CN109100741A - A kind of object detection method based on 3D laser radar and image data - Google Patents
A kind of object detection method based on 3D laser radar and image data Download PDFInfo
- Publication number
- CN109100741A CN109100741A CN201810594692.5A CN201810594692A CN109100741A CN 109100741 A CN109100741 A CN 109100741A CN 201810594692 A CN201810594692 A CN 201810594692A CN 109100741 A CN109100741 A CN 109100741A
- Authority
- CN
- China
- Prior art keywords
- interest
- point
- area
- radar
- target
- 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
- 238000001514 detection method Methods 0.000 title claims abstract description 29
- 238000000034 method Methods 0.000 claims abstract description 29
- 239000000284 extract Substances 0.000 claims abstract description 17
- 238000000605 extraction Methods 0.000 claims abstract description 13
- 239000011159 matrix material Substances 0.000 claims description 30
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000001914 filtration Methods 0.000 claims description 5
- 238000006243 chemical reaction Methods 0.000 claims description 4
- 239000012141 concentrate Substances 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 4
- 230000001629 suppression Effects 0.000 claims description 4
- 238000005259 measurement Methods 0.000 claims description 3
- 238000013519 translation Methods 0.000 claims description 3
- 238000003475 lamination Methods 0.000 claims description 2
- 239000010749 BS 2869 Class C1 Substances 0.000 claims 1
- 238000011897 real-time detection Methods 0.000 abstract 1
- 238000010586 diagram Methods 0.000 description 5
- 238000013527 convolutional neural network Methods 0.000 description 3
- 230000000717 retained effect Effects 0.000 description 3
- 238000012545 processing Methods 0.000 description 2
- 230000001953 sensory effect Effects 0.000 description 2
- 239000013589 supplement Substances 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000012549 training Methods 0.000 description 2
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 239000000047 product Substances 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- IEDVJHCEMCRBQM-UHFFFAOYSA-N trimethoprim Chemical compound COC1=C(OC)C(OC)=CC(CC=2C(=NC(N)=NC=2)N)=C1 IEDVJHCEMCRBQM-UHFFFAOYSA-N 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S17/00—Systems using the reflection or reradiation of electromagnetic waves other than radio waves, e.g. lidar systems
- G01S17/88—Lidar systems specially adapted for specific applications
- G01S17/89—Lidar systems specially adapted for specific applications for mapping or imaging
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S17/00—Systems using the reflection or reradiation of electromagnetic waves other than radio waves, e.g. lidar systems
- G01S17/86—Combinations of lidar systems with systems other than lidar, radar or sonar, e.g. with direction finders
Abstract
The invention discloses a kind of object detection method based on 3D laser radar and image data, this method obtains the 3D point cloud data and camera image of ambient enviroment using 3D laser radar and camera, and pre-processes to the 3D point cloud data;The ground point in 3D point cloud data is filtered out, space clustering is carried out to remaining non-ground points, extracts the 3D area-of-interest of target;The outer parameter of the coordinate of 3D laser radar and camera is demarcated, and the 3D area-of-interest of target is mapped in corresponding camera image according to calibrating parameters, extracts corresponding 2D area-of-interest in camera image;Feature extraction is carried out to the 2D area-of-interest using depth convolutional network, so that the target in 2D area-of-interest is positioned and be identified.The present invention takes full advantage of the complementarity among 3D laser radar and camera data, improves the precision and timeliness of the target positioning and Classification and Identification to scene, the target real-time detection that can be used in unmanned vehicle.
Description
Technical field
The present invention relates to multi-sensor information fusions, and in particular to one kind is based on 3D laser radar point cloud object candidate area
The object detection method with image convolution neural network classification is extracted, it is right for the important component of unmanned vehicle environment sensing
The detection accuracy of Vehicle target is improved, ensures that unmanned vehicle safety traffic is of great significance.
Background technique
Autonomous driving vehicle can fundamentally improve safety and the comfort of driving population, while reduce automobile to ring
The influence in border.In order to develop such vehicle, sensory perceptual system be vehicle analysis understand driving environment indispensable component it
One, the position including peripheral obstacle, direction and classification.
3D laser radar is one of the most popular sensor for autonomous vehicle sensory perceptual system, it has extensive view
Open country, the remote and Infravision in accurate depth information and target identification.In object detection task, since laser is swept
The space coordinate substantially comprising point cloud is retouched, so 3D laser radar is relative to posture and the shape tool for obtaining the object detected
There is certain advantage.However, with the increase with scanning center distance, the distribution of 3D laser radar point cloud becomes more and more diluter
It dredges, this makes 3D laser radar be difficult to detect specific object in classification.
Camera can provide high-definition picture for precise classification, in recent years in terms of field of image recognition, deep learning
Extensive research is obtained.These methods usually generate object candidate area using object candidate area generation method first,
Such as sliding window method, select search method and Multiscale combination method, then using convolutional neural networks model to candidate region into
Row feature extraction and target fixation and recognition.The shortcomings that these methods is to need to generate a large amount of candidate region, and real-time is very poor.This
Outside, camera is influenced by illumination change, lacks the position of 3D object, direction and geometry, and object candidate area is caused to be extracted not
Accurately.
Summary of the invention
The object of the present invention is to provide a kind of object detection method based on 3D laser radar and image data, makes full use of
3D laser radar can directly acquire high-precision target depth and target geometrical characteristic parameter and image object classification it is excellent
Gesture realizes mutual supplement with each other's advantages, overcome using single-sensor carry out target detection existing for precision it is low, poor robustness etc. is asked
Topic, has utmostly ensured unmanned vehicle in the safety traffic of complex condition.
Technical scheme is as follows:
A kind of object detection method based on 3D laser radar and image data, comprising the following steps:
Step 1, the 3D point cloud data and phase of ambient enviroment are obtained using the 3D laser radar and camera being installed on vehicle
Machine image, and the 3D point cloud data are pre-processed;
Step 2, the ground point in 3D point cloud data is filtered out, space clustering is carried out to remaining non-ground points, extracts target
3D area-of-interest;
Step 3, the outer parameter of the coordinate of 3D laser radar and camera is demarcated, and according to calibrating parameters by target
3D area-of-interest is mapped in corresponding camera image, extracts corresponding 2D area-of-interest in camera image;
Step 4, feature extraction is carried out to the 2D area-of-interest using depth convolutional network, thus interested in 2D
Target in region is positioned and is identified.
Further, the 3D point cloud data are pre-processed in step 1, comprising:
Step 1.1, the point cloud data that radar obtains is transformed under rectangular coordinate system
By the point set P of point cloud datarIt is transformed under rectangular coordinate system, calculates point and concentrate each scanning element in rectangular coordinate system
Under coordinate, obtain each scanning element piMulti-parameter representation:
pi=(γi,θi,φi,Ii,xi,yi,zi)
Wherein γiIndicate radial distance of the scanning element to radar, θi,φiIndicate water of the scanning element relative to spheric coordinate system
Gentle vertical angle, IiIndicate radar reflection intensity, xi,yi,ziFor scanning element piCoordinate under rectangular coordinate system;
The rectangular coordinate system is using the geometric center position of radar as coordinate origin, with the vertical axis direction of radar
For Z axis, using vehicle forward direction as Y-axis, and X-axis follows the right-hand rule by Z axis and Y-axis and determines;Conversion process is as follows:
Step 1.2, region filtering is carried out according to the rectangular co-ordinate, sets region of interest border, retained interested
Scanning element in region, it may be assumed that
Pf={ pi|-X<xi<X,Y<yi<Y,Z1<zi<Z2} (2)
As scanning element piCoordinate (xi,yi,zi) meet in region of interest border-X < xi<X,Y<yi<Y,Z1<zi<Z2It is interior
When, by scanning element piPoint set P is addedf, thus obtain the point set P of area-of-interest scanning elementf。
Further, the 3D point cloud data are pre-processed, further includes:
Step 1.3, noise spot filters out
For point set PfEach of scanning element pi, search in scanning element piNeighbor Points in radius R, if piIt is close
Adjoint point quantity, then will point p less than MiIt is labeled as noise spot and from point set PfMiddle removal;Traverse point set Pf, find all noises
Point and from point set PfMiddle removal obtains pretreated point set P.
