CN105389562B - A kind of double optimization method of the monitor video pedestrian weight recognition result of space-time restriction - Google Patents
A kind of double optimization method of the monitor video pedestrian weight recognition result of space-time restriction Download PDFInfo
- Publication number
- CN105389562B CN105389562B CN201510779639.9A CN201510779639A CN105389562B CN 105389562 B CN105389562 B CN 105389562B CN 201510779639 A CN201510779639 A CN 201510779639A CN 105389562 B CN105389562 B CN 105389562B
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- path
- probability
- monitoring point
- recognition result
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
- G06V20/42—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items of sport video content
Landscapes
- Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of double optimization methods of the monitor video pedestrian of space-time restriction weight recognition result, first, using pedestrian's weight recognizer based on human body appearance visual signature, obtain initial recognition result；Then, by the time series parameters lifted in video frame, computation vision matching probability, route matching probability and the two joint probability；Finally, select the pedestrian image under the maximum path of joint probability as output.The method of the present invention has been obviously improved the confidence level of pedestrian's weight recognition result under the premise of not reducing recall rate, effectively overcome the pedestrian of traditional view-based access control model feature again recognition methods to monitoring the defect of environment sensitive.
Description
Technical field
The invention belongs to Video Analysis Technology fields, are related to monitor video pedestrian weight discriminance analysis, and in particular to Yi Zhongshi
The double optimization method of the monitor video pedestrian weight recognition result of sky constraint.
Technical background
In recent years, as video monitoring system is largely popularized, video monitoring system is in fighting crime, safeguarding stability practice
Increasingly important role is just being played, video investigation has become the new tool of public security organ's solving criminal cases.It is answered in video investigation
In, the retrieval for specific suspected target (especially people) is important need.This process is mainly by being accomplished manually at present,
A large amount of manpower and materials and time are expended, efficiency of solving a case is influenced.The core key of specific objective monitor video retrieval is asked
Topic --- pedestrian identifies again, refers to judging whether the pedestrian image occurred under different monitoring camera belongs to same a group traveling together.With
Technology develops and the increase of application demand, this problem are just gradually developing into the hot spot of academic research and sector application.
For same a group traveling together of accurate match under multi-cam picture, related scholar is respectively in Deja Vu, biological characteristic
With development pedestrian weight Study of recognition on the basis of appearance visual signature.Although under the conditions of specific application the technologies such as recognition of face by with
In pedestrian's identification, however in the case where actual video monitors environment, there are the resolution ratio of monitor video picture poor, pedestrian's object
Scale is smaller and the factors such as the randomness of pedestrian's object gesture, and the biological characteristics such as face, gait is caused to be difficult to extract.Relative to
Deja Vu and biological characteristic, the appearance visual signature of pedestrian are easy to extract and with certain individual sense.And
In certain space-time unique, pedestrian's object tends not to change the outfit.Therefore, existing research mostly uses greatly the appearance spy of pedestrian image
Sign.
In existing pedestrian's weight Study of recognition based on appearance, researcher is developed around feature extraction and similarity measurement
Many methods.The former focuses on designing the reliable pedestrian image character representation model of robust, can distinguish different pedestrians, together
When can not be illuminated by the light influence with visual angle change；The latter focuses on the distance that study meets pedestrian image feature distribution characteristic
Function, to keep same pedestrian image characteristic distance smaller, different pedestrian image characteristic distances are larger.However, these methods are answered
Use in actual monitored business that there are still huge challenges.Be mainly shown as, pedestrian again the image credit in identification problem in not
With camera, due to the influence of the environment such as angle, illumination residing for different cameras, in the different pictures of the same pedestrian,
Macroscopic features has a degree of variation；Conversely, because the variation of pedestrian's posture and camera angle, in different cameras
In, the macroscopic features of different pedestrians may be more more like than the macroscopic features of same person.
In order to reduce the interference of uncontrollable monitoring environmental factor to the greatest extent, existing pedestrian's weight identification technology is had to each
One group of image is all provided in the recognition result of monitoring point to select for person, then identification knot is refined by interactive related feedback method
Fruit.This processing mode not only increases the workload manually studied and judged, the degree of automation for reducing video analysis, moreover, by
In visual angle and light differential, the macroscopic features of pedestrian can change a lot, the result that the sequence provided may be caused forward
It might not be more credible.
