CN104143194B - A kind of point cloud segmentation method and device - Google Patents
A kind of point cloud segmentation method and device Download PDFInfo
- Publication number
- CN104143194B CN104143194B CN201410410469.2A CN201410410469A CN104143194B CN 104143194 B CN104143194 B CN 104143194B CN 201410410469 A CN201410410469 A CN 201410410469A CN 104143194 B CN104143194 B CN 104143194B
- Authority
- CN
- China
- Prior art keywords
- cluster
- point
- ground
- current
- segmentation
- 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
Landscapes
- Measurement Of Velocity Or Position Using Acoustic Or Ultrasonic Waves (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of point cloud segmentation method and device, it is related to three-dimensional reconstruction field, the described method comprises the following steps:S1:Tested scene is scanned by column by depth transducer, to obtain the depth information of the tested scene, the depth information of the tested scene Coordinate Conversion is subjected to, to obtain three-dimensional information of the tested scene under local coordinate;S2:Ground point cloud is partitioned into from the three-dimensional information;S3:Cluster segmentation is carried out to non-ground points cloud by the cluster segmentation mode of adaptive threshold, the non-ground points cloud is other clouds in the three-dimensional information in addition to the ground point cloud.The present invention carries out cluster segmentation by the cluster segmentation mode of adaptive threshold to non-ground points cloud, effectively prevent the less divided and over-segmentation problem carried out using fixed threshold to non-ground points cloud caused by cluster segmentation.
Description
Technical field
The present invention relates to three-dimensional reconstruction field, more particularly to a kind of point cloud segmentation method and device.
Background technology
For the three-dimensional reconstruction of large scene, due to its weight in terms of three-dimensional city map, road upkeep, urban planning
Apply, receive very big concern.Fixed station is based on using depth transducer and position and attitude sensor or mobile platform is gathered
The three-dimensional information of surrounding environment, is widely adopted due to its efficient, real-time, high-precision characteristic.Due to the scene bag of scanning
Containing different types of object, such as ground, building, trees, vehicle, it is necessary to pass through a cloud before three-dimensional reconstruction is carried out
Segmentation is separated from one another by the corresponding cloud data of different types of object, to carry out a cloud modeling respectively to each object.
Current most of point cloud segmentation methods are that unordered, discrete point cloud is handled, without using data
The order that depth transducer is scanned successively by column in gatherer process.In point cloud segmentation method, cluster segmentation method especially base
In the cluster segmentation method of distance, due to its, method complexity is low, be easily achieved, and is most commonly used to the segmentation of spatial point cloud.But
Due to the presence of ground point cloud in outdoor large scene, and be used alone the clustering method based on distance be difficult to effectively ground-to-ground face and
Non- ground object is split, and this method is generally combined with other methods, first extracts the point cloud for belonging to ground, then to non-ground points
Cloud carries out cluster segmentation.
In non-ground points cloud cluster segmentation part, the cluster segmentation side based on fixed threshold radius of current more use
Method, the selection of threshold value influences very big to the result of segmentation, if threshold value chooses excessive, the nearer wisp of spacing distance may not
It can be separated (less divided);If threshold value chooses too small, the larger object of spacing distance (such as building) may be divided into many
Individual cluster (over-segmentation).
The content of the invention
In order to avoid being asked using fixed threshold the less divided caused by non-ground points cloud progress cluster segmentation and over-segmentation
Topic, the invention provides a kind of point cloud segmentation method, the described method comprises the following steps:
S1:Tested scene is scanned by column by depth transducer, to obtain the depth information of the tested scene,
The depth information of the tested scene is subjected to Coordinate Conversion, to obtain three-dimensional letter of the tested scene under local coordinate
Breath;
S2:Ground point cloud is partitioned into from the three-dimensional information;
S3:Cluster segmentation, the non-ground points are carried out to non-ground points cloud by the cluster segmentation mode of adaptive threshold
Cloud is other clouds in the three-dimensional information in addition to the ground point cloud.
Wherein, in step S2, ground point cloud is partitioned into from the three-dimensional information and is specifically included:
S201:To arrange the three-dimensional information is traveled through for unit, and using traverse one row as working as prostatitis;
S202:According to the direction traversal each point for working as prostatitis from bottom to up;
S203:Calculate the Z axis coordinate of current point traversed and the depth transducer position ground level it
The absolute value of difference, if the absolute value is less than height threshold, using the current point as first initial point Ps (1), is performed
Step S204, otherwise return to step 202;
S204:Detect the cut off Pe (k) after the initial point Ps (k), and by the initial point Ps (k) and cut off
Pe (k) is used as a pair;
S205:The initial point Ps (k+1) after cut off Pe (k) is detected, if detecting initial point Ps (k+1), k=is made
K+1, and return to step S204, otherwise by each pair initial point Ps (k) and cut off Pe (k) and each pair initial point Ps (k) and cut-off
Point between point Pe (k) performs step S206 as the ground point addition ground point cloud;
S206:Judge whether each row of the three-dimensional information are traversed to, if so, then performing step S3, otherwise return
Return step S202.
Wherein, in step S3, cluster segmentation tool is carried out to non-ground points cloud by the cluster segmentation mode of adaptive threshold
Body includes:
S301:Cluster segmentation is carried out to the non-ground points cloud using fixed initial threshold dth (0);
S302:Each cluster that traversal cluster segmentation is obtained;
S303:The adaptive threshold dth (w) of the current cluster traversed is calculated according to the cluster scale of construction, judges described adaptive
Answer whether threshold value dth (w) is less than the initial threshold dth (0), if so, then based on the adaptive threshold dth (w) to described
Current cluster is further split;
S304:Judge whether each cluster that the cluster segmentation is acquired is traversed to, if it is not, then return to step
S303。
Wherein, in step S303, had according to the adaptive threshold dth (w) that the cluster scale of construction calculates the current cluster traversed
Body includes:
S3031:Each point in the current cluster is traveled through, it is flat that the current point traversed is combined into the fitting of its neighborhood point
Face, and normal vector of the normal vector for obtaining plane as the current point will be fitted;
S3032:Calculate in the current cluster normal vector a little average value, and using the average value as described
The normal vector currently clustered;
S3033:Institute in the current cluster is a little projected to projection plane, chosen on the projection plane
The minimum rectangle of all subpoints is included, the projection plane is the plane vertical with the normal vector currently clustered;
S3034:The adaptive threshold dth (w) currently clustered is calculated by following formula,
Dth (w)=k1·L(w)+k2
Wherein, L (w) is the bond length of the minimum rectangle, k1And k2For constant.
Wherein, in step S3, cluster segmentation tool is carried out to non-ground points cloud by the cluster segmentation mode of adaptive threshold
Body includes:
S311:Each point in the non-ground points cloud is traveled through, the adaptive threshold of the current point traversed is calculated;
S312:Cluster minute is carried out to the non-ground points cloud according to the adaptive threshold each put in the non-ground points cloud
Cut.
Wherein, in step S311, the adaptive threshold dth of the current point traversed is calculated by following formula,
Wherein, L is the depth information of the current point, and A is the angular resolution of the depth transducer, and v is the depth
The translational speed of sensor in the horizontal direction, f is the scan frequency of the depth transducer, and α is plane where scan line and v
Angle between direction, plane where the scan line is the central point of the depth transducer and the T of the tested scene
Plane where row, the T is classified as the row including the current point, and β is plane where scan line and quilt in tested scene
The angle of body surface is surveyed, θ is the angle between the T row and horizontal direction of the tested scene, AwSwept for depth transducer
The effective angle retouched, Δ L is L error amount, and Δ v is v error amount, and Δ α is α error amount.
Wherein, after step S3, in addition to:
S4:Merged in the cluster that step S3 is obtained by the cluster of over-segmentation.
Wherein, step S4 is specifically included:
S401:Each cluster that traversal cluster segmentation is acquired;
S402:Each point in the current cluster q that traverses is traveled through, the current point traversed is combined into its neighborhood
Point fit Plane, and normal vector of the normal vector for obtaining plane as the current point will be fitted;
S403:Calculate in the current cluster q normal vector a little average value, and using the average value as described
The normal vector currently clustered;
S404:Institute in the current cluster q is a little projected to projection plane, calculates the current cluster q's
Projected area, the projection plane is the plane vertical with the normal vector of the current cluster q;
S405:Judge whether the projected area is more than preset area, if so, then using the current cluster q as substantially
Amount cluster;
S406:Judge whether each cluster that cluster segmentation is acquired is traversed to, if so, then performing step
S407, otherwise return to step S402;
S407:If the spacing distance between two big scale of construction cluster is less than pre-determined distance, by the corresponding two big scale of construction
Cluster merging is a cluster, until the spacing distance between the big scale of construction cluster of any two is all higher than the pre-determined distance, institute
It is the distance between nearest two points in two clusters to state the spacing distance between two big scale of construction cluster.
