CN109035305A - Indoor human body detection and tracking in the case of a kind of low visual angle based on RGB-D - Google Patents
Indoor human body detection and tracking in the case of a kind of low visual angle based on RGB-D Download PDFInfo
- Publication number
- CN109035305A CN109035305A CN201810908661.2A CN201810908661A CN109035305A CN 109035305 A CN109035305 A CN 109035305A CN 201810908661 A CN201810908661 A CN 201810908661A CN 109035305 A CN109035305 A CN 109035305A
- Authority
- CN
- China
- Prior art keywords
- cluster
- point
- tracking
- point cloud
- human body
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
- G06T7/251—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments involving models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10028—Range image; Depth image; 3D point clouds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30196—Human being; Person
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates in the case of a kind of low visual angle based on RGB-D indoor human body detection and tracking, belong to Human Detection field, specifically include following steps, point cloud acquisition is carried out using Asus Xtion Pro, obtain 3D point cloud, then noise reduction and down-sampled processing are carried out to 3D point cloud, and ground is detected and is removed, 3D cluster is carried out followed by the Euclidean distance between two o'clock, by the HOG feature for calculating each cluster, and it is fed to the soft classifier of SVM binary system trained in advance, being classified as people with high HOG characteristic, to achieve the purpose that human testing, the tracking of human body is finally realized using the colour consistency during data coherency and two likelihood probabilities constituted apart from consistency.Precision of the present invention is high, the indoor human body detection and tracking being widely used in the case of low visual angle.
Description
Technical field
The present invention relates in the case of a kind of low visual angle based on RGB-D indoor human body detection and tracking, belong to human body
Detection technique field.
Background technique
Human testing and tracking are the key that mobile robot executes task in environment indoors.Mobile robot is come
It says, it must be able to difference human body and other barriers, to adjust track according to its task.For example service robot must be in spy
Help is provided to specific people in fixed environment.
There are certain methods to realize human testing using RGB-D depth camera or radar range finding by simple at present
With tracking.With the appearance of RGB-D depth camera, for example, Microsoft Kinect or Asus Xtion Pro, can be with 30
The image of 640 × 480 pixel of frame/s rate capture, simultaneously because the advantages that its energy consumption is small, was widely used in recent years
The 3D of robot is perceived and the fields such as indoor positioning and identification.
There are some more successful human testing algorithms at present, however these algorithms are all based on the entire body of people,
The especially visible situation in head.In some cases, as in the case that robot is especially close far from object, the major part of object
It can be other than sensor sensing range.Meanwhile when RGB-D sensor is mounted on small scale robot, such as Turtlebot2.It
Sight be in close proximity to ground, therefore can only observe the lower part of the body of people or only leg is visible.At low visual angle
Human testing with tracking be it is very difficult, mainly some main features are all lost, and are distinguished without apparent feature
Human body and other objects, such as can not be distinguished the leg, chair and human body of desk using corresponding feature.When RGB-D pacifies
When on Turtlebot2, if people is very close from robot, (distance < 100cm), in most cases, from foot
To waist all as it can be seen that be based on this, we have proposed in the case of a kind of low visual angle human testing and track algorithm.
Summary of the invention
To solve technical problem of the existing technology, the present invention provides in the case of a kind of low visual angle based on RGB-D
Indoor human body detection and tracking, cluster the object on ground using in limited height, in conjunction with HOG feature
People or object are distinguished with SVM classifier, then the joint likelihood probability using colour consistency and apart from consistency composition is to human body
It is tracked, accuracy rate is higher.
