CN110232330A - A kind of recognition methods again of the pedestrian based on video detection - Google Patents
A kind of recognition methods again of the pedestrian based on video detection Download PDFInfo
- Publication number
- CN110232330A CN110232330A CN201910434555.XA CN201910434555A CN110232330A CN 110232330 A CN110232330 A CN 110232330A CN 201910434555 A CN201910434555 A CN 201910434555A CN 110232330 A CN110232330 A CN 110232330A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- video
- frame
- measured
- key
- 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
- G06T7/254—Analysis of motion involving subtraction of images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/269—Analysis of motion using gradient-based methods
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Human Computer Interaction (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of recognition methods again of the pedestrian based on video detection, which comprises the steps of: step S1 obtains the key frame of video to be measured using frame differential method；Step S2 extracts the critical depth feature in key frame based on detection network；Step S3 extracts non-key depth characteristic and corresponding manual feature in non-key frame based on light stream network；Step S4 carries out similarity calculation building pedestrian's weight identification model according to critical depth feature, non-key depth characteristic and manual feature；Step S5 analyzes each video to be measured by pedestrian's weight identification model, obtains location information and temporal information of the target pedestrian in each video to be measured, and be ranked up；Step S6 analyzes all videos to be measured by pedestrian's weight identification model, obtains the probability that target pedestrian occurs in each video to be measured, and be ranked up；Step S7 draws the track that target pedestrian occurs in predetermined monitoring scene according to ranking results.
Description
Technical field
The present invention relates to technical field of video monitoring more particularly to a kind of recognition methods again of the pedestrian based on video detection.
Background technique
Monitor video is widely used in subway, airport, traffic intersection, becomes the important tool of security protection, working principle
It is to be detected to emphasis pedestrian target in video, by the exact GPS positioning system of camera and time of occurrence point, is somebody's turn to do
Appearance track of the pedestrian target in entire scene.However in practical application scene, prevents in advance, verifies work often afterwards
It manually examines, low efficiency, the time is long, therefore realizes across the camera lens automatic identification to pedestrian, and then obtains each pedestrian's object
Appearance track in entire monitoring scene realizes that tracking is reasonably necessary.
Pedestrian identifies again to be referred in multiple and different camera videos, searches some specific pedestrian in which camera
In occurred, pedestrian detection work, the feature extraction work being related in scene, the characteristic similarity of two pedestrians measure work
Make.However identified in research work again in actual pedestrian, feature extraction and similarity measurement are generallyd use by manually marking
Or pedestrian's picture that detection algorithm obtains independently carries out with pedestrian detection as data set, is often very difficult to apply in actual view
It (can be with articles of reference Mengyue Geng, Yaowei Wang, Tao Xiang, Yonghong Tian.Deep in frequency scene
transfer learning for person reidentification[J].arXiv preprint arXiv:
1611.05244,2016.)。
Compared with picture detection, can have that motion blur, camera be out of focus in video or the strange posture of target object,
Phenomena such as seriously blocking, and the detection based on these phenomenons identifies the load for not only resulting in network again, but will substantially reduce mould
The accuracy rate of type generally extracts video frame as key frame using fixed step-length, other frames are as non-in order to solve these problems
Key frame infer using content of the timing information extracted based on light stream network to non-key frame and pixel supplement (can
With articles of reference: Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan.Deep Feature Flow for
Video Recognition.arXiv:1611.07715v2,2017), but inaccuracy and key frame step due to light stream network
Whether long difference, the supplement that can influence context information to a certain extent are complete.
Summary of the invention
To solve the above problems, present invention employs following technical solutions:
The present invention provides a kind of recognition methods again of the pedestrian based on video detection, for according to being clapped in predetermined monitoring scene
The multiple videos to be measured being made of picture frame taken the photograph identify the target pedestrian in predetermined monitoring scene, which is characterized in that
Include the following steps:
Step S1 reads the picture frame in video to be measured, is calculated using frame differential method picture frame, by picture frame
Key frame of the picture frame as video to be measured corresponding to middle differential intensity local maximum；
Step S2 extracts the feature of target pedestrian in key frame based on detection network, as critical depth feature；
Step S3 is extracted in non-key frame using remaining picture frame in picture frame as non-key frame based on light stream network
The correlated characteristic of target pedestrian, as non-key depth characteristic and corresponding manual feature；
Step S4, to critical depth feature, non-key depth characteristic and manual feature progress similarity calculation, and according to
The result of similarity calculation weighs identification model to construct pedestrian.