Wherein, also include after step S407:
S408:Each cluster that traversal step S407 is acquired;
S409:Calculate the difference DELTA H, the current cluster r between the current cluster r traversed height and ground level
Height for the current cluster r somewhat middle Z axis coordinate minimum value, the ground level of the current cluster r is described
The maximum of Z axis coordinate in first initial point of corresponding row is each put in current cluster r;
S410:Judge whether the difference DELTA H is more than preset difference value, if so, then that the current cluster r is poly- as suspending
Class;
S411:Whether each cluster that judgment step S407 is acquired is traversed to, if so, then performing step
S412, otherwise return to step S409;
S412:By each suspension Cluster merging to the cluster closest with it, until each cluster that suspends is incorporated into
In one non-cluster that suspends.
The invention also discloses a kind of point cloud segmentation device, described device includes:
Scan conversion module, it is described tested to obtain for being scanned by column by depth transducer to tested scene
The depth information of scene, carries out Coordinate Conversion, to obtain the tested scene local by the depth information of the tested scene
Three-dimensional information under coordinate system;
Module is split on ground, for being partitioned into ground point cloud from the three-dimensional information;
Module is split on non-ground, and cluster point is carried out to non-ground points cloud for the cluster segmentation mode by adaptive threshold
Cut, the non-ground points cloud is other clouds in the three-dimensional information in addition to the ground point cloud.
The present invention carries out cluster segmentation by the cluster segmentation mode of adaptive threshold to non-ground points cloud, effectively prevent
Less divided and over-segmentation problem caused by cluster segmentation is carried out to non-ground points cloud using fixed threshold.
Brief description of the drawings
When Fig. 1 is that the present invention is scanned to tested scene, the structured flowchart for the scanning system being based on;
Fig. 2 is the depth information of the tested scene collected according to Fig. 1 scanning system;
Fig. 3 is the flow chart of the point cloud segmentation method of one embodiment of the present invention;
Fig. 4 is the schematic diagram for the scan line that an embodiment of the present invention is gathered in depth transducer;
Fig. 5 is the schematic diagram of the point cloud in the scan line shown in Fig. 4;
Fig. 6 is schematic diagram of an embodiment of the present invention when ground is split;
Fig. 7 a are that first point in the scan line direction from bottom to up of an embodiment of the present invention belongs to the signal of ground point
Figure;
Fig. 7 b are that first point in the scan line direction from bottom to up of an embodiment of the present invention belongs to showing for non-ground points
It is intended to;
Fig. 8 is that the bond length of the minimum rectangle of an embodiment of the present invention obtains flow chart;
Fig. 9 is the schematic diagram of the cluster projection area of an embodiment of the present invention;
Figure 10 is the design sketch for being partitioned into ground point cloud of an embodiment of the present invention;
Figure 11 is an embodiment of the present invention by the design sketch after being merged by the cluster of over-segmentation;
Figure 12 is the target point A of an embodiment of the present invention1Neighbourhood model;
Figure 13 is the structured flowchart of the point cloud segmentation device of one embodiment of the present invention.
Embodiment
With reference to the accompanying drawings and examples, the embodiment to the present invention is described in further detail.Implement below
Example is used to illustrate the present invention, but is not limited to the scope of the present invention.
Reference picture 1, during to tested scene scanning, the scanning system being based on by storage control module, depth transducer and
Position and attitude sensor group is into depth transducer and the position and attitude sensor is both placed in above mobile platform.Storage control
Molding block provides order so that the depth information (reference picture 2) for the tested scene that depth transducer acquisition scans are arrived, meanwhile, position
Attitude transducer record is put in information such as the position of each moment mobile platform, posture and mobile platform gait of march.Will be in depth
The depth information obtained under degree sensor coordinate system, can be obtained under the attitude transducer coordinate system of position by coordinate transform
The three-dimensional information of scene is tested, then obtains being tested the three-dimensional information of scene under local coordinate by coordinate transform, due to collection
Outdoor scene be typically made up of substantial amounts of spatial point, the set of these spatial points is also referred to as a cloud.
In present embodiment, the depth transducer selects laser radar, and position and attitude sensor selects GPS, institute's rheme
Put attitude transducer and also can select GPS/IMU (inertial navigation) integrated navigation system.
Fig. 3 is the flow chart of the point cloud segmentation method of one embodiment of the present invention;Reference picture 3, methods described include with
Lower step:
S1:Tested scene is scanned by column by depth transducer, to obtain the depth information of the tested scene,
The depth information of the tested scene is subjected to Coordinate Conversion, to obtain three-dimensional letter of the tested scene under local coordinate
Breath;
S2:Ground point cloud is partitioned into from the three-dimensional information;
S3:Cluster segmentation, the non-ground points are carried out to non-ground points cloud by the cluster segmentation mode of adaptive threshold
Cloud is other clouds in the three-dimensional information in addition to the ground point cloud.
In step S1, when being scanned by column by depth transducer to tested scene, it can use from bottom to up or from upper
Scanning direction under is scanned by column, and it is scan line starting point to scan first obtained point, and last point is scanning
Line terminating point.
Depth transducer is scanned column by column during mobile platform is moved with fixed angular resolution, obtains correspondence ring
The depth information in border, every column data of scanning is also referred to as scan line.There is fixed acquisition order and points per column data, most
The depth information of the tested scene obtained eventually is by by showing all scanline groups of sequence into the corresponding data of each testee
It is to be collectively constituted by the partial dot in some scan lines.Because depth transducer takes the mode gathered by column, the points of each column
It is fixed with order, therefore the depth information obtained is with by the characteristic for showing sequence.Shown in Fig. 4 as under a given scenario,
The schematic diagram for the scan line that depth transducer on mobile platform is gathered in the course of the work, depth transducer is with fixed
Angular resolution is scanned from bottom to top, and ground, wisp, ground and building are arrived in scanning successively, obtain the corresponding depth of each point
Spend information.Triangle pair should be the data on the data on the ground collected, the circular non-ground object for corresponding to collect.
The information for the depth information binding site attitude transducer that depth transducer is collected, by coordinate transform, finally gives
The coordinate of each scanning element under local coordinate, three-dimensional information of the as described tested scene under local coordinate.
Initially face amount is determined to be accurate, while connecting in the segmentation having using the ground point on each column scan line, section
Continuous the characteristics of, extracts ground, to improve the accurate segmentation to ground point cloud, it is preferable that in step S2, from the three-dimensional information
Ground point cloud is partitioned into specifically include:
S201:To arrange the three-dimensional information is traveled through for unit, and using traverse one row as working as prostatitis;
S202:According to the direction traversal each point for working as prostatitis from bottom to up;
S203:Calculate the Z axis coordinate of current point traversed and the depth transducer position ground level it
The absolute value of difference, if the absolute value is less than height threshold, using the current point as first initial point Ps (1), is performed
Step S204, otherwise return to step 202;
S204:Detect the cut off Pe (k) after the initial point Ps (k), and by the initial point Ps (k) and cut off
Pe (k) is used as a pair;
S205:The initial point Ps (k+1) after cut off Pe (k) is detected, if detecting initial point Ps (k+1), k=is made
K+1, and return to step S204, otherwise by each pair initial point Ps (k) and cut off Pe (k) and each pair initial point Ps (k) and cut-off
Point between point Pe (k) performs step S206 as the ground point addition ground point cloud;
S206:Judge whether each row of the three-dimensional information are traversed to, if so, then performing step S3, otherwise return
Return step S202.
Two ways can be used by realizing the cluster segmentation of the non-ground points cloud of adaptive threshold:A kind of is oneself based on the scale of construction
Threshold segmentation mode is adapted to, another is the adaptive threshold fuzziness mode based on multi-parameter.
For ease of realizing the adaptive threshold fuzziness mode based on the scale of construction, it is preferable that in step S3, pass through adaptive threshold
Cluster segmentation mode to non-ground points cloud carry out cluster segmentation specifically include:
S301:Cluster segmentation is carried out to the non-ground points cloud using fixed initial threshold dth (0);
S302:Each cluster that traversal cluster segmentation is obtained;
S303:The adaptive threshold dth (w) of the current cluster traversed is calculated according to the cluster scale of construction, judges described adaptive
Answer whether threshold value dth (w) is less than the initial threshold dth (0), if so, then based on the adaptive threshold dth (w) to described
Current cluster is further split;
S304:Judge whether each cluster that the cluster segmentation is acquired is traversed to, if it is not, then return to step
S303。
For ease of being determined according to the scale of construction in adaptive threshold, step S303, according to cluster the scale of construction calculate traverse it is current
The adaptive threshold dth (w) of cluster is specifically included:
S3031:Each point in the current cluster is traveled through, it is flat that the current point traversed is combined into the fitting of its neighborhood point
Face, and normal vector of the normal vector for obtaining plane as the current point will be fitted;
S3032:Calculate in the current cluster normal vector a little average value, and using the average value as described
The normal vector currently clustered;
S3033:Institute in the current cluster is a little projected to projection plane, chosen on the projection plane
The minimum rectangle of all subpoints is included, the projection plane is the plane vertical with the normal vector currently clustered;
S3034:The adaptive threshold dth (w) currently clustered is calculated by following formula,
Dth (w)=k1·L(w)+k2
Wherein, L (w) is the bond length of the minimum rectangle, k1And k2For constant.
