matrix.
In the invention, all coordinates adopt a homogeneous coordinate notation; in addition, in the embodiment, a strategy of calculating all the pixel points is not adopted, only the random features of the randomly selected pixel points are calculated, and the calculated random features and the training labels are in one-to-one correspondence and combined into a queue for training a random forest.
This embodiment mentions a balance parameter QbThe method has the advantages that whether the subtrees are balanced or not can be deduced without really calculating the number of the nodes of the left subtrees and the right subtrees, the calculation time complexity is reduced, the training time is shortened, moreover, when the left subtrees and the right subtrees are approximately balanced, the data part over-fitting or under-fitting caused by overlarge depth difference of any subtrees can be avoided, finally, the random feature with the minimum calculation result is selected as the splitting parameter of the node splitting, and the point worthy of explanation is that the number of the nodes of the left subtrees and the right subtrees is larger than the set minimum number of the nodes.
In this embodiment, the method of minimizing spatial variance is used from the a +1 th layer to the B th layer, and when the number of layers is greater than a, we use this method to obtain the classification parameters. And the random forest generation method is a traditional random forest generation method by adopting a method of minimizing spatial variance to generate random forests. The invention improves the method, only uses the method from A +1 layer to B layer, and uses the tree balance method in the rest.
In this example, to better illustrate the tree balancing method, we take the following example:
assuming that a plurality of training features are obtained in the step (41), randomly extracting 10 training features for sorting, and presetting the number of layers adopting a tree balance method as 2 layers, when the number of layers is less than or equal to 2, adopting the tree balance method, and adopting a minimum spatial variance method for the rest layers. The training characteristics are shown in table 1 below:
table 1: randomly extracting 10 training feature ordered tables
| 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 |
(1) By using QbThe calculation formula of (2) calculates 10 training features to find out Q of the random feature 6bAt a minimum, 6 is used as the splitting parameter, the left data set is 1, 2, 3, 4, 5, the right data set is 8, 9, 10, the left data set is for the left sub-tree, the right data set is for the right sub-tree, and the first layer is complete because there is only one node.
(2) And (5) building a second layer: because the number of layers is also two, the left and right subtrees respectively use the feature sets distributed to themselves by the previous layer to perform the same operation, so that each node is selected, and if 3 selected by the left subtree and 8 selected by the right subtree are assumed, 4 sets are obtained, and are respectively distributed to the left and right subtrees of the node, and then the third layer is established.
(3) And a third layer: the number of layers is more than 2, so that the method of minimizing the space variance is used for finding the QvAt a minimum of QvThe corresponding training features are used as splitting parameters.
This is repeated until the threshold B is reached.
Example 2: referring to fig. 1 to 5, a method for recovering robot navigation tracking in a small scene based on a random forest according to the present embodiment includes the following steps:
(1) selecting a scene to be tracked, and randomly shooting the scene for M times by using an RGB-D camera to obtain M groups of images, wherein each group of images comprises an RGB image and a depth image; in the embodiment, an indoor image set kitchen in a public data set 4-Scenes of Stanford university is directly adopted, the size of an image is 640 × 480, a training picture 550 comprises an RGB-D picture, a test picture comprises 520 RGB-D pictures, and the indoor image set kitchen contains information of a transformation matrix and serves as a true value;
(2) (3) and (4) obtaining a random forest model T in the same way as the embodiment 1;
(5) and the robot starts tracking, and if the tracking fails, the RGB-D camera is adopted to shoot the scene for 1 time to obtain a group of images, wherein the group of images comprises an RGB image and a depth image. Here, to explain the effect of the present embodiment, we adopt the following means:
randomly extracting a plurality of test pictures from the indoor image set kitchen, sending the test pictures into the random forest model T obtained in the step (4) for testing, outputting a prediction transformation matrix of each group of pictures, and taking the prediction transformation matrix as an estimated value of an image transformation matrix;
and (3) analyzing an experimental result:
for comparison, the transformation matrix is analyzed into 6 values in the pitch angle, yaw angle, roll angle, X direction, Y direction and Z direction. FIG. 2 is a comparison graph of the true values and estimated values of the angles calculated by 520 groups of pictures in example 2 of the present invention, and FIG. 3 is a comparison graph of the true values and estimated values of the shifts calculated by 520 groups of pictures in example 2 of the present invention; since the similarity is particularly high, we supplement fig. 4 and fig. 5, in which fig. 4 is a graph of the angle true value and the estimation error in fig. 2, and fig. 5 is a graph of the shift true value and the estimation error in fig. 4.
As can be seen from fig. 2 to 5, the overlap ratio between the estimated value curve and the true value curve is high, which indicates that the similarity between the estimated pose and the true pose is high, the angle error value is mostly within the range of-5 ° to +5 °, the displacement error is mostly within the range of-0.05 m to +0.05m, the difference between the estimated value and the true value is small, the error range is also small, which indicates that the algorithm has better robustness, the output result of the algorithm operation is stable, and the estimated value is used for recovering the tracking, so that a better effect can be obtained.
Example 3: compared with other classical algorithms, the method has obvious application effect.