Invention content
In order to solve the above technical problem, the present invention provides a kind of monitor video pedestrian of space-time restriction weight recognition results
Double optimization method.
The technical solution adopted in the present invention is：A kind of two suboptimums of monitor video pedestrian's weight recognition result of space-time restriction
Change method, which is characterized in that include the following steps：
Step 1：The pedestrian of view-based access control model feature identifies again；
According to the monitor video video recording of N number of monitoring point on the pedestrian image of input and travel path, outside based on human body
Pedestrian's weight recognizer of looks visual signature, identifies the M width candidate rows that visual signature is most like in each monitoring point video recording one by one
People's image, and sort according to the sequence of visual identity probability from high to low, while recording the timestamp of every width pedestrian image and regarding
Feel probability；
Step 2：Time series parameters obtain；
According to the physical distance and the obtained pedestrian image timestamp of above-mentioned steps between monitoring point, adjacent prison two-by-two is calculated
Time difference in the distance between control point time and corresponding two groups of M width candidate images between arbitrary two images, for row
First monitoring point on inbound path, then relative to input picture from crime monitoring point calculate；
Step 3：Vision matching probability calculation；
According to the vision probability parameter for each image that step 1 preserves, for (M+1)N- a kind of possible combination of paths, meter
Calculate the vision matching probability P under each combinationv；
Step 4：Route matching probability calculation；
Time difference parameter of the distance between the monitoring point obtained according to step 2 between time and candidate image, for (M+1
)N- a kind of possible combination of paths calculates the route matching probability P under each combinationp；
Step 5：Joint probability calculation；
According to step 3 and 4 result of calculation, and given experience weighting coefficient, using calculated with weighted average method path-
Vision joint probability P；
Step 6：Secondary identification based on joint probability；
(M+1) calculated according to above-mentioned stepsNThe respective joint probability of-a kind of combination of paths, by sequence row from big to small
Sequence chooses the path that ranks the first as preferred path, the optimum results that the pedestrian image on path identifies again as pedestrian.
Preferably, pedestrian's weight recognizer described in step 1 is pedestrian's weight recognizer of multiple dimensioned study.
Preferably, the timestamp of pedestrian image described in step 1 be the original video frame where pedestrian image when
Between the average value that stabs.
Preferably, the distance between two adjacent monitoring points described in step 2 time li, pass through the object between monitoring point
Manage distance diDivided by be averaged gait of march v of pedestrian is calculated, i.e.,Wherein pedestrian is averaged gait of march v as empirical value；Step 2
Described in two images between time difference ti, front is subtracted by the timestamp of image in back monitoring point on travel path
The timestamp of image calculates in monitoring point.
Preferably, the vision matching probability under combination of paths described in step 3 is with all pedestrian images on path
Probability addition calculation, i.e.,Here PiFor the vision matching probability of pedestrian image on path, K is on path
Monitoring is counted out, and for the path of each determination, K is a fixed constant.
Preferably, the specific calculating process of the route matching probability described in step 4 includes following sub-step：
Step 4.1：Pass through formulaCalculate the path of two width pedestrian images in arbitrary neighborhood monitoring point on path
Deflection probability ei, l hereiAnd tiThe Distance Time and time difference parameter that respectively step 2 obtains；
Step 4.2：By the deflection probability e of all monitoring points on pathiIt is added, obtains overall path deflection probability Ep, i.e.,If Ep>1, then enable Ep=1, i.e., the appearance sequential of pedestrian is against traveling sequence；
Step 4.3：By formula Pp=1-EpRoute matching probability is calculated, P is worked aspWhen=0, i.e., the path is Invalid path.
Preferably, the calculation formula of path-vision joint probability P described in step 5 is：
P=wPp+(1-w)Pv,
Wherein w is a preset experience weighting coefficient.