For ease of realizing the adaptive threshold fuzziness mode based on multi-parameter, it is preferable that in step S3, pass through adaptive thresholding
The cluster segmentation mode of value carries out cluster segmentation to non-ground points cloud and specifically included:
S311:Each point in the non-ground points cloud is traveled through, the adaptive threshold of the current point traversed is calculated;
S312:Cluster minute is carried out to the non-ground points cloud according to the adaptive threshold each put in the non-ground points cloud
Cut.
For ease of being determined according to multi-parameter in adaptive threshold, step S311, the current point traversed is calculated by following formula
Adaptive threshold dth,
Wherein, L is the depth information of the current point, and A is the angular resolution of the depth transducer, and v is the depth
The translational speed of sensor in the horizontal direction, f is the scan frequency of the depth transducer, and α is plane where scan line and v
Angle between direction, plane where the scan line is the central point of the depth transducer and the T of the tested scene
Plane where row, the T is classified as the row including the current point, and β is plane where scan line and quilt in tested scene
The angle of body surface is surveyed, θ is the angle between the T row and horizontal direction of the tested scene, AwSwept for depth transducer
The effective angle retouched, Δ L is L error amount, and Δ v is v error amount, and Δ α is α error amount.
After the problem of to solve over-segmentation, step S3, in addition to:
S4:Merged in the cluster that step S3 is obtained by the cluster of over-segmentation.
To solve the problem of big scale of construction cluster is by over-segmentation, it is preferable that step S4 is specifically included:
S401:Each cluster that traversal cluster segmentation is acquired;
S402:Each point in the current cluster q that traverses is traveled through, the current point traversed is combined into its neighborhood
Point fit Plane, and normal vector of the normal vector for obtaining plane as the current point will be fitted;
S403:Calculate in the current cluster q normal vector a little average value, and using the average value as described
The normal vector currently clustered;
S404:Institute in the current cluster q is a little projected to projection plane, calculates the current cluster q's
Projected area, the projection plane is the plane vertical with the normal vector of the current cluster q;
S405:Judge whether the projected area is more than preset area, if so, then using the current cluster q as substantially
Amount cluster;
S406:Judge whether each cluster that cluster segmentation is acquired is traversed to, if so, then performing step
S407, otherwise return to step S402;
S407:If the distance between central point of two big scale of construction cluster is less than pre-determined distance, big by corresponding two
Scale of construction Cluster merging is a cluster, until the distance between central point of the big scale of construction cluster of any two is all higher than described preset
Distance.
The problem of for the cluster that solves to suspend by over-segmentation, it is preferable that also include after step S407:
S408:Each cluster that traversal step S407 is acquired;
S409:Calculate the difference DELTA H, the current cluster r between the current cluster r traversed height and ground level
Height for the current cluster r somewhat middle Z axis coordinate minimum value, the ground level of the current cluster r is described
The maximum of Z axis coordinate in first initial point of corresponding row is each put in current cluster r;
S410:Judge whether the difference DELTA H is more than preset difference value, if so, then that the current cluster r is poly- as suspending
Class;
S411:Whether each cluster that judgment step S407 is acquired is traversed to, if so, then performing step
S412, otherwise return to step S409;
S412:By each suspension Cluster merging to the cluster closest with it, until each cluster that suspends is incorporated into
In one non-cluster that suspends.
Embodiment 1
The data in scan line are analyzed, as shown in Figure 5, it can be seen that point p1To point, pointTo point
For the cloud data on ground, pointTo pointAnd pointTo point pnFor the cloud data on non-ground object.
It can be obtained with reference to most scene analysis, in a scan line, the cloud data for belonging to ground is segmented in presence and section continuously mostly,
Such as Fig. 5 midpoints p1To pointPoint cloud tract, pointTo pointPoint cloud tract be two sections and belong to ground
Point cloud section;Same section belong to ground point cloud, local relief is smaller between points, and elevation difference is also smaller.Divide with reference to more than
Analysis, is effective extraction ground point cloud, and detection column by column belongs to the point cloud tract on ground.Belong to the point cloud sequence on ground at every section
In, it is initial point Ps to define first point, and last point is cut off Pe, and initial point and cut off are distinguished in corresponding diagram 5
Point p1,And point,。
The present invention is illustrated with a specific embodiment below, but does not limit protection scope of the present invention.In above-mentioned reality
Apply on the basis of mode, ground point cloud is partitioned into from the three-dimensional information and is specifically included:
By row to be handled, the point cloud section for belonging to ground in every scan line is extracted column by column.Processed one is set
The initial point that kth section belongs to the point cloud tract on ground in bar scan line is Ps (k), and cut off is Pe (k), and first paragraph belongs to ground
The corresponding initial point of point cloud tract in face is Ps (1), and cut off is Pe (1), and the data to a scan line carry out ground distributor
The method cut mainly is made up of three links, is first initial point Ps (1) of detection, detection cut off Pe (k) and detection respectively
Initial point Ps (k+1) after cut off.Data to a row carry out flow such as Fig. 6 of ground segmentation, and specific steps include:
Step S2-1:Detect first initial point Ps (1);
As illustrated in figs. 7 a and 7b, triangular representation ground point, round dot represents non-ground points (wherein, ps1With above-mentioned Ps (1),
pe1With above-mentioned Pe (1), second initial point is ps2, pe2For the ps2Corresponding detection cut off, the expression of other marks
Mode ibid, is not being repeated again), first point in scan line direction from bottom to up both may belong to ground point, also may be used
To be non-ground points, in order to obtain the accurate segmentation of a scan line upper ground surface part and non-above ground portion it may first have to ensure
The correctness of first ground reference point.The detection of first initial point Ps (1) in scan line can be utilized based on mobile flat
The known element of platform, due to the position (X of the depth transducer scanning center under local coordinatelidar, Ylidar, Zlidar) it is real
When store, the height (abbreviation podium level) of depth transducer scanning center to mesa base, as shown in H in Fig. 4, Ke Yizuo
Use, install on a mobile platform after equipment for systematic parameter, platform can be drawn by being tested in the environment of wide flat
Height H, in the case where system is not dismounted again, podium level H is constant.Utilize the position peace of depth transducer scanning center
Platform height, with reference to the elevation information that measuring point is detected in scan line, can detect first initial point for belonging to ground in scan line
Ps (1), i.e. above-mentioned steps S2-1, comprise the following steps:
(1):To processed J row depth transducer depth informations, and set-point sequence number i (i=0,1 ..., n) and correspondingly
I-th point of local coordinate system (Xi, Yi, Zi), first point correspondence i=0, corresponding local seat in the row depth information are set
It is designated as (X0, Y0, Z0);
(2):Calculate the Z axis coordinate Z of lower i-th point of local coordinate systemiWith the ground level of depth transducer position
(Zlidar- H) difference DELTA hi, calculation formula is:Δhi=| (Zlidar-H)-Zi|;
(3):Judge corresponding difference in height Δ hiWhether threshold value Zth is less than, if Δ hi<Zth, then it is assumed that i-th point is to be somebody's turn to do
Belong to first initial point Ps (1) on ground in column scan point;Otherwise make i=i+1, return to step (2);
Situations such as threshold value Zth allows for measurement noise, ground local relief (including deceleration strip, traffic stud etc.), scan line
Patient difference in height between upper two ground point.China national standard path provides that the height of projection on the road surface such as deceleration strip is not
5 centimetres can be higher than, considering further that in the planarization on ground and the influence of measurement noise, the present embodiment can tolerate between two ground points
Difference in height Zth 15 centimetres are set to ensure enough allowances.
Step S2-2:Detect the cut off Pe (k) after initial point Ps (k);
Detect the initial point Ps (k) of kth section ground point cloud tract afterwards, it is necessary to further detect this section of point cloud sequence
Cut off Pe (k), and then kth section ground point cloud section could be extracted.In the method as proposed in the present invention, i-th is judged
After individual point is the initial point Ps (k) of kth section ground point cloud tract, using adjacent 2 points after initial point Ps (k) of lines with
The angle, θ of XOY plane formation judges whether the point after i points belongs to kth section ground point successively as the examination criteria of ground point
Cloud tract, the cut off Pe (k) until detecting kth section ground point cloud tract, it is thus regarded that to point by point Ps (k)
What Pe (k) ended belongs to a little ground.
Under local coordinate, the point u and v of difference two for belonging to ground in same scan line, their line is put down with XOY
The angle, θ that face is formedU, vFor characterizing the fluctuating quantity on ground between 2 points of u, v, the angle, θ as ground point examination criteria.