Step S5, by pedestrian weight identification model each video to be measured is analyzed, obtain target pedestrian it is each to
The location information and temporal information in video are surveyed, and location information and time letter to the target pedestrian in each video to be measured
Breath is ranked up；
Step S6 analyzes all videos to be measured by pedestrian's weight identification model, obtains mesh in each video to be measured
The probability that pedestrian occurs is marked, and video to be measured is ranked up according to the size of probability；
Step S7 draws target pedestrian according to the result to sort in step S5 and step S6 and occurs in predetermined monitoring scene
Track.
The present invention provides a kind of recognition methods again of the pedestrian based on video detection, can also have the feature that,
In, step S1 includes following sub-step:
Step S1-1 reads the picture frame in video to be measured；
Step S1-2 calculates the gray scale difference value of corresponding pixel between two adjacent picture frames；
Step S1-3 carries out binaryzation calculating to gray scale difference value, determines pixel according to the result that binaryzation calculates
Coordinate is prospect coordinate or background coordination；
Step S1-4 obtains the moving region in picture frame according to the result determined in step S1-3；
Step S1-5 carries out connectivity analysis to picture frame, when the area of the moving region in picture frame is greater than predetermined threshold
When value, then determine current picture frame for key frame.
The present invention provides a kind of recognition methods again of the pedestrian based on video detection, can also have the feature that,
In, include following sub-step in step S3:
Step S3-1 judges whether picture frame is key frame, if being judged as NO, using picture frame as non-key frame；
Step S3-2 carries out non-key frame and a upper key frame adjacent with non-key frame based on light stream algorithm for estimating
It calculates, obtains light stream figure；
Step S3-3, by the critical depth Character adjustment of key frame to spatial resolution identical with corresponding light stream figure
On propagated；
Step S3-4 extracts the non-key depth characteristic and corresponding craft in non-key frame according to the result of propagation
Feature.
The present invention provides a kind of recognition methods again of the pedestrian based on video detection, can also have the feature that,
In, critical depth feature is propagated using bilinear interpolation algorithm in step S-3.
The present invention provides a kind of recognition methods again of the pedestrian based on video detection, can also have the feature that,
In, the vector shift amount of the pixel in non-key frame is limited using time attention mechanism in step S-3.
The present invention provides a kind of recognition methods again of the pedestrian based on video detection, can also have the feature that,
In, step S4 includes following sub-step:
Step S4-1 carries out similarity calculation to critical depth feature, non-key depth characteristic and manual feature, obtains
Similarity matrix；
Similarity matrix fusion loss function is carried out parameter learning by step S4-2, to build pedestrian's weight identification model.
The present invention provides a kind of recognition methods again of the pedestrian based on video detection, can also have the feature that,
In, step S4-2 includes following sub-step:
Step S4-2-1 carries out classification learning to similarity matrix using Softmax loss function, to remove similarity
There is no the detection block of pedestrian in matrix；
Step S4-2-2 successively calculates manual feature and critical depth feature and non-pass using the method that COS distance is measured
The distance between key depth characteristic, and be ranked up according to the size of distance；
Step S4-2-3, it is based on sequence as a result, by OIM loss function in the way of multitask to similarity matrix after
It is continuous to carry out parameter learning, to build pedestrian's weight identification model.
Invention action and effect
Pedestrian based on video detection recognition methods again according to the present invention uses frame due to utilizing frame differential method
Between the mode that merges key frame is extracted from the picture frame of video, therefore can have preferably using relationship between picture frame
Effect ground mitigates since fuzzy frame is (i.e. because motion blur, camera be out of focus or the strange posture of target object, seriously blocks
Phenomenon causes picture frame fuzzy) bring network load and the negative effect to accuracy rate.Further, due to using light stream net
Network extracts the non-key depth characteristic of target pedestrian and manual feature in non-key frame, and allows key frame feature, non-key depth
Feature and manual feature point-to-point fusion on similarity matrix are spent, thus to existing apparent between adjacent picture frame
Contextual information is supplemented, so that the accuracy rate of pedestrian's weight identification model is higher, detects speed faster.