To achieve the above object, the technical scheme adopted by the invention is as follows a kind of room in the case of low visual angle based on RGB-D
Interior human testing and tracking, include the following steps,
The acquisition and pretreatment of step a. data
Point cloud acquisition is carried out using Asus Xtion Pro, obtains intensive 3D point cloud, by straight-through filter to intensive 3D point
Cloud carries out noise reduction process, and carries out down-sampled processing to the 3D point cloud after noise reduction by three-dimensional voxel lattice filter；
It step b. ground detection and filters out
Pass through the ground in the 3D initial point cloud of step a processing using the least square method detection based on RANSAC, for
For non-first frame data, the ground parameter that previous frame detects can be used as the initial parameter of next frame, according to initially giving
Fixed ground parameter and preset initial distance threshold value carry out ground detection to current frame data using RANSAC method, and
Index of each three-dimensional point in 3D point cloud according to ground, it is filtered out from 3D point cloud；It, can basis for the first frame data
The determining relativeness with ground of installation site of the RGB-D sensor in robot, thus the initial parameter on given ground；
Step c. cluster
The point cloud data within 130cm from the ground is obtained in the 3D point cloud for removing ground, then using between two o'clock
Euclidean distance carries out 3D cluster, and when the Euclidean distance between each pair of point is less than predefined distance threshold, this two o'clock is defined as
Belong to same class；Wherein, cluster needs to set two initial thresholds: belonging to the most narrow spacing between the two o'clock of the same cluster
From the minimal point required with one cluster of generation；
Step d.HOG+SVM classification
RGB image will be projected to through the point cloud in 3D cluster back boundary frame, using the image module of acquisition, calculates its HOG
Then obtained HOG describer is sent to preparatory trained SVM classifier, calculates the HOG confidence of each cluster by describer
Degree；When the HOG confidence level of calculating is higher than the threshold value of setting, then determine that the cluster for people, is not otherwise then；
Step e. tracking
Obtained human body is clustered into the input as tracking module, i.e., as the object to be tracked in next step, is then utilized
The human body cluster detected in each frame is matched with existing tracking object, specific using based on apart from consistency and color
The method that consistency is combined calculates the maximum likelihood probability between currently detected human body and known tracking object.
Preferably, in the step c, when being clustered, 1) the Kd tree for creating 3D point cloud first, as extracting later a little
Searching method used when cloud；2) the empty list C of setting cluster, and the empty queue of point cloud；3) for each of cloud
Point p, using preset distance threshold as radius, searches for all neighbor points in the ball centered on it, the neighbour obtained for search
Near point, first checks for whether it has been added other clusters, if it is not, queue Q is added in the Neighbor Points；4) as queue Q
In point it is all processed after, queue Q is added in cluster C；5) after point all in initial point cloud is all processed,
End of clustering.
Preferably, it in the step c, after the completion of cluster, needs further to locate cluster and deficient clustering problem excessively
Reason；
The processing for crossing cluster is as follows: the cluster C obtained for eachi, its central point is calculated first in XZ plane
Project piIf piWith cluster CjCentral projection point pjDistance be less than setting threshold value when, then it is assumed that cluster CiAnd CjBelong to
The same cluster, and then merged；
The processing for owing cluster is as follows: for each cluster, their geological information calculated, specifically includes width, depth,
Elevation information；If the geological information of some clusters is much larger than the threshold value set, further divided using colouring information, i.e., will
Point with same color is classified as same class；Those are put with very little cluster, is directly given up.
Preferably, in the step e,
Apart from conformance definition are as follows: a given human testing clusters Ci, by handling global arest neighbors data to find
Immediate tracking object TjIf their distance is less than threshold value, the cluster of detection is regarded as having connection with the cluster of tracking
System, then for tracking object TjEach of point pi,j, detection target C is found using Octree methodiEach of point pj,i,
Calculate the distance between its, point pi,jWith point pj,iThe distance between consistency definition of probability it is as follows:
Wherein α is weight vectors；
Colour consistency is defined as: cluster C when comparing current detectioniWith tracking object TjColouring information when, pass through calculating
Find nearest a pair of point < pj,i,pi,j> between colour consistency, colour consistency can be in RGB, HSV or other color spaces
Middle calculating；By taking HSV space as an example, point pi,jAnd pj,iBetween colour consistency definition of probability it is as follows:
Wherein ci,jAnd cj,iRespectively indicate pi,jAnd pj,iHSV information, β indicate weight；
pi,jAnd pj,iBetween joint consistency definition of probability are as follows:
L(pi,j,pj,i)=Ld(pi,j,pj,i)Lc(pi,j,pj,i)；
For each tracking object TjC is clustered with detectioniMaximum combined likelihood probability is defined as:
If L (j, i) is higher than the threshold value of setting, show currently to cluster CiWith tracking object TjThe same person, it is on the contrary then
It is not, if do not found and CiThe tracking object being associated then creates new tracking object.