Referring to table 2, the angle is 5 degrees, the displacement is 5cm, the test result of the method reaches 92.7% of accuracy, and the effect is obvious.
TABLE 2 comparison of the results of the different processes
In table 2, the method based on the keyframes uses a global keyframe matching method, the method based on the local features respectively uses two image features, namely ORB features and SIFT features, and the angle error is greater than 5 ° and the displacement error is greater than 50cm by using the deep learning method, so that the angle error is not listed in the table.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.
Claims (5)
1. A robot navigation tracking recovery method in a small scene based on random forests is characterized by comprising the following steps: the method comprises the following steps:
(1) selecting a scene to be tracked, and randomly shooting the scene for M times by using an RGB-D camera to obtain M groups of images, wherein each group of images comprises an RGB image and a depth image;
(2) obtaining a transformation matrix of each group of images by using a three-dimensional reconstruction algorithm to obtain M transformation matrices;
(3) obtaining a training label and training characteristics;
(31) randomly sampling a plurality of pixel points from an RGB image of a first group of images, calculating a 3D world coordinate and random characteristics corresponding to each pixel point, wherein the 3D world coordinate and the random characteristics of the pixel points are in one-to-one correspondence;
(32) obtaining randomly sampled pixel points, 3D world coordinates and random characteristics corresponding to the pixel points by using the remaining M-1 groups of images according to the method in the step (31), using the 3D world coordinates as a training label of the model, and using the random characteristics as training characteristics of the model;
(4) training a random forest model T by using the training labels and the training characteristics;
(41) sequencing all training features, randomly selecting a plurality of training features from the sequenced training features to form a random set, presetting the number of layers adopting a tree balance method as A layers, and presetting the depth threshold of the number as B layers, wherein A is less than B;
(42) acquiring splitting parameters of nodes from a1 st layer to an A th layer by adopting a tree balance method, and acquiring splitting parameters of nodes from an A +1 th layer to a B th layer by adopting a minimum spatial variance method;
(43) forming a random forest model T by the split nodes of each layer;
(5) the robot starts tracking, if the tracking fails, an RGB-D camera is adopted to shoot a scene for 1 time, and a group of images are obtained, wherein the group of images comprise an RGB image and a depth image;
(6) randomly sampling a plurality of pixel points from an RGB image of the group of images, calculating random characteristics corresponding to each pixel point, sending all the random characteristics into T, and outputting 3D world coordinates corresponding to all the random characteristics;
(7) obtaining a prediction transformation matrix according to the 3D world coordinates obtained in the step (6);
(8) and (5) the robot acquires the prediction transformation matrix in the step (7), determines the self pose, and completes navigation tracking recovery.
2. The method for recovering the robot navigation tracking under the small scene based on the random forest as claimed in claim 1, wherein: in the step (2), a Kinectfusion three-dimensional reconstruction method is adopted for each group of images to obtain a transformation matrix.
3. The method for recovering the robot navigation tracking under the small scene based on the random forest as claimed in claim 1, wherein: the step (31) is specifically as follows:
selecting a pixel point, calculating a 3D world coordinate X of the pixel point by adopting the following formula, and calculating a world coordinate W of the pixel point according to the 3D world coordinate X1；
X＝K-1*P*D(P)；
W1＝HX；
Wherein K is the internal reference matrix of the camera, K-1Expressing an inverse matrix of an internal reference matrix, wherein P is a vector of the pixel point, D (P) is a depth value in a depth map corresponding to the pixel point, and H is a transformation matrix corresponding to an RGB map where the pixel point is located;
calculating the random characteristic F of the pixel point by adopting the following formulap，
Wherein, c1And c2Representing two random channels of 3 image channels in an RGB image, I (P, c)1) Representing the channel c at vector P1And δ represents a 2D offset on the pixel coordinate.
4. The method for recovering the robot navigation tracking under the small scene based on the random forest as claimed in claim 1, wherein: in the step (42):
the method for obtaining the splitting parameters of the nodes from the 1 st layer to the A th layer by the tree balancing method specifically comprises the following steps:
A1. constructing nodes of a first layer;
calculating the balance parameter Q of each training feature in the random set by adopting the following formulabSelecting the smallest QbCorresponding training features are used as the first layerSplitting parameters of the node of (1);
in the formula, SLThe number of training features to the left of the current training feature, SRThe number of training features to the right of the current training feature;
A2. respectively forming a data set by using the left and right training features used as the training features of the splitting parameter of the node at the upper layer, and finding out the minimum Q in each data set by using the same method of A1bTaking the corresponding training characteristics as splitting parameters of the next layer of nodes;
A3. repeating the step A2 until a node of layer A is found;
the method for obtaining the splitting parameters of the nodes from the A +1 th layer to the B th layer by the minimum spatial variance method specifically comprises the following steps:
B1. respectively forming training characteristics at two ends of all splitting nodes on the A layer into a data set, and calculating the spatial variance Q of each training characteristic in the data set by adopting the following formulavFind Q in the data setvThe minimum training characteristic is used as a splitting parameter of the node of the A +1 layer;
wherein
Where n denotes the node index of the tree, m denotes the computed training labels, S denotes a randomly selected set of marked pixel points (p, m), S denotes a random selection of the set of marked pixel points (p, m)nRepresenting the complete set at node n, L representing the left sub-tree, R representing the right sub-tree,
representing a set of left or right sub-trees,
represents the mean value of m in S;
B2. according to the method of B1, the splitting parameter from the A +2 layer node to the B layer node is calculated.
