CN110335319A - Camera positioning and the map reconstruction method and system of a kind of semantics-driven - Google Patents
Camera positioning and the map reconstruction method and system of a kind of semantics-driven Download PDFInfo
- Publication number
- CN110335319A CN110335319A CN201910557726.8A CN201910557726A CN110335319A CN 110335319 A CN110335319 A CN 110335319A CN 201910557726 A CN201910557726 A CN 201910557726A CN 110335319 A CN110335319 A CN 110335319A
- Authority
- CN
- China
- Prior art keywords
- frame
- matching
- camera
- point
- key frame
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
- G06T17/05—Geographic models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/80—Analysis of captured images to determine intrinsic or extrinsic camera parameters, i.e. camera calibration
-
- 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
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Geometry (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Remote Sensing (AREA)
- Computer Graphics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of positioning of the camera of semantics-driven and map reconstruction method, belong to technical field of computer vision.The invention firstly uses carry out semantic segmentation to current frame image characteristic point;Later according to similarity and semantic classes, matching is carried out to all characteristic points in present frame and key frame using similar matching method and obtains matching pair;Camera posture is initialized by matchings all in present frame and key frame again;Semantic judgement is combined to update Feature Points Matching pair using tripleplane's method later;Posture is recycled to minimize to all Feature Points Matchings to being updated;Finally three-dimensional map is constructed using camera posture;The invention also achieves a kind of positioning of the camera of semantics-driven and map reconstruction system.Technical solution of the present invention has not only done multiple processing in camera positioning stage, and the constraint of some clouds has also been done in phase of regeneration, so that semantic segmentation and the combination that camera positions and reconstructing system is closer, obtain more accurate positioning result and more perfect reconstruction effect.
Description
Technical field
The invention belongs to technical field of computer vision, more particularly, to camera positioning and the ground of a kind of semantics-driven
Figure method for reconstructing.
Background technique
Currently, camera positioning with reconstruction technique with semantic segmentation technology do not combine or combine not closely.
For being not bound with the related algorithm of semantic segmentation, on the one hand, more difficult to deal with various environment such as dynamic
State scene, weak texture scene.On the other hand, these algorithms reconstruct the cartographic model come and are often made of either a cloud
Terrestrial reference composition, it is all based on the map of geological information, so the high-level understanding of any pair of ambient enviroment cannot be provided.
For combining the related algorithm of semantic segmentation, class label typically is sticked to identification object, has done one
The optimization that a little removal dynamic objects influence, but not adequately using semantic segmentation as a result, in turn that semantic segmentation is close
Involvement positioning and map reconstruction technological system in.
Summary of the invention
Aiming at the above defects or improvement requirements of the prior art, the present invention provides a kind of camera of semantics-driven positioning with
Map reconstruction method, its object is to be positioned with the matching of the characteristic point in reconstruction process, again using semantic information optimization camera
The optimization of projection error, the constraint of reconstruction point cloud, the detection of winding, the precision for thus allowing camera to position is higher, rebuilds comprising height
Level understand and it is more complete.
To achieve the above object, described the present invention provides a kind of positioning of the camera of semantics-driven and map reconstruction method
Method the following steps are included:
(1) extract the characteristic point of current frame image, and using the full convolutional neural networks put up to current frame image into
Row semantic segmentation, each characteristic point obtain corresponding semantic classes;
(2) according to similarity and semantic classes, features all in present frame and key frame are clicked through using similar matching method
Row matching, obtains Feature Points Matching pair;
The similar matching method specifically includes following sub-step:
(21) the same category object in present frame and key frame is obtained according to the semantic classes of characteristic point;
(22) the point cloud principal direction for calculating each object in same category object in present frame and key frame, if in present frame
The difference of the point cloud principal direction of some object is less than given threshold in some object and key frame, then described two objects are object
Matching pair;
(23) similarity mode is carried out to the characteristic point of two object regions of the object matches centering, obtained final
Feature Points Matching pair;
(3) camera posture is initialized by all Feature Points Matchings in present frame and key frame;
(4) the corresponding three-dimensional point of matching characteristic point d is obtained in present frame using camera Attitude Calculation, utilizes camera internal reference
Three-dimensional point is projected on present frame, judges subpoint whether where characteristic point d in object area, if not existing, in key
The new match point of characteristic point d is found using similar matching method in the non-matching characteristic point of frame, constitutes new matching pair;
(5) it utilizes step (4) to all Feature Points Matchings to being updated, minimizes following formula and update camera posture:
Wherein, exp (ξ∧) indicate camera posture representation of Lie algebra form;N indicates Feature Points Matching to quantity;uiIt indicates
Ith feature point is matched to image coordinate in the current frame;siIndicate i-th of scale factor;piIndicate i-th matching to
Image coordinate in key frame;
(6) three-dimensional map is constructed using new camera posture, and object is obtained according to the semantic classes of object in three-dimensional map
The resemblance of body will not meet the three-dimensional point deletion of object resemblance in object.
Further, the method also includes following steps:
(7) present frame is further judged using the semantic classes of object, point cloud quantity and point cloud principal direction in present frame
With the presence or absence of winding, and if it exists, then eliminate accumulated error using closed-loop optimization;
(8) optimize global crucial figure using the method that non-linear least square figure optimizes, finally carry out global optimization.
Further, the step (23) specifically includes:
If the set of characteristic points of two object regions of object matches centering are respectively as follows:
(231) a characteristic point ai is chosen from set A, successively calculate in characteristic point ai and set B all characteristic points it
Between similarity;If the similarity in set B between a characteristic point bj and characteristic point ai is maximum, and is greater than the similarity of setting
Threshold value, then bj and ai is characterized a matching pair,
(232) another characteristic point is chosen from set A, is repeated step (231), is found out all spies of set A until all
Levy the matching pair of point.