θ specific formula for calculation is as follows:
For certain section of ground point set G={ Ps (k) ..., Pe (k) } being partitioned into, due to the fluctuating between adjacent 2 points compared with
Small, this section of ground point concentrates the corresponding angle, θ of the point of arbitrary neighborhood two j, j+1J, j+1Certain threshold value is should be less than, i.e., should be met
θJ, j+1<θth.Thus ground sequence can further be detected by adjacent 2 points of corresponding angle information θ after initial point Ps (k),
Until finding cut off Pe (k), i.e. above-mentioned steps S2-2, following steps are specifically included:
(1) it is j points, j=i+1 to set first point after initial point Ps (k) (i-th point in correspondence scan line);
(2) under local coordinate, the line and the angle, θ of XOY plane of j points and j+1 points are calculatedJ, j+1, it is specific to calculate public
Formula is as follows:
Due to there is the noise that radar and GPS are produced, only by with adjacent a little corresponding angle, θJ, j+1Cut to detect
Stop Pe (k) method is excessively sensitive, to avoid next point (i.e. j+2 of j point consecutive points in influence of noise, the present embodiment
Point), also take in simultaneously, i.e., also calculate angle, θ simultaneouslyJ, j+2;
(3) θ is judgedJ, j+1、θJ, j+2Whether threshold θ is both greater thanth, if it is, thinking that the fluctuating of consecutive points is excessive, judge j
Point is cut off Pe (k);Conversely, j=j+1 is then made, into step (2);
Threshold θthTogether decided on by the fluctuating quantity and system noise on local ground, it is considered to above-mentioned various shadows that may be present
Threshold θ in the factor of sound, the present embodimentthIt is set as 17 ° to retain enough allowances.
Step S2-3:Detect the initial point Ps (k+1) after cut off Pe (k);
After the cut off Pe (k) for obtaining kth section ground point cloud tract, the point from point Ps (k) to point Pe (k) is judged to
It is set to the point for belonging to ground, set-point Pe (k) (i.e. in step S2-2 corresponding j-th point) local coordinate system is (Xj, Yj, Zj)。
, it is necessary to detect next section of ground point on the basis of the cut off Pe (k) for putting cloud sequence that the preceding paragraph belongs to ground has been obtained
Initial point Ps (k+1), and then continue to detect the next section of point cloud sequence for belonging to ground.
For the point m after cut off Pe (k) on the column scan line (i.e. j-th points of the row), 2 points of correspondences of j, m are used
Angle, θJ, mAnd elevation ZJ, mConstraint judge whether point m belongs to next initial point Ps (k+1), if meeting θJ, mIt is less than
Threshold θthAnd ZJ, mLess than threshold value Zth, then decision-point m is next initial point Ps (k+1), on the contrary then think that point m is not initial
Point, continuation is judged the point after m points.
Wherein 2 points of corresponding angle, θs of j, mJ, mIt is the angle for the line and XOY plane formation for giving directions j, m, definition and meter
Calculate angle, θ in formula and step S2-2U, vIt is identical;Elevation ZJ, mIt is point j and height differences of the point m under local coordinate, calculating
Formula is:ZJ, m=| Zj-Zm|;Threshold θthDefinition and value and step S2-2 in threshold θthIt is identical;Threshold value Zth definition and
Value is identical with the threshold value Zth in step S2-1.
Step S2-3 specific implementation step is as follows:
(1) it is m points, i.e. m=j+1 to set at i.e. j-th point at first point after cut off Pe (k) ();
(2) under local coordinate, cut off Pe (k) and point m difference in height Z are calculatedJ, m, ZJ, m=| Zj-Zm|;Calculate cut-off
The line of point Pe (k) and m points and the angle, θ of XOY planeJ, m;
(3) Z is judged whetherJ, mLess than threshold value Zth and θJ, mLess than threshold θth, if it is, thinking that m points are next section of ground
The initial point Ps (k+1) of sequence;It is on the contrary then make m=m+1, into step (2);
Next initial point Pss (k+1) of the step S2-3 after detecting cut off Pe (k) is performed, or to the row
Institute after midpoint Pe (k) is a little all judged, then carries out step S2-4.
Step S2-4:Judge whether detect new initial point Ps (k+1) after cut off Pe (k), if in cut off
New initial point Ps (k+1) is detected after Pe (k), then k=k+1 return to step S2-2 can be made to continue executing with;If in cut-off
New initial point Ps (k+1) is not detected after point Pe (k), then the institute in the scan line after cut off Pe (k) is somewhat equal
Non-ground points, it is possible to determine that the institute of the row cloud data is a little processed, all ground point cloud tracts of the row all by
Extract.
For next scan line, synchronize the processing of rapid S2-1 to S2-4 identicals, thus by column, each column it is right paragraph by paragraph
Ground point cloud is extracted, you can complete the segmentation of ground point cloud.
Embodiment 2
According to the ground dividing method based on scan line, it is necessary to non-after ground point cloud and non-ground points cloud are split
Ground point cloud carries out cluster segmentation, and the point cloud that different objects are belonged in non-ground points cloud is separated from each other.
In order to avoid the above-mentioned cluster over-segmentation occurred using same fixed threshold and less divided phenomenon, the present embodiment is adopted
With the relation between required threshold value when the adaptive threshold fuzziness mode based on the scale of construction, the i.e. scale of construction based on object and segmentation, lead to
For often, the spacing distance between the object (such as building) of the big scale of construction is larger, and threshold value should take when splitting the object of the big scale of construction
Obtain larger, and the spacing distance between the object (such as automobile) of small volume can be smaller, threshold value should during segmentation small volume object
Obtain smaller, it is, the threshold radius of cluster segmentation and the scale of construction of object, which exist, is incremented by relation.
The adaptive threshold fuzziness mode is comprised the following steps that:
Step S3-1:Using fixed initial threshold dth0, the base based on European clustering method is carried out to non-ground points cloud
Plinth is split;
Set and non-ground points are based on after fixed threshold dth0 clusters terminate, obtain W cluster, cluster numbering w, first
Cluster correspondence w=1;
S3-2:For carrying out clustering obtained cluster w based on fixed threshold dth0, calculate w based on the adaptive of the scale of construction
Threshold value dth (w);
For carrying out clustering obtained cluster w, it is necessary to calculate its corresponding adaptive threshold dth based on fixed threshold dth0
(w).Analyze as described above, spacing distance is larger between the object (such as building) of the big scale of construction, and the object (such as automobile) of small volume
Between spacing distance can be smaller.To characterize the incremental relation between threshold size and the object scale of construction, while complicated for calculating
Consideration in terms of degree, self-adaption cluster radial design be to the linear relationship that is directly proportional of cluster size, w-th cluster based on body
The calculation formula of the adaptive threshold of amount is shown below:
Dth (w)=k1·L(w)+k2
In the parameter that above formula characterizes linear relationship, coefficient k 1 is by analyzing distance and object body between general familiar object
Relation is obtained between amount, the object often occurred in such as building, automobile, bicycle, street lamp scene, and k1=is taken in the present embodiment
0.01m;Coefficient k 2 is gathered by vehicle-mounted mobile platform depth transducer in minimum distinguishable distance between two articles, the present embodiment
The resolution ratio of data is determined, k2=0.18m is taken in the present embodiment.
Length L (w) in formula is the length for clustering w, and its concrete operation process is as shown in Figure 8.Point in figure is represented
Cluster the point in w.As shown in (a1) in Fig. 8, for the point in cluster w, it is fitted by the point in the range of each surrounding neighbors
Plane, obtains the normal vector of the normal vector of the plane, the i.e. point.According to identical method, after the normal vector for obtaining each point, by w
In the normal vector n that middle normal vector a little is averaged as cluster w, such as Fig. 8 shown in (a2).By the normal vector with the cluster
Vertical plane elects projection plane as, by all spot projections in w to the perspective plane, and the point in Fig. 8 in (a3) on perspective plane is
The subpoint obtained after cluster projection.The minimum rectangle that all subpoints can be included is chosen in the projection plane,
In the both sides of the rectangle, shorter length of side L is defined as in current cluster w length L (w), such as Fig. 8 rectangle shorter edge in (a4)
Shown in long L.
S3-3:Judge whether the adaptive threshold dth (w) for clustering w is less than initial threshold radius dth0, if it is, base
Cluster w is further split in adaptive threshold dth (w), the object segmentation not separated during so as to by initial segmentation
Fully;It is on the contrary then think cluster w it is divided complete, no longer carry out the segmentation based on adaptive threshold.
S3-4:Judge whether all to have handled all W clusters, if it is, the cluster based on scale of construction adaptive threshold
Segmentation terminates;Conversely, w=w+1 is then made, return to step S3-2;
Embodiment 3
Carried out based on adaptive threshold radius to being tested non-ground points all in scene after cluster segmentation, because scanning hides
The reason such as gear, mobile platform swerve and testee is irregular, cluster segmentation result can have object showing by over-segmentation
As mainly having two class situations, one kind is for irregularly shaped objects such as trees, by taking trees as an example, because trunk, each branch etc. are closed
Connection is smaller, it is easy to be divided into multiple clusters, wherein only trunk is cluster on the ground, and other clusters are suspended in ground
On, referred to as suspend cluster.Also one kind be exactly for the big scale of construction object such as building, because threshold radius sets smaller,
Vehicle-mounted mobile platform swerves or scanned to deposit can cause big scale of construction cluster to be too cut into multiple clusters under occlusion
Situation.Therefore need to merge the cluster of over-segmentation processing.