Preferably, the specific implementation process of the secondary identification based on joint probability described in step 6 includes following sub-step
Suddenly：
Step 6.1：Combination of paths sorts；
By (M+1)N- 1 paths sort by joint probability, and the forward L paths of selected and sorted form path candidate collection；
Step 6.2：Path merges；
When the joint probability that certain short path is subordinated to an other long path and short path is not more than the joint in long path
When probability, short path is removed from Candidate Set, only retains long path；
Step 6.3：Pedestrian's figure in Candidate Set after merging from path corresponding to the path and path of output sequence first
Picture.
The method of the present invention propose route matching probability thinking be it is a kind of reduce search space effective measures, have compared with
Good promotional value, reference is all had to detection, tracking and the search problem of the suspected target in magnanimity monitor video big data
Effect.Compared to traditional pedestrian's recognition methods again based on human body appearance visual signature, the method for the present invention also has following excellent
Point and good effect：
1) the method for the present invention is by the ingenious temporal constraint relationship occurred using pedestrian in monitoring point on travel path, in original
On the basis of some vision matching probability, introducing path matching probability comes the various pedestrian paths of joint measure and is combined into existing possibility
Property, in this, as the preferred foundation to first time recognition result, it has been obviously improved the confidence level of pedestrian's weight recognition result；
2) the method for the present invention introduces the space-time restriction relationship of Route Dependence, and space-time restriction is not illuminated by the light, visual angle etc. is taken the photograph
As the influence of environment, effectively overcomes the pedestrian of traditional view-based access control model feature recognition methods is sensitive to imaging environment again and lack
It falls into；
3) the method for the present invention only retains pedestrian's figure the most reliable by the reasonable cleaning to a large amount of candidate recognition results
Picture is effectively reduced the workload that subsequent artefacts study and judge under the premise of not reducing recall rate.
Description of the drawings
The flow chart of Fig. 1 embodiment of the present invention.
Specific implementation mode
Understand for the ease of those of ordinary skill in the art and implement the present invention, with reference to the accompanying drawings and embodiments to this hair
It is bright to be described in further detail, it should be understood that implementation example described herein is merely to illustrate and explain the present invention, not
For limiting the present invention.
In fact, if the recognition result of each monitoring point on travel path treated as a whole, they it
Between should have strong space-time dependence.Such as, physical location can not possibly be appeared in apart in same time with a group traveling together
In different monitoring point, pedestrian, which appears in the time difference in different monitoring point also, must obtain at a distance between monitoring point and in common sense
Pedestrian's gait of march has rational relationship, the time that pedestrian appears in back monitoring point on travel path that should not be monitored earlier than front
The time of point.Therefore, it by the association analysis of the sequential relationship of the pedestrian image to being identified in different monitoring point, can reject big
Unreasonable recognition result is measured, only retains a small amount of result for meeting travel path restriction relation and is presented to the user.
Assuming that there is N number of monitoring point not covered mutually on travel path, for each monitoring point, pedestrian may occur in which or not go out
It is existing, then theoretically share 2NKind of combination, covers all monitoring points and all occurs, all monitoring points do not occur, appear in part
The medium various situations in monitoring point.But being not each combination has existing possibility, for example, for the two adjacent monitoring of A, B
Point, it is clearly unreasonable if the time of occurrence of the pedestrian detected in the two is identical, because people can not possibly be
The same time appears in two different places.In other words, the probability that pedestrian appears in different monitoring point must be by reality
The constraint of time-space relationship, pedestrian is only in different time can just appear in different monitoring point.More typically change, pedestrian appears in not
The pedestrian's gait of march that must be obtained at a distance between monitoring point and in common sense with the time difference in monitoring point has rational relationship.Pass through
Distance and pedestrian's speed between monitoring point can obtain the Distance Time between two monitoring points, if the real time difference detected with
This Distance Time is too wide in the gap, then is unreasonable；If it is considered that direction of travel, then the not only constraint of having time difference, also
There is the temporal constraint of sequencing, for example, from A to B, then the time occurred in B cannot be earlier than A.