Compared with prior art, the present invention has following technical effect that present invention is generally directed to when RGB-D sensor is liftoff
Face is lower, i.e., in the case where low visual angle, or when detection object is closer from sensor, only visible feelings of the human body lower part of the body
Condition proposes a kind of human testing algorithm based on human body lower half as main feature.The algorithm can effectively improve human body
The accuracy of detection is detected and is removed to the ground in scene first based on the common sense that human body moves on the ground,
The object on ground is clustered in limited height, the HOG feature by calculating them is then fed to training in advance
The soft classifier of SVM binary system, being classified as people with high HOG the value of the confidence are on the contrary then be classified as other.Then detection is obtained
Input of the human body result as tracking module, using colour consistency and the joint likelihood probability constituted apart from consistency,
It finds and currently detected human body clusters immediate tracking object, when maximum likelihood probability is greater than given threshold, it is believed that
Human body cluster and the tracking object that current detection obtains are same people.
Detailed description of the invention
Fig. 1 is flow chart of the invention.
Fig. 2 is that the present invention collects intensive 3D point cloud datagram.
Fig. 3 is that intensive 3D point cloud data pass through pretreated point cloud chart in the present invention.
Fig. 4 is the identification situation schematic diagram of the present invention under various circumstances.
Specific embodiment
In order to which technical problems, technical solutions and advantages to be solved are more clearly understood, tie below
Accompanying drawings and embodiments are closed, the present invention will be described in further detail.It should be appreciated that specific embodiment described herein is only
To explain the present invention, it is not intended to limit the present invention.
As shown in Figure 1, indoor human body detection and tracking in the case of a kind of low visual angle based on RGB-D, including it is following
Step,
The acquisition and processing of step a. data
As shown in Fig. 2, carrying out point cloud acquisition using Asus Xtion Pro, intensive 3D point cloud is obtained.Due to by different
The fact that discretization effect and camera in the influence of error source, especially depth measurement are calibrated in a certain range, adopts
There is very big noise in the Initial R GB-D point cloud collected, while the RGB-D point cloud that each frame obtains includes 307200 points,
Dimension 640 × 480 corresponding to its depth image.Therefore in order to improve the processing speed and precision of data, first with straight-through
Filter removes uninterested point, such as according to the parameter of Xtion pro, filters out the point for being greater than 5m on the direction z.Then three are used
It is down-sampled to cloud progress to tie up voxel grid filter, the algorithm is using the center of gravity of all the points in voxel come in approximate display voxel
Other points reduce operand to realize the down-sampled of point cloud data.The side length of voxel (i.e. three-dimensional cube) is dimensioned to
0.06m.The point cloud data that RGB-D sensor obtains can be compressed to an order of magnitude by down-sampled, to substantially reduce
The processing time of later period point cloud, at the same it is down-sampled after data density throughout it is identical.Treated point cloud data such as Fig. 3 institute
Show.
It step b. ground detection and filters out
This step is walked on the ground based on people it is assumed that therefore carrying out ground in our point clouds first after the filtering
Detection, and remove it.When being fixedly mounted in mobile robot due to RGB-D, it is approximate with the relative position on ground
It is known that being based on this point, initial ground parameter is set when handling first frame point cloud, is utilized on this basis based on RANSAC most
Ground in small square law detection 3D point cloud, obtains updated ground parameter, and using updated ground parameter as next
The initial parameter on the ground of frame.It in this way can be with parameter of the real-time update ground in robot coordinate system.Theoretically, ground is
Proper plane, relative to robot, or in robot coordinate system, its parameter is fixed, but due to machine
The vibration of camera and the small skew of actual ground in people's movement, cause ground parameter in different moments, in robot coordinate
Parameter is different in system.
Step c. cluster
The point cloud data within 130cm from the ground is obtained in the 3D point cloud for removing ground, then using between two o'clock
Euclidean distance carries out 3D cluster, and when the Euclidean distance between each pair of point is less than predefined distance threshold, this two o'clock is defined as
Belong to same class；Wherein, cluster needs to set two initial thresholds: belonging to the most narrow spacing between the two o'clock of the same cluster
From the minimal point required with one cluster of generation.
