CN107944392A - A kind of effective ways suitable for cell bayonet Dense crowd monitor video target mark - Google Patents

A kind of effective ways suitable for cell bayonet Dense crowd monitor video target mark Download PDF

Info

Publication number
CN107944392A
CN107944392A CN201711197741.3A CN201711197741A CN107944392A CN 107944392 A CN107944392 A CN 107944392A CN 201711197741 A CN201711197741 A CN 201711197741A CN 107944392 A CN107944392 A CN 107944392A
Authority
CN
China
Prior art keywords
video
image
boundary rectangle
foreground
algorithm
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
Application number
CN201711197741.3A
Other languages
Chinese (zh)
Inventor
周晓风
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xiao Feng Zhou
Original Assignee
Xiao Feng Zhou
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xiao Feng Zhou filed Critical Xiao Feng Zhou
Priority to CN201711197741.3A priority Critical patent/CN107944392A/en
Publication of CN107944392A publication Critical patent/CN107944392A/en
Pending legal-status Critical Current

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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10016—Video; Image sequence
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30232—Surveillance

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)

Abstract

The present invention provides a kind of effective ways suitable for cell bayonet Dense crowd monitor video target mark.It is main to include with lower part：Video is pre-processed, video frame is processed into and compares clearly image；The foreground target of video image is extracted using ASBL HL prospects modeling algorithm and carries out holes filling；The target object in non-bayonet region is filtered out using line design is stumbled；Carry out target positioning and target type discrimination using the regional area of SSD (the Single Shot Multibox Detector) video frame obtained to each foreground extraction, then by video regional area target detection to object coordinates be mapped in whole pictures.Eventually form and the target of whole video frame is marked.It is an advantage of the current invention that Utilization prospects extraction algorithm ASBL HL algorithms individually extract the local dynamic station content of image, and the high accuracy using SSD to small size picture target detection, realize advantage combination.Finally improve the precision of video frame object identification positioning.