When merging processing to the big scale of construction of over-segmentation cluster, tested cluster projection is hung down to the normal vector with the cluster
Straight plane, then extracts the profile on perspective plane to calculate the projected area of tested cluster.If the projected area of a cluster
More than empirical value Su, the cluster is judged as a big scale of construction cluster.If the spacing distance between the big scale of construction cluster of any two
Less than being required between buildings in general apart from lu, the two clusters are considered to belong to same building thing and need to be merged into one
Spacing distance between cluster, the big scale of construction cluster of two of which is the distance between two points nearest in the two clusters.Cross
The combining step of the big scale of construction cluster of segmentation is specific as follows:
(1) set and carried out based on adaptive threshold after cluster segmentation, non-ground points cloud is divided into Q cluster, set poly-
Class sequence number q (q=1,2 ..., Q), sets first cluster correspondence q=1 in segmentation result;
(2) for the point in cluster q, similar to the process in Fig. 8 in S2-2 shown in (a1) to (a3), each point week is passed through
The point fit Plane in contiguous range is enclosed, the normal vector of the normal vector of the plane, the i.e. point is obtained, according to identical method, obtains
After the normal vector of each point, institute's normal vector a little in q is averaged as the normal vector n for clustering q, will a little be thrown in q
Shadow is to the plane vertical with n.Extract the profiles that project on the projection surface of cluster q to calculate tested cluster by convex hull method afterwards
Q projected area S (q), as shown in Figure 9;
(3) judge whether cluster q projected area S (q) is more than threshold value Su, if it is, thinking that q clusters for the big scale of construction;
It is on the contrary then think that q is not belonging to big scale of construction cluster;
(4) determine whether to handle all Q clusters, if it is, into step (5);It is on the contrary then make q=q+
1, into step (2);
(5) processing is merged to the big scale of construction of over-segmentation cluster, if that is, the big scale of construction of any two cluster between away from
From less than requiring distance l between buildings in generalu, the two clusters are considered to belong to same building thing, the two are clustered
It is merged into a cluster.Until the distance between big scale of construction cluster of any two is all higher than luWhen, then the big scale of construction of over-segmentation is gathered
Class, which merges, to be terminated;Otherwise step (5) is continued executing with;
In the present embodiment, it is contemplated that tree and building is caused in itself blocks, the area of big scale of construction cluster is determined whether
Threshold value SuIt is set to empirical value 30m2, the general spacing l of two buildingsuIt is set to empirical value 10m.
When merging processing to the levitated object of over-segmentation, it is necessary first to which whether judge cluster is the cluster that suspends.On ground
Face point cloud segmentation link, obtains the ground point cloud and non-ground points cloud in every scan line by column, and then for each non-ground
Point can match corresponding ground level Zg, i.e., to each non-ground points on any one surface sweeping line j, all correspond to j row
First initial ground point of all non-ground points and the jth row for arranging jth in the ground level Zg of position, the present embodiment
Z axis coordinate value matchings of the Ps (1) under local coordinate, it is believed that point Ps (1) Z axis coordinate value reflects the non-ground of jth row
The ground level of point position;
Set and R cluster is had after terminating to the big scale of construction Cluster merging of over-segmentation, to wherein any one cluster r, need
The difference DELTA H between cluster r height and position and corresponding ground level is calculated, and then judges whether it suspends.Due to
Arbitrary point t have matched a corresponding ground level Zg (t) in cluster r, take cluster in r Z axis coordinate a little most
Small value is that min { Z (t) } is cluster r height and position, takes the maximum max in the ground level Zg that all-pair is answered in cluster r
{ Zg (t) } is the corresponding ground levels of cluster r, it is believed that the difference DELTA H between cluster r height and position and corresponding ground level
=| min { Z (t) }-max { Zg (t) } |.For any cluster r, if corresponding Δ H is more than threshold value Fth, then it is assumed that cluster r is outstanding
Floating cluster.For the cluster of suspension, the spacing distance between cluster is calculated, and the Cluster merging is clustered recently to it, until
The cluster that suspends all is ultimately to be incorporated into the non-suspension belonged on ground a cluster.Comprise the following steps that:
(1) set after the processing of big scale of construction Cluster merging, have R and cluster, setting cluster sequence number r (r=1,2 ...,
R), first cluster correspondence r=1;
(2) for cluster r, the height min { Z (t) } and corresponding ground level max { Zg (t) } of the cluster are calculated respectively,
And then calculate difference DELTA H, i.e. Δ H=between cluster r height and corresponding ground level | min { Z (t) }-max { Zg (t) }
|;
(3) judge that the corresponding Δ H of cluster r are no and be more than threshold value Fth, if it is, thinking that r clusters to suspend;It is on the contrary then recognize
The cluster that suspends is not belonging to for r;
In the present embodiment, it is considered to which the influence of error simultaneously stays certain allowance, given threshold Fth is 1m.
(4) determine whether all R clusters have all been carried out suspending to judge, if it is, into step (5);It is on the contrary then
R=r+1 is made, into step (2);
(5) e is clustered for any one suspension to be combined, calculates the spacer between the cluster of other in cluster e and R
From, and cluster e is merged into closest cluster with interval.If finally be all merged into one belongs to ground to the cluster that suspends
On non-suspension cluster on, then the merging treatment of the suspension cluster of over-segmentation is completed, on the contrary then continue executing with (5);Wherein, two
Spacing distance between the cluster that suspends refers to the distance of the central point of two clusters, and the coordinate of the central point of cluster is the cluster
Coordinate value a little be averaging.
Figure 10 show the design sketch being partitioned into after ground point cloud.To the ground segmentation proposed by the present invention based on scan line
When method carries out Performance Evaluation, for method accuracy from two class error evaluations:A) originally belong to ground point but be mistaken for non-
Ground point, b) originally belonging to non-ground points is mistaken for ground point.To assess above-mentioned error, we are counted manually from special scenes
Ground is counted and non-ground points number, and carries out comparing calculation with testing obtained result.Scan line is based on by proposed by the present invention
The ground dividing method (R.B.Rusu, " based on random sampling coherence method that proposes of ground dividing method and R.B.Rusu
Semantic 3D object maps for everyday manipulation in human living
Environments, " KI-K ü nstliche Intelligenz, vol.24, no.4, pp.345-348, Nov.2010.) carry out
Compare, by related experiment and calculating, draw set forth herein the ground dividing method error rate based on scan line be
The error rate of 0.674%, R.B.Rusu method is 1.52%.In terms of method time complexity, both of which is operated in
The computer platform of Intel E7400,2.8GHz, 2GB internal memory, method proposed by the present invention and R.B.Rusu method operation
Time is 0.028 second and 4.820 seconds respectively, illustrates method proposed by the present invention faster, and average each second can be handled up to
22358179 data.Specific performance comparison result such as following table:
Method | Error a (%) | Error b (%) | Overall error (%) | Run time (s) |
The inventive method | 0.401 | 0.273 | 0.674 | 0.028 |
R.B.Rusu methods | 0.919 | 0.623 | 1.52 | 4.820 |
From performance comparison result, the overall performance of the inventive method is higher than R.B.Rusu method.
Figure 11 is shown to carry out splitting based on scale of construction adaptive threshold to non-ground points cloud, then to the cluster of over-segmentation
The corresponding cloud data of different vehicles has different gray values in result after merging, Figure 11, has represented method
Different vehicles and separated through determining, it can be seen that method can carry out accurate to the large scene three-dimensional point cloud of acquisition
True segmentation, can split accurately even for the slope pavement in scene, automobile etc..
Embodiment 4
In order to avoid the above-mentioned cluster over-segmentation occurred using same fixed threshold and less divided phenomenon, the present embodiment is adopted
With the adaptive threshold fuzziness mode based on multi-parameter, the adaptive threshold fuzziness mode based on multi-parameter is specifically included:
First, the distance model of consecutive points in point cloud is set up:Based on mobile platform operation logic, as platform is to reach
Dynamic, depth transducer is scanned by column successively to tested scene, it is assumed that in subrange, where the point of proximity around a certain target point
Face is similar to plane, and collection schematic diagram is as shown in figure 12, to target point A1And its modeling of neighborhood point, the selection of neighborhood point is as in Figure 12
The point A of mark2, B1, B2Deng respective distances are followed successively by target point A1It is up and down between consecutive points apart from A in the scan line of place1A2,
Apart from A between the point nearest with it and the point in its adjacent two scan line1B1, it is nearest with it in its adjacent two scan line
Point B1Up and down 2 points with the point apart from A1B2.Based on consecutive points distance model, derive between consecutive points apart from A1A2、A1B2、A1B1's
Calculation expression.Due to A1B2>A1B1, and A1B2>A1A2, the consecutive points distance around each target point can pass through A1B2With reference to
It is determined that.In view of radar scanning center to measured point apart from L, mobile platform pace v, plane where radar scanning line with
Angle α is respectively present error delta L, Δ v and Δ α between car travel direction, the cluster segmentation threshold radius each put such as following formula institute
Show:
Wherein, L is the depth information of the current point, and A is the angular resolution of the depth transducer, and v is the depth
The translational speed of sensor in the horizontal direction, f is the scan frequency of the depth transducer, and α is plane where scan line and v
Angle between direction, plane where the scan line is the central point of the depth transducer and the T of the tested scene
Plane where row, the T is classified as the row including the current point, and β is plane where scan line and testee surface
Angle (because the parameter is a uncertain value, so the present embodiment using a default angle as the parameter value), θ is described
Angle between the T row and horizontal direction of tested scene, AwThe effective angle scanned for depth transducer, Δ L is L error
Value, Δ v is v error amount, and Δ α is α error amount.