Once ground is detected, so that it may remove the three-dimensional point for belonging to ground, therefore object and ground on ground
It no longer contacts, shows as the object being not connected with one by one.Due to our human body detecting method use the lower part of human body as
Feature description, therefore the point cloud data to be analyzed is limited for the point cloud data within 130cm from the ground, then using between two o'clock
Euclidean distance carry out 3D cluster.
When being clustered, 1) the Kd tree for creating 3D point cloud first, as extract later put cloud when searching method used；
2) the empty list C of setting cluster, and the empty queue of point cloud；3) for each of cloud point p, centered on it, with pre-
If distance threshold be radius, search for all neighbor points in the ball, for the obtained neighbor point of search, whether first check for it
Other clusters are added, if it is not, queue Q is added in the Neighbor Points；4) when the point in queue Q is all processed it
Afterwards, queue Q is added in cluster C；5) after point all in initial point cloud is all processed, end of clustering.
After the completion of Euclidean cluster, there is a problem of in actual operation two kinds it is typical, that is, cross and cluster and owe to cluster: (1) mistake
Cluster, i.e., the point cloud that should belong to the same object is divided into multiple clusters, this is primarily due to noise, blocks and lacks with some
Caused by the depth data of mistake.(2) cluster is owed, as its name suggests, two or more different objects should then be belonged to by owing cluster
Point cloud is classified as the same cluster by inappropriate.Such as when people and background, cabinet or desk lean on it is close when, people usually with back
Sight spot, desk are classified as same class.In our experiment, due to only considering the point on ground within 130 centimetres,
In cluster operation, the case where point of two different people is merged together, is not common.
In order to solve both of these problems, after carrying out initial European cluster, obtained cluster need to be further processed.
For crossing clustering problem, after the completion of cluster, the cluster C that is obtained for eachi, its central point is calculated first in XZ
Projection p in planeiIf piWith cluster CjCentral projection point pjDistance be less than setting threshold value when, then it is assumed that cluster Ci
And CjBelong to the same cluster, and then is merged.
For owing cluster, when a people and background are divided into one kind, separated using colouring information.This is the trousers based on people
Sub-color it is different with background color it is assumed that the color of such as jeans be it is blue, wall is white.Specific algorithm is as follows:
For each cluster, their geological information is calculated, width, depth, elevation information are specifically included；If some clusters is several
What information is much larger than the threshold value set, is further divided using colouring information, i.e., is classified as the point with same color same
A class；Very little cluster is put similarly for those in this operation, is directly given up.
Step d.HOG+SVM classification
RGB image will be projected to through the point cloud in 3D cluster back boundary frame, using the image module of acquisition, calculates its HOG
Then obtained HOG describer is sent to preparatory trained SVM classifier, calculates the HOG confidence of each cluster by describer
Degree；When the HOG confidence level of calculating is higher than the threshold value of setting, then determines that the cluster is people, be not otherwise；
Step e. tracking
Obtained human body is clustered into the input as tracking module, i.e., as the object to be tracked in next step, is then utilized
Human body cluster is matched with existing tracking object, specific using based on apart from the side that consistency and colour consistency are combined
Method calculates the maximum likelihood probability between currently detected human body and known tracking object.
Track algorithm is the track that each target is estimated using particle filter.Assuming that people moves on the ground, everyone
Status tracking amount show as 2D transformation, the i.e. position (x, y) of center of gravity and rotation angle θ.Wherein mobility model is set as constant speed fortune
It is dynamic, this is because the model is good at the complete coupled problem of processing.
When a cluster is classified as human body, it is meant that they are not in contact with known tracking object, at this moment create
New tracking object.Obviously, the human body detected in the first frame is both initialized to new tracking object.