Further, the ground point in 3D point cloud data is filtered out described in step 2, comprising:
Step 2.1, point set P is mapped in multi-dimensional matrix, the line number of matrix is equal to the item number of radar scanning line, matrix
Columns is equal to the points that a scan line includes;Scanning element p in point set PiBe mapped in matrix where row r and column c meter
Calculation mode is as follows:
R=(θi+180)/Δθ (3)
C=φi/Δφ (4)
In above formula, Δ θ, Δ φ respectively indicate the horizontal angular resolution and vertical angular resolution of radar, θi、φiRespectively
The level angle and vertical angle of scan line where indicating scanning element;
Step 2.2, b is usedr,cR row c column element in representing matrix calculates br,cIn point piDepth value pdepth i, meter
Calculation mode is as follows:
In above formula, xi,yiRespectively br,cCorresponding scanning element piSeat in the rectangular coordinate system relative to X-axis, Y-axis
Mark;
Step 2.3, calculating matrix element br,cCorresponding scanning element is the probability P (b of ground pointr,c), if probability is more than threshold
Value, then by br,cCorresponding scanning element is labeled as ground point;
Step 2.4, each element in Ergodic Matrices, ground point all in matrix is marked according to the method for step 2.3
And remove ground point from point set P, remaining non-ground points are denoted as point set Po。
Further, (the b of ground point probability P described in step 2.3r,c) calculating step are as follows:
Step 2.3.1, the adjacent element b of calculating matrix same rowr-1,cAnd br,cIn point between depth measurement difference Md
(br-1,c,br,c), calculation method is as follows:
Md(br-1,c,br,c)=| pdepth r,c-pdepth r-1,c| (6)
Step 2.3.2 estimates the adjacent member of matrix same row according to the distribution situation of radar point cloud data in the plane
Plain br-1,cAnd br,cIn point between depth difference Ed(br-1,c,br,c), circular is as follows:
Wherein, h indicates the mounting height of radar, and Δ φ indicates radar vertical angular resolution, φr-1And φrIt respectively indicates
The vertical angle of radar r-1 and the r articles scan line, γr-1Indicate element br-1,cRadial direction of the corresponding scanning element away from radar center
Distance value;
Step 2.3.3, then element br,cCorresponding scanning element piIt is the probability P (b of ground pointr,c) are as follows:
Wherein as probability P (br,c) when being greater than threshold value 0.8, then element br,cCorresponding scanning element piLabeled as ground point.
Further, space clustering is carried out to remaining non-ground points described in step 2, extracts the 3D region of interest of target
Domain, comprising:
Step 2.5.1 establishes first cluster C1, by non-ground points collection PoIn first scanning element p1It is divided into first
A cluster C1In;
Step 2.5.2, for point set PoIn other point pi∈Po, (i ≠ 1) calculates the cluster C nearest from itjIn scanning
It puts with its Euler apart from minimum value, it, will point p if minimum value is less than threshold value diIt is divided into cluster CjIn (j≤n), wherein n
Indicate current cluster numbers;Otherwise (n+1)th cluster C is re-createdn+1, and by piIt is divided into Cn+1In, until point set PoIn it is all
Scanning element is all divided into cluster;
Step 2.5.3 indicates cluster set with Γ, for each of cluster set Γ cluster Cj, utilize cluster CjInstitute
The spatial distribution for the scanning element for including calculates the minimum 3D axis alignment rectangular bounding box of the cluster, if the size of bounding box is big
In threshold size, then the cluster is marked into pseudo- target area, be otherwise labeled as candidate target region;
Step 2.5.4 retains all target 3D region of interest for marking the bounding box for being as extraction
Domain.
Further, the outer parameter of the coordinate of 3D laser radar and camera is demarcated, and according to calibrating parameters by mesh
Target 3D area-of-interest is mapped in corresponding camera image, extracts corresponding 2D area-of-interest in camera image, comprising:
Using gridiron pattern scaling board as target, characteristic point is marked on scaling board, while obtaining the point cloud data and phase of radar
Then the image data of machine calculates calibration ginseng in radar and magazine coordinate correspondence relationship according to the characteristic point on scaling board
Number, i.e., spin matrix and translation vector between radar fix system and camera coordinates system;
Finally the 3D area-of-interest of target is mapped in corresponding camera image according to calibrating parameters, extracts camera figure
Corresponding 2D axis alignment rectangular bounding box is as 2D area-of-interest as in.
Further, feature extraction is carried out to the 2D area-of-interest using depth convolutional network described in step 4,
To which the target in 2D area-of-interest is positioned and be identified, comprising:
The depth convolutional network uses VGG16, respectively by ' conv3 ' in model, ' conv4 ' and ' conv5 ' this three
The characteristic pattern that layer convolutional layer finally exports is normalized first, is then combined, so that final target signature has not
Same scale;1 × 1 convolution operation is carried out to the feature after combination, the feature vector obtained to the end is exported to the convolutional network most
The full articulamentum of two layers afterwards is to be positioned and be identified to the target in 2D area-of-interest.
Further, the method further include:
Step 5, it is optimized using result of the non-maxima suppression algorithm to step 4.
A kind of object detection system based on 3D laser radar and image data, including the acquisition of sequentially connected data with it is pre-
Processing module, 3D area-of-interest obtain module, 2D area-of-interest obtains module and positioning and identification module, in which:
The data acquisition obtains surrounding using the 3D laser radar and camera being installed on vehicle with preprocessing module
The 3D point cloud data and camera image of environment, and the 3D point cloud data are pre-processed;
The 3D area-of-interest obtains module and is used to filter out the ground point in 3D point cloud data, to remaining non-ground points
Space clustering is carried out, the 3D area-of-interest of target is extracted;
The 2D area-of-interest obtains module and is used to mark the outer parameter of the coordinate of 3D laser radar and camera
It is fixed, and the 3D area-of-interest of target is mapped in corresponding camera image according to calibrating parameters, it is right in camera image to extract
The 2D area-of-interest answered;
The positioning and identification module is used to carry out feature to the 2D area-of-interest using depth convolutional network
It extracts, so that the target in 2D area-of-interest is positioned and be identified.
The present invention has following technical characterstic compared with prior art:
1. the present invention extracts the candidate region 3D of target using the point cloud data of 3D laser radar, then according to radar and phase
Relationship is demarcated outside the coordinate of machine, the candidate region 3D in cloud is mapped in image space, and utilizes convolutional neural networks pair
Candidate region carries out feature extraction, target positioning and identification.When the method for the present invention can overcome single-sensor target detection
The shortcomings that precision is low, and environmental factor dependence is strong, poor robustness can satisfy unmanned vehicle in environment understanding to the essence of target detection
Degree, the requirement of real-time and adaptive capacity to environment.
2. the present invention is to obtain height in unmanned environment using the complementarity between 3D laser radar and camera data
Accurate object space and classification provide new approaches;By combining, the mutual supplement with each other's advantages of two sensors is realized, inspection is enhanced
The robustness of method of determining and calculating.
Detailed description of the invention
Fig. 1 is the flow chart of method of the invention;
(a) of Fig. 2 is the point cloud data of 3D laser radar acquisition, is (b) the corresponding camera image of point cloud data;
Fig. 3 is the geometrical principle figure of ground data reduction of the present invention;
(a) of Fig. 4 is the schematic diagram of the ground point cloud extracted;It (b) is to filter out showing for the non-ground points cloud after ground point cloud
It is intended to;
Fig. 5 is the schematic diagram of 3D area-of-interest;
Fig. 6 is that 3D area-of-interest is mapped to the 2D area-of-interest schematic diagram generated in camera image;
Fig. 7 is the convolutional network model that the present invention uses;
Fig. 8 is the actually detected result schematic diagram of the method for the present invention.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, once in conjunction with specific embodiments, such as Fig. 1
Shown, the method includes steps once:
Step 1, the 3D point cloud data and phase of ambient enviroment are obtained using the 3D laser radar and camera being installed on vehicle
Machine image, and the 3D point cloud data are pre-processed.
In an embodiment of the present invention, 3D laser radar (hereinafter referred to as radar) the model Velodyne HDL- of selection
32E, radar are mounted on the top of vehicle;In the present embodiment, vehicle is unmanned vehicle, mounting height 2.1m;It is travelled in unmanned vehicle
In the process, by radar scanning ambient enviroment to obtain 3D point cloud data, as shown in (a) of Fig. 2, to pass through radar scanning environment
The frame original point cloud data obtained;The radar by 32 single laser constitutions, 32 scan lines on a column, with
The frequency of 10Hz executes 360 ° of degree scannings, and the horizontal resolution of scanning angle is 0.16 °;The vertical angular resolution of radar is
1.33 °, cover -30.67 °~10.67 ° of pitch angle.
