CN109376677A - A kind of video behavior detection method merged based on shape-movement double fluid information - Google Patents
A kind of video behavior detection method merged based on shape-movement double fluid information Download PDFInfo
- Publication number
- CN109376677A CN109376677A CN201811298485.1A CN201811298485A CN109376677A CN 109376677 A CN109376677 A CN 109376677A CN 201811298485 A CN201811298485 A CN 201811298485A CN 109376677 A CN109376677 A CN 109376677A
- Authority
- CN
- China
- Prior art keywords
- behavior
- frame
- video
- behavioral
- behavioral chain
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of based on shape-movement double fluid information fusion video behavior detection method.This method extracts shape and motion information feature using depth network model, this two parts feature is carried out the fusion of depth by building convolutional network, carries out the video behavioral value of single frames on this basis；The link algorithm of dynamic increment formula a kind of is designed corresponding to the single frame detection results link of same moving target into complete action trail.Compared to current behavioral value algorithm, method detection accuracy proposed by the present invention is high, on the assessment data set announced at present, has reached leading detection level；It can be applied to the video not cut simultaneously, when there is multiple moving targets in video, there is very high detection efficiency.
Description
Technical field
The present invention relates to video behavior analysis technical fields, and in particular to is melted to one kind based on shape-movement double fluid information
The video behavior detection method of conjunction, this method utilize depth learning technology, pass through spatial domain behavioral value and time-domain behavior road
Diameter link, to realize the purpose of video behavior classification and behavior positioning.
Background technique
Video behavioral value is the research hotspot of computer vision field for a long time, has been obtained in recent years increasingly
More concerns, it establishes connection between image analysis and video understand, has potential application value in the life of reality.
Video behavioral value will usually answer two problems: being what behavior and where occur.In recent years, emerging with deep learning
Rise, the strategy that current research method is substantially walked in accordance with two steps: 1) single frames picture carries out behavioral value；2) it is advised with dynamic
Cost-effective method is single frames results link at effective behavioral chain.2015, Gkioxari et al. (G.Gkioxari, J.Malik,
“Finding action tubes”,IEEE Conference on Computer Vision and Pattern
Recognition, pp.759-768) propose a kind of two-part video behavior detection method based on deep learning.They
There is following two defects for model: 1) shape and motion information carry out feature extraction in two independent channels respectively, then
They are connected carry out motion detection, does not account for complementarity between the two；2) single frame detection results link algorithm is at criticizing
Reason mode, until video terminates just provide processing result, it cannot achieve the on-line checking of video behavior.In addition, they
Algorithm can only handle the video (that is, behavior continues up to video since video to be terminated) cut, and can not handle not
The video (that is, behavior can be since any one frame in video, in subsequent any one frame end) of cutting.
Summary of the invention
The object of the present invention is to provide a kind of based on shape-movement double fluid information fusion video behavior detection method.It should
Method realizes the on-line checking task of behavior by dynamic growth algorithm on the basis of single frame video behavioral value.This method Shandong
Stick is strong, on multiple video detection data collection, can behavior accurately be classified and is precisely located simultaneously.
Method proposed by the present invention has the main improvement of two o'clock: 1) shape-proposed by the present invention compared with the existing methods
Motion information fusion method is the depth characteristic fusion based on convolutional network, i.e., in the shape mode of image block fortune corresponding with its
Between dynamic model formula establish association, rather than existing method pass through frequently with result fusion, feature connection etc. shallow hierarchies merge；2)
Behavioral chain dynamic growth algorithm proposed by the present invention be it is online, the video not cut can be handled, and existing
Path link algorithm based on Dynamic Programming can only handle the video cut.Further it is proposed that algorithm can be to view
A plurality of behavioral chain present in frequency is handled simultaneously, and existing method can only be handled one by one.
The principle of the present invention is: 1) deep learning model extraction single-frame images shape and motion information abstract characteristics are constructed,
It recycles depth convolutional network to merge these two types of information, is that frame extracts and behavior divides in the enterprising every trade of the result of fusion
Class realizes single frame video behavioral value；2) single frame detection knot is based on based on classification score, positional relationship and characteristic similarity building
The dynamic growth behavior link algorithm of fruit detected the action trail for corresponding to same moving target.
Technical solution provided by the invention is as follows:
Video behavior detection method proposed by the present invention includes two parts: extracting the shape and motion information of single-frame images
Depth expressing feature, building convolutional network merge this two parts information, then extract behavior in fused feature
Propose, using more sorter networks and position Recurrent networks carry out classification to behavior proposal and position adjusts, and obtain behavioral value knot
Fruit；The link algorithm that building dynamic increases gets up to constitute row the multiframe behavioral value results link for corresponding to same moving target
For path.From one section of video input to testing result, output includes following several steps:
A kind of video behavior detection method merged based on shape-movement double fluid information, comprising the following steps:
Step 1: light stream image being calculated to present frame, extracts the depth expressing feature of RGB image and light stream image, specifically
Light stream image is extracted to current video frame, building depth convolutional network calculates separately the expressing feature of RGB image and light stream image；
Step 2: RGB and light stream depth characteristic are merged, it may be assumed that using convolutional network that obtained RGB and light stream is deep
Degree feature is merged；
Step 3: single frames behavioral value is carried out, behavior classification score and motion frame position are obtained, is that selection behavior is proposed, into
Every trade is that classification and behavior position return, and obtains single frames behavioral value result；
Step 4: the link algorithm building increased using dynamic corresponds to the behavior path of same moving target.It is using dynamic
The link algorithm building that state increases belongs to the behavior path of same moving target, and final detection knot is determined by behavioral chain score
Fruit.