In order to the people detected in present frame and known tracking object match, it is calculated based on color and distance
Consistency, specific as follows shown:
Apart from conformance definition are as follows: a given human testing clusters Ci, by handling global arest neighbors data to find
Immediate human body tracking object TjIf their distance is less than threshold value, the cluster of detection is regarded as the object with tracking
It is related, then for tracking object TjEach of point pi,j, detection target C is found using Octree methodiIn each point
pj,i, calculate the distance between its, point pi,jWith point pj,iThe distance between consistency definition of probability it is as follows:
Wherein α is weight vectors；
Colour consistency is defined as: cluster C when comparing current detectioniWith tracking object TjColouring information when, pass through calculating
Find nearest a pair of point < pj,i,pi,j> between colour consistency.Colour consistency can be in RGB, HSV or other color spaces
Middle calculating.By taking HSV space as an example, point pi,jAnd pj,iBetween colour consistency definition of probability it is as follows:
Wherein ci,jAnd cj,iRespectively indicate pi,jAnd pj,iHSV information, β indicate weight；
pi,jAnd pj,iBetween joint consistency definition of probability are as follows:
L(pi,j,pj,i)=Ld(pi,j,pj,i)Lc(pi,j,pj,i)；
For each tracking object TjC is clustered with detectioniMaximum combined likelihood probability then is defined as:
If L (j, i) is higher than the threshold value of setting, show currently to cluster CiWith tracking object TjThe same person, it is on the contrary then
It is not.If do not found and CiThe tracking object being associated then creates new tracking object.
In order to verify the algorithm of proposition, carries out following experiment: Asus Xtion Pro being installed on Turtlebot 2 and is swashed
Optical radar, wherein Xtion Pro is used to collect the original RGB-D point cloud data of environment, and laser radar is then used to avoidance.
As shown in figure 4, reality has been carried out under three kinds of different scenes respectively in order to verify method proposed in the present invention
It tests:
Simple environment: not in the environment of barrier, sensor is fixed, and two people are moved with identical track, in Fig. 4
A.
Moderate environment: not in the environment of barrier, sensor is fixed, more than two people's random walks and movement rail
There is intersection between mark, such as the b in Fig. 4.
Difficult circumstances: having a barrier, robot motion, between three or more people's random walks and its motion profile mutually
Intersect, such as the c in Fig. 4.
For each scene, for video sequence in about 250 frames, total test set includes 798 frames.One has 2698 people's
Example, they by hand labeled on RGB image, using as truthful data.
Human detection result:
In order to verify the performance of method proposed by the invention, using the measurement based on frame, precision (p), recall rate are utilized
(r) and f1 score value three measures the superiority and inferiority of proposed method.Three are respectively defined as: Wherein TP indicates kidney-Yang rate, and FP indicates false positive rate, and FN indicates False-Negative Rate.
As shown in figure 4, the human body detected is enclosed with green box, it can be, it is evident that for first two feelings from figure
Condition, under especially simple environment, Comparison of experiment results is good.It is as a result just good without first two however under difficult scene.This is very
It is readily appreciated that there is barrier in difficult scene.It in some cases, is highly difficult by geological information and HOG+SVM classification
's.As a result some barriers are identified as human body by mistake in.The human body under three kinds of different environment is had recorded in Table 1
The performance of detection.
In order to reduce rate of false alarm, -2.2 are strictly set by HOG confidence threshold.This helps to reduce false positive rate, but also can
Lead to the increase of False-Negative Rate.
Table one: the performance parameter of human testing model under three kinds of varying environments
| Precision | Recall rate | F1 scoring | |
| Simple environment | 0.97 | 0.91 | 0.94 |
| Moderate environment | 0.92 | 0.88 | 0.89 |
| Difficult circumstances | 0.82 | 0.78 | 0.80 |
Human body tracking result:
We have evaluated our tracking result with regard to false positive rate and False-Negative Rate, and table two has recorded in three kinds of different environment
Under tracking performance, result is preferable however in the case where difficulty in the case where simple and medium, FP and FN ratio is somewhat high,
It is 5.8% and 5.2% respectively.This is mainly due in this scenario, the movement speed of people is very fast, is blocked or people by other people
Other than camera fields of view range.
Table two: the performance parameter of human body tracking model under three kinds of varying environments
| FP | FN | |
| Simple environment | 2.4% | 1.8% |
| Moderate environment | 4.6% | 4.4% |
| Difficult circumstances | 5.8% | 5.2% |
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all wrap within the scope of the present invention within mind and principle.