In an embodiment of the present invention, to obtain the model Basler acA640-100gc of the camera of image data
Colored CCD 658*492 camera.Radar and camera are by outer calibration progress unified coordinate system, and camera and radar are using identical
Frame per second, and synchronous acquisition, i.e., the frame camera image in image data correspond to the frame point cloud number in 3D point cloud data
According to.
Each frame point cloud data that radar scanning obtains is the point set P by scanning elementrIt constitutes, and point set PrIn it is each
A scanning element piIt is to be described by spherical coordinate:
Pr={ pi|γi,θi,φi,Ii}
Wherein γiIndicate radial distance of the scanning element to radar, θi,φiIndicate water of the scanning element relative to spheric coordinate system
Gentle vertical angle, IiIndicate radar reflection intensity, i indicates that point concentrates the number of scanning element.In single sweep operation (i.e. radar scanning
One week) in generate about 70,000 point, can describe targets and detecting distance around all is more than 70 meters of object, Fig. 2
(a) and (b) be original radar point cloud data and its corresponding camera image.For subsequent processing, need to calculate each sweep
Coordinate of the described point in rectangular coordinate system.
Since the data coverage that radar scanning generates is wide, data are big, if all carrying out processing is unable to satisfy unmanned vehicle
Requirement to algorithm real-time.Therefore it needs to retain the data of area-of-interest, and filters out the data of inactive area.In addition, by
There are some isolated noise spots in the data that radar is returned due to physical cause, if may influence candidate without filtering out
The precision of algorithm, can be by being filtered out noise spot based on a method for cloud density.
It includes that the conversion of coordinate system, noise spot filter out that 3D point cloud data shown in step 1, which carry out pretreatment, specific wrap
Include following steps:
Step 1.1, the point cloud data that radar obtains is transformed under rectangular coordinate system
By the point set P of point cloud datarIt is transformed under rectangular coordinate system, calculates point and concentrate each scanning element in rectangular coordinate system
Under coordinate, obtain each scanning element piMulti-parameter representation:
pi=(γi,θi,φi,Ii,xi,yi,zi)
The rectangular coordinate system is using the geometric center position of radar as coordinate origin, with the vertical axis direction of radar
For Z axis, using vehicle forward direction as Y-axis, and X-axis follows the right-hand rule by Z axis and Y-axis and determines.Conversion process is as follows:
Step 1.2, region filtering is carried out according to the rectangular co-ordinate, sets region of interest border, retained interested
Scanning element in region, it may be assumed that
Pf={ pi|-X<xi<X,Y<yi<Y,Z1<zi<Z2} (2)
That is, working as scanning element piCoordinate (xi,yi,zi) meet in region of interest border-X < xi<X,Y<yi<Y,Z1<zi<Z2
When interior, by scanning element piPoint set P is addedf, thus obtain the point set P of area-of-interest scanning elementf。
In the present embodiment, the X of the region of interest border of unmanned vehicle takes 15m, Y to take 50m, Z1Take -2.1, Z2Take 0.5m.
Step 1.3, noise spot filters out
For point set PfEach of scanning element pi, search in scanning element piNeighbor Points in radius R, if piIt is close
Adjoint point quantity, then will point p less than MiIt is labeled as noise spot and from point set PfMiddle removal.
Traverse point set Pf, find all noise spots and from point set PfMiddle removal obtains pretreated point set P.This implementation
In example, radius R takes 0.5m, M to take 3.
Step 2, the ground point in 3D point cloud data is filtered out, space clustering is carried out to remaining non-ground points, extracts target
3D area-of-interest.
In the point cloud data of radar, the scanning element in all targets is connected together by ground point, it is difficult to point
It opens, so needing first to filter out the ground point in scanning element, and filters out ground point to obtain non-ground points.
Step 2.1, point set P is mapped in multi-dimensional matrix, the line number of matrix is equal to the item number of radar scanning line, matrix
Columns is equal to the points that a scan line includes;Radar scanning line in the present embodiment is 32, the point that a scan line includes
It is 2250.Scanning element p in point set PiBe mapped in matrix where row r and column c calculation it is as follows:
R=(θi+180)/Δθ (3)
C=φi/Δφ (4)
In above formula, Δ θ, Δ φ respectively indicate the horizontal angular resolution and vertical angular resolution of radar, θi、φiRespectively
The level angle and vertical angle of scan line where indicating scanning element.
Step 2.2, b is usedr,cR row c column element in representing matrix calculates br,cIn point piDepth value pdepth i, meter
Calculation mode is as follows:
In above formula, xi,yiRespectively br,cCorresponding scanning element piSeat in the rectangular coordinate system relative to X-axis, Y-axis
Mark.
Step 2.3, calculating matrix element br,cCorresponding scanning element is the probability P (b of ground pointr,c), if probability is more than threshold
Value, then by br,cCorresponding scanning element is labeled as ground point;If Fig. 3 is the geometrical principle figure of ground data reduction of the present invention.
Ground point probability P (the br,c) calculating step are as follows:
Step 2.3.1, the adjacent element b of calculating matrix same rowr-1,cAnd br,cIn point between depth measurement difference Md
(br-1,c,br,c), calculation method is as follows:
Md(br-1,c,br,c)=| pdepth r,c-pdepth r-1,c| (6)
Step 2.3.2 estimates the adjacent member of matrix same row according to the distribution situation of radar point cloud data in the plane
Plain br-1,cAnd br,cIn point between depth difference Ed(br-1,c,br,c), circular is as follows:
Wherein, h indicates the mounting height of radar, and Δ φ indicates radar vertical angular resolution, φr-1And φrIt respectively indicates
The vertical angle of radar r-1 and the r articles scan line, γr-1Indicate element br-1,cRadial direction of the corresponding scanning element away from radar center
Distance value.
Step 2.3.3, then element br,cCorresponding scanning element piIt is the probability P (b of ground pointr,c) are as follows:
Wherein as probability P (br,c) when being greater than threshold value 0.8, then element br,cCorresponding scanning element piLabeled as ground point.Fig. 4's
(b) it is non-ground points cloud after filtering out ground point cloud.
Step 2.4, each element in Ergodic Matrices, ground point all in matrix is marked according to the method for step 2.3
(such as (a) of Fig. 4) and ground point is removed from point set P, remaining non-ground points are denoted as point set Po, as shown in (b) of Fig. 4.
Step 2.5, space clustering is carried out to non-ground points, to obtain the geometrical characteristic information of target, to obtain target
3D area-of-interest, specific steps include:
Step 2.5.1 establishes first cluster C1, by non-ground points collection PoIn first scanning element p1It is divided into first
A cluster C1In;
Step 2.5.2, for point set PoIn other point pi∈Po, (i ≠ 1) calculates the cluster C nearest from itjIn scanning
It puts with its Euler apart from minimum value, it, will point p if minimum value is less than threshold value diIt is divided into cluster CjIn (j≤n), wherein n
Indicate current cluster numbers;Otherwise (n+1)th cluster C is re-createdn+1, and by piIt is divided into Cn+1In, until point set PoIn it is all
Scanning element is all divided into cluster;
Step 2.5.3 indicates cluster set with Γ, for each of cluster set Γ cluster Cj, utilize cluster CjInstitute
The spatial distribution for the scanning point set for including calculates the minimum 3D axis alignment rectangular bounding box OBB (oriented of the cluster
Bounding box), if the size of bounding box is greater than threshold size, which is marked into pseudo- target area, is otherwise marked
For candidate target region;Threshold size length is 10m, width 5m in the present embodiment, is highly 3m.
Step 2.5.4 retains all target 3D region of interest for marking the bounding box for being as extraction
Domain, such as the 3D area-of-interest of Fig. 5 target finally retained.
Step 3, the outer parameter of the coordinate of 3D laser radar and camera is demarcated, and according to calibrating parameters by target
3D area-of-interest is mapped in corresponding camera image, and the 2D for extracting (3D area-of-interest) corresponding in camera image is interested
Region.