In the parameter that above-mentioned formula is related to, L, A, Aw, α, f has determined as constant, gathered data mistake before acquisition
Constant in journey, depth transducer range error Δ L is 0.03 meter in the present embodiment, and angular resolution A is the radian of 0.5 °, i.e. π/360,
Scan effective range AwFor 90 °, i.e. pi/2 radian, α is 45 ° and 135 °, and scan frequency is 75 hertz.β values are one and preset
Value, takes β=15 °.L, v, Δ v and Δ α are the real time data that is obtained in scanning process, and wherein radar scanning center is to being scanned
The distance between point L is obtained by depth transducer, mobile platform travel speed v, velocity error Δ v, scan line angular error Δ α
Obtained by position and attitude sensor.
The N number of non-ground points arranged for jth, set-point sequence number I (I=1,2 ..., N), according to above-mentioned based on multi-parameter
Cluster segmentation threshold expression, calculates each self-corresponding cluster segmentation threshold value of each non-ground points respectively, sets i-th point correspondence
Threshold value be dthI;First non-ground points correspondence I=1 that jth is arranged is set, for target point I, according to the corresponding threshold value of point
Radius dthI, judge whether non-ground points to target point I that the row of jth -1 and jth are arranged distance are less than threshold value dthIIf,
It is, then it is assumed that corresponding point is target point I neighbor point, neighbor point and target point I is gathered for a class;Repeat the above steps, directly
To the cluster segmentation processing all carried out to all non-ground points based on multi-parameter adaptive threshold, then non-ground points are based on
The cluster segmentation of multi-parameter adaptive threshold terminates.
Even if not carrying out the merging treatment of over-segmentation cluster also, the cluster segmentation method based on multi-parameter adaptive threshold
Obtained segmentation result is very fully and very close to the result finally clustered, simultaneously as this method is directly right
All non-ground points cloud carries out the cluster segmentation once based on oneself corresponding threshold radius of each point, and based on the scale of construction from
A based fragmentation will first be carried out by adapting to the cluster segmentation method of threshold value, then based fragmentation result is carried out to be based on adaptive threshold
Cluster segmentation, therefore the cluster segmentation method time complexity of the adaptive threshold based on multi-parameter is lower.
Embodiment 5
It is different from depth transducer in embodiment 1-4 and uses fixed effective angle scope Aw(AwLess than 360 degree) carry out by
Depth transducer is by the way of 360 degree of scannings in column scan, the present embodiment, i.e., the mistake that depth transducer is moved in system platform
Gather the depth information of surrounding scene in journey by 360 degree of rotation sweeps with fixed angular resolution, depth transducer often rotates
One circle correspondence obtains one and encloses depth information.Under normal circumstances, can more efficient, more fully it be obtained by the way of 360 degree of scannings
The depth information of tested scene is taken, for example, when movement is scanned system platform on road;According in embodiment 1-4
Scan mode, i.e., depth transducer is with fixed effective angle scope Aw(AwLess than 360 degree) tested scene is swept by column
When retouching, depth transducer is tended not to while the corresponding depth information of the scene for collecting both sides of the road;If depth transducer is adopted
With the mode of 360 degree of scannings, because the angle of scanning is sufficiently large, depth transducer can be scanned in the range of 360 degree simultaneously
The depth information of tested scene.
Depth transducer is actually by the effective angle of depth transducer in embodiment 1-4 by the way of 360 degree of scannings
Spend scope Aw360 degree have been increased to, obtain each is scanned by column in each circle depth information for collecting correspondence embodiment 1-4
Row depth information, if thinking, the circle depth information in embodiment 5 is " row " data, and 360 degree of scannings mode is obtained
The characteristics of data, is consistent with the data characteristicses handled in embodiment 1-4, and point cloud segmentation method proposed by the invention is applied to deep
Spend the situation of sensor 360 degree of scannings.
The invention also discloses a kind of point cloud segmentation device, reference picture 13, described device includes:
Scan conversion module, it is described tested to obtain for being scanned by column by depth transducer to tested scene
The depth information of scene, carries out Coordinate Conversion, to obtain the tested scene local by the depth information of the tested scene
Three-dimensional information under coordinate system;
Module is split on ground, for being partitioned into ground point cloud from the three-dimensional information;
Module is split on non-ground, and cluster point is carried out to non-ground points cloud for the cluster segmentation mode by adaptive threshold
Cut, the non-ground points cloud is other clouds in the three-dimensional information in addition to the ground point cloud.
The system is corresponding with the above method, in addition to other are to realize the module of the above method, submodule, unit, son
Unit etc., to avoid redundancy, so do not illustrate herein.
Embodiment of above is merely to illustrate the present invention, and not limitation of the present invention, about the common of technical field
Technical staff, without departing from the spirit and scope of the present invention, can also make a variety of changes and modification, therefore all
Equivalent technical scheme falls within scope of the invention, and scope of patent protection of the invention should be defined by the claims.
Claims (9)
1. a kind of point cloud segmentation method, it is characterised in that the described method comprises the following steps:
S1:Tested scene is scanned by column by depth transducer, to obtain the depth information of the tested scene, by institute
The depth information for stating tested scene carries out Coordinate Conversion, to obtain three-dimensional information of the tested scene under local coordinate;
S2:Ground point cloud is partitioned into from the three-dimensional information;
S3:Cluster segmentation is carried out to non-ground points cloud by the cluster segmentation mode of adaptive threshold, the non-ground points cloud is
Other clouds in the three-dimensional information in addition to the ground point cloud,
Wherein, in step S2, ground point cloud is partitioned into from the three-dimensional information and is specifically included:
S201:To arrange the three-dimensional information is traveled through for unit, and using traverse one row as working as prostatitis;
S202:According to the direction traversal each point for working as prostatitis from bottom to up;
S203:Calculate the difference of the Z axis coordinate of the current point traversed and the ground level of the depth transducer position
Absolute value, if the absolute value is less than height threshold, using the current point as first initial point Ps (1), performs step
S204, otherwise return to step S 202;
S204:The cut off Pe (k) after initial point Ps (k) is detected, and the initial point Ps (k) and cut off Pe (k) are made
For a pair;
S205:The initial point Ps (k+1) after cut off Pe (k) is detected, if detecting initial point Ps (k+1), k=k+1 is made,
And return to step S204, otherwise by each pair initial point Ps (k) and cut off Pe (k) and each pair initial point Ps (k) and cut off Pe
(k) point between performs step S206 as the ground point addition ground point cloud;
S206:Judge whether each row of the three-dimensional information are traversed to, if so, step S3 is then performed, otherwise will be current
The next column of row as working as prostatitis, and return to step S202,
Wherein, the line and the angle of XOY plane formation of the cut off and next consecutive points are more than angle threshold.
2. the method as described in claim 1, it is characterised in that in step S3, passes through the cluster segmentation mode of adaptive threshold
Cluster segmentation is carried out to non-ground points cloud to specifically include:
S301:Cluster segmentation is carried out to the non-ground points cloud using fixed initial threshold dth (0);
S302:Each cluster that traversal cluster segmentation is obtained;
S303:The adaptive threshold dth (w) of the current cluster traversed is calculated according to the cluster scale of construction, the adaptive thresholding is judged
Whether value dth (w) is less than the initial threshold dth (0), if so, then based on the adaptive threshold dth (w) to described current
Cluster is further split;
S304:Judge whether each cluster that the cluster segmentation is acquired is traversed to, if it is not, then will currently cluster
Next cluster is clustered as current, and return to step S303.
3. method as claimed in claim 2, it is characterised in that in step S303, according to working as that cluster scale of construction calculating is traversed
The adaptive threshold dth (w) of preceding cluster is specifically included:
S3031:Each point in the current cluster is traveled through, the current point traversed is combined into its neighborhood point fit Plane, and
The normal vector of the plane obtained will be fitted as the normal vector of the current point;
S3032:Calculate in the current cluster normal vector a little average value, and using the average value as described current
The normal vector of cluster;
S3033:Institute in the current cluster is a little projected to projection plane, chooses and includes on the projection plane
The minimum rectangle of all subpoints, the projection plane is the plane vertical with the normal vector currently clustered;
S3034:The adaptive threshold dth (w) currently clustered is calculated by following formula,
Dth (w)=k1·L(w)+k2
Wherein, L (w) is the bond length of the minimum rectangle, k1And k2For constant.