How probability existing for quantum chemical method each combination of pathsPedestrian occurs between 2 points of arbitrary neighborhood in the paths
Real time difference tiWith Distance Time liCloser, then this 2 points of possibilities being comprised in path are bigger, therefore, can be with
Route matching degree is weighed with the error of the two parameters.In view of the error of the permissions bigger of Distance Time length, with error and
The ratio of Distance Time is more suitable to weigh, i.e.,Here eiPractical is exactly the deflection probability of route matching.It will go
The deflection probability between adjacent monitoring point is cumulative two-by-two on inbound path, i.e.,The deflection probability in entire path is just obtained,
Here K is that the monitoring on path is counted out, and for determining path, K is fixed value.Since error may compare duration parameters
Also big, the deflection probability calculated at this time is likely larger than 1, therefore, should limit E and be not more than 1.Correspondingly, Pp=1-E is exactly path
Matching probability, abbreviation path probability.2 are calculated one by oneNThe path probability that each in kind of combination of paths combines, and to sort result,
The path probability possibility existing for forward path that sorts is bigger.
Above-mentioned thinking can be used in the secondary identification optimization of pedestrian's weight recognition result of view-based access control model characteristic.It is similarly assumed that
The path monitoring points Shang YouNGe, for a certain input pedestrian image, if recognition methods has been presented for each monitoring point to pedestrian again
M width identification sequence image, the pedestrian of each monitoring point has (M+1) kind selection (comprising this special circumstances are not selected), then
Shared (M+1)N(it is not occur this do not meet in each monitoring point to remove pedestrian to subtract 1 to-a kind of pedestrian's combination of paths
Assuming that the case where).The route matching probability of each combination and above-mentioned calculation are just the same, i.e.,K is same
Count out for the monitoring on path, but be different from the former, due to input picture natively come from a monitoring point, institute in the hope of
With when K do not have to subtract 1.It selects the pedestrian that path probability sorts on forward path as output, is equivalent to view-based access control model spy
Property pedestrian weight recognition result carried out space time correlation constraint it is preferred.
In fact, two it is less preferred in can also further combined with vision sort probability.In vision identifies again, to each
M recognition results can all assign a vision matching probability, vision sequence is based on this probability to realize.Secondary identification
In should not ignore this probability completely, but the confidence level of recognition result should be weighed jointly using it and path probability.
The pedestrian of view-based access control model characteristic identifies again in, the identification of each monitoring point in combination of paths is independently carried out, therefore vision
Probability PiIt can be considered independent observation variable, therefore, the probability read group total of all the points on total vision matching probability available path,
I.e.Here K is to monitor to count out on path, and for a kind of combination of paths of determination, K is fixed.In turn,
Total vision matching probability and route matching probability are weighted averagely, path-vision joint probability is obtained, is referred to as combined general
Rate namely joint probability P=wPp+(1-w)Pv, w is experience weighting coefficient here.In principle, joint probability is science conjunction the most
The secondary basis of characterization of reason.
Based on above-mentioned thought, the present invention provides a kind of two suboptimums of monitor video pedestrian's weight recognition result of space-time restriction
Change method, referring to Fig.1, the specific implementation of the present invention includes the following steps：
Step 1：The pedestrian of view-based access control model feature identifies again.According to N (this implementation on the pedestrian image of input and travel path
Example N=5) the monitor video video recording of a monitoring point known one by one using pedestrian's weight recognizer based on human body appearance visual signature
Most like M (the present embodiment M=10) width candidate's pedestrian image of visual signature in not each monitoring point video recording, and know according to vision
The sequence sequence of other probability from high to low, while recording the timestamp and vision probability of every width pedestrian image；
As a kind of specific implementation, pedestrian identifies the pedestrian's weight recognizer for selecting multiple dimensioned study again, and algorithm is from each
Identify that pedestrian image of the M width by visual similarity rankings, common property give birth to NxM width candidate images in monitoring point video recording；
The implementation procedure of pedestrian's weight recognizer of multiple dimensioned study is as follows：
1) pedestrian's appearance type-collection, according to vision difference of the pedestrian between different images under multi-cam, to pedestrian's number
According to being clustered, per the pedestrian that class data markers are same appearance type；
2) the scale learning algorithm based on stochastical sampling carries out stochastical sampling to overall training sample, and learns to obtain just
The mahalanobis distance function of beginning carries out scale learning to every class data on this basis and updates mahalanobis distance function；
3) k nearest neighbor of input pedestrian image is searched, and mahalanobis distance function is selected by ballot；
4) the two-way content similarities and neighbour's similitude, automatic re-arrangement for calculating input pedestrian and pedestrian to be measured are initially tied
Fruit.