Claims (4)
1. indoor human body detection and tracking in the case of a kind of low visual angle based on RGB-D, which is characterized in that including following step
Suddenly,
The acquisition and pretreatment of step a. data
Carry out point cloud acquisition using Asus Xtion Pro, obtain intensive 3D point cloud, by straight-through filter to intensive 3D point cloud into
Row noise reduction process, and down-sampled processing is carried out to the 3D point cloud after noise reduction by three-dimensional voxel lattice filter；
It step b. ground detection and filters out
Using the ground in 3D initial point cloud of the least square method detection by step a processing based on RANSAC, for non-the
For one frame data, the ground parameter that previous frame detects can be used as the initial parameter of next frame, according to what is initially given
Ground parameter and preset initial distance threshold value carry out ground detection, and foundation to current frame data using RANSAC method
Index of each three-dimensional point on ground in 3D point cloud, it is filtered out from 3D point cloud；It, can be according to RGB-D for the first frame data
The determining relativeness with ground of installation site of the sensor in robot, thus the initial parameter on given ground；
Step c. cluster
The point cloud data within 130cm from the ground is obtained in the 3D point cloud for removing ground, then utilizes the Euclidean between two o'clock
Distance carries out 3D cluster, and when the Euclidean distance between each pair of point is less than predefined distance threshold, this two o'clock is defined as belonging to
Same class；Wherein, cluster needs to set two initial thresholds: belong to minimum range between the two o'clock of the same cluster and
Minimal point needed for generating a cluster；
Step d.HOG+SVM classification
RGB image will be projected to through the point cloud in 3D cluster back boundary frame, using the image module of acquisition, calculates its HOG description
Then obtained HOG describer is sent to preparatory trained SVM classifier, calculates the HOG confidence level of each cluster by device；
When the HOG confidence level of calculating is higher than the threshold value of setting, then determine that the cluster for people, is not otherwise then；
Step e. tracking
Obtained human body is clustered into the input as tracking module, i.e., as the object to be tracked in next step, then utilizes each frame
The middle obtained human body cluster that detects is matched with existing tracking object, specific using based on apart from consistency and solid colour
The method that property is combined, calculates the maximum likelihood probability between currently detected human body and known tracking object.
2. indoor human body detection and tracking in the case of a kind of low visual angle based on RGB-D according to claim 1,
It is characterized in that, in the step c, when being clustered, 1) the Kd tree that creates 3D point cloud first, as extraction point Yun Shisuo later
Searching method；2) the empty list C of setting cluster, and the empty queue of point cloud；3) for each of cloud point p, with
Centered on it, using preset distance threshold as radius, all neighbor points in the ball are searched for, for searching for obtained neighbor point,
First check for whether it has been added other clusters, if it is not, queue Q is added in the Neighbor Points；4) when the point in queue Q
After being all processed, queue Q is added in cluster C；5) after point all in initial point cloud is all processed, cluster knot
Beam.
3. indoor human body detection and tracking in the case of a kind of low visual angle based on RGB-D according to claim 2,
It is characterized in that, in the step c, after the completion of cluster, needs that cluster and deficient clustering problem excessively is further processed；
The processing for crossing cluster is as follows: the cluster C obtained for eachi, projection of its central point in XZ plane is calculated first
piIf piWith cluster CjCentral projection point pjDistance be less than setting threshold value when, then it is assumed that cluster CiAnd CjBelong to same
A cluster, and then merged；
The processing for owing cluster is as follows: for each cluster, calculating their geological information, specifically includes width, depth, height
Information；If the geological information of some clusters is much larger than the threshold value set, is further divided using colouring information, i.e., will had
The point of same color is classified as same class；Those are put with very little cluster, is directly given up.