Further, the step (3) specifically:
(31) essential matrix E is calculated using 8 methods;
(32) by SVD (Singular Value Decomposition, singular value decomposition), essential matrix is decomposed,
Obtain four kinds of possible solutions, i.e. camera posture;
(33) according to every kind of possible camera posture and Feature Points Matching pair, three-dimensional point cloud is calculated, if the position of point cloud
It sets and meets camera imaging model, then corresponding camera posture is initialization camera posture.
Further, judge that present frame specifically includes following sub-step with the presence or absence of winding in the step (7):
(41) candidate winding frame is detected by bag of words (Bag of word, BOW);
(42) the candidate winding frame that will test out carries out the comparison of semantic classes with present frame again, finds out with identical number
The candidate winding frame of amount and identical semantic classes;
(43) the reconstruction point cloud quantity for comparing above-mentioned candidate winding frame again, saves the candidate that similarity is greater than given threshold
Winding frame;
(44) finally by the principal direction for comparing the point cloud that present frame is reconstructed with each candidate winding frame participation, retain
Greater than the candidate winding frame of similarity threshold, as winding;
(45) accumulated error is eliminated using closed-loop optimization.
Further, accumulated error is eliminated using closed-loop optimization in the step (45) specifically include following sub-step:
(451) by calculating the matching pair between current key frame and winding key frame, the transformation between two frames is solved;
(452) Closed-cycle correction is carried out if Feature Points Matching is to correction threshold is met, and calculates each using propagation algorithm
The correct posture of key frame.
Further, the step (8) specifically:
(81) using the posture of each key frame and point cloud as vertex;
(82) establish the binding side between vertex, binding side be relative motion estimation between two pose nodes, point cloud with
The semantic constraint between mappings constraint, point cloud between camera;
(83) vertex is solved full as optimized variable, Bian Zuowei bound term using Gauss-Newton method (GAUSS-NEWTON)
The best vertex of the above-mentioned constraint of foot, that is, camera posture and point cloud position after finding out optimization.
Further, the selection standard of the key frame are as follows: if first meet condition creates key frame:
Nth frame after last round of map reconstruction is set to new key frame;
It has passed through N frame after the insertion of a upper key frame and be set to new key frame;
The Feature Points Matching that present frame traces into is less than number and refers to key frame Feature Points Matching logarithm purpose percent
90, then settled previous frame is new key frame.
It is another aspect of this invention to provide that the present invention provides a kind of positioning of the camera of semantics-driven and map reconstruction system
System, the system comprises following parts:
First module, for extracting the characteristic point of current frame image, and using the full convolutional neural networks put up to working as
Prior image frame carries out semantic segmentation, and each characteristic point obtains corresponding semantic classes;
Second module is used for according to similarity and semantic classes, using similar matching method to institute in present frame and key frame
There is characteristic point to be matched, obtains Feature Points Matching pair;
Include similar matching unit in second module, similar matching unit specifically includes following part:
First subelement, for obtaining the same category object in present frame and key frame according to the semantic classes of characteristic point
Body;
Second subelement, for calculating the point Yun Zhufang of each object in same category object in present frame and key frame
To described if the difference of the point cloud principal direction of some object is less than given threshold in some object and key frame in present frame
Two objects are object matches pair;
Third subelement carries out similarity for the characteristic point to two object regions of the object matches centering
Match, obtains final Feature Points Matching pair;
Third module, for being initialized by all Feature Points Matchings in present frame and key frame to camera posture;
4th module, for obtaining in present frame the corresponding three-dimensional point of matching characteristic point d, benefit using camera Attitude Calculation
Three-dimensional point is projected on present frame with camera internal reference, judges subpoint whether where characteristic point d in object area, if not
The new match point of characteristic point d is then being found using similar matching method in the non-matching characteristic point of key frame, is constituting new matching
It is right;
5th module, for using the 4th module to all Feature Points Matchings to being updated, later minimize following formula more
New camera posture:
Wherein, exp (ξ∧) indicate camera posture representation of Lie algebra form;N indicates Feature Points Matching to quantity;uiIt indicates
Ith feature point is matched to image coordinate in the current frame;siIndicate i-th of scale factor;piIndicate i-th matching to
Image coordinate in key frame;
6th module, for constructing three-dimensional map using new camera posture, and according to the semanteme of object in three-dimensional map
Classification obtains the resemblance of object, and the three-dimensional point deletion of object resemblance will not be met in object.
Further, the system also includes:
7th module, for further using the semantic classes of object, point cloud quantity and point cloud principal direction in present frame
Judge present frame with the presence or absence of winding, and if it exists, then to eliminate accumulated error using closed-loop optimization;
8th module, the global crucial figure of method optimization for being optimized using non-linear least square figure, is finally carried out complete
Office's optimization.
In general, through the invention it is contemplated above technical scheme is compared with the prior art, have below beneficial to effect
Fruit:
(1) the method for the present invention uses the similar matching process based on semantic segmentation and carries out Feature Points Matching, this method benefit
With the semantic label information of each frame image, the direction of object after semantic classification and classification is matched as constraint, is increased
The constraint condition in matching process is added, so that the range of Feature Points Matching is reduced, to save Feature Points Matching
Time, and reduce the Feature Points Matching pair of many mistakes, improve matched precision, the estimation for camera posture provides
Good calculating environment;
(2) the method for the present invention uses semantic-based re-projection optimization method, and this method combines the language of each frame image
Adopted information increases the constraint condition of re-projection point, the re-projection point of a part of mistake has been filtered, so that re-projection optimization efficiency
Be improved, and the re-projection point due to eliminating a part of mistake so that the precision of camera pose refinement obtained into
The raising of one step, so that camera tracking is more accurate, it is not easy to drift about since error is excessive;
(3) method that the method for the present invention uses the optimization of semantic-based figure, the object that this method utilizes semantic segmentation to go out
Geological information so that not only according to the mapping between the transformation and point cloud and camera posture between camera posture come the phase that optimizes
Machine posture and point cloud, also pass through the position between geometrical constraint obligatory point Yun Yudian cloud, also affect the excellent of camera posture indirectly
Change adjustment, to obtain more accurate camera posture and point cloud
(4) the method for the present invention uses semantic-based winding detection method, and this method is by each frame image, semantic label
Categorical measure as bound term, further judge after finding a series of candidate winding frames by BOW, so that looking into
The winding frame and present frame looked for are increasingly similar, and the accuracy of winding is higher, so that winding optimization elimination error is more accurate.