As a kind of specific implementation, the timestamp of pedestrian image is defined as the time of the original video frame where pedestrian image
The average value of stamp, timestamp are stored in the array that length is NxM.
Step 2：Time series parameters obtain.According between monitoring point physical distance and above-mentioned steps obtain pedestrian figure
As timestamp, arbitrary two width in the distance between adjacent monitoring point time two-by-two and corresponding two groups of M width candidate images is calculated
Time difference between image, for first monitoring point on travel path, then relative to input picture from crime monitoring
Point calculates；
The distance between wherein two adjacent monitoring points time li, pass through the physical distance d between monitoring pointiDivided by pedestrian is flat
Equal gait of march v is calculated, i.e.,Here pedestrian is averaged gait of march v as empirical value；Distance Time is stored in length
In the array of N；Time difference t between two imagesi, before being subtracted by the timestamp of image in back monitoring point on travel path
The timestamp of image calculates in the monitoring point of face；Often the data length of the time difference two-by-two between 2 groups of adjacent monitoring point M width images is
MxM shares N number of monitoring point, therefore the array length of holding time difference parameter is NxMxM.
Step 3：Vision matching probability calculation.According to the vision probability parameter for each image that step 1 preserves, for (M+
1)N- a kind of possible combination of paths calculates the vision matching probability P under each combinationv；
As a kind of specific implementation, vision matching probability calculation is as follows：
The probability addition calculation of all pedestrian images on vision matching probability available path under combination of paths, i.e.,Here PiFor the vision matching probability of pedestrian image on path, K is that the monitoring on path is counted out, for every
The path of a determination, K are a fixed constant.
Step 4：Route matching probability calculation.The distance between the monitoring point obtained according to step 2 time and candidate image
Between time difference parameter, for (M+1)N- a kind of possible combination of paths calculates the route matching probability P under each combinationp；
As a kind of specific implementation, the calculating of route matching probability includes following sub-step：
Step 4.1：Pass through formulaCalculate the road of two width pedestrian images in arbitrary neighborhood monitoring point on path
Diameter deflection probability ei, l hereiAnd tiThe Distance Time and time difference parameter that respectively step 2 obtains；
Step 4.2：By the deflection probability e of all monitoring points on pathiIt is added, obtains overall path deflection probability Ep, i.e.,If Ep>1 enables Ep=1, the appearance sequential of pedestrian is meant against traveling sequence；
Step 4.3：By formula Pp=1-EpRoute matching probability is calculated, P is worked aspWhen=0, it is meant that this paths is invalid road
Diameter.
Step 5：According to step 3 and 4 result of calculation, and given experience weighting coefficient, using weighted mean method meter
Calculate path-vision joint probability P；
As a kind of specific implementation, path-vision joint probability P presses formula P=wPp+(1-w)PvIt calculates, w is one here
Preset experience weighting coefficient (the present embodiment w=0.8).
Step 6：Secondary identification based on joint probability.(M+1) calculated according to above-mentioned stepsN- a kind of combination of paths is respectively
Joint probability choose the path that ranks the first by sequence sequence from big to small and be used as preferred path, the pedestrian on path schemes
As the optimum results identified as pedestrian again.
As a kind of specific implementation, secondary identification includes specifically following sub-step：
Step 6.1：Paths ordering, by (M+1)N- 1 paths sort by joint probability, the forward L paths of selected and sorted
Form path candidate collection；
Step 6.2：Path merges, when certain short path is subordinated to the joint probability of an other long path and short path not
More than long path joint probability when, short path is removed from Candidate Set, only retains long path；
N number of monitoring point, each point M width images, then share NxM width images, to these image number consecutivelies, path can express
At the set of number, short path collection shares SsIt indicates, long set of paths SlIt indicates, works as Ss∈SlWhen, then short path is subordinated to length
Path.
Step 6.3：Pedestrian's figure in Candidate Set after merging from path corresponding to the path and path of output sequence first
Picture.
It should be understood that the part that this specification does not elaborate belongs to the prior art.