Detailed description of the invention
Fig. 1 is the implementation flow chart of pedestrian in present example based on video detection recognition methods again；
Fig. 2 is the work flow diagram of pedestrian of the embodiment of the present invention based on video detection recognition methods again；
Fig. 3 is the implementation flow chart of the non-key frame feature extraction in present example based on light stream figure；
Fig. 4 is the work flow diagram of the non-key frame feature extraction in present example based on light stream figure；
Fig. 5 is finally obtained pedestrian movement's track schematic diagram in present example.
Specific embodiment
In order to be easy to understand the technical means, the creative features, the aims and the efficiencies achieved by the present invention, tie below
Attached drawing is closed to be specifically addressed the pedestrian of the invention based on video detection again recognition methods.
<embodiment>
Network model is carried out using pytorch deep learning frame in the present embodiment to build, and is applied in model training
Mars data set, the data set share 6 cameras, 1261 pedestrians and 1,191,003 callout box, using CUHK03 data
Collection is tested, which shares 2 cameras, 1360 pedestrians.Test method is the video shot in a camera
The pedestrian target that middle interception needs to retrieve knows the pedestrian target in the video of another or the shooting of multiple cameras again
Not, camera position information and temporal information are returned to according to weight recognition result, all search results of single video is carried out
Sequence calculates a possibility that corresponding searched targets occur for all videos to be detected, video is ranked up.
It should be noted that the part not elaborated in the present invention belongs to the prior art.
Fig. 1 is the specific implementation flow chart of pedestrian in present example based on video detection recognition methods again, and Fig. 2 is this
The work flow diagram of pedestrian of the inventive embodiments based on video detection recognition methods again.
As depicted in figs. 1 and 2, the recognition methods again of the pedestrian based on video detection in the present embodiment, for according to predetermined
Multiple videos to be measured being made of picture frame of shooting identify the target pedestrian in predetermined monitoring scene in monitoring scene,
Include the following steps:
Step S1 reads the picture frame in video to be measured, is calculated using frame differential method picture frame, by picture frame
Key frame of the picture frame as video to be measured corresponding to middle differential intensity local maximum, specifically includes following sub-step:
Step S1-1 reads the picture frame in each video to be measured；
Step S1-2 calculates the gray scale difference value of corresponding pixel between two adjacent picture frames, it is assumed that ft(i, j) and
ft-1(i, j) is respectively the t frame and t-1 frame of a certain image sequence, then their difference image indicates are as follows:
Dt=| ft(i,j)-ft-1(i,j)|
Wherein, (i, j) indicates discrete picture coordinate.
Step S1-3 carries out binaryzation calculating to gray scale difference value, determines pixel according to the result that binaryzation calculates
Coordinate is prospect coordinate or background coordination, works as DtProspect coordinate is then considered greater than predetermined threshold T, is otherwise background coordination.
Step S1-4 obtains the moving region R in picture frame according to the result determined in step S1-3t(i, j), the movement
Region indicates are as follows:
Step S1-5 carries out connectivity analysis to the picture frame after binaryzation, when the area of the moving region in picture frame
When greater than predetermined threshold, then determine that current picture frame for key frame, is defined as Ik。
Step S2 extracts the feature of target pedestrian in key frame based on detection network, as critical depth feature fk, this reality
Applying detection network used by example is Faster-RCNN, but is simultaneously only not limited to Faster-RCNN network.
Step S3, using remaining picture frame in picture frame as non-key frame, based on target in light stream network non-key frame
The feature of pedestrian extracts, so that the correlated characteristic of feature alignment is obtained, as non-key depth characteristic and corresponding
Manual feature.
Fig. 3 is the specific implementation flow chart of the non-key frame feature extraction in present example based on light stream figure, and Fig. 4 is this
The work flow diagram of non-key frame feature extraction in invention example based on light stream figure.