Detailed description of the invention
Fig. 1 is the general flow chart of the method for the present invention;
Fig. 2 is to decompose essential matrix using SVD in the method for the present invention, obtains four kinds of possible camera posture schematic diagrames;
Fig. 3 is three-dimensional point perspective view in the method for the present invention;
Fig. 4 is the schematic diagram of global optimization in the method for the present invention.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that described herein, specific examples are only used to explain the present invention, not
For limiting the present invention.As long as in addition, technical characteristic involved in the various embodiments of the present invention described below that
Not constituting conflict between this can be combined with each other.
As shown in Figure 1, the method for the present invention the following steps are included:
(1) extract the characteristic point of current frame image, and using the full convolutional neural networks put up to current frame image into
Row semantic segmentation, each characteristic point obtain corresponding semantic classes;
(2) according to similarity and semantic classes, features all in present frame and key frame are clicked through using similar matching method
Row matching, obtains Feature Points Matching pair;
The similar matching method specifically includes following sub-step:
(21) the same category object in present frame and key frame is obtained according to the semantic classes of characteristic point;
(22) the point cloud principal direction for calculating each object in same category object in present frame and key frame, if in present frame
The difference of the point cloud principal direction of some object is less than given threshold in some object and key frame, then described two objects are object
Matching pair;
(23) similarity mode is carried out to the characteristic point of two object regions of the object matches centering, obtained final
Feature Points Matching pair;
(3) camera posture is initialized by all Feature Points Matchings in present frame and key frame;
(4) the corresponding three-dimensional point of matching characteristic point d is obtained in present frame using camera Attitude Calculation, utilizes camera internal reference
Three-dimensional point is projected on present frame, judges subpoint whether where characteristic point d in object area, if not existing, in key
The new match point of characteristic point d is found using similar matching method in the non-matching characteristic point of frame, constitutes new matching pair;
(5) it uses step (4) to all Feature Points Matchings to being updated, minimizes following formula and update camera posture:
Wherein, exp (ξ∧) indicate camera posture representation of Lie algebra form;N indicates Feature Points Matching to quantity;uiIt indicates
Ith feature point is matched to image coordinate in the current frame;siIndicate i-th of scale factor;piIndicate i-th matching to
Image coordinate in key frame;
(6) three-dimensional map is constructed using new camera posture, and object is obtained according to the semantic classes of object in three-dimensional map
The resemblance of body will not meet the three-dimensional point deletion of object resemblance in object;
(7) present frame is further judged using the semantic classes of object, point cloud quantity and point cloud principal direction in present frame
With the presence or absence of winding, and if it exists, then eliminate accumulated error using closed-loop optimization;
(8) optimize global crucial figure using the method that non-linear least square figure optimizes, finally carry out Global B A optimization.
The method of the present invention is introduced now in conjunction with one embodiment of the present of invention:
1. semantic segmentation: extracting the characteristic point of current frame image, and using the full convolutional neural networks put up to current
Frame image carries out semantic segmentation, and each characteristic point obtains corresponding semantic classes.
2. tracking: by finding corresponding relationship as much as possible between present frame and local map, Lai Youhua present frame
Pose estimation.Specifically:
A.ORB feature extraction and semantic segmentation: setting present frame for input frame, extracts ORB characteristic point and corresponding
Present frame is put into segmentation network by ORB Feature Descriptor, waits until its prediction result.
B. estimate camera motion: similar by the similarity of ORB Feature Descriptor and semantic classes information, use first
The characteristic point of matching method matches present frame and previous frame, specific practice are as follows:
(b1) position of the same category object in present frame and previous frame is obtained according to the semantic classes of characteristic point;
(b2) between the characteristic point in same category of object space, calculate two-by-two between the sub- similarity of description, every acquisition
One group of highest similarity, then as final Feature Points Matching to preserving;
Then camera posture is predicted using motor pattern.Motion model assumes camera uniform motion, passes through previous frame camera
Pose, two interframe Feature Points Matching relationship and speed estimate the pose of present frame, if Feature Points Matching is low to number
When threshold value, it is changed to Key Mode.It attempts to carry out Feature Points Matching with nearest key frame, matching process is same as above, if working as
A number of pairs for previous frame and arest neighbors key frame is still below threshold value, then matches present frame with global all key frames, seek
The highest key frame of a number of pairs is looked for, solves camera pose using PnP algorithm, specific practice is as follows:
(bb1) essential matrix E is calculated using 8 methods;
(bb2) it is decomposed by SVD, essential matrix is decomposed, four kinds of possible solutions (spin matrix, translation matrix) are obtained,
That is posture;
(bb3) according to every kind of possible posture and Feature Points Matching pair, three-dimensional point cloud is calculated, passes through judgement point cloud
Position determines which solution chosen, that is, calculates the posture of camera, as shown in Figure 2.
Then the re-projection based on semantic segmentation is carried out using pose of the matched characteristic point to previous frame to optimize, worked as
The posture of previous frame, specific practice are as follows:
If project to the characteristic point in image fallen on it is not identical as the classification of characteristic point matching in original image
Place, then it is assumed that this is unqualified to characteristic point re-projection, weed out this to matching pair, be not involved in the optimization of objective function.Such as
Shown in figure, the corresponding two-dimensional image point of P spatial point is p1, in characteristic matching stage, by previous frame characteristic point p1 and present frame
Characteristic point p2 matching together, it is believed that P should project to the position of p2, however because the mistake of camera Attitude estimation
Difference, drop point have fallen on p ' usually not in the position of p2.If the semantic label of p ' pixel and p1 be not identical, then it is assumed that p1
It is no longer participate in the estimation of camera motion, as shown in Figure 3 to weed out this to match point with p2 matching error.