It should be understood that the above-mentioned description for preferred embodiment is more detailed, can not therefore be considered to this
The limitation of invention patent protection range, those skilled in the art under the inspiration of the present invention, are not departing from power of the present invention
Profit requires under protected ambit, can also make replacement or deformation, each fall within protection scope of the present invention, this hair
It is bright range is claimed to be determined by the appended claims.
Claims (8)
1. a kind of double optimization method of the monitor video pedestrian weight recognition result of space-time restriction, which is characterized in that including following
Step：
Step 1：The pedestrian of view-based access control model feature identifies again；
According to the monitor video video recording of N number of monitoring point on the pedestrian image of input and travel path, regarded using based on human body appearance
The pedestrian's weight recognizer for feeling feature, identifies the M width candidate pedestrians figure that visual signature is most like in each monitoring point video recording one by one
Picture, and sorting according to visual identity probability sequence from high to low, at the same record every width pedestrian image timestamp and vision it is general
Rate；
Step 2：Time series parameters obtain；
According to the physical distance and the obtained pedestrian image timestamp of above-mentioned steps between monitoring point, adjacent monitoring point two-by-two is calculated
The distance between time difference in time and corresponding two groups of M width candidate images between arbitrary two images, for traveling road
First monitoring point on diameter, then relative to input picture from crime monitoring point calculate；Wherein, the two adjacent monitoring
The distance between the point time is denoted as li, the time difference between two images described in step 2 is denoted as ti；
Step 3：Vision matching probability calculation；
According to the vision probability parameter for each image that step 1 preserves, for (M+1)N- a kind of possible combination of paths calculates every
Vision matching probability P under kind combinationv；
Step 4：Route matching probability calculation；
Time difference parameter of the distance between the monitoring point obtained according to step 2 between time and candidate image, for (M+1)N- a kind
Possible combination of paths calculates the route matching probability P under each combinationp；
Step 5：Joint probability calculation；
According to step 3 and 4 result of calculation, and given experience weighting coefficient, using calculated with weighted average method path-vision
Joint probability P；
Step 6：Secondary identification based on joint probability；
(M+1) calculated according to above-mentioned stepsNThe respective joint probability of-a kind of combination of paths, by sequence sequence from big to small, choosing
Take the path that ranks the first as preferred path, the optimum results that the pedestrian image on path identifies again as pedestrian.
2. the double optimization method of the monitor video pedestrian weight recognition result of space-time restriction according to claim 1, special
Sign is：Pedestrian's weight recognizer described in step 1 is pedestrian's weight recognizer of multiple dimensioned study.
3. the double optimization method of the monitor video pedestrian weight recognition result of space-time restriction according to claim 1, special
Sign is：The timestamp of pedestrian image described in step 1 is being averaged for the timestamp of the original video frame where pedestrian image
Value.
4. the double optimization method of the monitor video pedestrian weight recognition result of space-time restriction according to claim 1, special
Sign is：The distance between two adjacent monitoring points described in step 2 time li, pass through the physical distance d between monitoring pointiIt removes
It is calculated be averaged gait of march v of pedestrian, i.e.,Wherein pedestrian is averaged gait of march v as empirical value；Two described in step 2
Time difference t between width imagei, subtracted in the monitoring point of front and schemed by the timestamp of image in back monitoring point on travel path
The timestamp of picture calculates.
5. the double optimization method of the monitor video pedestrian weight recognition result of space-time restriction according to claim 1, special
Sign is：Vision matching probability under combination of paths described in step 3 is added with the probability of all pedestrian images on path
It calculates, i.e.,Here PiFor the vision matching probability of pedestrian image on path, K is that the monitoring on path is counted out,
For the path of each determination, K is a fixed constant.
6. the double optimization method of the monitor video pedestrian weight recognition result of space-time restriction according to claim 1, special
Sign is that the specific calculating process of the route matching probability described in step 4 includes following sub-step：
Step 4.1：Pass through formulaCalculate the path offset of two width pedestrian images in arbitrary neighborhood monitoring point on path
Probability ei, l hereiAnd tiThe Distance Time and time difference parameter that respectively step 2 obtains；
Step 4.2：By the deflection probability e of all monitoring points on pathiIt is added, obtains overall path deflection probability Ep, i.e.,If Ep>1, then enable Ep=1, i.e., the appearance sequential of pedestrian is against traveling sequence；Wherein, K is the monitoring on path
It counts out, for the path of each determination, K is a fixed constant；
Step 4.3：By formula Pp=1-EpRoute matching probability is calculated, P is worked aspWhen=0, i.e., the path is Invalid path.