5. The method for recovering the robot navigation tracking under the small scene based on the random forest as claimed in claim 1, wherein: in step (7), the specific method for obtaining a prediction transformation matrix is as follows:
and calculating the 3D world coordinates by using a Kabsch method to obtain a pose matrix, and optimizing the pose matrix by using a RANSAC algorithm to finally obtain a prediction transformation matrix.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910593421.2A CN110400349B (en) | 2019-07-03 | 2019-07-03 | Robot navigation tracking recovery method in small scene based on random forest |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910593421.2A CN110400349B (en) | 2019-07-03 | 2019-07-03 | Robot navigation tracking recovery method in small scene based on random forest |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110400349A CN110400349A (en) | 2019-11-01 |
| CN110400349B true CN110400349B (en) | 2022-04-15 |
Family
ID=68322729
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910593421.2A Expired - Fee Related CN110400349B (en) | 2019-07-03 | 2019-07-03 | Robot navigation tracking recovery method in small scene based on random forest |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110400349B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113255490A (en) * | 2021-05-15 | 2021-08-13 | 成都理工大学 | Unsupervised pedestrian re-identification method based on k-means clustering and merging |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108898623A (en) * | 2018-05-24 | 2018-11-27 | 北京飞搜科技有限公司 | Method for tracking target and equipment |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9940553B2 (en) * | 2013-02-22 | 2018-04-10 | Microsoft Technology Licensing, Llc | Camera/object pose from predicted coordinates |
| US9626766B2 (en) * | 2014-02-28 | 2017-04-18 | Microsoft Technology Licensing, Llc | Depth sensing using an RGB camera |
| US9613298B2 (en) * | 2014-06-02 | 2017-04-04 | Microsoft Technology Licensing, Llc | Tracking using sensor data |
-
2019
- 2019-07-03 CN CN201910593421.2A patent/CN110400349B/en not_active Expired - Fee Related
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108898623A (en) * | 2018-05-24 | 2018-11-27 | 北京飞搜科技有限公司 | Method for tracking target and equipment |
Non-Patent Citations (2)
| Title |
|---|
| 基于树结构分层随机森林在非约束环境下的头部姿态估计;刘袁缘等;《电子与信息学报》;20150315(第03期);第39-47页全文 * |
| 基于深度优先随机森林分类器的目标检测;马娟娟等;《中国惯性技术学报》;20180815(第04期);第104-109页全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110400349A (en) | 2019-11-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109387204B (en) | Mobile robot synchronous positioning and composition method facing indoor dynamic environment | |
| Wang et al. | Matchformer: Interleaving attention in transformers for feature matching | |
| Shotton et al. | Scene coordinate regression forests for camera relocalization in RGB-D images | |
| CN108038420B (en) | Human behavior recognition method based on depth video | |
| CN114782691A (en) | Robot target identification and motion detection method based on deep learning, storage medium and equipment | |
| CN111862213A (en) | Positioning method and device, electronic equipment and computer readable storage medium | |
| Blanton et al. | Extending absolute pose regression to multiple scenes | |
| CN108171249B (en) | RGBD data-based local descriptor learning method | |
| KR20030062313A (en) | Image conversion and encoding techniques | |
| CN111583340B (en) | Method for reducing monocular camera pose estimation error rate based on convolutional neural network | |
| Meng et al. | Exploiting points and lines in regression forests for RGB-D camera relocalization | |
| WO2011131029A1 (en) | Method for detecting similar units based on outline belt graph | |
| CN114842033B (en) | Image processing method for intelligent AR equipment | |
| CN113361542A (en) | Local feature extraction method based on deep learning | |
| CN115761734A (en) | Object pose estimation method based on template matching and probability distribution | |
| CN111797692A (en) | Depth image gesture estimation method based on semi-supervised learning | |
| CN112465021A (en) | Pose track estimation method based on image frame interpolation method | |
| CN116188825A (en) | Efficient feature matching method based on parallel attention mechanism | |
| CN113673354A (en) | Human body key point detection method based on context information and combined embedding | |
| CN110400349B (en) | Robot navigation tracking recovery method in small scene based on random forest | |
| CN114663880A (en) | Three-dimensional target detection method based on multi-level cross-modal self-attention mechanism | |
| Yu et al. | Learning bipartite graph matching for robust visual localization | |
| EP3995993A1 (en) | Process for the extension of deep-learning based 2d-object-detectors for the 6d-pose-estimation | |
| CN113011359A (en) | Method for simultaneously detecting plane structure and generating plane description based on image and application | |
| KR102186764B1 (en) | Apparatus and method for estimating optical flow and disparity via cycle consistency |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20220415 |