4. the method as described in claim 1, it is characterised in that in step S3, passes through the cluster segmentation mode of adaptive threshold
Cluster segmentation is carried out to non-ground points cloud to specifically include:
S311:Each point in the non-ground points cloud is traveled through, the adaptive threshold of the current point traversed is calculated;
S312:Cluster segmentation is carried out to the non-ground points cloud according to the adaptive threshold each put in the non-ground points cloud.
5. method as claimed in claim 4, it is characterised in that in step S311, the current point traversed is calculated by following formula
Adaptive threshold dth,
Wherein, L is the depth information of the current point, and A is the angular resolution of the depth transducer, and v is the depth sensing
The translational speed of device in the horizontal direction, f is the scan frequency of the depth transducer, and α is the direction of plane and v where scan line
Between angle, plane where the scan line is the central point of the depth transducer and the T row institute of the tested scene
Plane, the T is classified as the row including the current point, and β is plane where scan line and measured object in tested scene
The angle in body surface face, AwThe effective angle scanned for depth transducer, Δ L is L error amount, and Δ v is v error amount, and Δ α is
α error amount.
6. the method as described in claim 1, it is characterised in that after step S3, in addition to:
S4:Merged in the cluster that step S3 is obtained by the cluster of over-segmentation.
7. method as claimed in claim 6, it is characterised in that step S4 is specifically included:
S401:Each cluster that traversal cluster segmentation is acquired;
S402:Each point in the current cluster q that traverses is traveled through, the current point traversed is intended with reference to its neighborhood point
Plane is closed, and the normal vector of the plane obtained will be fitted as the normal vector of the current point;
S403:Calculate in the current cluster q normal vector a little average value, and using the average value as described current
The normal vector of cluster;
S404:Institute in the current cluster q is a little projected to projection plane, the projection of the current cluster q is calculated
Area, the projection plane is the plane vertical with the normal vector of the current cluster q;
S405:Judge whether the projected area is more than preset area, if so, then gathering the current cluster q as the big scale of construction
Class;
S406:Judge whether each cluster that cluster segmentation is acquired is traversed to, if so, then execution step S407, no
Then the next cluster currently clustered is clustered as current, and return to step S402;
S407:If the spacing distance between two big scale of construction cluster is less than pre-determined distance, the corresponding two big scale of construction is clustered
A cluster is merged into, until the spacing distance between the big scale of construction cluster of any two is all higher than the pre-determined distance, described two
Spacing distance between individual big scale of construction cluster is the distance between nearest two points in two clusters.
8. method as claimed in claim 7, it is characterised in that also include after step S407:
S408:Each cluster that traversal step S407 is acquired;
S409:Calculate the difference DELTA H between the current cluster r traversed height and ground level, the height of the current cluster r
Spend for the current cluster r somewhat middle Z axis coordinate minimum value, the ground level of the current cluster r is described current
Cluster the maximum of Z axis coordinate in first initial point of the corresponding row of each point in r;
S410:Judge whether the difference DELTA H is more than preset difference value, if so, then clustering the current cluster r as suspending;
S411:Whether each cluster that judgment step S407 is acquired is traversed to, if so, then execution step S412, no
Then the next cluster currently clustered is clustered as current, and return to step S409;
S412:By each suspension Cluster merging to the cluster closest with it, until each cluster that suspends is incorporated into one
In the non-cluster that suspends.
9. a kind of point cloud segmentation device, it is characterised in that described device includes:
Scan conversion module, for being scanned by column by depth transducer to tested scene, to obtain the tested scene
Depth information, the depth information of the tested scene is subjected to Coordinate Conversion, to obtain the tested scene in local coordinate system
Three-dimensional information under system;
Module is split on ground, for being partitioned into ground point cloud from the three-dimensional information;
Module is split on non-ground, and cluster segmentation is carried out to non-ground points cloud for the cluster segmentation mode by adaptive threshold,
The non-ground points cloud is other clouds in the three-dimensional information in addition to the ground point cloud,
Wherein, the ground segmentation module, specifically for:
S201:To arrange the three-dimensional information is traveled through for unit, and using traverse one row as working as prostatitis;
S202:According to the direction traversal each point for working as prostatitis from bottom to up;
S203:Calculate the difference of the Z axis coordinate of the current point traversed and the ground level of the depth transducer position
Absolute value, if the absolute value is less than height threshold, using the current point as first initial point Ps (1), performs step
S204, otherwise return to step S 202;
S204:The cut off Pe (k) after initial point Ps (k) is detected, and the initial point Ps (k) and cut off Pe (k) are made
For a pair;
S205:The initial point Ps (k+1) after cut off Pe (k) is detected, if detecting initial point Ps (k+1), k=k+1 is made,
And return to step S204, otherwise by each pair initial point Ps (k) and cut off Pe (k) and each pair initial point Ps (k) and cut off Pe
(k) point between performs step S206 as the ground point addition ground point cloud;
S206:Judge whether each row of the three-dimensional information are traversed to, if so, then performing the non-ground segmentation mould
Block, otherwise using when the next column in prostatitis is as working as prostatitis, and return to step S202,
Wherein, the line and the angle of XOY plane formation of the cut off and next consecutive points are more than angle threshold.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410410469.2A CN104143194B (en) | 2014-08-20 | 2014-08-20 | A kind of point cloud segmentation method and device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410410469.2A CN104143194B (en) | 2014-08-20 | 2014-08-20 | A kind of point cloud segmentation method and device |
Publications (2)
Publication Number | Publication Date |
---|---|
CN104143194A CN104143194A (en) | 2014-11-12 |
CN104143194B true CN104143194B (en) | 2017-09-08 |
Family
ID=51852362
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201410410469.2A Active CN104143194B (en) | 2014-08-20 | 2014-08-20 | A kind of point cloud segmentation method and device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104143194B (en) |
Families Citing this family (45)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105139376B (en) * | 2015-07-16 | 2018-02-02 | 武汉体育学院 | A kind of shooting method of counting based on shooting counting device |
CN109033920B (en) * | 2017-06-08 | 2021-10-08 | 株式会社理光 | Recognition method and device for grabbed target and computer readable storage medium |
CN107369161B (en) * | 2017-07-19 | 2020-09-11 | 无锡信捷电气股份有限公司 | Scattered workpiece point cloud segmentation method based on improved Euclidean clustering |
CN107657621B (en) * | 2017-10-20 | 2021-04-13 | 南京林业大学 | Two-dimensional laser point cloud sequence real-time segmentation method based on linear region growth |
US20190197167A1 (en) * | 2017-12-21 | 2019-06-27 | Ford Global Technologies, Llc | Depth-data segmentation |
CN110349158A (en) * | 2018-04-04 | 2019-10-18 | 北京京东尚科信息技术有限公司 | A kind of method and apparatus handling point cloud data |
CN110363834B (en) * | 2018-04-10 | 2023-05-30 | 北京京东尚科信息技术有限公司 | Point cloud data segmentation method and device |
US10705534B2 (en) * | 2018-04-19 | 2020-07-07 | Faraday&Future Inc. | System and method for ground plane detection |
CN108596860A (en) * | 2018-05-10 | 2018-09-28 | 芜湖航飞科技股份有限公司 | A kind of ground point cloud dividing method based on three-dimensional laser radar |
CN108710845A (en) * | 2018-05-11 | 2018-10-26 | 北京旷视科技有限公司 | The correlating method of target object and article, apparatus and system |
CN108734120B (en) * | 2018-05-15 | 2022-05-10 | 百度在线网络技术(北京)有限公司 | Method, device and equipment for labeling image and computer readable storage medium |
CN109100741B (en) * | 2018-06-11 | 2020-11-20 | 长安大学 | Target detection method based on 3D laser radar and image data |
CN108986162B (en) * | 2018-06-28 | 2022-02-22 | 杭州吉吉知识产权运营有限公司 | Dish and background segmentation method based on inertial measurement unit and visual information |
CN109033989B (en) * | 2018-07-02 | 2020-11-24 | 深圳辰视智能科技有限公司 | Target identification method and device based on three-dimensional point cloud and storage medium |