As shown in Figure 3 and Figure 4, the non-key frame feature extraction in step S3 includes following sub-step:
Step S3-1 judges whether picture frame is key frame IkIf being judged as NO, using picture frame as non-key frame Ii；
Step S3-2, be based on above-mentioned detection network, extract with non-key frame IiAn adjacent upper key frame IkIn it is opposite
The critical depth feature f answeredk；
Step S3-3, based on light stream algorithm for estimating to non-key frame IiWith with non-key frame IiAn adjacent upper key frame
IkIt is calculated, obtains light stream figure, detailed process are as follows:
Enable Mi→kFor two-dimensional flow field, light stream algorithm for estimating is based on (such as by a non-key frame upper key frame adjacent thereto
FlowNet, but be not limited only to FlowNet network) light stream figure F is obtained, wherein Mi→k=F (Ik, Ii)。
Step S3-4, by the critical depth Character adjustment of key frame to spatial resolution identical with corresponding light stream figure
On propagated, the position P in current non-key frame i is projected into the position p+ δ p in key frame k in communication process, wherein
δ p=Mi→k(p), following two steps are specifically included:
1) δ p under normal conditions is decimal, and pixel respective coordinates are integer, therefore uses bilinear interpolation
Algorithm realizes the propagation operation of feature, the formula of the propagation are as follows:
Wherein, c is characterized the channel in figure f, and p, q enumerate all spatial position coordinates in characteristic pattern, and G is two-dimensional
The kernel function of bilinear interpolation.
2) to eliminate the inaccuracy because of light stream network, it is corresponding that non-key frame is further limited using time attention mechanism
The vector shift amount of pixel coordinate in characteristic pattern, the time formula of attention mechanism are as follows:
Wherein, ftThe feature obtained after network (detection network or light stream network as described above) for t frame image
Figure, e represent ftE-th of channel, p represents the corresponding coordinate position of each pixel on characteristic pattern.
Step S3-4 extracts the non-key depth characteristic and corresponding craft in non-key frame according to the result of propagation
Feature.
Step S4, to critical depth feature, non-key depth characteristic and manual feature progress similarity calculation, and according to
The result of similarity calculation weighs identification model to construct pedestrian.Including following sub-step:
Step S4-1 carries out similarity calculation to critical depth feature, non-key depth characteristic and manual feature, obtains
Similarity matrix；
Similarity matrix fusion loss function is carried out parameter learning by step S4-2, so that pedestrian's weight identification model is built,
Including following sub-step:
Step S4-2-1 carries out classification learning to similarity matrix using Softmax loss function, to remove similarity
There is no the detection block of pedestrian in matrix；
The specific calculating process of softmax loss function is as follows in step S4-2-1:
1) detection block of no pedestrian is screened out, it is assumed that pedestrian's classification number is N, and output layer is [Z1,Z2,...ZN], normalization is every
A pedestrian's probability are as follows:
2) using cross entropy as loss function:
Wherein PiIndicate the softmax value found out, tiFor true value.
Step S4-2-2 successively calculates manual feature and critical depth feature and non-pass using the method that COS distance is measured
The distance between key depth characteristic, and be ranked up according to the size of distance；
Step S4-2-3, it is based on sequence as a result, by OIM loss function in the way of multitask to similarity matrix after
It is continuous to carry out parameter learning, to build pedestrian's weight identification model.
The example and unlabeled of the labeled identity in training data are only considered in step S4-2-3
The example of identity minimizes the pedestrian target of same ID, the otherness between the pedestrian target of different ID is maximized, with more
Task state counterweight identification network continues parameter learning, and the specific calculating process of OIM loss function is as follows:
1) characteristic probability that feature vector f is considered as i class pedestrian indicates are as follows:
Wherein, L is the pedestrian's feature list marked, the pedestrian's characteristic series for having detected but not marked of Q storage
Table.V is the feature vector marked, and u represents the feature vector not marked in pedestrian detection, and T is the gentle factor, general for controlling
The gradual degree of rate distribution.
2) loss function is calculated:
L=Ex[log Pt]
Wherein, t is the tag along sort of target pedestrian.