7. the double optimization method of the monitor video pedestrian weight recognition result of space-time restriction according to claim 1, special
Sign is that the calculation formula of the path described in step 5-vision joint probability P is：
P=wPp+(1-w)Pv,
Wherein w is a preset experience weighting coefficient.
8. the double optimization method of the monitor video pedestrian weight recognition result of space-time restriction according to claim 1, special
Sign is that the specific implementation process of the secondary identification based on joint probability described in step 6 includes following sub-step：
Step 6.1：Combination of paths sorts；
By (M+1)N- 1 paths sort by joint probability, and the forward L paths of selected and sorted form path candidate collection；
Step 6.2：Path merges；
When the joint probability that certain short path is subordinated to an other long path and short path is not more than the joint probability in long path
When, short path is removed from Candidate Set, only retains long path；
Step 6.3：Pedestrian image in Candidate Set after merging from path corresponding to the path and path of output sequence first.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510779639.9A CN105389562B (en) | 2015-11-13 | 2015-11-13 | A kind of double optimization method of the monitor video pedestrian weight recognition result of space-time restriction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510779639.9A CN105389562B (en) | 2015-11-13 | 2015-11-13 | A kind of double optimization method of the monitor video pedestrian weight recognition result of space-time restriction |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105389562A CN105389562A (en) | 2016-03-09 |
| CN105389562B true CN105389562B (en) | 2018-08-21 |
Family
ID=55421834
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510779639.9A Active CN105389562B (en) | 2015-11-13 | 2015-11-13 | A kind of double optimization method of the monitor video pedestrian weight recognition result of space-time restriction |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105389562B (en) |
Families Citing this family (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107016374A (en) * | 2017-04-12 | 2017-08-04 | 电子科技大学 | Intelligent Measurement tracking and the generation method of space-time track towards specific objective |
| CN107862240B (en) * | 2017-09-19 | 2021-10-08 | 中科(深圳)科技服务有限公司 | Multi-camera collaborative face tracking method |
| CN107545256B (en) * | 2017-09-29 | 2020-03-10 | 上海交通大学 | Camera network pedestrian re-identification method combining space-time and network consistency |
| CN108171184B (en) * | 2018-01-03 | 2020-04-10 | 南京理工大学 | Method for re-identifying pedestrians based on Simese network |
| CN108664920B (en) * | 2018-05-10 | 2022-12-20 | 东方网力科技股份有限公司 | Real-time large-scale cascading face clustering method and device |
| CN109101917A (en) * | 2018-08-01 | 2018-12-28 | 北京旷视科技有限公司 | Mask method, training method, the apparatus and system identified again for pedestrian |
| CN109472191B (en) * | 2018-09-17 | 2020-08-11 | 西安电子科技大学 | Pedestrian re-identification and tracking method based on space-time context |
| CN110414441B (en) * | 2019-07-31 | 2022-05-10 | 浙江大学 | Pedestrian track analysis method and system |
| CN111061825B (en) * | 2019-12-10 | 2020-12-18 | 武汉大学 | Method for identifying matching and correlation of space-time relationship between mask and reloading camouflage identity |
| CN111291611A (en) * | 2019-12-20 | 2020-06-16 | 长沙千视通智能科技有限公司 | Pedestrian re-identification method and device based on Bayesian query expansion |
| CN111444758A (en) * | 2019-12-26 | 2020-07-24 | 珠海大横琴科技发展有限公司 | Pedestrian re-identification method and device based on spatio-temporal information |
| CN112270241B (en) * | 2020-10-22 | 2021-12-10 | 珠海大横琴科技发展有限公司 | Pedestrian re-identification method and device, electronic equipment and computer readable storage medium |
| CN112989911A (en) * | 2020-12-10 | 2021-06-18 | 奥比中光科技集团股份有限公司 | Pedestrian re-identification method and system |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103810476A (en) * | 2014-02-20 | 2014-05-21 | 中国计量学院 | Method for re-identifying pedestrians in video monitoring network based on small-group information correlation |
| CN104200206A (en) * | 2014-09-09 | 2014-12-10 | 武汉大学 | Double-angle sequencing optimization based pedestrian re-identification method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9396412B2 (en) * | 2012-06-21 | 2016-07-19 | Siemens Aktiengesellschaft | Machine-learnt person re-identification |
-
2015
- 2015-11-13 CN CN201510779639.9A patent/CN105389562B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103810476A (en) * | 2014-02-20 | 2014-05-21 | 中国计量学院 | Method for re-identifying pedestrians in video monitoring network based on small-group information correlation |
| CN104200206A (en) * | 2014-09-09 | 2014-12-10 | 武汉大学 | Double-angle sequencing optimization based pedestrian re-identification method |
Non-Patent Citations (2)
| Title |
|---|
| 智能监控系统中活动相关性分析与行人再识别研究;张磊;《中国优秀硕士学位论文全文数据库 信息科技辑》;20150415;第2015年卷(第04期);第I140-464页 * |
| 面向监控视频的行人重识别技术研究;王亦民;《中国博士学位论文全文数据库 信息科技辑》;20150615;第2015年卷(第06期);第I138-36页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105389562A (en) | 2016-03-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105389562B (en) | A kind of double optimization method of the monitor video pedestrian weight recognition result of space-time restriction | |
| CN111259850B (en) | Pedestrian re-identification method integrating random batch mask and multi-scale representation learning | |
| CN109583342B (en) | Human face living body detection method based on transfer learning | |
| CN108921051B (en) | Pedestrian attribute identification network and technology based on cyclic neural network attention model | |
| CN104881637B (en) | Multimodal information system and its fusion method based on heat transfer agent and target tracking | |
| CN104091176B (en) | Portrait comparison application technology in video | |
| CN111783576B (en) | Pedestrian re-identification method based on improved YOLOv3 network and feature fusion | |
| Ko et al. | Spatiotemporal bag-of-features for early wildfire smoke detection | |
| Salimi et al. | Visual-based trash detection and classification system for smart trash bin robot | |
| CN111126360A (en) | Cross-domain pedestrian re-identification method based on unsupervised combined multi-loss model | |
| CN110378259A (en) | A kind of multiple target Activity recognition method and system towards monitor video | |
| CN109800624A (en) | A kind of multi-object tracking method identified again based on pedestrian | |
| CN105303152B (en) | A kind of human body recognition methods again | |
| CN106022220A (en) | Method for performing multi-face tracking on participating athletes in sports video | |
| CN102811343A (en) | Intelligent video monitoring system based on behavior recognition | |
| CN113762009B (en) | Crowd counting method based on multi-scale feature fusion and double-attention mechanism | |
| CN113111767A (en) | Fall detection method based on deep learning 3D posture assessment | |
| CN108960047A (en) | Face De-weight method in video monitoring based on the secondary tree of depth | |
| CN109492534A (en) | A kind of pedestrian detection method across scene multi-pose based on Faster RCNN | |
| CN109902681B (en) | User group relation determining method, device, equipment and storage medium | |
| CN110414336A (en) | A kind of depth complementation classifier pedestrian's searching method of triple edge center loss | |
| CN112541403A (en) | Indoor personnel falling detection method utilizing infrared camera | |
| CN113011359B (en) | Method for simultaneously detecting plane structure and generating plane description based on image and application | |
| Miao et al. | Abnormal behavior learning based on edge computing toward a crowd monitoring system | |
| Lyu et al. | EFP-YOLO: a quantitative detection algorithm for marine benthic organisms |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20200930 Address after: 519000 unit 1, No. 33, Haihe street, Hengqin New District, Zhuhai City, Guangdong Province Patentee after: ZHUHAI DAHENGQIN TECHNOLOGY DEVELOPMENT Co.,Ltd. Address before: 430072 Hubei Province, Wuhan city Wuchang District of Wuhan University Luojiashan Patentee before: WUHAN University |
|
| TR01 | Transfer of patent right |