4. indoor human body detection and tracking in the case of a kind of low visual angle based on RGB-D according to claim 1,
It is characterized in that, in the step e,
Apart from conformance definition are as follows: a given human testing clusters Ci, closest to find by handling global arest neighbors data
Tracking object TjIf their distance is less than threshold value, the cluster of detection is regarded as being related with the cluster of tracking, then
For tracking object TjEach of point pi,j, detection target C is found using Octree methodiEach of point pj,i, calculate its it
Between distance, point pi,jWith point pj,iThe distance between consistency definition of probability it is as follows:
Wherein α is weight vectors；
Colour consistency is defined as: cluster C when comparing current detectioniWith tracking object TjColouring information when, by calculate find
Nearest a pair of point < pj,i,pi,j> between colour consistency, colour consistency can fall into a trap in RGB, HSV or other color spaces
It calculates；By taking HSV space as an example, point pi,jAnd pj,iBetween colour consistency definition of probability it is as follows:
Wherein ci,jAnd cj,iRespectively indicate pi,jAnd pj,iHSV information, β indicate weight；
pi,jAnd pj,iBetween joint consistency definition of probability are as follows:
L(pi,j,pj,i)=Ld(pi,j,pj,i)Lc(pi,j,pj,i)；
For each tracking object TjC is clustered with detectioniMaximum combined likelihood probability is defined as:
If L (j, i) is higher than the threshold value of setting, show currently to cluster CiWith tracking object TjThe same person, it is on the contrary then not
It is, if do not found and CiThe tracking object being associated then creates new tracking object.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810908661.2A CN109035305B (en) | 2018-08-10 | 2018-08-10 | Indoor human body detection and tracking method based on RGB-D low-visual-angle condition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810908661.2A CN109035305B (en) | 2018-08-10 | 2018-08-10 | Indoor human body detection and tracking method based on RGB-D low-visual-angle condition |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109035305A true CN109035305A (en) | 2018-12-18 |
| CN109035305B CN109035305B (en) | 2021-06-25 |
Family
ID=64632680
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810908661.2A Active CN109035305B (en) | 2018-08-10 | 2018-08-10 | Indoor human body detection and tracking method based on RGB-D low-visual-angle condition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109035305B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110110687A (en) * | 2019-05-15 | 2019-08-09 | 江南大学 | Fruit automatic identifying method on tree based on colouring information and three-D profile information |
| CN110175528A (en) * | 2019-04-29 | 2019-08-27 | 北京百度网讯科技有限公司 | Human body tracing method and device, computer equipment and readable medium |
| CN110456308A (en) * | 2019-07-08 | 2019-11-15 | 广西工业职业技术学院 | A kind of three dimension location method for fast searching |
| CN111582352A (en) * | 2020-04-30 | 2020-08-25 | 上海高仙自动化科技发展有限公司 | Object-based sensing method and device, robot and storage medium |
| CN112070840A (en) * | 2020-09-11 | 2020-12-11 | 上海幻维数码创意科技有限公司 | Human body space positioning and tracking method with integration of multiple depth cameras |
| CN113033481A (en) * | 2021-04-20 | 2021-06-25 | 湖北工业大学 | Method for detecting hand-held stick combined with aspect ratio-first order fully-convolved object detection (FCOS) algorithm |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130156264A1 (en) * | 2011-12-15 | 2013-06-20 | Linus Mårtensson | Minimizing drift using depth camera images |
| CN106384079A (en) * | 2016-08-31 | 2017-02-08 | 东南大学 | RGB-D information based real-time pedestrian tracking method |
| CN107016373A (en) * | 2017-04-12 | 2017-08-04 | 广东工业大学 | The detection method and device that a kind of safety cap is worn |
| CN107491712A (en) * | 2016-06-09 | 2017-12-19 | 北京雷动云合智能技术有限公司 | A kind of human body recognition method based on RGB D images |
| CN107833270A (en) * | 2017-09-28 | 2018-03-23 | 浙江大学 | Real-time object dimensional method for reconstructing based on depth camera |
-
2018
- 2018-08-10 CN CN201810908661.2A patent/CN109035305B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130156264A1 (en) * | 2011-12-15 | 2013-06-20 | Linus Mårtensson | Minimizing drift using depth camera images |
| CN107491712A (en) * | 2016-06-09 | 2017-12-19 | 北京雷动云合智能技术有限公司 | A kind of human body recognition method based on RGB D images |
| CN106384079A (en) * | 2016-08-31 | 2017-02-08 | 东南大学 | RGB-D information based real-time pedestrian tracking method |