Step S5, by pedestrian weight identification model each video to be measured is calculated, obtain target pedestrian it is each to
The location information and temporal information in video are surveyed, and location information and time letter to the target pedestrian in each video to be measured
Breath is ranked up；
Step S6 calculates all videos to be measured by pedestrian's weight identification model, obtains mesh in each video to be measured
The probability that pedestrian occurs is marked, and video to be measured is ranked up according to the size of probability；
Fig. 3 is finally obtained pedestrian movement's track schematic diagram in present example.
Step S7 draws target pedestrian according to the result to sort in step S5 and step S6 and occurs in predetermined monitoring scene
Track, as shown in figure 3, motion profile of the target pedestrian under specified camera as in the present embodiment.
Embodiment action and effect
It is used according to the recognition methods again of the pedestrian based on video detection of the present embodiment due to utilizing frame differential method
The mode of interframe fusion extracts key frame from the picture frame of video, therefore can preferably utilize relationship between picture frame,
Effectively mitigate since fuzzy frame is (i.e. because motion blur, camera be out of focus or the strange posture of target object, seriously blocks
Phenomena such as cause picture frame fuzzy) bring network load and the negative effect to accuracy rate.Further, due to using light stream
Network extracts the non-key depth characteristic of target pedestrian and manual feature in non-key frame, and allows key frame feature, non-key
Depth characteristic and manual feature point-to-point fusion on similarity matrix, thus to existing obvious between adjacent picture frame
Contextual information supplemented so that pedestrian weight identification model accuracy rate it is higher, detection speed faster.
For the non-key frame obscured caused by by reasons such as multiple light courcess, blocking property, noise and the transparencys, when due to using
Between attention mechanism the vector shift amount of the pixel in non-key frame is limited so that by streamer network it is obtained
The timing information of non-key frame feature and manual feature is more acurrate, therefore, the picture of non-key frame corresponding to these timing informations
Vegetarian refreshments can preferably supplement an adjacent upper key frame, not only the more conducively parameter training of similarity matrix, together
When also significantly reduce network load.
Due to being propagated using bilinear interpolation algorithm critical depth feature, so that being obtained by streamer figure
The non-key depth characteristic and manual feature taken is in feature aligned condition, to be more advantageous to the parameter instruction of similarity matrix
Practice, so that the analysis result of pedestrian's weight identification model is more accurate.
For pedestrian ID in video to be measured in actual use is very little and each picture frame in only several pedestrian ID the case where,
By in this present embodiment utilize OIM loss function, only consider training data in labeled identity example and
The example of unlabeled identity minimizes the pedestrian target of same ID, between the pedestrian target for maximizing different ID
Otherness, the method measured by COS distance calculate manual feature and critical depth feature and non-key depth characteristic one by one
Distance, and be ranked up according to the size of distance, to effectively reduce the calculation amount of pedestrian's weight identification model, avoided
The problem of model occurred in parameter learning can not restrain.
It is by pedestrian detection and traditional pedestrian by the pedestrian based on video detection in this present embodiment again recognition methods
Weight identification mission is combined together to form one-stage pedestrian's recognition methods again of end-to-end, therefore, under everyday scenes
Pedestrian under the image scene that gets, video scene is identified again with more real value.
The preferred embodiment of the present invention has been described in detail above.It should be appreciated that the ordinary skill of this field is without wound
The property made labour, which according to the present invention can conceive, makes many modifications and variations.Therefore, all technician in the art
Pass through the available technology of logical analysis, reasoning, or a limited experiment on the basis of existing technology under this invention's idea
Scheme, all should be within the scope of protection determined by the claims.
Claims (7)
1. a kind of recognition methods again of the pedestrian based on video detection, for multiple by image according to what is shot in predetermined monitoring scene
The video to be measured that frame is constituted identifies the target pedestrian in the predetermined monitoring scene, which is characterized in that including walking as follows
It is rapid:
Step S1 is read the described image frame in the video to be measured, is calculated using frame differential method described image frame,
Using picture frame corresponding to differential intensity local maximum in described image frame as the key frame of the video to be measured；
Step S2 extracts the feature of target pedestrian described in the key frame based on detection network, as critical depth feature；
Step S3 is extracted described non-key using remaining picture frame in described image frame as non-key frame based on light stream network
The feature of target pedestrian described in frame, as non-key depth characteristic and corresponding manual feature；
Step S4 carries out similarity meter to the critical depth feature, the non-key depth characteristic and the manual feature
It calculates, and constructs pedestrian's weight identification model according to the result of the similarity calculation；
Step S5 analyzes each video to be measured by pedestrian's weight identification model, obtains the target pedestrian every
Location information and temporal information in a video to be measured, and to the target pedestrian's in each video to be measured
Location information and temporal information are ranked up；
Step S6 analyzes all videos to be measured by pedestrian weight identification model, obtains each described to be measured
The probability that target pedestrian described in video occurs, and the video to be measured is ranked up according to the size of the probability；
Step S7 draws the target pedestrian in the predetermined monitoring field according to the result of sequence described in step S5 and step S6
The track occurred in scape.
2. the recognition methods again of the pedestrian based on video detection according to claim 1, it is characterised in that:
Wherein, the step S1 includes following sub-step:
Step S1-1 reads the described image frame in the video to be measured；
Step S1-2 calculates the gray scale difference value of corresponding pixel between two adjacent described image frames；
Step S1-3 carries out binaryzation calculating to the gray scale difference value, is determined according to the result that the binaryzation calculates described
The coordinate of pixel is prospect coordinate or background coordination；
Step S1-4 obtains the moving region in described image frame according to the result of judgement described in step S1-3；
Step S1-5 carries out connectivity analysis to described image frame, when the area of the moving region in described image frame is big
When predetermined threshold, then determine that current described image frame is the key frame.
3. the recognition methods again of the pedestrian based on video detection according to claim 1, it is characterised in that:
Wherein, include following sub-step in the step S3:
Step S3-1 judges whether described image frame is key frame, if being judged as NO, using described image frame as the non-pass
Key frame；
Step S3-2, based on light stream algorithm for estimating to the non-key frame and the upper pass adjacent with the non-key frame
Key frame is calculated, and light stream figure is obtained；
Step S3-3, by the critical depth Character adjustment of the key frame to sky identical with the corresponding light stream figure
Between propagated in resolution ratio；
Step S3-4 extracts the non-key depth characteristic in the non-key frame and opposite according to the result of the propagation
The manual feature answered.
4. the recognition methods again of the pedestrian based on video detection according to claim 3, it is characterised in that:
Wherein, the critical depth feature is propagated using bilinear interpolation algorithm in the step S-3.
5. the recognition methods again of the pedestrian based on video detection according to claim 3, it is characterised in that:
Wherein, inclined to the vector of the pixel in the non-key frame using time attention mechanism in the step S-3
Shifting amount is limited.
6. the recognition methods again of the pedestrian based on video detection according to claim 1, it is characterised in that:
Wherein, the step S4 includes following sub-step:
Step S4-1 carries out similarity to the critical depth feature, the non-key depth characteristic and the manual feature
It calculates, obtains similarity matrix；
Similarity matrix fusion loss function is carried out parameter learning, identified again to build the pedestrian by step S4-2
Model.
7. the recognition methods again of the pedestrian based on video detection according to claim 6, it is characterised in that:
Wherein, the step S4-2 includes following sub-step:
Step S4-2-1 carries out classification learning to the similarity matrix using Softmax loss function, to remove the phase
There is no the detection block of pedestrian like spending in matrix；
Step S4-2-2, using the method that COS distance is measured successively calculate the manual feature and the critical depth feature and
The distance between described non-key depth characteristic, and be ranked up according to the size of the distance；
Step S4-2-3, it is based on the sequence as a result, by OIM loss function in the way of multitask to the similarity moment
Battle array continues parameter learning, to build pedestrian's weight identification model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910434555.XA CN110232330B (en) | 2019-05-23 | 2019-05-23 | Pedestrian re-identification method based on video detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910434555.XA CN110232330B (en) | 2019-05-23 | 2019-05-23 | Pedestrian re-identification method based on video detection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110232330A true CN110232330A (en) | 2019-09-13 |
| CN110232330B CN110232330B (en) | 2020-11-06 |
Family
ID=67861545
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910434555.XA Active CN110232330B (en) | 2019-05-23 | 2019-05-23 | Pedestrian re-identification method based on video detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110232330B (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111008558A (en) * | 2019-10-30 | 2020-04-14 | 中山大学 | Picture/video important person detection method combining deep learning and relational modeling |
| CN111291633A (en) * | 2020-01-17 | 2020-06-16 | 复旦大学 | Real-time pedestrian re-identification method and device |
| CN111310728A (en) * | 2020-03-16 | 2020-06-19 | 中国科学技术大学 | Pedestrian re-identification system based on monitoring camera and wireless positioning |
| CN111738362A (en) * | 2020-08-03 | 2020-10-02 | 成都睿沿科技有限公司 | Object recognition method and device, storage medium and electronic equipment |
| CN111832457A (en) * | 2020-07-01 | 2020-10-27 | 济南浪潮高新科技投资发展有限公司 | Stranger intrusion detection method based on cloud edge cooperation |
| CN112200067A (en) * | 2020-10-09 | 2021-01-08 | 宁波职业技术学院 | Intelligent video event detection method, system, electronic equipment and storage medium |
| CN112801020A (en) * | 2021-02-09 | 2021-05-14 | 福州大学 | Pedestrian re-identification method and system based on background graying |
| CN113343810A (en) * | 2021-05-28 | 2021-09-03 | 国家计算机网络与信息安全管理中心 | Pedestrian re-recognition model training and recognition method and device based on time sequence diversity and correlation |
| CN113743387A (en) * | 2021-11-05 | 2021-12-03 | 中电科新型智慧城市研究院有限公司 | Video pedestrian re-identification method and device, electronic equipment and readable storage medium |
| CN114897762A (en) * | 2022-02-18 | 2022-08-12 | 众信方智(苏州)智能技术有限公司 | Automatic positioning method and device for coal mining machine on coal mine working face |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103475935A (en) * | 2013-09-06 | 2013-12-25 | 北京锐安科技有限公司 | Method and device for retrieving video segments |
| CN104077605A (en) * | 2014-07-18 | 2014-10-01 | 北京航空航天大学 | Pedestrian search and recognition method based on color topological structure |
| CN107832672A (en) * | 2017-10-12 | 2018-03-23 | 北京航空航天大学 | A kind of pedestrian's recognition methods again that more loss functions are designed using attitude information |
-
2019
- 2019-05-23 CN CN201910434555.XA patent/CN110232330B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103475935A (en) * | 2013-09-06 | 2013-12-25 | 北京锐安科技有限公司 | Method and device for retrieving video segments |
| CN104077605A (en) * | 2014-07-18 | 2014-10-01 | 北京航空航天大学 | Pedestrian search and recognition method based on color topological structure |
| CN107832672A (en) * | 2017-10-12 | 2018-03-23 | 北京航空航天大学 | A kind of pedestrian's recognition methods again that more loss functions are designed using attitude information |
Non-Patent Citations (1)
| Title |
|---|
| XIZHOU ZHU: "Deep Feature Flow for Video Recognition", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111008558B (en) * | 2019-10-30 | 2023-05-30 | 中山大学 | Picture/video important person detection method combining deep learning and relational modeling |
| CN111008558A (en) * | 2019-10-30 | 2020-04-14 | 中山大学 | Picture/video important person detection method combining deep learning and relational modeling |
| CN111291633A (en) * | 2020-01-17 | 2020-06-16 | 复旦大学 | Real-time pedestrian re-identification method and device |
| CN111310728B (en) * | 2020-03-16 | 2022-07-15 | 中国科学技术大学 | Pedestrian re-identification system based on monitoring camera and wireless positioning |
| CN111310728A (en) * | 2020-03-16 | 2020-06-19 | 中国科学技术大学 | Pedestrian re-identification system based on monitoring camera and wireless positioning |
| CN111832457A (en) * | 2020-07-01 | 2020-10-27 | 济南浪潮高新科技投资发展有限公司 | Stranger intrusion detection method based on cloud edge cooperation |
| CN111832457B (en) * | 2020-07-01 | 2022-07-22 | 山东浪潮科学研究院有限公司 | Stranger intrusion detection method based on cloud edge cooperation |
| CN111738362A (en) * | 2020-08-03 | 2020-10-02 | 成都睿沿科技有限公司 | Object recognition method and device, storage medium and electronic equipment |
| CN111738362B (en) * | 2020-08-03 | 2020-12-01 | 成都睿沿科技有限公司 | Object recognition method and device, storage medium and electronic equipment |
| CN112200067A (en) * | 2020-10-09 | 2021-01-08 | 宁波职业技术学院 | Intelligent video event detection method, system, electronic equipment and storage medium |
| CN112200067B (en) * | 2020-10-09 | 2024-02-02 | 宁波职业技术学院 | Intelligent video event detection method, system, electronic equipment and storage medium |
| CN112801020A (en) * | 2021-02-09 | 2021-05-14 | 福州大学 | Pedestrian re-identification method and system based on background graying |
| CN112801020B (en) * | 2021-02-09 | 2022-10-14 | 福州大学 | Pedestrian re-identification method and system based on background graying |
| CN113343810A (en) * | 2021-05-28 | 2021-09-03 | 国家计算机网络与信息安全管理中心 | Pedestrian re-recognition model training and recognition method and device based on time sequence diversity and correlation |
| CN113343810B (en) * | 2021-05-28 | 2023-03-21 | 国家计算机网络与信息安全管理中心 | Pedestrian re-recognition model training and recognition method and device based on time sequence diversity and correlation |
| CN113743387A (en) * | 2021-11-05 | 2021-12-03 | 中电科新型智慧城市研究院有限公司 | Video pedestrian re-identification method and device, electronic equipment and readable storage medium |
| CN114897762A (en) * | 2022-02-18 | 2022-08-12 | 众信方智(苏州)智能技术有限公司 | Automatic positioning method and device for coal mining machine on coal mine working face |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110232330B (en) | 2020-11-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110232330A (en) | A kind of recognition methods again of the pedestrian based on video detection | |
| CN110837778B (en) | Traffic police command gesture recognition method based on skeleton joint point sequence | |
| CN108805083B (en) | Single-stage video behavior detection method | |
| CN103208008B (en) | Based on the quick adaptive method of traffic video monitoring target detection of machine vision | |
| CN109102547A (en) | Robot based on object identification deep learning model grabs position and orientation estimation method | |
| CN110378259A (en) | A kind of multiple target Activity recognition method and system towards monitor video | |
| CN110147743A (en) | Real-time online pedestrian analysis and number system and method under a kind of complex scene | |
| CN112836640B (en) | Single-camera multi-target pedestrian tracking method | |
| CN104680559B (en) | The indoor pedestrian tracting method of various visual angles based on motor behavior pattern | |
| CN107545582A (en) | Video multi-target tracking and device based on fuzzy logic | |
| CN104517095B (en) | A kind of number of people dividing method based on depth image | |
| CN104794737B (en) | A kind of depth information Auxiliary Particle Filter tracking | |
| CN105869178A (en) | Method for unsupervised segmentation of complex targets from dynamic scene based on multi-scale combination feature convex optimization | |
| CN105760831A (en) | Pedestrian tracking method based on low-altitude aerial photographing infrared video | |
| CN110288627A (en) | One kind being based on deep learning and the associated online multi-object tracking method of data | |
| CN109063549A (en) | High-resolution based on deep neural network is taken photo by plane video moving object detection method | |
| CN112861808B (en) | Dynamic gesture recognition method, device, computer equipment and readable storage medium | |
| CN112446882A (en) | Robust visual SLAM method based on deep learning in dynamic scene | |
| CN110390294A (en) | Target tracking method based on bidirectional long-short term memory neural network | |
| CN115115672B (en) | Dynamic vision SLAM method based on target detection and feature point speed constraint | |
| CN111680560A (en) | Pedestrian re-identification method based on space-time characteristics | |
| CN103150552A (en) | Driving training management method based on people counting | |
| CN113033468A (en) | Specific person re-identification method based on multi-source image information | |
| Zhang et al. | Body localization in still images using hierarchical models and hybrid search | |
| Gopal et al. | Tiny object detection: Comparative study using single stage CNN object detectors |
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 |