Description

It is a kind of to be suitable for the effective of cell bayonet Dense crowd monitor video target mark Method
Technical field
The present invention relates to deep learning, pattern-recognition, image processing field, more particularly to one kind to be suitable for cell bayonet height The effective ways of density population surveillance video object mark.
Technical background
With the development of society, a large amount of populations are poured in city, construction and the treatment status in city become particularly severe, greatly The criminal activity of amount and public security destabilizing factor are flooded with each corner of society.So that the safety and happiness of people, community Public security, which is stablized, receives great challenge.Each crucial bayonet of the security administration departments at different levels in community installs monitoring camera, To monitor community policy threat.The prevention implemented to the criminal offence for endangering social security in advance becomes very with subsequent tracing It is important.Monitoring system software and hardware is quickly grown, and the specialized company in large quantities of safety-security areas establishes, and forms in the world the most It is huge, technology security-protection management system the most complete.
In numerous security protection scenes, the most common monitoring system for surely belonging to cell bayonet, it is with the daily life work of people Make closely bound up with study.So ensureing that the monitor video of cell bayonet accurately efficiently works becomes extremely important.Traditional Computer vision field is powerless to object segmentation and target mark performance.Nowadays, the depth to grow up like the mushrooms after rain Study causes extensive concern in academia and industrial quarters, and deep learning is almost flooded with each corner of Chinese society, One learning climax is worldwide started.It is applied to the skill in the image field such as classification, image recognition, target positioning more Art achieves higher achievement, is subject to come from global praise.
Problem is marked for the target in cell bayonet security protection scene.Its feature is that the target classification of cell bayonet is answered It is miscellaneous, mobility is strong, dbjective state is diversified.Traditional machine learning is not enough to deal with these scenes.The mesh of deep learning method Calibration position identification technology compensate for deficiency in this respect.So that the work of Multi-target position identification becomes simple.SSD(Single Shot Multibox Detector) as state-of-art target positioning with sorting technique order is showed in bayonet scene People attractes attention.But have naturally to wisp in video frame larger in monitoring image and globe detection identification for SSD300 Weakness, often there are the situations such as missing inspection.Therefore, the Small object in bayonet scene and intensive target detection precision are urgently carried Rise.
The content of the invention
The present invention provides a kind of effective ways suitable for cell bayonet Dense crowd monitor video target mark. When carrying out SSD marks to larger image, its difficult point marked is become to the mark of Small object and intensive target.The present invention It is to reduce sight for foreground extraction with reference to the purpose of foreground extraction and SSD (Single Shot Multibox Detector) Examine the advantage that region and SSD are positioned, classify over small areas, formation one can also obtain larger image that precision is higher to determine Position categorizing system.The technical solution is as follows：
A kind of effective ways suitable for cell bayonet Dense crowd monitor video target mark, comprise the following steps：
Step 1：A frame of video is read, the enhancing pretreatment of defogging clarity is carried out to picture, is mainly realized to the greasy weather Image is strengthened.
Step 2：Delimitation is stumbled line, is drawn at cell entrance and is taken line of stumbling.
Step 3：Use ASBL-HF (Adaptive Selective Background Learning-Hole Filling Algorithm) algorithm realizes the extraction of display foreground.Initialisation image first, recycles the background model of former frame and current Frame calculates the model of present frame, and finally cavity in picture is filled using empty filling algorithm.Formation is contained at this time The bianry image of foreground and background, white are represented with 1, and black is represented with 0, and are less than setting threshold to image area using threshold value The image of value is filtered.
Step 4：All foreground areas in obtaining step 3, and these regions are confined using boundary rectangle.And And judge whether the position of the boundary rectangle of foreground area with line of stumbling has overlapping region, if overlapping region then retains the area Domain.
Step 5：Each area of step 4 acquisition is directed to using SSD (Single Shot Multibox Detector) algorithm Domain carries out the object boundary rectangle coordinate setting of subgraph, with the other identification of object type.
Step 6：The boundary rectangle of all objects of subgraph mark is mapped in entire image.Form final view picture The fixation and recognition result of image.
In some embodiments, wherein step 1 carries out defogging clarity enhancing to each frame in video, uses enhancing Algorithm strengthens the cell monitoring video for having mist.
Its advantage is：It is clearer that the video for having mist in video using defogging sharpness enhancement is reached into image Effect.
In some embodiments, in its step 2, setting bayonet is stumbled line so that the endpoint for line both sides of stumbling is the two of bayonet Side.
It has the advantages that：The object in other regions in video is ignored, ensures that the detection interest region of this algorithm exists At bayonet.
In some embodiments, in its step 3, ASBL-HF (Adaptive Selective Background are used Learning-Hole Filling Algorithm) algorithm, obtain video image prospect.Initialize picture first:
Wherein, B represents background, and l represents video length, and z represents the frame in video, and formula represents to seek all video frame Average value
Here this two field picture of t moment is given in the background model model that its coordinate is (i, j), such as following formula：
Bt(i, j)=(1- α) Bt(i,j)+αZt(i,j)
Wherein BtRepresent the background model of t moment, α represents learning rate, takes 0 or 1 here.Formula represents present frame background mould The renewal of type is together decided on by the background model and present frame of former frame.
Finally plus an empty filling algorithm, empty algorithm is gone the background area around target to be got rid of.This Process is also noise remove, such as following formula：
In above formula, A represents the border set of a connection, X0It is a subset of A, B is a criss-cross 3x3 filtering Device.C represents supplementary set.The purpose of this formula is short covering hole.
ASBL-HF initializes all pictures first, and for a certain frame therein, its background model is former frame background mould Type is together decided on present frame, finally uses empty filling algorithm so that aperture inside picture is eliminated.It is secondary complete to form one Whole prospect background and chart.
Acquired foreground target is screened, foregrounding region area then retains this more than a given threshold value Block region；And filter out the foreground area less than given threshold.Its minimum area threshold value is arranged to：
Athreshold=500
It has the advantages that：ASBL-HF algorithms can obtain the black and white binary map without cavity of prospect and background separation Picture.And usable floor area threshold filtering is fallen less than defined threshold.
In some embodiments, in its step 4, its boundary rectangle is obtained to acquired foreground area.And to being obtained The boundary rectangle obtained is screened, and pair boundary rectangle for having intersection with line of stumbling retains.Equipped with a boundary rectangle, its upper left corner Coordinate is (x1,y1), bottom right angular coordinate is (x2,y2).The left end coordinate of line of stumbling is (xl,yl), the right end coordinate for line of stumbling is (xr, yr), then retain the boundary rectangle if boundary rectangle meets the following conditions：
yl=yr∈(y1,y2) and x0∈(xl,xr)∪x1∈(xl,xr)
It has the advantages that：By stumbling, line filters the boundary rectangle for being not belonging to bayonet.Ensure only detection bayonet The target object at place.
In some embodiments, SSD (Single Shot Multibox Detector) is used in its step 5 to institute The subgraph of acquisition is labeled respectively.Four vertex of object boundary rectangle of acquisition are mapped in entire image.If subgraph As the coordinate of foreground target upper left side is (x, y).
The boundary rectangle confined equipped with SSD in a subgraph, its top-left coordinates are (xlu,ylu), lower right coordinate is (xrd, yrd), then the top-left coordinates that subgraph is mapped to entire image are (x+xlu,y+ylu), lower right side coordinate is (x+xrd,y+yrd)。
It has the advantages that：Each target boundary rectangle fixed-point computation of all subgraphs is come out by SSD algorithms, And identify its type.And target in subgraph is mapped to entire image.
After coordinate and classification are respectively mapped to entire image, that is, form each object coordinates and classification in final image.
The beneficial effects of the present invention are：Cleverly combine the high-precision of positioning and identification of the SSD to sub- objects in images Degree and background extracting obtain the local region of interest in big image.Make figure by strengthening pretreatment to the defogging clarity of picture As being apparent from.By acquisition of the ASBL-HF foreground extractions algorithm to moving target at bayonet in video, sense is greatly reduced The region of interest so that the strong point of SSD is played.And use bundle of lines of stumbling wants detection zone to be limited at cell bayonet. With reference to any of the above algorithm and skill, relatively good bayonet object mark effect is finally reached.It can meet more accurate small The target mark of area's bayonet monitor video.
Brief description of the drawings：
Fig. 1 is the algorithm steps flow chart of the present invention；
Fig. 2 is the algorithm flow chart of foreground extraction algorithm in the present invention.
Embodiment
With reference to specific implementation, the present invention is described further, but protection scope of the present invention is not limited in This：
Embodiment is as shown in drawings, a kind of to have efficacious prescriptions suitable for cell bayonet Dense crowd monitor video target mark Method.It comprises the following steps：
(1) video image defogging clarity strengthens：
Each two field picture in cell bayonet video is obtained, and defogging clarity enhancing processing is carried out to each two field picture.
(2) setting bayonet is stumbled line：
Video is opened, and is parked on the first frame of video, takes one end of bayonet in video using mouse point, and pull to another End.Form a line of stumbling lain across before bayonet.
(3) object in extraction prospect：
Use ASBL-HF (Adaptive Selective Background Learning-Hole Filling Algorithm) algorithm carries out foreground extraction to video frame, initializes picture first:
Wherein, B represents background, and l represents video length, and z represents the frame in video, and formula represents to seek all video frame Average value
Here this frame picture of t moment is given in the background model model that coordinate is (i, j), such as following formula：
Bt(i, j)=(1- α) Bt(i,j)+αZt(i,j)
Wherein BtRepresent the background model of t moment, α represents learning rate, takes 0 or 1 here.Formula represents present frame background mould The renewal of type is together decided on by the background model and present frame of former frame.
Finally plus an empty filling algorithm, empty algorithm is gone the background area around target to be got rid of.This Process is also noise remove, such as following formula：
In above formula, A represents the border set of a connection, X0It is a subset of A, B is a criss-cross 3x3 filtering Device.C represents supplementary set.The purpose of this formula is short covering hole.
To acquired foreground target, prospect error caused by tiny change is filtered out first by a threshold value, This threshold value value is 500 pixels.Draw the foreground target of picture in its entirety later using pixel threshold.
Its boundary rectangle is obtained to acquired foreground area.And the boundary rectangle to being obtained screens, with line of stumbling The boundary rectangle for having intersection then retains.Equipped with a boundary rectangle, its top left co-ordinate is (x1,y1), bottom right angular coordinate is (x2, y2).The left end coordinate of line of stumbling is (xl,yl), the right end coordinate for line of stumbling is (xr,yr), then if boundary rectangle meets the following conditions Then retain the boundary rectangle：
yl=yr∈(y1,y2) and x0∈(xl,xr)∪x1∈(xl,xr)。
(4) using VOC data training SSD, and subgraph is detected with SSD
SSD networks are trained using VOC2007 and VOC2012 data sets.VOC data are mainly comprising common in life Object.Such as：Pedestrian, vehicle, bicycle, motorcycle, truck, bus etc..Meet wanting for cell bayonet target detection Ask.
Training parameter is set：Adjustment picture size is 300x300, and basic learning rate is 0.0004*20, and learning strategy is Multi-step, total iterations are 120000 times, multiply learning rate when 8000,100000,120000 iteration With 1/10.And the training on the work station of gpu model GTX TITAN X.Draw last caffemodel training patterns.
The model come is trained to acquired subgraph using SSD (Single Shot Multibox Detector) The coordinate setting of object boundary rectangle and the identification of object classification are carried out respectively.The object boundary rectangle and classification of acquisition are mapped to whole In width image.If the coordinate of subgraph foreground target upper left side is (x, y).
(5) map the coordinate on subgraph and class is clipped to entire image
The boundary rectangle confined equipped with a SSD, its top-left coordinates are (xlu,ylu), lower right coordinate is (xrd,yrd), then it is sub The lower right coordinate that image is mapped to entire image is (x+xlu,y+ylu), and lower right side coordinate is (x+xrd,y+yrd), the two structure Into total coordinate.
Draw the target boundary rectangle coordinate gone out in entire image bayonet and target classification at this time.
The foregoing is merely illustrative of the preferred embodiments of the present invention.It is not intended to limit the invention, all spirit of the invention With all any modification, equivalent and improvement done within principle etc., it should be included within the scope of the present invention.

Claims (8)

  1. A kind of 1. effective ways suitable for cell bayonet Dense crowd monitor video target mark.It is characterized by comprising with Lower step：
    Step 1：A frame video is read, this video frame is pre-processed, main realize increases the image in greasy weather into line definition By force；
    Step 2：Delimitation is stumbled line, is drawn in cell entrance and is taken line of stumbling；
    Step 3：Use ASBL-HF (Adaptive Selective Background Learning-Hole Filling Algorithm) algorithm realizes the extraction of display foreground and fills the hole region of foreground image.The foreground and background shape extracted Into a bianry image, white is represented with 1, and black is represented with 0, is less than the very small region of setting value to image area using threshold value Filtered；
    Step 4：All foreground areas in obtaining step 3, and these regions are confined using boundary rectangle.And sentence Whether the position of the boundary rectangle of disconnected foreground area with line of stumbling has intersection, and the boundary rectangle is then preserved if overlapping；
    Step 5：Each external square of step 4 acquisition is directed to using SSD (Single Shot Multibox Detector) algorithm Subgraph determined by shape carries out subgraph object boundary rectangle and the classification of subgraph object continues to position and identifies；
    Step 6：Object boundary rectangle coordinate and classification determined by subgraph are mapped in entire image.Form final pair The annotation results of entire image.
  2. 2. it is according to claim 1 it is a kind of suitable for cell bayonet Dense crowd monitor video target mark have efficacious prescriptions Method, it is characterised in that in the step 1, strengthen definition algorithm using defogging, to there is the video of mist to carry out defogging enhancing, increase Add image definition.
  3. 3. it is according to claim 1 it is a kind of suitable for cell bayonet Dense crowd monitor video target mark have efficacious prescriptions Method, it is characterised in that in the step 2, set the line of stumbling being buckled out, video card is disposed across using 2 points of one determined line segments At mouthful.
  4. 4. it is according to claim 1 it is a kind of suitable for cell bayonet Dense crowd monitor video target mark have efficacious prescriptions Method, it is characterised in that in the step 3, use ASBL-HF (Adaptive Selective Background Learning- Hole Filling Algorithm) algorithm, obtain video frame prospect.Image is initialized first:
    <mrow> <mi>B</mi> <mo>=</mo> <mfrac> <mn>1</mn> <mi>l</mi> </mfrac> <msubsup> <mi>&amp;Sigma;</mi> <mrow> <mi>t</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>l</mi> </msubsup> <msub> <mi>z</mi> <mi>t</mi> </msub> </mrow>
    Wherein, B represents background, and l represents video length, and z represents the frame in video, which represents to seek the flat of all video frame Average
    Here this frame picture of t moment is given in the background model model that coordinate is (i, j), such as following formula：
    Bt(i, j)=(1- α) Bt(i,j)+αZt(i,j)
    Wherein BtRepresent the background model of t moment, α represents learning rate, takes 0 or 1 here.Formula represents present frame background model Renewal is together decided on by the background model and present frame of former frame.
    It is finally an empty filling algorithm, goes empty algorithm the background area around target can be got rid of.This process Also noise remove, such as following formula are：
    <mrow> <msub> <mi>X</mi> <mi>n</mi> </msub> <mo>=</mo> <mrow> <mo>(</mo> <msub> <mi>X</mi> <mrow> <mi>n</mi> <mo>-</mo> <mn>1</mn> </mrow> </msub> <mo>&amp;CirclePlus;</mo> <mi>B</mi> <mo>)</mo> </mrow> <mo>&amp;cap;</mo> <msup> <mi>A</mi> <mi>c</mi> </msup> <mo>,</mo> <mi>n</mi> <mo>=</mo> <mn>1</mn> <mo>,</mo> <mn>2</mn> <mo>,</mo> <mn>3</mn> <mo>,</mo> <mo>...</mo> </mrow>
    In above formula, A represents the border set of a connection, X0It is a subset of A, B is a criss-cross 3x3 wave filter.c Represent supplementary set.The purpose of this formula is short covering hole.
  5. 5. it is according to claim 1 it is a kind of suitable for cell bayonet Dense crowd monitor video target mark have efficacious prescriptions Method, it is characterised in that in the step 3, screened to acquired foreground target, foregrounding region area is more than one Given threshold value then retains this block region；And the foreground area less than the given threshold is filtered out, which is arranged to 500 pictures Plain (whole picture size is 856x480).
  6. 6. it is according to claim 1 it is a kind of suitable for cell bayonet Dense crowd monitor video target mark have efficacious prescriptions Method, it is characterised in that in the step 4, its boundary rectangle is obtained to acquired foreground area.And the external square to being obtained Shape is screened, and pair boundary rectangle for having intersection with line of stumbling retains.Equipped with a boundary rectangle, its top-left coordinates is (x1, y1), lower right coordinate is (x2,y2).The left end coordinate of line of stumbling is (xl,yl), the right end coordinate for line of stumbling is (xr,yr), then it is if outer Connect rectangle and meet the following conditions：
    yl=yr∈(y1,y2) and x0∈(xl,xr)∪x1∈(xl,xr),
    Then retain this boundary rectangle, remaining boundary rectangle abandons.
  7. 7. it is according to claim 1 it is a kind of suitable for cell bayonet Dense crowd monitor video target mark have efficacious prescriptions Method, it is characterised in that in the step 5, using SSD (Single Shot Multibox Detector) to acquired son Image carries out the positioning of object object boundary rectangle and type identification respectively.Acquired object boundary rectangle is mapped to view picture figure As in.If subgraph foreground target boundary rectangle top left co-ordinate is (x, y), the subgraph equipped with a SSD institutes fixation and recognition Boundary rectangle, its top-left coordinates are (xlu,ylu), lower right coordinate is (xrd,yrd), then subgraph is mapped to the upper left of entire image Coordinate is (x+xlu,y+ylu), lower right side coordinate is (x+xrd,y+yrd)。
  8. 8. it is according to claim 1 it is a kind of suitable for cell bayonet Dense crowd monitor video target mark have efficacious prescriptions Method, it is characterised in that Utilization prospects extraction algorithm ASBL-HF obtains subgraph and carries out SSD to subgraph, is finally mapped to whole Width image.Mainly include the following steps that：
    8.1 are strengthened video image using defogging algorithm.
    8.2 setting bayonets are stumbled line.
    8.3 use ASBL-HF algorithms, obtain the bianry image containing foreground and background.
    Filtered using given threshold in 8.4 pairs of less regions of prospect.
    8.5 obtain the boundary rectangle of bianry image, and filtering is defined using line of stumbling.
    8.6 obtain the SSD annotation results of subgraph, and map coordinates to entire image.Form each thing in entire image Body coordinate and classification.
CN201711197741.3A 2017-11-25 2017-11-25 A kind of effective ways suitable for cell bayonet Dense crowd monitor video target mark Pending CN107944392A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201711197741.3A CN107944392A (en) 2017-11-25 2017-11-25 A kind of effective ways suitable for cell bayonet Dense crowd monitor video target mark