CN108885791B (en) * | 2018-07-06 | 2022-04-08 | 达闼机器人有限公司 | Ground detection method, related device and computer readable storage medium |
CN109118500B (en) * | 2018-07-16 | 2022-05-10 | 重庆大学产业技术研究院 | Image-based three-dimensional laser scanning point cloud data segmentation method |
CN109242855B (en) * | 2018-07-19 | 2020-08-11 | 中国科学院自动化研究所 | Multi-resolution three-dimensional statistical information-based roof segmentation method, system and equipment |
CN110858415B (en) * | 2018-08-24 | 2021-04-02 | 北京图森智途科技有限公司 | Method and device for labeling object in 3D point cloud data |
CN110148144B (en) * | 2018-08-27 | 2024-02-13 | 腾讯大地通途(北京)科技有限公司 | Point cloud data segmentation method and device, storage medium and electronic device |
CN109190573B (en) | 2018-09-12 | 2021-11-12 | 阿波罗智能技术(北京)有限公司 | Ground detection method applied to unmanned vehicle, electronic equipment and vehicle |
CN110909569B (en) * | 2018-09-17 | 2022-09-23 | 深圳市优必选科技有限公司 | Road condition information identification method and terminal equipment |
CN109359614B (en) * | 2018-10-30 | 2021-06-11 | 百度在线网络技术(北京)有限公司 | Laser point cloud plane identification method, device, equipment and medium |
CN109613553B (en) * | 2018-12-18 | 2021-05-18 | 歌尔光学科技有限公司 | Method, device and system for determining number of objects in scene based on laser radar |
CN111352106B (en) * | 2018-12-24 | 2022-06-14 | 珠海一微半导体股份有限公司 | Sweeping robot slope identification method and device, chip and sweeping robot |
CN109872329A (en) * | 2019-01-28 | 2019-06-11 | 重庆邮电大学 | A kind of ground point cloud fast partition method based on three-dimensional laser radar |
CN110033457B (en) * | 2019-03-11 | 2021-11-30 | 北京理工大学 | Target point cloud segmentation method |
CN109961440B (en) * | 2019-03-11 | 2021-06-18 | 重庆邮电大学 | Three-dimensional laser radar point cloud target segmentation method based on depth map |
CN111724429B (en) * | 2019-03-21 | 2024-03-01 | 北京京东乾石科技有限公司 | Ground characteristic extraction method and device |
JP7147651B2 (en) * | 2019-03-22 | 2022-10-05 | トヨタ自動車株式会社 | Object recognition device and vehicle control system |
TWI706841B (en) * | 2019-03-22 | 2020-10-11 | 達明機器人股份有限公司 | Device and method for calibrating coordinate of 3d camera and robot arm |
CN112634181B (en) * | 2019-09-24 | 2024-06-14 | 阿波罗智能技术(北京)有限公司 | Method and device for detecting ground point cloud points |
CN111192284B (en) * | 2019-12-27 | 2022-04-05 | 吉林大学 | Vehicle-mounted laser point cloud segmentation method and system |
CN113366532B (en) * | 2019-12-30 | 2023-03-21 | 深圳元戎启行科技有限公司 | Point cloud based segmentation processing method and device, computer equipment and storage medium |
CN111227444A (en) * | 2020-01-17 | 2020-06-05 | 泉州装备制造研究所 | 3D sole glue spraying path planning method based on k nearest neighbor |
CN111462141B (en) * | 2020-05-19 | 2024-01-12 | 北京爱笔科技有限公司 | Point cloud plane acquisition method and device, equipment and computer readable storage medium |
DE102020206815A1 (en) * | 2020-05-29 | 2021-12-02 | Robert Bosch Gesellschaft mit beschränkter Haftung | Method for classifying measuring points in a point cloud |
CN112213735A (en) * | 2020-08-25 | 2021-01-12 | 上海主线科技有限公司 | Laser point cloud noise reduction method for rainy and snowy weather |
CN112508970B (en) * | 2020-12-16 | 2023-08-25 | 北京超星未来科技有限公司 | Point cloud data segmentation method and device |
CN112785596B (en) * | 2021-02-01 | 2022-06-10 | 中国铁建电气化局集团有限公司 | Dot cloud picture bolt segmentation and height measurement method based on DBSCAN clustering |
CN113450315B (en) * | 2021-06-08 | 2024-06-14 | 北京伟景智能科技有限公司 | Bar counting method, bar counting device and steel separating system |
CN113537311B (en) * | 2021-06-30 | 2023-08-04 | 北京百度网讯科技有限公司 | Spatial point clustering method and device and electronic equipment |
CN113587930B (en) * | 2021-10-08 | 2022-04-05 | 广东省科学院智能制造研究所 | Indoor and outdoor navigation method and device of autonomous mobile robot based on multi-sensor fusion |
CN113607166B (en) * | 2021-10-08 | 2022-01-07 | 广东省科学院智能制造研究所 | Indoor and outdoor positioning method and device for autonomous mobile robot based on multi-sensor fusion |
CN114782467A (en) * | 2022-04-14 | 2022-07-22 | 电子科技大学 | Point cloud ground segmentation method based on region division and self-adaptive threshold |
CN115375699B (en) * | 2022-10-25 | 2023-03-24 | 杭州华橙软件技术有限公司 | Point cloud segmentation method, mobile robot and computer-readable storage medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN201218726Y (en) * | 2008-04-23 | 2009-04-08 | 哈尔滨理工大学 | Cultural relics three-dimensional reconstruction apparatus based on colorful structured light |
CN101738172A (en) * | 2008-11-05 | 2010-06-16 | 哈尔滨理工大学 | Method for three-dimensional measurement of high sampling density color structured light based on green stripe segmentation |
CN102881047A (en) * | 2012-08-01 | 2013-01-16 | 桂林电子科技大学 | Automatic non-closed implicit curved surface reconstruction method |
CN103886312A (en) * | 2014-03-21 | 2014-06-25 | 华东师范大学 | Method for rapidly selecting and displaying interested blood vessel in magnetic resonance blood vessel image |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102419794A (en) * | 2011-10-31 | 2012-04-18 | 武汉大学 | Method for quickly filtering airborne laser point cloud data |
CN103020637B (en) * | 2012-11-22 | 2016-01-27 | 北京航空航天大学 | A kind of buildings end face point cloud data segmentation method based on K-plane algorithm |
CN103268609B (en) * | 2013-05-17 | 2016-04-20 | 清华大学 | A kind of some cloud dividing method of orderly extraction ground |
CN103514598B (en) * | 2013-07-20 | 2016-08-10 | 北京航空航天大学 | A kind of automatic division method building SFM point cloud |
-
2014
- 2014-08-20 CN CN201410410469.2A patent/CN104143194B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN201218726Y (en) * | 2008-04-23 | 2009-04-08 | 哈尔滨理工大学 | Cultural relics three-dimensional reconstruction apparatus based on colorful structured light |
CN101738172A (en) * | 2008-11-05 | 2010-06-16 | 哈尔滨理工大学 | Method for three-dimensional measurement of high sampling density color structured light based on green stripe segmentation |
CN102881047A (en) * | 2012-08-01 | 2013-01-16 | 桂林电子科技大学 | Automatic non-closed implicit curved surface reconstruction method |
CN103886312A (en) * | 2014-03-21 | 2014-06-25 | 华东师范大学 | Method for rapidly selecting and displaying interested blood vessel in magnetic resonance blood vessel image |
Non-Patent Citations (1)
Title |
---|
图像阈值分割算法研究;郭臻 等;《中国传媒大学学报自然科学版》;20080630;第15卷(第2期);第77-82页 * |
Also Published As
Publication number | Publication date |
---|---|
CN104143194A (en) | 2014-11-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104143194B (en) | A kind of point cloud segmentation method and device | |
CN106997049B (en) | Method and device for detecting barrier based on laser point cloud data | |
CN106370189B (en) | Indoor navigation device and method based on multi-sensor fusion | |
CN103136393B (en) | A kind of areal coverage computing method based on stress and strain model | |
CN106441292B (en) | A kind of building indoor plane figure method for building up based on crowdsourcing IMU inertial guidance data | |
CN108195377B (en) | Reflector matching algorithm based on triangular perimeter matching | |
CN103778429B (en) | Automatic extraction method for road information in a kind of Vehicle-borne Laser Scanning point cloud | |
CN109060821B (en) | Tunnel disease detection method and tunnel disease detection device based on laser detection | |
CN107792115B (en) | It is a kind of to automatically extract both wired rail crest level methods using three-dimensional laser point cloud | |
CN109410586A (en) | A kind of Traffic State Detection Method based on multivariate data fusion | |
CN109144072A (en) | A kind of intelligent robot barrier-avoiding method based on three-dimensional laser | |
CN107657636A (en) | A kind of method that route topography figure elevational point is automatically extracted based on mobile lidar data | |
CN106199558A (en) | Barrier method for quick | |
CN103605978A (en) | Urban illegal building identification system and method based on three-dimensional live-action data | |
CN106887020A (en) | A kind of road vertical and horizontal section acquisition methods based on LiDAR point cloud | |
CN103226833A (en) | Point cloud data partitioning method based on three-dimensional laser radar | |
CN105761305B (en) | A kind of vehicle traveling detection method and system | |
CN102096072B (en) | Method for automatically measuring urban parts | |
CN104657968B (en) | Automatic vehicle-mounted three-dimensional laser point cloud facade classification and outline extraction method | |
JP6465421B1 (en) | Structural deformation detector | |
CN106842193A (en) | The processing method of Road Detection information, device and system | |
CN101425128A (en) | People stream detecting method based on laser sensor depth image | |
CN106017476A (en) | Method for generating indoor positioning and navigating map model | |
CN110057362A (en) | The method for planning path for mobile robot of finite elements map | |
Lehtomäki et al. | Performance analysis of a pole and tree trunk detection method for mobile laser scanning data |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |