CN110210436A - A kind of vehicle-mounted camera line walking image-recognizing method - Google Patents
A kind of vehicle-mounted camera line walking image-recognizing method Download PDFInfo
- Publication number
- CN110210436A CN110210436A CN201910497061.6A CN201910497061A CN110210436A CN 110210436 A CN110210436 A CN 110210436A CN 201910497061 A CN201910497061 A CN 201910497061A CN 110210436 A CN110210436 A CN 110210436A
- Authority
- CN
- China
- Prior art keywords
- line walking
- vehicle
- mounted camera
- camera line
- image
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- 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
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/49—Segmenting video sequences, i.e. computational techniques such as parsing or cutting the sequence, low-level clustering or determining units such as shots or scenes
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Computing Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of vehicle-mounted camera line walking image-recognizing methods, pass through the image interception to vehicle-mounted camera line walking video, image information per second is obtained to be marked, form corresponding data set, key point triple detection object method is recycled to be trained, corresponding data model is obtained, realizes the identification to line walking image.Method provided by the present invention has the ability of perception interior of articles information, so as to effectively inhibit erroneous detection;Identification is horizontal horizontal consistent with naked eyes identification, it is ensured that the abnormal conditions of aerial optical cable are accurately differentiated.
Description
Technical field
The present invention relates to visual target tracking technical fields, and in particular to a kind of vehicle-mounted camera line walking image recognition side
Method.
Background technique
Visual target tracking is an important branch of computer vision field, and main task is by the figure to intake
Picture or video carry out analytical calculation, achieve the purpose that carry out recognition and tracking to the target in scene.Based on multiple camera shooting units
At camera network also produce massive video data while expanding monitoring range, to video transmission, storage and
Real-time modeling method application, brings great challenge.If video is all uploaded cloud center, cloud center will face choosing for data mighty torrent
War.It needs nearby to handle video, and existing camera shooting generator terminal computing resource is insufficient.Therefore, extensive video camera net need to be directed to
Network, to reduce traffic load, computation burden and improve algorithm real-time as target, proposition is suitble to target following under extensive environment
Scheme.
For example most representative CornerNet model of traditional object detection method based on key point passes through detectable substance
The upper left angle point and bottom right angle point of body determines target, but during determining target, can not efficiently use the inside of object
Feature, i.e., can not perceive the information of interior of articles, produce many erroneous detections (false target frame) so as to cause such method.
Summary of the invention
The object of the present invention is to provide a kind of vehicle-mounted camera line walking image-recognizing methods, to realize perception interior of articles letter
The ability of breath so as to effectively inhibit erroneous detection, and reduces traffic load, computation burden and improves algorithm real-time, proposes to be suitble to
The scheme of target following under extensive environment.
In order to achieve the above objectives, the present invention provides a kind of vehicle-mounted camera line walking image-recognizing methods, by vehicle-mounted
The image interception of camera line walking video obtains image information per second and is marked, and forms corresponding data set, recycles and closes
Key point triple detection object method is trained, and obtains corresponding data model, realizes the identification to line walking image.
Above-mentioned vehicle-mounted camera line walking image-recognizing method, wherein the key point triple detection object method is logical
Three central point, upper left angle point and bottom right angle point key points are crossed to determine a target.
Above-mentioned vehicle-mounted camera line walking image-recognizing method, wherein generated using upper left and the angle point of bottom right two initial
Target frame defines a central area to each prediction block, then judges whether the central area of each target frame contains center
Point retains the target frame if having, if without the target frame is deleted.
Above-mentioned vehicle-mounted camera line walking image-recognizing method a, wherein phase is defined when the scale of prediction block is larger
To lesser central area, a relatively large central area is predicted when the scale of prediction block is smaller.
Above-mentioned vehicle-mounted camera line walking image-recognizing method, wherein central point level side is extracted by central point pondization
To with the maximum value of vertical direction and be added, the information other than present position is provided with this to central point.
Above-mentioned vehicle-mounted camera line walking image-recognizing method, wherein in a horizontal direction being maximized operation by
Left pool area and right pool areaization are realized by series connection, similarly, are maximized operation by upper region in a vertical direction
Pool areaization is realized by series connection under Chi Huahe.
Above-mentioned vehicle-mounted camera line walking image-recognizing method, wherein extract object edge first by cascading angle point pondization
Then boundary's maximum value continues internally to propose maximum value at the maximum value of boundary, and is added with boundary maximum value, give angle with this
Point feature provides associate semantic information.
Above-mentioned vehicle-mounted camera line walking image-recognizing method, wherein cascade angle point pond passes through the angle on different directions
The combination in point pond is realized.
Compared with the existing technology, the invention has the following advantages:
(1) present invention is directed to extensive camera network, to reduce traffic load, computation burden and improve algorithm real-time
For target, the scheme for being suitble to target following under extensive environment is proposed;
(2) present invention localizes data processing, advantageously accounts for higher transmission cost, bandwidth demand and longer sound
It should postpone, solve the short slab in extensive real-time tracking at present;
(3) method provided by the present invention has the ability of perception interior of articles information, so as to effectively inhibit erroneous detection;
Identification is horizontal horizontal consistent with naked eyes identification, it is ensured that the abnormal conditions of aerial optical cable are accurately differentiated.
Detailed description of the invention
Fig. 1 is the flow diagram of vehicle-mounted camera line walking image-recognizing method of the present invention.
Specific embodiment
Below in conjunction with attached drawing, by specific embodiment, the invention will be further described, these embodiments are merely to illustrate
The present invention is not limiting the scope of the invention.
As shown in Figure 1, the present invention provides a kind of vehicle-mounted camera line walking image-recognizing methods, by vehicle-mounted camera
The image interception of line walking video obtains image information per second and is marked, and forms corresponding data set, recycles key point three
Tuple detection object method is trained, and obtains corresponding data model, realizes the identification to line walking image.
Edge calculations are a kind of by that will calculate the control of application, data and service from certain central nodes (" core ") band
To the cloud computing system optimization method of the internet end (" edge ") of another connection physical world.Data are via various sensings
Device is incoming from physical world, takes action by various outputs and controller and changes physical state, by executing analysis at edge
And knowledge formation reduces the communication bandwidth between controlled system and data center.Edge calculations can be set using associated physics
The degree of approach and possible relationship between standby.Recently, edge calculations model provides new thinking for the solution of such problem, will
Data processing localization, advantageously accounts for higher transmission cost, bandwidth demand and longer operating lag, solves big rule at present
Short slab in mould real-time tracking.
Increase the hardware cell of video processing in collection terminal, collected video information is replicated, then carries out pre-
Processing, by the image interception to vehicle-mounted camera line walking video, obtains image information per second and is marked, and is formed corresponding
Data set, recycle key point triple detection object method be trained, obtain corresponding data model, to line walking video into
Row identification.After obtaining target position, useful information is transferred to its neighbor node by video camera, while receiving neighbor node hair
The information brought, the node by the information received and itself measure information carry out information fusion, to extract useful letter
Breath, then fused information is sent to neighbor node.Similar information transmission is completed repeatedly in adjacent moment, so that it may so that
It obtains whole network information and reaches consistent.Therefore, the information that can merge each video camera realizes global letter with distributed mode
Breath is shared.The Target Tracking System of last use state algorithm for estimating composition.
The present invention utilizes key point triple, that is, central point, three key points of upper left angle point and bottom right angle point rather than two
For point to determine a target, the cost for making network take very little just has the ability for perceiving interior of articles information, so as to
Effectively inhibit erroneous detection.
Inhibit the principle of erroneous detection to be based on following inference: if target frame is accurately, region can be examined in its center
The probability for measuring target's center's point will be very high, and vice versa.Therefore, it is generated first with upper left and the angle point of bottom right two initial
Target frame defines a central area to each prediction block, then judges whether the central area of each target frame contains center
Point retains the target frame if having, if without the target frame is deleted.
It is not only that the one-stage detection method based on key point can not perceive interior of articles information in fact, it is nearly all
One-stage detection method all there are problems that this.Firstly, CornerNet model determines target by detecting angle point, without
It is that target is determined by the recurrence of initial candidate frame anchor point (anchor), due to the limitation without anchor point, so that any two
As soon as angle point can form a target frame, this to judge two angle points whether belong to same object algorithm requirement it is very high, one
But poor accuracy is a bit, will generate many false target frames.Secondly, exactly this algorithm is defective.Because algorithm is being sentenced thus
When whether disconnected two angle points belong to same object, lack the auxiliary of global information, therefore is easy to script not be same object
Two angle points regard a pair of of angle point as, therefore produce many false target frames.Finally, the feature of angle point is quicker to edge
Sense, it is equally very sensitive to the edge of background that this leads to many angle points, therefore the angle point of mistake has been also detected that at background.To sum up
Reason, so that CornerNet produces many erroneous detections.
It is next exactly to find out solution, critical issue is that network is allowed to have after the problem of having analyzed CornerNet
Fully feel the ability for knowing interior of articles information.One method for being easier to expect is CornerNet to be become a two-stage inspection
Survey method extracts prediction block using area-of-interest pond (RoI pooling) or area-of-interest matching (RoI align)
Internal information, to obtain sensing capability.But it is very big to do so expense, therefore we have proposed examined with key point triple
Survey target, so that our method can obtain the ability of perception interior of articles information under the premise of one-stage.And
And expense is smaller, because we need to only pay close attention to the center of object, pays close attention to object so as to avoid RoI pooling or RoI align
The all information in internal portion.
Network passes through central point pond (center pooling) and cascade angle point pond (cascade corner
Pooling central point thermal map (center heatmap) and angle point thermal map (corner heatmap) are respectively obtained) for predicting
The position of key point.After obtaining position and the classification of angle point, the position of angle point is mapped to by input figure by offset (offsets)
Then the corresponding position of piece judges which two angle point belongs to the same object by being embedded in (embedings), to form one
Detection block.It is described as discussed above, due to lacking the auxiliary from target area internal information in anabolic process, so as to cause a large amount of
Erroneous detection.In order to solve this problem, the algorithm that we use not only predicts angle point, also prediction central point.We are to each pre-
An adopted central area is confined in survey, by judging whether the central area of each target frame contains central point, is retained if having, and
And point centered on the confidence level (confidence) of this time-frame, the confidence level of upper left angle point and bottom right angle point are averaged, if without if
Removal, so that network has the ability of perception target area internal information, it can be effectively except the target frame of mistake.
It was found that the scale of central area will affect wrong frame removal effect.The too small mistake for leading to many small scales in central area
Accidentally target frame can not be removed, and the excessive false target frame for leading to many large scales in central area can not be removed, therefore I
Propose the adjustable central area of scale and define method.This method can define one relatively when the scale of prediction block is larger
Lesser central area predicts a relatively large central area when the scale of prediction block is smaller.
Central point pond (Center pooling): the center of an object might not be easily distinguishable containing very strong
In the semantic information of other classifications.For example, very strong, the easily distinguishable semanteme letter in other classifications is contained on the head of a people
Breath, but its center is often positioned in the middle part of people.The invention proposes central point Chi Hualai to enrich center point feature.Central point pond
Change the maximum value extracted central point horizontally and vertically and addition, with this to the letter other than central point offer present position
Breath.This operation makes central point have an opportunity to obtain the semantic information for being easier to distinguish over other classifications.Central point pondization can pass through
The combination in the angle point pond (corner pooling) on different directions is realized.The operation that is maximized in one horizontal direction can
It is realized by left pool area (left pooling) and right pool area (right pooling) by series connection, similarly, one hangs down
The upward operation that is maximized of histogram can be by upper pool area (top pooling) and lower pool area (bottom pooling)
It is realized by series connection.
Cascade angle point pond (cascade corner pooling): angle point is located at outside object under normal circumstances, locating
Position and the semantic information for not containing associate, this brings difficulty for the detection of angle point.Traditional method, referred to as angle point pond
(corner pooling).It extracts object boundary maximum value and is added, and this method can only provide associate edge semanteme letter
Breath, for interior of articles semantic information more abundant, then it is difficult to extract arrive.Cascade angle point pond principle: it extracts object first
Then boundary maximum value continues internally to propose maximum value at the maximum value of boundary, and is added with boundary maximum value, given with this
Corner feature provides associate semantic information more abundant.Cascade angle point pondization can also pass through the angle point pond on different directions
The combination of change is realized.
With the continuous maturation of software and hardware product, monitored using respective depth learning framework aerial optical cable laying state at
For may, which greatly improves production efficiency, belongs to edge calculations, artificial intelligence, Internet technology to fusion
Software and hardware integration product (can also establish complete system), and device configuration includes:
1, software frame: using based on CenterNet frame model under Pytorch;
2, artificial intelligence model training: the typical scene based on city status, autonomous training are completed;
3, hardware forms:
(1) edge calculations product: the Jetson series or corresponding product of NviDIA.The modules such as GPS, 4G communication can be integrated;
(2) video acquisition: the hard of equipment and edge calculations can be acquired by vehicle-mounted camera, mobile phone, law-enforcing recorder etc.
The connection of part server, can also install camera module in Jetson system;
(3) GPS geography information: it may be from the GPS information of the equipment such as automobile data recorder, mobile phone, law-enforcing recorder, pass through
The hardware server interactive information of interface and edge calculations can also install additional GPS module in Jetson system.
4, internet return path: calculating passback abnormal point image/video information and gps coordinate in real time using wireless public network,
Complete video of making an inspection tour is offline preservation filing.
It by the image interception to vehicle-mounted camera line walking video, obtains image information per second and is marked, form phase
The data set answered recycles key point triple detection object method to be trained, obtains corresponding data model, realizes to patrolling
The identification of line video.
In conclusion method provided by the present invention has the ability of perception interior of articles information, so as to effectively press down
Erroneous detection processed;Identification is horizontal horizontal consistent with naked eyes identification, it is ensured that the abnormal conditions of aerial optical cable are accurately differentiated.
It is discussed in detail although the contents of the present invention have passed through above preferred embodiment, but it should be appreciated that above-mentioned
Description is not considered as limitation of the present invention.After those skilled in the art have read above content, for of the invention
A variety of modifications and substitutions all will be apparent.Therefore, protection scope of the present invention should be limited to the appended claims.
Claims (8)
1. a kind of vehicle-mounted camera line walking image-recognizing method, which is characterized in that pass through the figure to vehicle-mounted camera line walking video
It as interception, obtains image information per second and is marked, form corresponding data set, recycle key point triple detection object
Method is trained, and obtains corresponding data model, realizes the identification to line walking image.
2. vehicle-mounted camera line walking image-recognizing method as described in claim 1, which is characterized in that the key point triple
Detection object method determines a target by three central point, upper left angle point and bottom right angle point key points.
3. vehicle-mounted camera line walking image-recognizing method as claimed in claim 2, which is characterized in that utilize upper left and bottom right two
A angle point generates initial target frame, defines a central area to each prediction block, then judges the center of each target frame
Whether domain contains central point, and the target frame is retained if having, if without the target frame is deleted.
4. vehicle-mounted camera line walking image-recognizing method as claimed in claim 3, which is characterized in that prediction block scale compared with
A relatively small central area is defined when big, and a relatively large center is predicted when the scale of prediction block is smaller
Domain.
5. vehicle-mounted camera line walking image-recognizing method as claimed in claim 4, which is characterized in that pass through central point Chi Huati
Central point maximum value horizontally and vertically and addition are taken, the information other than present position is provided with this to central point.
6. vehicle-mounted camera line walking image-recognizing method as claimed in claim 5, which is characterized in that in a horizontal direction
It is maximized operation and is realized by left pool area and right pool areaization by series connection, similarly, take maximum in a vertical direction
Value Operations are realized by upper pool area and lower pool areaization by series connection.
7. vehicle-mounted camera line walking image-recognizing method as claimed in claim 5, which is characterized in that by cascading angle point pond
Object boundary maximum value is extracted first, then continues internally to propose maximum value at the maximum value of boundary, and maximum with boundary
Value is added, and provides associate semantic information to corner feature with this.