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201711197741.3A CN107944392A (en) 2017-11-25 2017-11-25 A kind of effective ways suitable for cell bayonet Dense crowd monitor video target mark

Publications (1)

Publication Number Publication Date
CN107944392A true CN107944392A (en) 2018-04-20

Family

ID=61949803

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201711197741.3A Pending CN107944392A (en) 2017-11-25 2017-11-25 A kind of effective ways suitable for cell bayonet Dense crowd monitor video target mark

Country Status (1)

Country Link
CN (1) CN107944392A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108710897A (en) * 2018-04-24 2018-10-26 江苏科海智能系统有限公司 A kind of online general target detecting system in distal end based on SSD-T
CN109285181A (en) * 2018-09-06 2019-01-29 百度在线网络技术（北京）有限公司 The method and apparatus of image for identification
CN109726691A (en) * 2018-12-30 2019-05-07 杭州铭智云教育科技有限公司 A kind of monitoring method
CN110490073A (en) * 2019-07-15 2019-11-22 浙江省北大信息技术高等研究院 Object detection method, device, equipment and storage medium
CN110740231A (en) * 2019-09-27 2020-01-31 浙江省北大信息技术高等研究院 Video data labeling method and device, electronic equipment and medium

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1607835A (en) * 2003-09-30 2005-04-20 英特尔公司 Regular-shape motion search
CN102170563A (en) * 2011-03-24 2011-08-31 杭州海康威视软件有限公司 Intelligent person capture system and person monitoring management method
CN104331905A (en) * 2014-10-31 2015-02-04 浙江大学 Surveillance video abstraction extraction method based on moving object detection
CN105245853A (en) * 2015-10-27 2016-01-13 太原市公安局 Video monitoring method
CN106126680A (en) * 2016-06-29 2016-11-16 北京互信互通信息技术有限公司 A kind of video image reconnaissance method and system
CN106600581A (en) * 2016-12-02 2017-04-26 北京航空航天大学 Train operation fault automatic detection system and method based on binocular stereoscopic vision

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1607835A (en) * 2003-09-30 2005-04-20 英特尔公司 Regular-shape motion search
CN102170563A (en) * 2011-03-24 2011-08-31 杭州海康威视软件有限公司 Intelligent person capture system and person monitoring management method
CN104331905A (en) * 2014-10-31 2015-02-04 浙江大学 Surveillance video abstraction extraction method based on moving object detection
CN105245853A (en) * 2015-10-27 2016-01-13 太原市公安局 Video monitoring method
CN106126680A (en) * 2016-06-29 2016-11-16 北京互信互通信息技术有限公司 A kind of video image reconnaissance method and system
CN106600581A (en) * 2016-12-02 2017-04-26 北京航空航天大学 Train operation fault automatic detection system and method based on binocular stereoscopic vision

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
MACHMUD R ALHAMIDI ET AL.: "An Adaptive Selective Background Learning-Hole Filling Algorithm to Improve Vehicle Detection", 《2015 INTERNATIONAL CONFERENCE ON ADVANCED COMPUTER SCIENCE AND INFORMATION SYSTEMS (ICACSIS)》 *
乔延军: "面向户外增强现实的地理实体目标识别与检测", 《中国优秀硕士学位论文全文数据库 信息科技辑》 *
杨阿丽 等: "用于交通治安卡口的全天候视频车辆检测方法", 《合肥工业大学学报（自然科学版）》 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108710897A (en) * 2018-04-24 2018-10-26 江苏科海智能系统有限公司 A kind of online general target detecting system in distal end based on SSD-T
CN109285181A (en) * 2018-09-06 2019-01-29 百度在线网络技术（北京）有限公司 The method and apparatus of image for identification
CN109726691A (en) * 2018-12-30 2019-05-07 杭州铭智云教育科技有限公司 A kind of monitoring method
CN109726691B (en) * 2018-12-30 2020-12-04 安徽润谷科技有限公司 Monitoring method
CN110490073A (en) * 2019-07-15 2019-11-22 浙江省北大信息技术高等研究院 Object detection method, device, equipment and storage medium
CN110740231A (en) * 2019-09-27 2020-01-31 浙江省北大信息技术高等研究院 Video data labeling method and device, electronic equipment and medium

Similar Documents

Publication Publication Date Title
CN107944392A (en) A kind of effective ways suitable for cell bayonet Dense crowd monitor video target mark
CN108171104B (en) Character detection method and device
WO2018145470A1 (en) Image detection method and device
CN104299008B (en) Vehicle type classification method based on multi-feature fusion
CN106600625A (en) Image processing method and device for detecting small-sized living thing
CN107122777A (en) A kind of vehicle analysis system and analysis method based on video file
CN103034852B (en) The detection method of particular color pedestrian under Still Camera scene
CN109918971B (en) Method and device for detecting number of people in monitoring video
CN105404857A (en) Infrared-based night intelligent vehicle front pedestrian detection method
CN107622258A (en) A kind of rapid pedestrian detection method of combination static state low-level image feature and movable information
CN107680095A (en) The electric line foreign matter detection of unmanned plane image based on template matches and optical flow method
CN108960185A (en) Vehicle target detection method and system based on YOLOv2
CN111832461B (en) Method for detecting wearing of non-motor vehicle riding personnel helmet based on video stream
CN104751142A (en) Natural scene text detection algorithm based on stroke features
CN102663354A (en) Face calibration method and system thereof
CN105740945A (en) People counting method based on video analysis
CN104915642B (en) Front vehicles distance measuring method and device
CN104463134B (en) A kind of detection method of license plate and system
CN105869174B (en) A kind of Sky Scene image partition method
CN105956552A (en) Face black list monitoring method
CN102819733A (en) Rapid detection fuzzy method of face in street view image
CN109255326A (en) A kind of traffic scene smog intelligent detecting method based on multidimensional information Fusion Features
CN112115878B (en) Forest fire smoke root node detection method based on smoke area density
CN107578048A (en) A kind of long sight scene vehicle checking method based on vehicle rough sort
CN112270330A (en) Intelligent detection method for concerned target based on Mask R-CNN neural network

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
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20180420

WD01 Invention patent application deemed withdrawn after publication