| CN107016373A (en) * | 2017-04-12 | 2017-08-04 | 广东工业大学 | The detection method and device that a kind of safety cap is worn |
| CN107833270A (en) * | 2017-09-28 | 2018-03-23 | 浙江大学 | Real-time object dimensional method for reconstructing based on depth camera |
Non-Patent Citations (2)
| Title |
|---|
| 居青: "基于RGB-D传感器的移动服务机器人功能软件开发及应用", 《中国优秀硕士学位论文全文数据库》 * |
| 张松: "基于RGB-D传感器的地面移动机器人目标检测与跟踪", 《中国优秀硕士学位论文全文数据库》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110175528A (en) * | 2019-04-29 | 2019-08-27 | 北京百度网讯科技有限公司 | Human body tracing method and device, computer equipment and readable medium |
| CN110175528B (en) * | 2019-04-29 | 2021-10-26 | 北京百度网讯科技有限公司 | Human body tracking method and device, computer equipment and readable medium |
| CN110110687A (en) * | 2019-05-15 | 2019-08-09 | 江南大学 | Fruit automatic identifying method on tree based on colouring information and three-D profile information |
| CN110110687B (en) * | 2019-05-15 | 2020-11-17 | 江南大学 | Method for automatically identifying fruits on tree based on color information and three-dimensional contour information |
| CN110456308A (en) * | 2019-07-08 | 2019-11-15 | 广西工业职业技术学院 | A kind of three dimension location method for fast searching |
| CN111582352A (en) * | 2020-04-30 | 2020-08-25 | 上海高仙自动化科技发展有限公司 | Object-based sensing method and device, robot and storage medium |
| CN111582352B (en) * | 2020-04-30 | 2023-06-27 | 上海高仙自动化科技发展有限公司 | Object-based perception method, object-based perception device, robot and storage medium |
| CN112070840A (en) * | 2020-09-11 | 2020-12-11 | 上海幻维数码创意科技有限公司 | Human body space positioning and tracking method with integration of multiple depth cameras |
| CN112070840B (en) * | 2020-09-11 | 2023-10-10 | 上海幻维数码创意科技股份有限公司 | Human body space positioning and tracking method fused by multiple depth cameras |
| CN113033481A (en) * | 2021-04-20 | 2021-06-25 | 湖北工业大学 | Method for detecting hand-held stick combined with aspect ratio-first order fully-convolved object detection (FCOS) algorithm |
| CN113033481B (en) * | 2021-04-20 | 2023-06-02 | 湖北工业大学 | Handheld stick detection method based on first-order full convolution target detection algorithm |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109035305B (en) | 2021-06-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109035305A (en) | Indoor human body detection and tracking in the case of a kind of low visual angle based on RGB-D | |
| US11367219B2 (en) | Video analysis apparatus, person retrieval system, and person retrieval method | |
| Datta et al. | Person-on-person violence detection in video data | |
| Baysal et al. | Sentioscope: a soccer player tracking system using model field particles | |
| CN105940430B (en) | Personnel's method of counting and its device | |
| Mittal et al. | M2tracker: A multi-view approach to segmenting and tracking people in a cluttered scene using region-based stereo | |
| Van Oosterhout et al. | Head detection in stereo data for people counting and segmentation | |
| CN106682573B (en) | A kind of pedestrian tracting method of single camera | |
| CN107657244B (en) | Human body falling behavior detection system based on multiple cameras and detection method thereof | |
| CN102262725A (en) | Analysis Of Three-dimensional Scenes | |
| Ren et al. | Multi-camera video surveillance for real-time analysis and reconstruction of soccer games | |
| Teachabarikiti et al. | Players tracking and ball detection for an automatic tennis video annotation | |
| CN104616006A (en) | Surveillance video oriented bearded face detection method | |
| CN110020618A (en) | A kind of crowd's abnormal behaviour monitoring method can be used for more shooting angle | |
| Nguyen et al. | Single camera based fall detection using motion and human shape features | |
| CN103577832A (en) | People flow statistical method based on spatio-temporal context | |
| Saisan et al. | Multi-view classifier swarms for pedestrian detection and tracking | |
| Li et al. | Evaluating the performance of systems for tracking football players and ball | |
| Girisha et al. | Tracking humans using novel optical flow algorithm for surveillance videos | |
| CN106023252A (en) | Multi-camera human body tracking method based on OAB algorithm | |
| Pane et al. | A people counting system for business analytics | |
| Hung et al. | Detecting fall incidents of the elderly based on human-ground contact areas | |
| Lo et al. | Vanishing point-based line sampling for real-time people localization | |
| Zeng et al. | Human detection using multi-camera and 3D scene knowledge | |
| Yu | Automatic basketball tracking in broadcast basketball video |
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 |