8. vehicle-mounted camera line walking image-recognizing method as claimed in claim 7, which is characterized in that cascade angle point pond passes through
The combination in the angle point pond on different directions is realized.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910497061.6A CN110210436A (en) | 2019-06-10 | 2019-06-10 | A kind of vehicle-mounted camera line walking image-recognizing method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910497061.6A CN110210436A (en) | 2019-06-10 | 2019-06-10 | A kind of vehicle-mounted camera line walking image-recognizing method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110210436A true CN110210436A (en) | 2019-09-06 |
Family
ID=67791821
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910497061.6A Pending CN110210436A (en) | 2019-06-10 | 2019-06-10 | A kind of vehicle-mounted camera line walking image-recognizing method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110210436A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110765906A (en) * | 2019-10-12 | 2020-02-07 | 上海雪湖科技有限公司 | Pedestrian detection algorithm based on key points |
CN112347864A (en) * | 2020-10-21 | 2021-02-09 | 北京市地铁运营有限公司地铁运营技术研发中心 | Method, device, equipment and system for sensing and inducing rail transit passenger flow |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102147971A (en) * | 2011-01-14 | 2011-08-10 | 赵秀江 | Traffic information acquisition system based on video image processing technology |
CN107729958A (en) * | 2017-09-06 | 2018-02-23 | 新华三技术有限公司 | A kind of method for sending information and device |
-
2019
- 2019-06-10 CN CN201910497061.6A patent/CN110210436A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102147971A (en) * | 2011-01-14 | 2011-08-10 | 赵秀江 | Traffic information acquisition system based on video image processing technology |
CN107729958A (en) * | 2017-09-06 | 2018-02-23 | 新华三技术有限公司 | A kind of method for sending information and device |
Non-Patent Citations (1)
Title |
---|
KAIWEN DUAN 等: "CenterNet: Keypoint Triplets for Object Detection", 《ARXIV》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110765906A (en) * | 2019-10-12 | 2020-02-07 | 上海雪湖科技有限公司 | Pedestrian detection algorithm based on key points |
CN112347864A (en) * | 2020-10-21 | 2021-02-09 | 北京市地铁运营有限公司地铁运营技术研发中心 | Method, device, equipment and system for sensing and inducing rail transit passenger flow |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111126152B (en) | Multi-target pedestrian detection and tracking method based on video | |
CN102542289B (en) | Pedestrian volume statistical method based on plurality of Gaussian counting models | |
WO2021170030A1 (en) | Method, device, and system for target tracking | |
CN104517095B (en) | A kind of number of people dividing method based on depth image | |
CN110516636A (en) | A kind of monitoring method of process, device, computer equipment and storage medium | |
CN109460709A (en) | The method of RTG dysopia analyte detection based on the fusion of RGB and D information | |
US11748896B2 (en) | Object tracking method and apparatus, storage medium, and electronic device | |
CN103824070A (en) | Rapid pedestrian detection method based on computer vision | |
CN104601964A (en) | Non-overlap vision field trans-camera indoor pedestrian target tracking method and non-overlap vision field trans-camera indoor pedestrian target tracking system | |
CN105306892B (en) | A kind of generation of ship video of chain of evidence form and display methods | |
CN101022505A (en) | Method and device for automatically detecting moving target under complex background | |
CN105760846A (en) | Object detection and location method and system based on depth data | |
CN113284144B (en) | Tunnel detection method and device based on unmanned aerial vehicle | |
CN113920097B (en) | Power equipment state detection method and system based on multi-source image | |
CN112434566B (en) | Passenger flow statistics method and device, electronic equipment and storage medium | |
CN112132157B (en) | Gait face fusion recognition method based on raspberry pie | |
CN116434159A (en) | Traffic flow statistics method based on improved YOLO V7 and Deep-Sort | |
CN112132873A (en) | Multi-lens pedestrian recognition and tracking based on computer vision | |
CN110210436A (en) | A kind of vehicle-mounted camera line walking image-recognizing method | |
CN112465854A (en) | Unmanned aerial vehicle tracking method based on anchor-free detection algorithm | |
KR20150042544A (en) | Mobile terminal for providing location information, method and system for measuring the location information | |
CN113780145A (en) | Sperm morphology detection method, sperm morphology detection device, computer equipment and storage medium | |
CN109711232A (en) | Deep learning pedestrian recognition methods again based on multiple objective function | |
KR101313879B1 (en) | Detecting and Tracing System of Human Using Gradient Histogram and Method of The Same | |
CN113743380B (en) | Active tracking method based on video image dynamic monitoring |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190906 |