The specific method of this step is to mark characteristic point on scaling board using gridiron pattern scaling board as target, obtain simultaneously
The point cloud data of radar and the image data of camera, then according to the characteristic point on scaling board in radar and magazine coordinate pair
It should be related to and calculate calibrating parameters, i.e., the spin matrix and translation vector between radar fix system and camera coordinates system;
The 3D area-of-interest for the target that step 2 obtains finally is mapped to corresponding camera image according to calibrating parameters
In, it extracts corresponding 2D axis in camera image and is aligned rectangular bounding box as 2D area-of-interest;If Fig. 6 is will to put cloud to be mapped to
The 2D area-of-interest generated in camera image.
Step 4, feature extraction is carried out to the 2D area-of-interest using depth convolutional network, thus interested in 2D
Target in region is positioned and is identified.
Used in this programme depth convolutional neural networks in camera image 2D area-of-interest carry out feature extraction, with
And bounding box recurrence and identification are carried out to the target in region;The boundary for surrounding target position is accurately positioned by way of recurrence
Frame, and target is identified.
In order to improve this method to the detection accuracy of Small object, VGG16 convolutional network mould that the present invention as shown in Figure 7 uses
Type, respectively by ' conv3 ' in model, the characteristic pattern that ' conv4 ' and ' conv5 ' this three-layer coil lamination finally export uses L2 first
Method for normalizing is normalized, and is then combined, so that final target signature has different scale;To the spy after combination
Sign carries out 1 × 1 convolution operation, obtains feature vector to the end and exports to two layers last of full articulamentum of the convolutional network with right
Target in 2D area-of-interest positioned and identified, the result schematic diagram of the method for the present invention detection as shown in Figure 8.
This programme is with 80% pedestrian, bicycle and the vehicle sample in KITTI public target detection training set to the volume
Product network is trained, and is obtained for detecting vehicle, pedestrian and the model parameter of bicycle.In test in addition to using KITTI
20% pedestrian, bicycle and vehicle sample in target detection data set in remaining training set test, and also utilize me
The survey of detection pedestrian, bicycle and vehicle is carried out using the data that HDL-32E 3D laser radar and Balser camera acquire
Examination, obtains position, the classification of the 2D rectangular bounding box of each target, to realize the positioning and identification to target.
In order to advanced optimize as a result, this programme further include: step 5, using non-maxima suppression algorithm to the knot of step 4
Fruit optimizes, and obtains the range information of final target bezel locations, classification and target.
In order to reduce the pseudo- target of detection, detection accuracy is improved, is lower than 0.5 using non-maxima suppression algorithm removal probability
2D rectangular bounding box, obtain the range information of final target bezel locations and classification and target.
It is demonstrated experimentally that this method still can be detected effectively in real time in front of unmanned vehicle in different traffic scenes
Each target.
Claims (10)
1. a kind of object detection method based on 3D laser radar and image data, comprising the following steps:
Step 1, the 3D point cloud data and camera figure of ambient enviroment are obtained using the 3D laser radar and camera being installed on vehicle
Picture, and the 3D point cloud data are pre-processed;
Step 2, the ground point in 3D point cloud data is filtered out, space clustering is carried out to remaining non-ground points, extracts the 3D sense of target
Interest region;
Step 3, the outer parameter of the coordinate of 3D laser radar and camera is demarcated, and is felt the 3D of target according to calibrating parameters
Interest area maps extract corresponding 2D area-of-interest in camera image into corresponding camera image;
Step 4, feature extraction is carried out to the 2D area-of-interest using depth convolutional network, thus to 2D area-of-interest
In target positioned and identified.
2. the object detection method based on 3D laser radar and image data as described in claim 1, which is characterized in that step
The 3D point cloud data are pre-processed in 1, comprising:
Step 1.1, the point cloud data that radar obtains is transformed under rectangular coordinate system
By the point set P of point cloud datarIt is transformed under rectangular coordinate system, calculates point and concentrate each scanning element under rectangular coordinate system
Coordinate obtains each scanning element piMulti-parameter representation:
pi=(γi,θi,φi,Ii,xi,yi,zi)
Wherein γiIndicate radial distance of the scanning element to radar, θi,φiIndicate scanning element relative to spheric coordinate system level and
Vertical angle, IiIndicate radar reflection intensity, xi,yi,ziFor scanning element piCoordinate under rectangular coordinate system;
The rectangular coordinate system is using the geometric center position of radar as coordinate origin, is Z with the vertical axis direction of radar
Axis, using vehicle forward direction as Y-axis, and X-axis follows the right-hand rule by Z axis and Y-axis and determines;Conversion process is as follows:
Step 1.2, region filtering is carried out according to the rectangular co-ordinate, sets region of interest border, retain area-of-interest
In scanning element, it may be assumed that
Pf={ pi|-X<xi<X,Y<yi<Y,Z1<zi<Z2} (2)
As scanning element piCoordinate (xi,yi,zi) meet in region of interest border-X < xi<X,Y<yi<Y,Z1<zi<Z2It, will when interior
Scanning element piPoint set P is addedf, thus obtain the point set P of area-of-interest scanning elementf。
3. the object detection method based on 3D laser radar and image data as described in claim 1, which is characterized in that described
3D point cloud data pre-processed, further includes:
Step 1.3, noise spot filters out
For point set PfEach of scanning element pi, search in scanning element piNeighbor Points in radius R, if piNeighbour points
Amount, then will point p less than MiIt is labeled as noise spot and from point set PfMiddle removal;Traverse point set Pf, find all noise spots and from
Point set PfMiddle removal obtains pretreated point set P.
4. the object detection method based on 3D laser radar and image data as described in claim 1, which is characterized in that step
The ground point in 3D point cloud data is filtered out described in 2, comprising:
Step 2.1, point set P is mapped in multi-dimensional matrix, the line number of matrix is equal to the item number of radar scanning line, matrix column number
The points for including equal to a scan line;Scanning element p in point set PiBe mapped in matrix where row r and the column calculating side c
Formula is as follows:
R=(θi+180)/Δθ (3)
C=φi/Δφ (4)
In above formula, Δ θ, Δ φ respectively indicate the horizontal angular resolution and vertical angular resolution of radar, θi、φiIt respectively indicates
The level angle and vertical angle of scan line where scanning element;
Step 2.2, b is usedr,cR row c column element in representing matrix calculates br,cIn point piDepth value pdepth i, calculating side
Formula is as follows:
In above formula, xi,yiRespectively br,cCorresponding scanning element piCoordinate in the rectangular coordinate system relative to X-axis, Y-axis;
Step 2.3, calculating matrix element br,cCorresponding scanning element is the probability P (b of ground pointr,c), if probability is more than threshold value,
By br,cCorresponding scanning element is labeled as ground point;
Step 2.4, each element in Ergodic Matrices marks ground point all in matrix according to the method for step 2.3 and incites somebody to action
Ground point is removed from point set P, and remaining non-ground points are denoted as point set Po。
5. the object detection method based on 3D laser radar and image data as claimed in claim 4, which is characterized in that step
Ground point probability P (b described in 2.3r,c) calculating step are as follows:
Step 2.3.1, the adjacent element b of calculating matrix same rowr-1,cAnd br,cIn point between depth measurement difference Md
(br-1,c,br,c), calculation method is as follows:
Md(br-1,c,br,c)=| pdepth r,c-pdepth r-1,c| (6)
Step 2.3.2 estimates the adjacent element of matrix same row according to the distribution situation of radar point cloud data in the plane
br-1,cAnd br,cIn point between depth difference Ed(br-1,c,br,c), circular is as follows:
Wherein, h indicates the mounting height of radar, and Δ φ indicates radar vertical angular resolution, φr-1And φrRespectively indicate radar
The vertical angle of r-1 and the r articles scan line, γr-1Indicate element br-1,cRadial distance of the corresponding scanning element away from radar center
Value;
Step 2.3.3, then element br,cCorresponding scanning element piIt is the probability P (b of ground pointr,c) are as follows:
Wherein as probability P (br,c) when being greater than threshold value 0.8, then element br,cCorresponding scanning element piLabeled as ground point.