To institute's re-projection point with a grain of salt, subpoint is calculated at a distance from characteristic point matching in same image, most
Smallization all distances update camera posture:
Wherein, exp (ξ∧) indicate camera posture representation of Lie algebra form;N indicates Feature Points Matching to quantity;uiIt indicates
I-th pair is matched to image coordinate in the current frame;siIndicate i-th of scale factor;piI-th pair matching pair is indicated, in present frame
Image coordinate in tracking frame;
C. track local map: find the key frame for having common three-dimensional space point in local map with present frame and and this
The adjacent key frame of a little key frames.The three-dimensional point for projecting to three-dimensional space corresponding in above-mentioned key frame is projected in present frame,
Be updated and match with the characteristic point in present frame, finally with it is all matching to again optimize camera pose, optimal way with it is upper
One step is identical.
D. key frame judges: if first meet condition creates key frame: apart from upper one 15 frames of global reorientation,
It has passed through 15 frames after the insertion of a upper key frame, the characteristic point that present frame traces into is less than number and matches with reference to key frame
Logarithm purpose 90 percent.(being the key frame for having most common observation three-dimensional point with present frame with reference to key frame) is if be unsatisfactory for
Condition then carries out boundling adjustment, optimizes the posture of a upper key frame.
3. semantic label merges: after creation key frame, visual range degree will be total in local map with current key frame and be higher than centainly
The key frame of threshold value is used to update the corresponding semantic label probability of each pixel in current key frame.Visual range degree is by two interframe altogether
Matching observed to number and jointly three-dimensional space same point number determine.
4. figure is built in part: after semantic label fusion, current key frame being inserted into local map, and filter out redundancy
Three-dimensional space point and key frame finally carry out local boundling adjustment.
A. it is inserted into key frame: pose figure is added using the posture of key frame as node, addition has altogether with current key frame
With the optimization side of the key frame of observation three-dimensional space point.
B. local boundling adjustment: by current key frame, adjacent key frame, the key frame for having common observation three-dimensional point and
Corresponding three-dimensional space point, which is put into pose figure, to be optimized.Each key frame is checked, if there is 90 percent feature
Point is all exceeded three other key frames and observes, just weeds out this key frame.
5. winding detects: detecting if the number of key frames in map is less than 10 without winding, if more than 10, exist first
The key frame for having a common BoW word with current key frame is found in map, then counts shared most with the BoW of current key frame
More word numbers are used as threshold value for 80 the percent of this number, the key frame that word number is greater than threshold value are found, as candidate key
Frame.A series of candidate winding frames that will test out carry out the comparison of semantic classes with present frame again, find out with identical quantity,
The candidate winding frame of identical semantic classes compares the reconstruction point cloud quantity of above-mentioned candidate winding frame again, saves similarity and be greater than
The candidate winding frame of certain threshold value, finally by the master for comparing the point cloud that present frame is reconstructed with each candidate winding frame participation
Direction remains larger than the candidate winding frame of certain similarity, as winding, rear by calculating current key frame and winding key frame
Between matching pair, solve two frames between transformation Closed-cycle correction is carried out if Feature Points Matching is to number is met, use propagation
Algorithm calculates the correct transformed value of each key frame.
6. finally carrying out figure optimization and global optimization.
(1) using the posture of each key frame and point cloud as vertex;
(2) establish the binding side between vertex, binding side be relative motion between two pose nodes estimate (being denoted as T),
The semantic constraint (as shown in the figure) between mappings constraint (being denoted as M), point cloud between point cloud and camera;
(3) vertex solves the best top for meeting above-mentioned constraint using L-M method as optimized variable, Bian Zuowei bound term
Point, that is, camera posture and point cloud position after finding out optimization, as shown in Figure 4.
The camera positioning and map reconstruction system of a kind of semantics-driven, the system are first further illustrated in conjunction with specific embodiments
System includes following part:
First module, for extracting the characteristic point of current frame image, and using the full convolutional neural networks put up to working as
Prior image frame carries out semantic segmentation, and each characteristic point obtains corresponding semantic classes;
Second module is used for according to similarity and semantic classes, using similar matching method to institute in present frame and key frame
There is characteristic point to be matched, obtains Feature Points Matching pair;
Include similar matching unit in second module, similar matching unit specifically includes following part:
First subelement, for obtaining the same category object in present frame and key frame according to the semantic classes of characteristic point
Body;
Second subelement, for calculating the point Yun Zhufang of each object in same category object in present frame and key frame
To described if the difference of the point cloud principal direction of some object is less than given threshold in some object and key frame in present frame
Two objects are object matches pair;
Third subelement carries out similarity for the characteristic point to two object regions of the object matches centering
Match, obtains final Feature Points Matching pair;
Third module, for being initialized by all Feature Points Matchings in present frame and key frame to camera posture;
4th module, for obtaining in present frame the corresponding three-dimensional point of matching characteristic point d, benefit using camera Attitude Calculation
Three-dimensional point is projected on present frame with camera internal reference, judges subpoint whether where characteristic point d in object area, if not
The new match point of characteristic point d is then being found using similar matching method in the non-matching characteristic point of key frame, is constituting new matching
It is right;
5th module, for using the 4th module to all Feature Points Matchings to being updated, later minimize following formula more
New camera posture:
Wherein, exp (ξ∧) indicate camera posture representation of Lie algebra form;N indicates Feature Points Matching to quantity;uiIt indicates
Ith feature point is matched to image coordinate in the current frame;siIndicate i-th of scale factor;piIndicate i-th matching to
Image coordinate in key frame;
6th module, for constructing three-dimensional map using new camera posture, and according to the semanteme of object in three-dimensional map
Classification obtains the resemblance of object, and the three-dimensional point deletion of object resemblance will not be met in object.
7th module, for further using the semantic classes of object, point cloud quantity and point cloud principal direction in present frame
Judge present frame with the presence or absence of winding, and if it exists, then to eliminate accumulated error using closed-loop optimization;
8th module, the global crucial figure of method optimization for being optimized using non-linear least square figure, is finally carried out complete
Office's optimization.
The above content as it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention,
It is not intended to limit the invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention,
It should all be included in the protection scope of the present invention.
Claims (10)
1. camera positioning and the map reconstruction method of a kind of semantics-driven, which is characterized in that the method specifically includes following step
It is rapid:
(1) characteristic point of current frame image is extracted, and language is carried out to current frame image using the full convolutional neural networks put up
Justice segmentation, each characteristic point obtain corresponding semantic classes;
(2) according to similarity and semantic classes, using similar matching method to characteristic points all in present frame and key frame progress
Match, obtains Feature Points Matching pair;
The similar matching method specifically includes following sub-step:
(21) the same category object in present frame and key frame is obtained according to the semantic classes of characteristic point;
(22) the point cloud principal direction for calculating each object in same category object in present frame and key frame, if some in present frame
The difference of the point cloud principal direction of some object is less than given threshold in object and key frame, then described two objects are object matches
It is right;
(23) similarity mode is carried out to the characteristic point of two object regions of the object matches centering, obtains final feature
Point matching pair;
(3) camera posture is initialized by all Feature Points Matchings in present frame and key frame;
(4) the corresponding three-dimensional point of matching characteristic point d is obtained in present frame using camera Attitude Calculation, using camera internal reference by three
Dimension point projects on present frame, whether judges subpoint where characteristic point d in object area, if not existing, key frame not
The new match point of characteristic point d is found using similar matching method in matching characteristic point, constitutes new matching pair;
(5) it utilizes step (4) to all Feature Points Matchings to being updated, minimizes following formula and update camera posture:
Wherein, exp (ξ ^) indicates the representation of Lie algebra form of camera posture;N indicates Feature Points Matching to quantity;uiIt indicates i-th
Feature Points Matching is to image coordinate in the current frame;siIndicate i-th of scale factor;piIndicate i-th of matching in key frame
In image coordinate;
(6) three-dimensional map is constructed using new camera posture, and object is obtained according to the semantic classes of object in three-dimensional map
Resemblance will not meet the three-dimensional point deletion of object resemblance in object.
2. camera positioning and the map reconstruction method of a kind of semantics-driven according to claim 1, which is characterized in that described
Method is further comprising the steps of:
(7) further whether present frame is judged using the semantic classes of object, point cloud quantity and point cloud principal direction in present frame
There are winding, and if it exists, then eliminates accumulated error using closed-loop optimization;
(8) optimize global crucial figure using the method that non-linear least square figure optimizes, finally carry out global optimization.
3. camera positioning and the map reconstruction method of a kind of semantics-driven according to claim 1, which is characterized in that described
Step (23) specifically includes:
If the set of characteristic points of two object regions of object matches centering are respectively as follows:
(231) a characteristic point ai is chosen from set A, is successively calculated in characteristic point ai and set B between all characteristic points
Similarity;If the similarity in set B between a characteristic point bj and characteristic point ai is maximum, and is greater than the similarity threshold of setting
Value, then bj and ai is characterized a matching pair,
(232) another characteristic point is chosen from set A, is repeated step (231), is found out all characteristic points of set A until all
Matching pair.
4. camera positioning and the map reconstruction method of a kind of semantics-driven according to claim 1, which is characterized in that described
Step (3) specifically:
(31) essential matrix E is calculated using 8 methods;
(32) by SVD (Singular Value Decomposition, singular value decomposition), essential matrix is decomposed, is obtained
Four kinds of possible solutions, i.e. camera posture;
(33) according to every kind of possible camera posture and Feature Points Matching pair, three-dimensional point cloud is calculated, if the position symbol of point cloud
Camera imaging model is closed, then corresponding camera posture is initialization camera posture.
5. camera positioning and the map reconstruction method of a kind of semantics-driven according to claim 2, which is characterized in that described
Judge that present frame specifically includes following sub-step with the presence or absence of winding in step (7):
(41) candidate winding frame is detected by bag of words (Bag of word, BOW);
(42) the candidate winding frame that will test out carries out the comparison of semantic classes with present frame again, find out with identical quantity and
The candidate winding frame of identical semantic classes;
(43) the reconstruction point cloud quantity for comparing above-mentioned candidate winding frame again, saves the candidate winding that similarity is greater than given threshold
Frame;
(44) it finally by the principal direction for comparing the point cloud that present frame is reconstructed with each candidate winding frame participation, remains larger than
The candidate winding frame of similarity threshold, as winding;
(45) accumulated error is eliminated using closed-loop optimization.
6. camera positioning and the map reconstruction method of a kind of semantics-driven according to claim 5, which is characterized in that described
Accumulated error, which is eliminated, using closed-loop optimization in step (45) specifically includes following sub-step:
(451) by calculating the matching pair between current key frame and winding key frame, the transformation between two frames is solved;
(452) Closed-cycle correction is carried out if Feature Points Matching is to correction threshold is met, and calculates each key using propagation algorithm
The correct posture of frame.
7. camera positioning and the map reconstruction method of a kind of semantics-driven according to claim 1, which is characterized in that described
Step (8) specifically:
(81) using the posture of each key frame and point cloud as vertex;
(82) binding side between vertex is established, binding side is relative motion estimation, point cloud and the camera between two pose nodes
Between mappings constraint, point cloud between semantic constraint;
(83) vertex is solved using Gauss-Newton method (GAUSS-NEWTON) and is met as optimized variable, Bian Zuowei bound term
The best vertex for stating constraint, that is, camera posture and point cloud position after finding out optimization.
8. the camera positioning and map reconstruction method, feature of a kind of any semantics-driven according to claim 1~7
It is, the selection standard of the key frame are as follows: if first meet condition creates key frame:
Nth frame after last round of map reconstruction is set to new key frame;
It has passed through N frame after the insertion of a upper key frame and be set to new key frame;
The Feature Points Matching that present frame traces into is less than number and refers to key frame Feature Points Matching logarithm purpose 90 percent,
Then settled previous frame is new key frame.
9. camera positioning and the map reconstruction system of a kind of semantics-driven, which is characterized in that the system comprises following parts:
First module, for extracting the characteristic point of current frame image, and using the full convolutional neural networks put up to present frame
Image carries out semantic segmentation, and each characteristic point obtains corresponding semantic classes;
Second module is used for according to similarity and semantic classes, using similar matching method to all spies in present frame and key frame
Sign point is matched, and Feature Points Matching pair is obtained;
Include similar matching unit in second module, similar matching unit specifically includes following part:
First subelement, for obtaining the same category object in present frame and key frame according to the semantic classes of characteristic point;
Second subelement, for calculating the point cloud principal direction of each object in same category object in present frame and key frame, if
The difference of the point cloud principal direction of some object is less than given threshold in some object and key frame in present frame, then described two objects
Body is object matches pair;
Third subelement carries out similarity mode for the characteristic point to two object regions of the object matches centering,
Obtain final Feature Points Matching pair;
Third module, for being initialized by all Feature Points Matchings in present frame and key frame to camera posture;
4th module utilizes phase for obtaining in present frame the corresponding three-dimensional point of matching characteristic point d using camera Attitude Calculation
Machine internal reference projects to three-dimensional point on present frame, whether judges subpoint where characteristic point d in object area, if not existing,
The new match point of characteristic point d is found using similar matching method in the non-matching characteristic point of key frame, constitutes new matching pair;
5th module, for using the 4th module to all Feature Points Matchings to being updated, later minimize following formula more cenotype
Machine posture:
Wherein, exp (ξ ^) indicates the representation of Lie algebra form of camera posture;N indicates Feature Points Matching to quantity;uiIt indicates i-th
Feature Points Matching is to image coordinate in the current frame;siIndicate i-th of scale factor;piIndicate i-th of matching in key frame
In image coordinate;
6th module, for constructing three-dimensional map using new camera posture, and according to the semantic classes of object in three-dimensional map
The resemblance for obtaining object, will not meet the three-dimensional point deletion of object resemblance in object.
10. being existed according to a kind of camera positioning of semantics-driven any in claim 9 and map reconstruction method, feature
In, the system also includes:
7th module, for further being judged using the semantic classes of object, point cloud quantity and point cloud principal direction in present frame
Present frame whether there is winding, and if it exists, then eliminate accumulated error using closed-loop optimization;
8th module, the global crucial figure of method optimization for being optimized using non-linear least square figure, is finally carried out global excellent
Change.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910557726.8A CN110335319B (en) | 2019-06-26 | 2019-06-26 | Semantic-driven camera positioning and map reconstruction method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910557726.8A CN110335319B (en) | 2019-06-26 | 2019-06-26 | Semantic-driven camera positioning and map reconstruction method and system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110335319A true CN110335319A (en) | 2019-10-15 |
CN110335319B CN110335319B (en) | 2022-03-18 |
Family
ID=68142729
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910557726.8A Active CN110335319B (en) | 2019-06-26 | 2019-06-26 | Semantic-driven camera positioning and map reconstruction method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110335319B (en) |
Cited By (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110910389A (en) * | 2019-10-30 | 2020-03-24 | 中山大学 | Laser SLAM loop detection system and method based on graph descriptor |
CN111046125A (en) * | 2019-12-16 | 2020-04-21 | 视辰信息科技(上海)有限公司 | Visual positioning method, system and computer readable storage medium |
CN111310654A (en) * | 2020-02-13 | 2020-06-19 | 北京百度网讯科技有限公司 | Map element positioning method and device, electronic equipment and storage medium |
CN111311708A (en) * | 2020-01-20 | 2020-06-19 | 北京航空航天大学 | Visual SLAM method based on semantic optical flow and inverse depth filtering |
CN111311742A (en) * | 2020-03-27 | 2020-06-19 | 北京百度网讯科技有限公司 | Three-dimensional reconstruction method, three-dimensional reconstruction device and electronic equipment |
CN111325842A (en) * | 2020-03-04 | 2020-06-23 | Oppo广东移动通信有限公司 | Map construction method, repositioning method and device, storage medium and electronic equipment |
CN111368759A (en) * | 2020-03-09 | 2020-07-03 | 河海大学常州校区 | Monocular vision-based semantic map construction system for mobile robot |
CN111427373A (en) * | 2020-03-24 | 2020-07-17 | 上海商汤临港智能科技有限公司 | Pose determination method, device, medium and equipment |
CN111429517A (en) * | 2020-03-23 | 2020-07-17 | Oppo广东移动通信有限公司 | Relocation method, relocation device, storage medium and electronic device |
CN111815687A (en) * | 2020-06-19 | 2020-10-23 | 浙江大华技术股份有限公司 | Point cloud matching method, positioning method, device and storage medium |
CN112085026A (en) * | 2020-08-26 | 2020-12-15 | 的卢技术有限公司 | Closed loop detection method based on deep neural network semantic segmentation |
CN112288817A (en) * | 2020-11-18 | 2021-01-29 | Oppo广东移动通信有限公司 | Three-dimensional reconstruction processing method and device based on image |
CN112419512A (en) * | 2020-10-13 | 2021-02-26 | 南昌大学 | Air three-dimensional model repairing system and method based on semantic information |
CN112507056A (en) * | 2020-12-21 | 2021-03-16 | 华南理工大学 | Map construction method based on visual semantic information |
CN112585946A (en) * | 2020-03-27 | 2021-03-30 | 深圳市大疆创新科技有限公司 | Image shooting method, image shooting device, movable platform and storage medium |
CN112927269A (en) * | 2021-03-26 | 2021-06-08 | 深圳市无限动力发展有限公司 | Map construction method and device based on environment semantics and computer equipment |
CN113591865A (en) * | 2021-07-28 | 2021-11-02 | 深圳甲壳虫智能有限公司 | Loop detection method and device and electronic equipment |
CN114639006A (en) * | 2022-03-15 | 2022-06-17 | 北京理工大学 | Loop detection method and device and electronic equipment |
Citations (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101645170A (en) * | 2009-09-03 | 2010-02-10 | 北京信息科技大学 | Precise registration method of multilook point cloud |
CN102308320A (en) * | 2009-02-06 | 2012-01-04 | 香港科技大学 | Generating three-dimensional models from images |
CN104361627A (en) * | 2014-11-07 | 2015-02-18 | 武汉科技大学 | SIFT-based (scale-invariant feature transform) binocular vision three-dimensional image reconstruction method of asphalt pavement micro-texture |
CN107392964A (en) * | 2017-07-07 | 2017-11-24 | 武汉大学 | The indoor SLAM methods combined based on indoor characteristic point and structure lines |
CN107610175A (en) * | 2017-08-04 | 2018-01-19 | 华南理工大学 | The monocular vision SLAM algorithms optimized based on semi-direct method and sliding window |
CN107833236A (en) * | 2017-10-31 | 2018-03-23 | 中国科学院电子学研究所 | Semantic vision positioning system and method are combined under a kind of dynamic environment |
CN108230337A (en) * | 2017-12-31 | 2018-06-29 | 厦门大学 | A kind of method that semantic SLAM systems based on mobile terminal are realized |
CN108596053A (en) * | 2018-04-09 | 2018-09-28 | 华中科技大学 | A kind of vehicle checking method and system based on SSD and vehicle attitude classification |
CN109272577A (en) * | 2018-08-30 | 2019-01-25 | 北京计算机技术及应用研究所 | A kind of vision SLAM method based on Kinect |
CN109544629A (en) * | 2018-11-29 | 2019-03-29 | 南京人工智能高等研究院有限公司 | Camera pose determines method and apparatus and electronic equipment |
CN109658449A (en) * | 2018-12-03 | 2019-04-19 | 华中科技大学 | A kind of indoor scene three-dimensional rebuilding method based on RGB-D image |
CN109815847A (en) * | 2018-12-30 | 2019-05-28 | 中国电子科技集团公司信息科学研究院 | A kind of vision SLAM method based on semantic constraint |
CN109816686A (en) * | 2019-01-15 | 2019-05-28 | 山东大学 | Robot semanteme SLAM method, processor and robot based on object example match |
-
2019
- 2019-06-26 CN CN201910557726.8A patent/CN110335319B/en active Active
Patent Citations (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102308320A (en) * | 2009-02-06 | 2012-01-04 | 香港科技大学 | Generating three-dimensional models from images |
CN101645170A (en) * | 2009-09-03 | 2010-02-10 | 北京信息科技大学 | Precise registration method of multilook point cloud |
CN104361627A (en) * | 2014-11-07 | 2015-02-18 | 武汉科技大学 | SIFT-based (scale-invariant feature transform) binocular vision three-dimensional image reconstruction method of asphalt pavement micro-texture |
CN107392964A (en) * | 2017-07-07 | 2017-11-24 | 武汉大学 | The indoor SLAM methods combined based on indoor characteristic point and structure lines |
CN107610175A (en) * | 2017-08-04 | 2018-01-19 | 华南理工大学 | The monocular vision SLAM algorithms optimized based on semi-direct method and sliding window |
CN107833236A (en) * | 2017-10-31 | 2018-03-23 | 中国科学院电子学研究所 | Semantic vision positioning system and method are combined under a kind of dynamic environment |
CN108230337A (en) * | 2017-12-31 | 2018-06-29 | 厦门大学 | A kind of method that semantic SLAM systems based on mobile terminal are realized |
CN108596053A (en) * | 2018-04-09 | 2018-09-28 | 华中科技大学 | A kind of vehicle checking method and system based on SSD and vehicle attitude classification |
CN109272577A (en) * | 2018-08-30 | 2019-01-25 | 北京计算机技术及应用研究所 | A kind of vision SLAM method based on Kinect |
CN109544629A (en) * | 2018-11-29 | 2019-03-29 | 南京人工智能高等研究院有限公司 | Camera pose determines method and apparatus and electronic equipment |
CN109658449A (en) * | 2018-12-03 | 2019-04-19 | 华中科技大学 | A kind of indoor scene three-dimensional rebuilding method based on RGB-D image |
CN109815847A (en) * | 2018-12-30 | 2019-05-28 | 中国电子科技集团公司信息科学研究院 | A kind of vision SLAM method based on semantic constraint |
CN109816686A (en) * | 2019-01-15 | 2019-05-28 | 山东大学 | Robot semanteme SLAM method, processor and robot based on object example match |
Non-Patent Citations (2)
Title |
---|
MOHAMMAD NOORUDDIN等: ""Improved 3D Reconstruction for Images having Moving Object using Semantic Image Segmentation and Binary Masking"", 《2018 4TH INTERNATIONAL CONFERENCE ON ELECTRICAL ENGINEERING AND INFORMATION & COMMUNICATION TECHNOLOGY (ICEEICT)》 * |
于金山等: ""基于云的语义库设计及机器人语义地图构建"", 《机器人 ROBOT》 * |
Cited By (28)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110910389A (en) * | 2019-10-30 | 2020-03-24 | 中山大学 | Laser SLAM loop detection system and method based on graph descriptor |
CN110910389B (en) * | 2019-10-30 | 2021-04-09 | 中山大学 | Laser SLAM loop detection system and method based on graph descriptor |
CN111046125A (en) * | 2019-12-16 | 2020-04-21 | 视辰信息科技(上海)有限公司 | Visual positioning method, system and computer readable storage medium |
CN111311708A (en) * | 2020-01-20 | 2020-06-19 | 北京航空航天大学 | Visual SLAM method based on semantic optical flow and inverse depth filtering |
CN111310654A (en) * | 2020-02-13 | 2020-06-19 | 北京百度网讯科技有限公司 | Map element positioning method and device, electronic equipment and storage medium |
CN111310654B (en) * | 2020-02-13 | 2023-09-08 | 北京百度网讯科技有限公司 | Map element positioning method and device, electronic equipment and storage medium |
CN111325842B (en) * | 2020-03-04 | 2023-07-28 | Oppo广东移动通信有限公司 | Map construction method, repositioning method and device, storage medium and electronic equipment |
CN111325842A (en) * | 2020-03-04 | 2020-06-23 | Oppo广东移动通信有限公司 | Map construction method, repositioning method and device, storage medium and electronic equipment |
EP4113451A4 (en) * | 2020-03-04 | 2023-07-19 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Map construction method and apparatus, repositioning method and apparatus, storage medium, and electronic device |
WO2021175022A1 (en) * | 2020-03-04 | 2021-09-10 | Oppo广东移动通信有限公司 | Map construction method and apparatus, repositioning method and apparatus, storage medium, and electronic device |
CN111368759A (en) * | 2020-03-09 | 2020-07-03 | 河海大学常州校区 | Monocular vision-based semantic map construction system for mobile robot |
CN111368759B (en) * | 2020-03-09 | 2022-08-30 | 河海大学常州校区 | Monocular vision-based mobile robot semantic map construction system |
CN111429517A (en) * | 2020-03-23 | 2020-07-17 | Oppo广东移动通信有限公司 | Relocation method, relocation device, storage medium and electronic device |
CN111427373A (en) * | 2020-03-24 | 2020-07-17 | 上海商汤临港智能科技有限公司 | Pose determination method, device, medium and equipment |
CN111427373B (en) * | 2020-03-24 | 2023-11-24 | 上海商汤临港智能科技有限公司 | Pose determining method, pose determining device, medium and pose determining equipment |
CN111311742A (en) * | 2020-03-27 | 2020-06-19 | 北京百度网讯科技有限公司 | Three-dimensional reconstruction method, three-dimensional reconstruction device and electronic equipment |
CN112585946A (en) * | 2020-03-27 | 2021-03-30 | 深圳市大疆创新科技有限公司 | Image shooting method, image shooting device, movable platform and storage medium |
CN111815687A (en) * | 2020-06-19 | 2020-10-23 | 浙江大华技术股份有限公司 | Point cloud matching method, positioning method, device and storage medium |
CN112085026A (en) * | 2020-08-26 | 2020-12-15 | 的卢技术有限公司 | Closed loop detection method based on deep neural network semantic segmentation |
CN112419512A (en) * | 2020-10-13 | 2021-02-26 | 南昌大学 | Air three-dimensional model repairing system and method based on semantic information |
CN112288817A (en) * | 2020-11-18 | 2021-01-29 | Oppo广东移动通信有限公司 | Three-dimensional reconstruction processing method and device based on image |
CN112288817B (en) * | 2020-11-18 | 2024-05-07 | Oppo广东移动通信有限公司 | Three-dimensional reconstruction processing method and device based on image |
CN112507056A (en) * | 2020-12-21 | 2021-03-16 | 华南理工大学 | Map construction method based on visual semantic information |
CN112927269A (en) * | 2021-03-26 | 2021-06-08 | 深圳市无限动力发展有限公司 | Map construction method and device based on environment semantics and computer equipment |
CN113591865A (en) * | 2021-07-28 | 2021-11-02 | 深圳甲壳虫智能有限公司 | Loop detection method and device and electronic equipment |
CN113591865B (en) * | 2021-07-28 | 2024-03-26 | 深圳甲壳虫智能有限公司 | Loop detection method and device and electronic equipment |
CN114639006A (en) * | 2022-03-15 | 2022-06-17 | 北京理工大学 | Loop detection method and device and electronic equipment |
CN114639006B (en) * | 2022-03-15 | 2023-09-26 | 北京理工大学 | Loop detection method and device and electronic equipment |
Also Published As
Publication number | Publication date |
---|---|
CN110335319B (en) | 2022-03-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110335319A (en) | Camera positioning and the map reconstruction method and system of a kind of semantics-driven | |
CN110533722B (en) | Robot rapid repositioning method and system based on visual dictionary | |
CN109631855B (en) | ORB-SLAM-based high-precision vehicle positioning method | |
Eade et al. | Monocular graph SLAM with complexity reduction | |
CN106909877B (en) | Visual simultaneous mapping and positioning method based on dotted line comprehensive characteristics | |
CN113012212B (en) | Depth information fusion-based indoor scene three-dimensional point cloud reconstruction method and system | |
CN107392964B (en) | The indoor SLAM method combined based on indoor characteristic point and structure lines | |
CN110889243B (en) | Aircraft fuel tank three-dimensional reconstruction method and detection method based on depth camera | |
CN110223348A (en) | Robot scene adaptive bit orientation estimation method based on RGB-D camera | |
CN110378997B (en) | ORB-SLAM 2-based dynamic scene mapping and positioning method | |
Tang et al. | ESTHER: Joint camera self-calibration and automatic radial distortion correction from tracking of walking humans | |
CN113205595B (en) | Construction method and application of 3D human body posture estimation model | |
CN112562081B (en) | Visual map construction method for visual layered positioning | |
CN111311708B (en) | Visual SLAM method based on semantic optical flow and inverse depth filtering | |
CN107832786B (en) | A kind of recognition of face classification method dictionary-based learning | |
CN112509044A (en) | Binocular vision SLAM method based on dotted line feature fusion | |
CN113532420B (en) | Visual inertial odometer method integrating dotted line characteristics | |
CN110807828A (en) | Oblique photography three-dimensional reconstruction matching method | |
CN111709331A (en) | Pedestrian re-identification method based on multi-granularity information interaction model | |
Moghaddam et al. | A statistical variable selection solution for RFM ill-posedness and overparameterization problems | |
CN112767546A (en) | Binocular image-based visual map generation method for mobile robot | |
CN114067128A (en) | SLAM loop detection method based on semantic features | |
CN114495170A (en) | Pedestrian re-identification method and system based on local self-attention inhibition | |
Yang et al. | Probabilistic projective association and semantic guided relocalization for dense reconstruction | |
CN117036404A (en) | Monocular thermal imaging simultaneous positioning and mapping method and system |
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 |