CN110296705B - Visual SLAM loop detection method based on distance metric learning - Google Patents
Visual SLAM loop detection method based on distance metric learning Download PDFInfo
- Publication number
- CN110296705B CN110296705B CN201910575905.4A CN201910575905A CN110296705B CN 110296705 B CN110296705 B CN 110296705B CN 201910575905 A CN201910575905 A CN 201910575905A CN 110296705 B CN110296705 B CN 110296705B
- Authority
- CN
- China
- Prior art keywords
- scene
- frame
- picture
- loop
- training
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
- G01C21/00—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00
- G01C21/20—Instruments for performing navigational calculations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Remote Sensing (AREA)
- Radar, Positioning & Navigation (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Automation & Control Theory (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a visual SLAM loop detection method based on distance metric learning, which comprises the following steps: pre-training the CNN model, and optimizing by using a training set; averagely dividing pictures of a training set into k groups, and simultaneously inputting k CNN pre-training models which share parameters; constructing a multi-tuple for training by using a multi-tuple construction method; after constructing the multi-element groups of all scenes, optimizing a CNN model according to the distance relation among the multi-element groups; when no proper multi-tuple is constructed in the fourth step, directly finishing the training and entering a testing stage; and (5) using the optimized CNN model to perform actual robot loop detection application. The invention solves the technical problem of robustness to the appearance change and the visual angle change at the same time, and reduces the computation amount of similarity measurement.
Description
Technical Field
The invention discloses a visual SLAM loop detection method based on distance metric learning, and relates to the technical field of robot mobile positioning.
Background
Visual SLAM (simultaneous localization and mapping) is a key technology in the field of mobile robots. In a SLAM system, the robot will model the surrounding environment and estimate its motion trajectory synchronously. A typical visual SLAM system is generally made up of several modules: visual odometry, rear-end optimization, loop detection and graph building. The loop detection module is used for automatically detecting whether the robot has arrived at a certain place or not. If the loop is successfully detected, the robot can provide an additional constraint term for the optimization of the back end and reduce the optimization error.
During the actual navigation, the environment around the robot may change. These changes can be classified into changes in appearance and changes in viewing angle. The change in appearance may be due to changes in lighting, weather, and shadows. Meanwhile, pictures shot by the robot at different angles in the same place in the moving process look different. Therefore, loop detection needs to be robust to changes in appearance and viewing angle.
In addition to being robust to changes in appearance and viewing angle, another important issue for loop detection is real-time. The robot needs to determine whether the current position has passed or not within a short time. However, loop back detection cannot occupy much computing resources, since other modules of the visual SLAM also consume a large amount of memory.
At present, most of loop detection methods for mainstream visual SLAM systems are bag-of-word models. The bag-of-words model requires manual design of features and composition of the features into a dictionary, and the distance between the target picture and the features in the dictionary is calculated, so as to judge whether a loop is formed. However, this method of artificially designing features is highly susceptible to interference from environmental factors such as lighting, shooting angles, and dynamic objects.
Recently, many researchers have used Convolutional Neural Networks (CNNs) to solve the loop detection problem. Compared with the traditional method, the CNN can extract picture features with higher quality, and the CNN model pre-trained on the ImageNet data set has strong generalization and can be used for solving various different visual tasks. Research has found that features extracted from the middle layer of the CNN are robust to appearance changes, and features extracted from the high layer of the CNN are robust to view angle changes. However, how to have robustness to both appearance change and viewing angle change remains a key technical challenge. Meanwhile, the similarity measure of pictures is usually the euclidean distance between features of the comparison pictures. However, the image feature dimension extracted by the CNN model is very high, so that the similarity measurement requires a large amount of computation, which is a great challenge to real-time performance.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: aiming at the defects of the prior art, the visual SLAM loop detection method based on distance metric learning is provided, the technical problem that robustness is provided for appearance change and visual angle change at the same time is solved, and the calculation amount of similarity metric is reduced.
The invention adopts the following technical scheme for solving the technical problems:
a visual SLAM loop detection method based on distance metric learning, the method comprising the steps of:
firstly, pre-training a CNN model, and optimizing by using a training set;
averagely dividing the pictures of the training set into k groups, and simultaneously inputting k CNN pre-training models which share parameters;
step three, constructing a multi-tuple by using a multi-tuple construction method for training;
fourthly, after the multi-element groups of all scenes are constructed, optimizing the CNN model according to the distance relation among the multi-element groups;
step five, when no proper multi-tuple is constructed in the step four, directly finishing the training and entering a testing stage;
and step six, using the optimized CNN model to perform actual robot loop detection application.
As a further preferred aspect of the present invention, in the first step, the training set is a NewCollege, City Center and TUM data set; the New college and City Center datasets are collected by two cameras on the left and right sides of the mobile robot and provide GPS information and real loopback picture markers; the TUM data set is an indoor data set collected by a depth camera and provides true trajectory information.
As a further preferable scheme of the present invention, in the second step, the pre-training model is VGG-16, and the picture feature is extracted by using the first full-connected layer of the VGG-16.
As a further preferable scheme of the present invention, the step three specifically includes:
3.1, taking all picture features of the first scene, and calculating the central points of the feature vectors;
let the feature vector O belong to the ith sceneiComprises the following steps:
wherein n is the number of the feature vectors;
wherein a is the a-th feature vector belonging to the i-th scene;
3.2, calculating Euclidean distances from all the characteristic vectors of the scene to a central point, and if the characteristic vectors larger than a threshold value d exist, directly entering the step 3.4 to construct a multi-element group for the scene; if all the Euclidean distances from the feature vectors to the central point are smaller than the threshold value d, the step 3.3 is carried out to continue judging;
3.3, calculating Euclidean distances from all the characteristic vectors which do not belong to the scene to the central point, if the characteristic vectors smaller than a threshold value d exist, entering step 3.4, constructing a multi-element group for the scene, and if the characteristic vectors are not smaller than the threshold value d, directly entering step 3.5, and judging whether the scene is traversed completely;
3.4, constructing a multi-element group for the scene, taking A picture characteristics belonging to the scene and B picture characteristics not belonging to the scene, wherein the A picture characteristics and the B picture characteristics are from different scenes, and the multi-element group UiCan be expressed as:
wherein A is the number of the selected characteristic vectors belonging to the ith scene, and B is the number of the characteristic vectors not belonging to the ith scene;
and 3.5, judging whether the scene is traversed or not, if so, starting to perform reverse propagation, and otherwise, returning to the step 3.1 to take the picture characteristics of the next scene.
As a further preferred embodiment of the present invention, in step four, in the optimization process, the distance loss function L of the multi-tuple pictures of 5 scenes and the multi-tuple picture of the p-th scene are input each timemulti-constraintunit(O) is calculated as follows:
wherein the content of the first and second substances,is the feature vector center point of the p-th scene,is the feature vector in the p-th scene that is farthest from the center point,is the feature vector of the non-p-th scene closest to the center point,is different from random selection in the tuplesIs a predefined constant parameter, representing aToA distance betweenToThe minimum boundary of the distance between, beta is a predefined constant parameter, representingToA distance betweenToThe minimum boundary of the distance between.
As a further preferable scheme of the present invention, the step six specifically includes:
6.1, inputting continuous k frames of pictures into the optimized k shared CNN models simultaneously, wherein k is A + B, and extracting picture characteristics of each picture;
6.2, carrying out similarity measurement on the extracted feature vector of each frame of picture and the feature vector in the historical picture features;
6.3, supposing that the similarity measurement is carried out on the i-frame picture, selecting a T-frame picture closest to the (i-1) frame picture, taking the T-frame picture as a center, and selecting Num frame pictures before and after the T-frame picture as reference candidate frames, so that at most T (Num +1) reference candidate frames can be obtained;
comparing Euclidean distances between the ith frame picture and each reference candidate frame, and if the Euclidean distances are smaller than a threshold value theta, taking the reference candidate frame as an undetermined loop of the ith frame picture so as to reduce the similarity measurement range;
6.4, once a loop is formed between two frames, the two frames are listed as pending loops;
and further judging whether a loop exists between the front frame and the rear frame of the two frames, and if the loop is formed between the front frame and the rear frame, determining the front frame and the rear frame as loop frames.
As a further preferable aspect of the present invention, the method for determining whether a loop exists between two frames before and after the two frames includes:
calculating the Euclidean distance between the feature vector extracted from the Mth frame and the feature vector extracted from the Nth frame, and recording the 1 st result as a loop if the distance is less than a threshold value theta;
respectively calculating Euclidean distances between the feature vectors extracted from the two frames before and after the Mth frame and the feature vectors extracted from the two frames before and after the Nth frame, and recording the Euclidean distances as a W result, wherein W is greater than 1, if a loop exists in the W result, recording the Mth frame and the Nth frame as a loop group, otherwise, recording the Mth frame and the Nth frame as a non-loop group.
Compared with the prior art, the invention adopting the technical scheme has the following technical effects: the invention solves the technical problem of robustness to the appearance change and the visual angle change at the same time, and reduces the computation amount of similarity measurement.
Drawings
FIG. 1 is a system flow diagram of the method of the present invention.
Fig. 2 is a schematic diagram of a training process in the method of the present invention.
FIG. 3 is a schematic diagram of a multi-element construction method in the method of the present invention.
Detailed Description
Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the accompanying drawings are illustrative only for the purpose of explaining the present invention, and are not to be construed as limiting the present invention.
The technical scheme of the invention is further explained in detail by combining the attached drawings:
the system flow chart of the method of the invention is shown in fig. 1, and the method for detecting visual SLAM loop based on distance metric learning disclosed by the invention comprises the following specific steps:
step one, on the basis of a pre-trained CNN model, a training set is used for optimization, and then picture features are extracted. The type of training set and the structure of the CNN are two important considerations for feature extraction. The New college and City Center datasets are two outdoor datasets that are widely used in the field of visual SLAM closed loop detection. They are collected by the mobile robot at the left and right two cameras and provide GPS information and real loop back picture markers. The TUM is an indoor data set acquired by a depth camera and provides true trajectory information. Therefore, the selection is trained on the NewCollege, City Center, and TUM datasets.
Secondly, VGG-16 is a classification model trained on a large object recognition dataset ImageNet. The device has simple and effective structure and strong generalization capability to other data sets. Therefore, VGG-16 is selected as the pre-training model. Meanwhile, studies have shown that features extracted from different layers of CNN exhibit different properties in image tasks. The middle layer characteristics of the CNN have better generalization capability than the high layer characteristics, and the higher the number of the CNN full-connection layers is, the worse the generalization capability is. In order to ensure that the extracted features have low dimensionality and good generalization capability, the first full-link layer of VGG-16 is selected to be used for extracting the picture features.
Step two, the training process is as shown in fig. 2, before starting the training, firstly, the pictures of the training set are averagely divided into k groups, and k CNN pre-training models are simultaneously input, and the k CNN pre-training models share parameters. In this way, the feature vector of each picture can be extracted.
And step three, constructing the multi-tuple by using a multi-tuple construction method for training, as shown in fig. 3.
And 3.1, taking all picture features of the first scene, and calculating the central points of the feature vectors. Let the feature vector belonging to the ith scene be:n is the number of feature vectors. Then the feature vector center point of the ith scene is:
And 3.2, calculating Euclidean distances from all the feature vectors of the scene to a central point, and if the feature vectors larger than a threshold value d exist, directly entering the step 3.4 to construct a multi-element group for the scene. And if the Euclidean distances from all the feature vectors to the central point are smaller than the threshold value d, continuing to judge in the next step.
And 3.3, calculating Euclidean distances from all the feature vectors which do not belong to the scene to the central point, and if the feature vectors smaller than a threshold value d exist, entering the next step and constructing a multi-element group for the scene. Otherwise, directly entering the step 3.5 and judging whether the scene is traversed.
And 3.4, constructing a multi-tuple for the scene. And taking A picture features belonging to the scene and B picture features not belonging to the scene, wherein the A picture features and the B picture features are from different scenes. The multivariate group can be represented as:
And 3.5, judging whether the scene is traversed or not, and starting to perform reverse propagation if the scene is traversed. Otherwise, returning to the step 3.1, and taking the picture characteristics of the next scene.
And fourthly, after the multi-element groups of all scenes are constructed, optimizing the CNN model according to the distance relation among the multi-element groups. In the optimization process, each time 5 scene multi-tuple pictures are input, the distance loss function of the p scene multi-tuple picture is defined as follows:
wherein the content of the first and second substances,is the feature vector center point of the p-th scene,is the feature vector in the p-th scene that is farthest from the center point,is the feature vector of the non-p-th scene closest to the center point,is different from random selection in the tuplesIs not the p-th scene. α is a predefined constant parameter, representingToA distance betweenToThe minimum boundary of the distance between, beta is a predefined constant parameter, representingToA distance betweenToThe minimum boundary of the distance between.
And step five, if no proper multi-element group is constructed in the step four, directly finishing the training and entering a testing stage.
And step six, using the optimized CNN model to perform actual robot loop detection application, wherein the specific implementation process is as follows:
and 6.1, simultaneously inputting continuous k frames of pictures into the optimized k shared CNN models, wherein k is A + B, and extracting picture characteristics of each picture.
And 6.2, performing similarity measurement on the extracted feature vector of each frame of picture and the feature vector in the historical picture features.
And 6.3, assuming that similarity measurement is carried out on the i-frame picture, selecting a T-frame picture closest to the (i-1) frame picture, taking the T-frame picture as a center, and selecting Num frame pictures before and after the T-frame picture as reference candidate frames, so that at most T (Num +1) reference candidate frames can be obtained. And comparing Euclidean distances between the picture of the ith frame and each reference candidate frame, and if the Euclidean distances are smaller than a threshold value theta, taking the reference candidate frame as a loop to be determined of the picture of the ith frame. Thereby reducing the range of similarity metrics.
6.4, in order to avoid detecting false loops as much as possible, the following optimization strategies are adopted: once a loop is detected between two frames, they are first classified as pending loops. The robot has continuity in the process of traveling, whether a loop exists between the two frames before and after the two frames can be judged, and if the loop is formed between the two frames before and after, the two frames are determined as loop frames.
Namely: and calculating the Euclidean distance between the feature vector extracted from the Mth frame and the feature vector extracted from the Nth frame, and if the distance is smaller than a threshold value theta, marking the 1 st result as a loop. Respectively calculating Euclidean distances between the feature vectors extracted from two frames before and after the Mth frame and the feature vectors extracted from two frames before and after the Nth frame, and recording as the Wth result (wherein W is more than 1). And if the loop exists in the W results, recording the Mth frame and the Nth frame as a loop group, and otherwise, recording the Mth frame and the Nth frame as a non-loop group.
The embodiments of the present invention have been described in detail with reference to the drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the gist of the present invention. Although the present invention has been described with reference to a preferred embodiment, it should be understood that various changes, substitutions and alterations can be made herein without departing from the spirit and scope of the invention as defined by the appended claims.
Claims (5)
1. A visual SLAM loop detection method based on distance metric learning is characterized by comprising the following steps:
firstly, pre-training a CNN model, and optimizing by using a training set;
averagely dividing the pictures of the training set into k groups, and simultaneously inputting k CNN pre-training models which share parameters;
step three, constructing a multi-tuple by using a multi-tuple construction method for training;
fourthly, after the multi-element groups of all scenes are constructed, optimizing the CNN model according to the distance relation among the multi-element groups;
step five, when no proper multi-tuple is constructed in the step four, directly finishing the training and entering a testing stage;
step six, using the optimized CNN model to perform actual robot loop detection application;
the third step specifically comprises:
3.1, taking all picture features of the first scene, and calculating the central points of the feature vectors;
let the feature vector O belong to the ith sceneiComprises the following steps:
wherein n is the number of the feature vectors;
wherein a is the a-th feature vector belonging to the i-th scene;
3.2, calculating Euclidean distances from all the characteristic vectors of the scene to a central point, and if the characteristic vectors larger than a threshold value d exist, directly entering the step 3.4 to construct a multi-element group for the scene; if all the Euclidean distances from the feature vectors to the central point are smaller than the threshold value d, the step 3.3 is carried out to continue judging;
3.3, calculating Euclidean distances from all the characteristic vectors which do not belong to the scene to the central point, if the characteristic vectors smaller than a threshold value d exist, entering step 3.4, constructing a multi-element group for the scene, and if the characteristic vectors are not smaller than the threshold value d, directly entering step 3.5, and judging whether the scene is traversed completely;
3.4, constructing a multi-element group for the scene, taking A picture characteristics belonging to the scene and B picture characteristics not belonging to the scene, wherein the A picture characteristics and the B picture characteristics are from different scenes, and the multi-element group UiCan be expressed as:
wherein A is the number of the selected characteristic vectors belonging to the ith scene, and B is the number of the characteristic vectors not belonging to the ith scene;
3.5, judging whether the scene is traversed or not, if so, starting to perform reverse propagation, otherwise, returning to the step 3.1, and taking the picture characteristics of the next scene;
in step four, in the optimization process, the distance loss function L of the multi-tuple pictures of 5 scenes and the multi-tuple picture of the p-th scene are input each timemulti-constraintunit(O) is calculated as follows:
wherein the content of the first and second substances,is the feature vector center point of the p-th scene,is the feature vector in the p-th scene that is farthest from the center point,is the feature vector of the non-p-th scene closest to the center point,is different from random selection in the tuplesIs a predefined constant parameter, representing aToA distance betweenToThe minimum boundary of the distance between, beta is a predefined constant parameter, representingToA distance betweenToThe minimum boundary of the distance between.
2. The visual SLAM loop detection method based on distance metric learning of claim 1, characterized in that: in the first step, the training set is a NewCollege, City Center and TUM data set; the New college and City Center datasets are collected by two cameras on the left and right sides of the mobile robot and provide GPS information and real loopback picture markers; the TUM data set is an indoor data set collected by a depth camera and provides true trajectory information.
3. The visual SLAM loop detection method based on distance metric learning of claim 1, characterized in that: in the second step, the pre-training model is VGG-16, and the picture features are extracted by using the first full connection layer of the VGG-16.
4. The visual SLAM loop detection method based on distance metric learning of claim 1, wherein the sixth step specifically comprises:
6.1, inputting continuous k frames of pictures into the optimized k shared CNN models simultaneously, wherein k is A + B, and extracting picture characteristics of each picture;
6.2, carrying out similarity measurement on the extracted feature vector of each frame of picture and the feature vector in the historical picture features;
6.3, supposing that the similarity measurement is carried out on the i-frame picture, selecting a T-frame picture closest to the (i-1) frame picture, taking the T-frame picture as a center, and selecting Num frame pictures before and after the T-frame picture as reference candidate frames, so that at most T (Num +1) reference candidate frames can be obtained;
comparing Euclidean distances between the ith frame picture and each reference candidate frame, and if the Euclidean distances are smaller than a threshold value theta, taking the reference candidate frame as an undetermined loop of the ith frame picture so as to reduce the similarity measurement range;
6.4, once a loop is formed between two frames, the two frames are listed as pending loops;
and further judging whether a loop exists between the front frame and the rear frame of the two frames, and if the loop is formed between the front frame and the rear frame, determining the front frame and the rear frame as loop frames.
5. The method as claimed in claim 4, wherein the method for determining whether a loop exists between two frames before and after the two frames is as follows:
calculating the Euclidean distance between the feature vector extracted from the Mth frame and the feature vector extracted from the Nth frame, and recording the 1 st result as a loop if the distance is less than a threshold value theta;
respectively calculating Euclidean distances between the feature vectors extracted from the two frames before and after the Mth frame and the feature vectors extracted from the two frames before and after the Nth frame, and recording the Euclidean distances as a W result, wherein W is greater than 1, if a loop exists in the W result, recording the Mth frame and the Nth frame as a loop group, otherwise, recording the Mth frame and the Nth frame as a non-loop group.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910575905.4A CN110296705B (en) | 2019-06-28 | 2019-06-28 | Visual SLAM loop detection method based on distance metric learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910575905.4A CN110296705B (en) | 2019-06-28 | 2019-06-28 | Visual SLAM loop detection method based on distance metric learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110296705A CN110296705A (en) | 2019-10-01 |
CN110296705B true CN110296705B (en) | 2022-01-25 |
Family
ID=68029496
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910575905.4A Active CN110296705B (en) | 2019-06-28 | 2019-06-28 | Visual SLAM loop detection method based on distance metric learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110296705B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111325712B (en) * | 2020-01-20 | 2024-01-23 | 北京百度网讯科技有限公司 | Method and device for detecting image validity |
CN111882663A (en) * | 2020-07-03 | 2020-11-03 | 广州万维创新科技有限公司 | Visual SLAM closed-loop detection method achieved by fusing semantic information |
CN113033555B (en) * | 2021-03-25 | 2022-12-23 | 天津大学 | Visual SLAM closed loop detection method based on metric learning |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106203533A (en) * | 2016-07-26 | 2016-12-07 | 厦门大学 | The degree of depth based on combined training study face verification method |
CN108986168A (en) * | 2018-06-13 | 2018-12-11 | 深圳市感动智能科技有限公司 | A kind of robot winding detection method and device combining bag of words tree-model based on depth measure study |
CN109101602A (en) * | 2018-08-01 | 2018-12-28 | 腾讯科技(深圳)有限公司 | Image encrypting algorithm training method, image search method, equipment and storage medium |
CN109215117A (en) * | 2018-09-12 | 2019-01-15 | 北京航空航天大学青岛研究院 | Flowers three-dimensional rebuilding method based on ORB and U-net |
CN109341703A (en) * | 2018-09-18 | 2019-02-15 | 北京航空航天大学 | A kind of complete period uses the vision SLAM algorithm of CNNs feature detection |
CN109871803A (en) * | 2019-02-18 | 2019-06-11 | 清华大学 | Robot winding detection method and device |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10945641B2 (en) * | 2011-10-09 | 2021-03-16 | The Medical Research, Infrastructure and Health Services Fund of the Tel Aviv Medical Center | Virtual reality for movement disorder diagnosis and/or treatment |
CN105512273A (en) * | 2015-12-03 | 2016-04-20 | 中山大学 | Image retrieval method based on variable-length depth hash learning |
CN109255364B (en) * | 2018-07-12 | 2021-06-08 | 杭州电子科技大学 | Scene recognition method for generating countermeasure network based on deep convolution |
-
2019
- 2019-06-28 CN CN201910575905.4A patent/CN110296705B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106203533A (en) * | 2016-07-26 | 2016-12-07 | 厦门大学 | The degree of depth based on combined training study face verification method |
CN108986168A (en) * | 2018-06-13 | 2018-12-11 | 深圳市感动智能科技有限公司 | A kind of robot winding detection method and device combining bag of words tree-model based on depth measure study |
CN109101602A (en) * | 2018-08-01 | 2018-12-28 | 腾讯科技(深圳)有限公司 | Image encrypting algorithm training method, image search method, equipment and storage medium |
CN109215117A (en) * | 2018-09-12 | 2019-01-15 | 北京航空航天大学青岛研究院 | Flowers three-dimensional rebuilding method based on ORB and U-net |
CN109341703A (en) * | 2018-09-18 | 2019-02-15 | 北京航空航天大学 | A kind of complete period uses the vision SLAM algorithm of CNNs feature detection |
CN109871803A (en) * | 2019-02-18 | 2019-06-11 | 清华大学 | Robot winding detection method and device |
Non-Patent Citations (2)
Title |
---|
Loop Closure Detection in SLAM by Combining Visual CNN Features and Submaps;Hao Qin,等;《2018 4th International Conference on Control, Automation and Robotics》;20180630;全文 * |
基于学习方法的高精度SLAM算法研究;冯爱迪;《中国优秀硕士学位论文全文数据库信息科技辑》;20190115(第01期);全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN110296705A (en) | 2019-10-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109961034B (en) | Video target detection method based on convolution gating cyclic neural unit | |
CN107274433B (en) | Target tracking method and device based on deep learning and storage medium | |
CN109800689B (en) | Target tracking method based on space-time feature fusion learning | |
CN110296705B (en) | Visual SLAM loop detection method based on distance metric learning | |
CN111222574B (en) | Ship and civil ship target detection and classification method based on multi-model decision-level fusion | |
CN109800692B (en) | Visual SLAM loop detection method based on pre-training convolutional neural network | |
CN107872644A (en) | Video frequency monitoring method and device | |
CN106683118B (en) | Unmanned aerial vehicle target tracking method based on hierarchical model | |
CN111105439B (en) | Synchronous positioning and mapping method using residual attention mechanism network | |
CN109544592B (en) | Moving object detection algorithm for camera movement | |
CN110781790A (en) | Visual SLAM closed loop detection method based on convolutional neural network and VLAD | |
CN110009665A (en) | A kind of target detection tracking method blocked under environment | |
CN110766723B (en) | Unmanned aerial vehicle target tracking method and system based on color histogram similarity | |
CN111104831B (en) | Visual tracking method, device, computer equipment and medium | |
CN114913386A (en) | Training method of multi-target tracking model and multi-target tracking method | |
CN110968711A (en) | Autonomous unmanned system position identification and positioning method based on sequence image characteristics | |
CN111709301A (en) | Method for estimating motion state of curling ball | |
KR20200010971A (en) | Apparatus and method for detecting moving object using optical flow prediction | |
Yin | Object Detection Based on Deep Learning: A Brief Review | |
Mao et al. | Ngel-slam: Neural implicit representation-based global consistent low-latency slam system | |
Li et al. | The integration adjacent frame difference of improved ViBe for foreground object detection | |
Gopal et al. | Tiny object detection: Comparative study using single stage CNN object detectors | |
CN112233141B (en) | Moving target tracking method and system based on unmanned aerial vehicle vision in electric power scene | |
CN110458867B (en) | Target tracking method based on attention circulation network | |
CN113129336A (en) | End-to-end multi-vehicle tracking method, system and computer readable medium |
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 |