6. the object detection method based on 3D laser radar and image data as described in claim 1, which is characterized in that step
Space clustering is carried out to remaining non-ground points described in 2, extracts the 3D area-of-interest of target, comprising:
Step 2.5.1 establishes first cluster C1, by non-ground points collection PoIn first scanning element p1First is divided into gather
Class C1In;
Step 2.5.2, for point set PoIn other point pi∈Po, (i ≠ 1) calculates the cluster C nearest from itjIn scanning element with
Its Euler is apart from minimum value, will point p if minimum value is less than threshold value diIt is divided into cluster CjIn (j≤n), wherein n is indicated
Current cluster numbers;Otherwise (n+1)th cluster C is re-createdn+1, and by piIt is divided into Cn+1In, until point set PoIn all scanning
Point is all divided into cluster;
Step 2.5.3 indicates cluster set with Γ, for each of cluster set Γ cluster Cj, utilize cluster CjIncluded
Scanning element spatial distribution, calculate the cluster minimum 3D axis alignment rectangular bounding box, if the size of bounding box be greater than threshold
It is worth size, then the cluster is marked into pseudo- target area, be otherwise labeled as candidate target region;
Step 2.5.4 retains all target 3D area-of-interests for marking the bounding box for being as extraction.
7. the object detection method based on 3D laser radar and image data as described in claim 1, which is characterized in that step
The outer parameter of the coordinate of 3D laser radar and camera is demarcated described in 3, and according to calibrating parameters that the 3D sense of target is emerging
Interesting area maps extract corresponding 2D area-of-interest in camera image into corresponding camera image, comprising:
Using gridiron pattern scaling board as target, characteristic point is marked on scaling board, while obtaining the point cloud data and camera of radar
Then image data calculates calibrating parameters in radar and magazine coordinate correspondence relationship according to the characteristic point on scaling board,
Spin matrix and translation vector i.e. between radar fix system and camera coordinates system;
Finally the 3D area-of-interest of target is mapped in corresponding camera image according to calibrating parameters, is extracted in camera image
Corresponding 2D axis alignment rectangular bounding box is as 2D area-of-interest.
8. the object detection method based on 3D laser radar and image data as described in claim 1, which is characterized in that step
Feature extraction is carried out to the 2D area-of-interest using depth convolutional network described in 4, thus in 2D area-of-interest
Target positioned and identified, comprising:
The depth convolutional network uses VGG16, respectively by ' conv3 ' in model, ' conv4 ' and ' conv5 ' this three-layer coil
The characteristic pattern that lamination finally exports is normalized first, is then combined, so that final target signature has different rulers
Degree;The convolution operation that 1 × 1 is carried out to the feature after combination, the feature vector for obtaining to the end export last to the convolutional network
Two layers of full articulamentum is to be positioned and be identified to the target in 2D area-of-interest.
9. the object detection method based on 3D laser radar and image data as described in claim 1, which is characterized in that described
Method further include:
Step 5, it is optimized using result of the non-maxima suppression algorithm to step 4.
10. a kind of object detection system based on 3D laser radar and image data, which is characterized in that including sequentially connected number
Module is obtained with preprocessing module, 3D area-of-interest according to acquisition, 2D area-of-interest obtains module and positioning and identification mould
Block, in which:
The data acquisition obtains ambient enviroment using the 3D laser radar and camera being installed on vehicle with preprocessing module
3D point cloud data and camera image, and the 3D point cloud data are pre-processed;
The 3D area-of-interest obtains module and is used to filter out the ground point in 3D point cloud data, carries out to remaining non-ground points
Space clustering extracts the 3D area-of-interest of target;
The 2D area-of-interest obtains module and is used to demarcate the outer parameter of the coordinate of 3D laser radar and camera, and
The 3D area-of-interest of target is mapped in corresponding camera image according to calibrating parameters, extracts corresponding 2D in camera image
Area-of-interest;
The positioning and identification module is used to carry out feature extraction to the 2D area-of-interest using depth convolutional network,
To which the target in 2D area-of-interest is positioned and be identified.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810594692.5A CN109100741B (en) | 2018-06-11 | 2018-06-11 | Target detection method based on 3D laser radar and image data |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810594692.5A CN109100741B (en) | 2018-06-11 | 2018-06-11 | Target detection method based on 3D laser radar and image data |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109100741A true CN109100741A (en) | 2018-12-28 |
CN109100741B CN109100741B (en) | 2020-11-20 |
Family
ID=64796805
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810594692.5A Active CN109100741B (en) | 2018-06-11 | 2018-06-11 | Target detection method based on 3D laser radar and image data |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109100741B (en) |
Cited By (57)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109934228A (en) * | 2019-03-18 | 2019-06-25 | 上海盎维信息技术有限公司 | 3D point cloud processing method and processing device based on artificial intelligence |
CN109949372A (en) * | 2019-03-18 | 2019-06-28 | 北京智行者科技有限公司 | A kind of laser radar and vision combined calibrating method |
CN109948661A (en) * | 2019-02-27 | 2019-06-28 | 江苏大学 | A kind of 3D vehicle checking method based on Multi-sensor Fusion |
CN109978954A (en) * | 2019-01-30 | 2019-07-05 | 杭州飞步科技有限公司 | The method and apparatus of radar and camera combined calibrating based on cabinet |
CN109978955A (en) * | 2019-03-11 | 2019-07-05 | 武汉环宇智行科技有限公司 | A kind of efficient mask method for combining laser point cloud and image |
CN110008891A (en) * | 2019-03-29 | 2019-07-12 | 厦门金龙旅行车有限公司 | A kind of pedestrian detection localization method, device, cart-mounted computing device and storage medium |
CN110018470A (en) * | 2019-03-01 | 2019-07-16 | 北京纵目安驰智能科技有限公司 | Based on example mask method, model, terminal and the storage medium merged before multisensor |
CN110033457A (en) * | 2019-03-11 | 2019-07-19 | 北京理工大学 | A kind of target point cloud dividing method |
CN110080326A (en) * | 2019-04-29 | 2019-08-02 | 北京拓疆者智能科技有限公司 | A kind of discharge method, controller, excavator, electronic equipment and storage medium |
CN110146865A (en) * | 2019-05-31 | 2019-08-20 | 阿里巴巴集团控股有限公司 | Target identification method and device for radar image |
CN110161464A (en) * | 2019-06-14 | 2019-08-23 | 成都纳雷科技有限公司 | A kind of Radar Multi Target clustering method and device |
CN110276793A (en) * | 2019-06-05 | 2019-09-24 | 北京三快在线科技有限公司 | A kind of method and device for demarcating three-dimension object |
CN110363158A (en) * | 2019-07-17 | 2019-10-22 | 浙江大学 | A kind of millimetre-wave radar neural network based cooperates with object detection and recognition method with vision |
CN110361717A (en) * | 2019-07-31 | 2019-10-22 | 苏州玖物互通智能科技有限公司 | Laser radar-camera combined calibration target and combined calibration method |
CN110458780A (en) * | 2019-08-14 | 2019-11-15 | 上海眼控科技股份有限公司 | 3D point cloud data de-noising method, apparatus, computer equipment and readable storage medium storing program for executing |
CN110472553A (en) * | 2019-08-12 | 2019-11-19 | 北京易航远智科技有限公司 | Target tracking method, computing device and the medium of image and laser point cloud fusion |
CN110503093A (en) * | 2019-07-24 | 2019-11-26 | 中国航空无线电电子研究所 | Area-of-interest exacting method based on disparity map DBSCAN cluster |
CN110531340A (en) * | 2019-08-22 | 2019-12-03 | 吴文吉 | A kind of identifying processing method based on deep learning of laser radar point cloud data |
CN110632617A (en) * | 2019-09-29 | 2019-12-31 | 北京邮电大学 | Laser radar point cloud data processing method and device |
CN110909656A (en) * | 2019-11-18 | 2020-03-24 | 中电海康集团有限公司 | Pedestrian detection method and system with integration of radar and camera |
CN111047901A (en) * | 2019-11-05 | 2020-04-21 | 珠海格力电器股份有限公司 | Parking management method, parking management device, storage medium and computer equipment |
CN111079652A (en) * | 2019-12-18 | 2020-04-28 | 北京航空航天大学 | 3D target detection method based on point cloud data simple coding |
CN111077506A (en) * | 2019-12-12 | 2020-04-28 | 苏州智加科技有限公司 | Method, device and system for calibrating millimeter wave radar |
CN111179331A (en) * | 2019-12-31 | 2020-05-19 | 智车优行科技(上海)有限公司 | Depth estimation method, depth estimation device, electronic equipment and computer-readable storage medium |
CN111427032A (en) * | 2020-04-24 | 2020-07-17 | 森思泰克河北科技有限公司 | Room wall contour recognition method based on millimeter wave radar and terminal equipment |
WO2020168464A1 (en) * | 2019-02-19 | 2020-08-27 | SZ DJI Technology Co., Ltd. | Local sensing based autonomous navigation, and associated systems and methods |
CN111598770A (en) * | 2020-05-15 | 2020-08-28 | 弗徕威智能机器人科技(上海)有限公司 | Object detection method and device based on three-dimensional data and two-dimensional image |
CN111626288A (en) * | 2019-02-28 | 2020-09-04 | 深圳市速腾聚创科技有限公司 | Data processing method, data processing device, computer equipment and storage medium |
CN111638499A (en) * | 2020-05-08 | 2020-09-08 | 上海交通大学 | Camera-laser radar relative external reference calibration method based on laser radar reflection intensity point characteristics |
CN111736114A (en) * | 2020-08-21 | 2020-10-02 | 武汉煜炜光学科技有限公司 | Method for improving data transmission speed of laser radar and laser radar |
CN111950543A (en) * | 2019-05-14 | 2020-11-17 | 北京京东尚科信息技术有限公司 | Target detection method and device |
CN111964673A (en) * | 2020-08-25 | 2020-11-20 | 一汽解放汽车有限公司 | Unmanned vehicle positioning system |
CN112034432A (en) * | 2019-06-03 | 2020-12-04 | 华为技术有限公司 | Radar target clustering method and related device |
CN112219206A (en) * | 2019-07-25 | 2021-01-12 | 北京航迹科技有限公司 | System and method for determining pose |
CN112270694A (en) * | 2020-07-07 | 2021-01-26 | 中国人民解放军61540部队 | Method for detecting urban environment dynamic target based on laser radar scanning pattern |
CN112543877A (en) * | 2019-04-03 | 2021-03-23 | 华为技术有限公司 | Positioning method and positioning device |
WO2021052121A1 (en) * | 2019-09-20 | 2021-03-25 | 于毅欣 | Object identification method and apparatus based on laser radar and camera |
CN112614189A (en) * | 2020-12-09 | 2021-04-06 | 中国北方车辆研究所 | Combined calibration method based on camera and 3D laser radar |
WO2021097807A1 (en) * | 2019-11-22 | 2021-05-27 | 深圳市大疆创新科技有限公司 | Method and device for calibrating external parameters of detection device, and mobile platform |
CN112989877A (en) * | 2019-12-13 | 2021-06-18 | 阿里巴巴集团控股有限公司 | Method and device for labeling object in point cloud data |
CN113064179A (en) * | 2021-03-22 | 2021-07-02 | 上海商汤临港智能科技有限公司 | Point cloud data screening method and vehicle control method and device |
CN113095324A (en) * | 2021-04-28 | 2021-07-09 | 合肥工业大学 | Classification and distance measurement method and system for cone barrel |
CN113255560A (en) * | 2021-06-09 | 2021-08-13 | 深圳朗道智通科技有限公司 | Target detection system based on image and laser data under automatic driving scene |
WO2021166912A1 (en) * | 2020-02-18 | 2021-08-26 | 株式会社デンソー | Object detection device |
CN113359148A (en) * | 2020-02-20 | 2021-09-07 | 百度在线网络技术(北京)有限公司 | Laser radar point cloud data processing method, device, equipment and storage medium |
CN113436273A (en) * | 2021-06-28 | 2021-09-24 | 南京冲浪智行科技有限公司 | 3D scene calibration method, calibration device and calibration application thereof |
WO2021189375A1 (en) * | 2020-03-26 | 2021-09-30 | Baidu.Com Times Technology (Beijing) Co., Ltd. | A point cloud feature-based obstacle filter system |
CN113671458A (en) * | 2020-05-13 | 2021-11-19 | 华为技术有限公司 | Target object identification method and device |
CN113711273A (en) * | 2019-04-25 | 2021-11-26 | 三菱电机株式会社 | Motion amount estimation device, motion amount estimation method, and motion amount estimation program |
CN113761999A (en) * | 2020-09-07 | 2021-12-07 | 北京京东乾石科技有限公司 | Target detection method and device, electronic equipment and storage medium |
CN113838140A (en) * | 2021-08-16 | 2021-12-24 | 中国矿业大学(北京) | Monocular video pedestrian three-dimensional positioning method based on three-dimensional map assistance |
CN113947639A (en) * | 2021-10-27 | 2022-01-18 | 北京斯年智驾科技有限公司 | Self-adaptive online estimation calibration system and method based on multi-radar-point cloud line characteristics |
CN114820953A (en) * | 2022-06-29 | 2022-07-29 | 深圳市镭神智能系统有限公司 | Data processing method, device, equipment and storage medium |
WO2023028774A1 (en) * | 2021-08-30 | 2023-03-09 | 华为技术有限公司 | Lidar calibration method and apparatus, and storage medium |
CN116839499A (en) * | 2022-11-03 | 2023-10-03 | 上海点莘技术有限公司 | Large-visual-field micro-size 2D and 3D measurement calibration method |
CN117008122A (en) * | 2023-08-04 | 2023-11-07 | 江苏苏港智能装备产业创新中心有限公司 | Method and system for positioning surrounding objects of engineering mechanical equipment based on multi-radar fusion |
CN116839499B (en) * | 2022-11-03 | 2024-04-30 | 上海点莘技术有限公司 | Large-visual-field micro-size 2D and 3D measurement calibration method |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101581575A (en) * | 2009-06-19 | 2009-11-18 | 南昌航空大学 | Three-dimensional rebuilding method based on laser and camera data fusion |
CN101975951A (en) * | 2010-06-09 | 2011-02-16 | 北京理工大学 | Field environment barrier detection method fusing distance and image information |
CN102944224A (en) * | 2012-11-09 | 2013-02-27 | 大连理工大学 | Automatic environmental perception system for remotely piloted vehicle and work method for automatic environmental perception system |
CN103226833A (en) * | 2013-05-08 | 2013-07-31 | 清华大学 | Point cloud data partitioning method based on three-dimensional laser radar |
CN103455144A (en) * | 2013-08-22 | 2013-12-18 | 深圳先进技术研究院 | Vehicle-mounted man-machine interaction system and method |
CN104143194A (en) * | 2014-08-20 | 2014-11-12 | 清华大学 | Point cloud partition method and device |
CN106407947A (en) * | 2016-09-29 | 2017-02-15 | 百度在线网络技术(北京)有限公司 | Target object recognition method and device applied to unmanned vehicle |
CN106530380A (en) * | 2016-09-20 | 2017-03-22 | 长安大学 | Ground point cloud segmentation method based on three-dimensional laser radar |
CN107192994A (en) * | 2016-03-15 | 2017-09-22 | 山东理工大学 | Multi-line laser radar mass cloud data is quickly effectively extracted and vehicle, lane line characteristic recognition method |
-
2018
- 2018-06-11 CN CN201810594692.5A patent/CN109100741B/en active Active
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101581575A (en) * | 2009-06-19 | 2009-11-18 | 南昌航空大学 | Three-dimensional rebuilding method based on laser and camera data fusion |
CN101975951A (en) * | 2010-06-09 | 2011-02-16 | 北京理工大学 | Field environment barrier detection method fusing distance and image information |
CN102944224A (en) * | 2012-11-09 | 2013-02-27 | 大连理工大学 | Automatic environmental perception system for remotely piloted vehicle and work method for automatic environmental perception system |
CN103226833A (en) * | 2013-05-08 | 2013-07-31 | 清华大学 | Point cloud data partitioning method based on three-dimensional laser radar |
CN103455144A (en) * | 2013-08-22 | 2013-12-18 | 深圳先进技术研究院 | Vehicle-mounted man-machine interaction system and method |
CN104143194A (en) * | 2014-08-20 | 2014-11-12 | 清华大学 | Point cloud partition method and device |
CN107192994A (en) * | 2016-03-15 | 2017-09-22 | 山东理工大学 | Multi-line laser radar mass cloud data is quickly effectively extracted and vehicle, lane line characteristic recognition method |
CN106530380A (en) * | 2016-09-20 | 2017-03-22 | 长安大学 | Ground point cloud segmentation method based on three-dimensional laser radar |
CN106407947A (en) * | 2016-09-29 | 2017-02-15 | 百度在线网络技术(北京)有限公司 | Target object recognition method and device applied to unmanned vehicle |
Non-Patent Citations (1)
Title |
---|
孙朋朋 等: "基于3D激光雷达城市道路边界鲁棒检测算法", 《浙江大学学报(工学版)》 * |
Cited By (80)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109978954A (en) * | 2019-01-30 | 2019-07-05 | 杭州飞步科技有限公司 | The method and apparatus of radar and camera combined calibrating based on cabinet |
WO2020168464A1 (en) * | 2019-02-19 | 2020-08-27 | SZ DJI Technology Co., Ltd. | Local sensing based autonomous navigation, and associated systems and methods |
CN109948661A (en) * | 2019-02-27 | 2019-06-28 | 江苏大学 | A kind of 3D vehicle checking method based on Multi-sensor Fusion |
CN109948661B (en) * | 2019-02-27 | 2023-04-07 | 江苏大学 | 3D vehicle detection method based on multi-sensor fusion |
CN111626288B (en) * | 2019-02-28 | 2023-12-01 | 深圳市速腾聚创科技有限公司 | Data processing method, device, computer equipment and storage medium |
CN111626288A (en) * | 2019-02-28 | 2020-09-04 | 深圳市速腾聚创科技有限公司 | Data processing method, data processing device, computer equipment and storage medium |
CN110018470A (en) * | 2019-03-01 | 2019-07-16 | 北京纵目安驰智能科技有限公司 | Based on example mask method, model, terminal and the storage medium merged before multisensor |
CN110033457A (en) * | 2019-03-11 | 2019-07-19 | 北京理工大学 | A kind of target point cloud dividing method |
CN109978955A (en) * | 2019-03-11 | 2019-07-05 | 武汉环宇智行科技有限公司 | A kind of efficient mask method for combining laser point cloud and image |
CN109978955B (en) * | 2019-03-11 | 2021-03-19 | 武汉环宇智行科技有限公司 | Efficient marking method combining laser point cloud and image |
CN110033457B (en) * | 2019-03-11 | 2021-11-30 | 北京理工大学 | Target point cloud segmentation method |
CN109934228A (en) * | 2019-03-18 | 2019-06-25 | 上海盎维信息技术有限公司 | 3D point cloud processing method and processing device based on artificial intelligence |
CN109949372A (en) * | 2019-03-18 | 2019-06-28 | 北京智行者科技有限公司 | A kind of laser radar and vision combined calibrating method |
CN109934228B (en) * | 2019-03-18 | 2023-02-10 | 上海盎维信息技术有限公司 | 3D point cloud processing method and device based on artificial intelligence |
CN110008891A (en) * | 2019-03-29 | 2019-07-12 | 厦门金龙旅行车有限公司 | A kind of pedestrian detection localization method, device, cart-mounted computing device and storage medium |
CN112543877B (en) * | 2019-04-03 | 2022-01-11 | 华为技术有限公司 | Positioning method and positioning device |
CN112543877A (en) * | 2019-04-03 | 2021-03-23 | 华为技术有限公司 | Positioning method and positioning device |
CN113711273A (en) * | 2019-04-25 | 2021-11-26 | 三菱电机株式会社 | Motion amount estimation device, motion amount estimation method, and motion amount estimation program |
CN110080326A (en) * | 2019-04-29 | 2019-08-02 | 北京拓疆者智能科技有限公司 | A kind of discharge method, controller, excavator, electronic equipment and storage medium |
CN110080326B (en) * | 2019-04-29 | 2021-11-16 | 北京拓疆者智能科技有限公司 | Unloading method, controller, excavator, electronic equipment and storage medium |
CN111950543A (en) * | 2019-05-14 | 2020-11-17 | 北京京东尚科信息技术有限公司 | Target detection method and device |
CN110146865A (en) * | 2019-05-31 | 2019-08-20 | 阿里巴巴集团控股有限公司 | Target identification method and device for radar image |
CN112034432A (en) * | 2019-06-03 | 2020-12-04 | 华为技术有限公司 | Radar target clustering method and related device |
CN110276793A (en) * | 2019-06-05 | 2019-09-24 | 北京三快在线科技有限公司 | A kind of method and device for demarcating three-dimension object |
CN110161464B (en) * | 2019-06-14 | 2023-03-10 | 成都纳雷科技有限公司 | Radar multi-target clustering method and device |
CN110161464A (en) * | 2019-06-14 | 2019-08-23 | 成都纳雷科技有限公司 | A kind of Radar Multi Target clustering method and device |
CN110363158A (en) * | 2019-07-17 | 2019-10-22 | 浙江大学 | A kind of millimetre-wave radar neural network based cooperates with object detection and recognition method with vision |
CN110363158B (en) * | 2019-07-17 | 2021-05-25 | 浙江大学 | Millimeter wave radar and visual cooperative target detection and identification method based on neural network |
CN110503093A (en) * | 2019-07-24 | 2019-11-26 | 中国航空无线电电子研究所 | Area-of-interest exacting method based on disparity map DBSCAN cluster |
CN110503093B (en) * | 2019-07-24 | 2022-11-04 | 中国航空无线电电子研究所 | Region-of-interest extraction method based on disparity map DBSCAN clustering |
CN112219206A (en) * | 2019-07-25 | 2021-01-12 | 北京航迹科技有限公司 | System and method for determining pose |
WO2021012245A1 (en) * | 2019-07-25 | 2021-01-28 | Beijing Voyager Technology Co., Ltd. | Systems and methods for pose determination |
CN110361717A (en) * | 2019-07-31 | 2019-10-22 | 苏州玖物互通智能科技有限公司 | Laser radar-camera combined calibration target and combined calibration method |
CN110472553B (en) * | 2019-08-12 | 2022-03-11 | 北京易航远智科技有限公司 | Target tracking method, computing device and medium for fusion of image and laser point cloud |
CN110472553A (en) * | 2019-08-12 | 2019-11-19 | 北京易航远智科技有限公司 | Target tracking method, computing device and the medium of image and laser point cloud fusion |
CN110458780A (en) * | 2019-08-14 | 2019-11-15 | 上海眼控科技股份有限公司 | 3D point cloud data de-noising method, apparatus, computer equipment and readable storage medium storing program for executing |
CN110531340B (en) * | 2019-08-22 | 2023-01-13 | 吴文吉 | Identification processing method of laser radar point cloud data based on deep learning |
CN110531340A (en) * | 2019-08-22 | 2019-12-03 | 吴文吉 | A kind of identifying processing method based on deep learning of laser radar point cloud data |
WO2021052121A1 (en) * | 2019-09-20 | 2021-03-25 | 于毅欣 | Object identification method and apparatus based on laser radar and camera |
CN110632617B (en) * | 2019-09-29 | 2021-11-02 | 北京邮电大学 | Laser radar point cloud data processing method and device |
CN110632617A (en) * | 2019-09-29 | 2019-12-31 | 北京邮电大学 | Laser radar point cloud data processing method and device |
CN111047901A (en) * | 2019-11-05 | 2020-04-21 | 珠海格力电器股份有限公司 | Parking management method, parking management device, storage medium and computer equipment |
CN110909656B (en) * | 2019-11-18 | 2023-10-13 | 中电海康集团有限公司 | Pedestrian detection method and system integrating radar and camera |
CN110909656A (en) * | 2019-11-18 | 2020-03-24 | 中电海康集团有限公司 | Pedestrian detection method and system with integration of radar and camera |
WO2021097807A1 (en) * | 2019-11-22 | 2021-05-27 | 深圳市大疆创新科技有限公司 | Method and device for calibrating external parameters of detection device, and mobile platform |
CN111077506A (en) * | 2019-12-12 | 2020-04-28 | 苏州智加科技有限公司 | Method, device and system for calibrating millimeter wave radar |
CN112989877A (en) * | 2019-12-13 | 2021-06-18 | 阿里巴巴集团控股有限公司 | Method and device for labeling object in point cloud data |
CN111079652B (en) * | 2019-12-18 | 2022-05-13 | 北京航空航天大学 | 3D target detection method based on point cloud data simple coding |
CN111079652A (en) * | 2019-12-18 | 2020-04-28 | 北京航空航天大学 | 3D target detection method based on point cloud data simple coding |
CN111179331B (en) * | 2019-12-31 | 2023-09-08 | 智车优行科技(上海)有限公司 | Depth estimation method, depth estimation device, electronic equipment and computer readable storage medium |
CN111179331A (en) * | 2019-12-31 | 2020-05-19 | 智车优行科技(上海)有限公司 | Depth estimation method, depth estimation device, electronic equipment and computer-readable storage medium |
WO2021166912A1 (en) * | 2020-02-18 | 2021-08-26 | 株式会社デンソー | Object detection device |
CN113359148A (en) * | 2020-02-20 | 2021-09-07 | 百度在线网络技术(北京)有限公司 | Laser radar point cloud data processing method, device, equipment and storage medium |
US11609333B2 (en) | 2020-03-26 | 2023-03-21 | Baidu Usa Llc | Point cloud feature-based obstacle filter system |
WO2021189375A1 (en) * | 2020-03-26 | 2021-09-30 | Baidu.Com Times Technology (Beijing) Co., Ltd. | A point cloud feature-based obstacle filter system |
CN111427032A (en) * | 2020-04-24 | 2020-07-17 | 森思泰克河北科技有限公司 | Room wall contour recognition method based on millimeter wave radar and terminal equipment |
CN111638499B (en) * | 2020-05-08 | 2024-04-09 | 上海交通大学 | Camera-laser radar relative external parameter calibration method based on laser radar reflection intensity point characteristics |
CN111638499A (en) * | 2020-05-08 | 2020-09-08 | 上海交通大学 | Camera-laser radar relative external reference calibration method based on laser radar reflection intensity point characteristics |
CN113671458A (en) * | 2020-05-13 | 2021-11-19 | 华为技术有限公司 | Target object identification method and device |
CN111598770B (en) * | 2020-05-15 | 2023-09-19 | 汇智机器人科技(深圳)有限公司 | Object detection method and device based on three-dimensional data and two-dimensional image |
CN111598770A (en) * | 2020-05-15 | 2020-08-28 | 弗徕威智能机器人科技(上海)有限公司 | Object detection method and device based on three-dimensional data and two-dimensional image |
CN112270694A (en) * | 2020-07-07 | 2021-01-26 | 中国人民解放军61540部队 | Method for detecting urban environment dynamic target based on laser radar scanning pattern |
CN111736114A (en) * | 2020-08-21 | 2020-10-02 | 武汉煜炜光学科技有限公司 | Method for improving data transmission speed of laser radar and laser radar |
CN111964673A (en) * | 2020-08-25 | 2020-11-20 | 一汽解放汽车有限公司 | Unmanned vehicle positioning system |
CN113761999B (en) * | 2020-09-07 | 2024-03-05 | 北京京东乾石科技有限公司 | Target detection method and device, electronic equipment and storage medium |
CN113761999A (en) * | 2020-09-07 | 2021-12-07 | 北京京东乾石科技有限公司 | Target detection method and device, electronic equipment and storage medium |
CN112614189A (en) * | 2020-12-09 | 2021-04-06 | 中国北方车辆研究所 | Combined calibration method based on camera and 3D laser radar |
CN113064179A (en) * | 2021-03-22 | 2021-07-02 | 上海商汤临港智能科技有限公司 | Point cloud data screening method and vehicle control method and device |
CN113095324A (en) * | 2021-04-28 | 2021-07-09 | 合肥工业大学 | Classification and distance measurement method and system for cone barrel |
CN113255560A (en) * | 2021-06-09 | 2021-08-13 | 深圳朗道智通科技有限公司 | Target detection system based on image and laser data under automatic driving scene |
CN113436273A (en) * | 2021-06-28 | 2021-09-24 | 南京冲浪智行科技有限公司 | 3D scene calibration method, calibration device and calibration application thereof |
CN113838140B (en) * | 2021-08-16 | 2023-07-18 | 中国矿业大学(北京) | Monocular video pedestrian three-dimensional positioning method based on three-dimensional map assistance |
CN113838140A (en) * | 2021-08-16 | 2021-12-24 | 中国矿业大学(北京) | Monocular video pedestrian three-dimensional positioning method based on three-dimensional map assistance |
WO2023028774A1 (en) * | 2021-08-30 | 2023-03-09 | 华为技术有限公司 | Lidar calibration method and apparatus, and storage medium |
CN113947639B (en) * | 2021-10-27 | 2023-08-18 | 北京斯年智驾科技有限公司 | Self-adaptive online estimation calibration system and method based on multi-radar point cloud line characteristics |
CN113947639A (en) * | 2021-10-27 | 2022-01-18 | 北京斯年智驾科技有限公司 | Self-adaptive online estimation calibration system and method based on multi-radar-point cloud line characteristics |
CN114820953A (en) * | 2022-06-29 | 2022-07-29 | 深圳市镭神智能系统有限公司 | Data processing method, device, equipment and storage medium |
CN116839499A (en) * | 2022-11-03 | 2023-10-03 | 上海点莘技术有限公司 | Large-visual-field micro-size 2D and 3D measurement calibration method |
CN116839499B (en) * | 2022-11-03 | 2024-04-30 | 上海点莘技术有限公司 | Large-visual-field micro-size 2D and 3D measurement calibration method |
CN117008122A (en) * | 2023-08-04 | 2023-11-07 | 江苏苏港智能装备产业创新中心有限公司 | Method and system for positioning surrounding objects of engineering mechanical equipment based on multi-radar fusion |
Also Published As
Publication number | Publication date |
---|---|
CN109100741B (en) | 2020-11-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109100741A (en) | A kind of object detection method based on 3D laser radar and image data | |
CN107341453B (en) | Lane line extraction method and device | |
CN107067415B (en) | A kind of object localization method based on images match | |
CN104574393B (en) | A kind of three-dimensional pavement crack pattern picture generates system and method | |
CN109444911A (en) | A kind of unmanned boat waterborne target detection identification and the localization method of monocular camera and laser radar information fusion | |
CN110415342A (en) | A kind of three-dimensional point cloud reconstructing device and method based on more merge sensors | |
CN102032875B (en) | Image-processing-based cable sheath thickness measuring method | |
CN106969706A (en) | Workpiece sensing and three-dimension measuring system and detection method based on binocular stereo vision | |
CN109472831A (en) | Obstacle recognition range-measurement system and method towards road roller work progress | |
CN109711288A (en) | Remote sensing ship detecting method based on feature pyramid and distance restraint FCN | |
CN102609701B (en) | Remote sensing detection method based on optimal scale for high-resolution SAR (synthetic aperture radar) | |
CN110297232A (en) | Monocular distance measuring method, device and electronic equipment based on computer vision | |
CN108007388A (en) | A kind of turntable angle high precision online measuring method based on machine vision | |
CN105335973A (en) | Visual processing method for strip steel processing production line | |
CN104764407B (en) | A kind of fine measuring method of thickness of cable sheath | |
CN113470090A (en) | Multi-solid-state laser radar external reference calibration method based on SIFT-SHOT characteristics | |
CN114089329A (en) | Target detection method based on fusion of long and short focus cameras and millimeter wave radar | |
CN109816051B (en) | Hazardous chemical cargo feature point matching method and system | |
CN108573280B (en) | Method for unmanned ship to autonomously pass through bridge | |
Palenichka et al. | Multi-scale segmentation of forest areas and tree detection in LiDAR images by the attentive vision method | |
CN110298271A (en) | Seawater method for detecting area based on critical point detection network and space constraint mixed model | |
CN116844147A (en) | Pointer instrument identification and abnormal alarm method based on deep learning | |
CN107767366B (en) | A kind of transmission line of electricity approximating method and device | |
CN107345814A (en) | A kind of mobile robot visual alignment system and localization method | |
Liu et al. | Outdoor camera calibration method for a GPS & camera based surveillance 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 | ||
TR01 | Transfer of patent right | ||
TR01 | Transfer of patent right |
Effective date of registration: 20231228 Address after: 100193 area a, building 12, Zhongguancun Software Park, Haidian District, Beijing Patentee after: BEIJING WANJI TECHNOLOGY Co.,Ltd. Address before: 710064 No. 126 central section of South Ring Road, Yanta District, Xi'an, Shaanxi Patentee before: CHANG'AN University |