Preferably, fusion is carried out to RGB and light stream depth characteristic described in the step 2 and refers to building convolution fusion
Network, specifically:
Setting convolution kernel is 1*1, and convolutional layer number is the depth convolutional network for inputting RGB or Optical-flow Feature number of layers, is used
It is merged in resemblance and motion feature are set ground depth convolution by turn, obtains shape-motion information fusion feature；
Preferably, the module of single frames behavioral value described in the step 3, specifically includes behavior sorter network and movement
Frame position Recurrent networks, in which:
Behavior sorter network refers to and proposes behavior to carry out more classification marking, including to all behavior classifications and background
Class；
Motion frame position Recurrent networks refer to that the center propose behavior and wide, height are adjusted, and adjustment amount
Operation is carried out in log space.
Compared with prior art, the beneficial effects of the present invention are:
Using technical solution provided by the invention, when carrying out single frames behavioral value, shape and motion information are carried out deep
The fusion of degree ground, connects specific elementary area sports immunology corresponding with its, merges compared to traditional shallow hierarchy
The accuracy that processing method, behavior classification and position return is improved；The dynamic proposed in the present invention increases link algorithm
It can handle the video not cut.Compared to the method for current batch-type, when handling multiple target behavioral value, have more
High realization efficiency.
Detailed description of the invention
Fig. 1 is flow chart of the invention
Fig. 2 is the network structure of model proposed by the invention
Fig. 3 is behavioral chain dynamic growth algorithm flow chart
In attached drawing:
1-input video present frame, 2-convolutional network 1-4 layers of RGB branches, 3-the 5th layer of RGB branch convolutional networks, 4-
RGB convolution feature, 5-present frame light stream images, 6-light stream branch convolutional networks, 7-tired folded RGB and Optical-flow Feature, 8-
Fused feature, 9-ROI regions, 10-full articulamentums, 11-behavioral value networks, 12-behaviors classification output, 13-positions
Adjustment output is set, 14-, which calculate single frame detection result, accumulates score, 15-regeneration behavior chain candidate pools, 16-backtracking behavioral chain inspections
Survey result
Specific embodiment:
Fig. 1 is flow chart of the invention, and wherein s1-s6 is corresponding in turn in specific implementation step 1)-6).As shown in Figure 1,
It is a kind of integrated operation process based on shape-movement double fluid information fusion video behavior detection method of the invention, existing division
It is as follows:
1) Fig. 2 is the network frame figure of model proposed by the invention.The network include: RGB branch convolutional network 1-4 layer 2,
The 5th layer 3 of RGB branch convolutional network, light stream branch convolutional network 6, ROI region 9, full articulamentum 10, behavioral value network 11 etc..
Given test video present frame Ft, input video present frame 1 as shown in Figure 2 calculates its each two field pictures in front and back
Light stream sequence, is denoted as Ot′, t '=t-2, t-1 ..., t+2, present frame light stream image 5 as shown in Figure 2；
2) image FtIt is input to RGB branch convolutional network and carries out depth characteristic extraction.Shown in Fig. 2, RGB branch convolution net
The 5th layer of 3, RGB convolution feature 4 of 1-4 layers of 2, RGB branch convolutional network before network.The output feature for remembering the 5th layer isIts feature
Number of layers is denoted as?Network RPN is proposed in upper one behavior of building, proposes for generating behavior.The cunning of one 3*3 of RPN
Dynamic window existsOn do convolution algorithm, there are two outputs: current location include behavior score, be denoted as Sr, and to current location
Adjustment amount.Position after amendment comprising probable behavior can be calculated by adjustment amount, be denoted as Pr.Light stream sequence Ot′It is input to light
Flow branching convolutional network carries out depth characteristic extraction.The output of 1-5 layer 6 of the convolutional network of light stream branch shown in Fig. 2, the 5th layer of note is special
Sign isIts feature number of layers is
3) willWithIt is stacked up, as shown in Figure 2 tires out folded RGB and Optical-flow Feature 7, constructs convolutional network on it
Shape-movement two parts information is merged.
Convolutional network has the convolution kernel of 1*1, and convolutional layer number isBy convolution operation, the feature merged, note
For fro, it is illustrated in figure 2 fused feature 8；
4) by the output S of RPN networkrAnd PrConstruct area-of-interest.Sr≥δr(δrFor specified threshold, it is taken as 0.6) institute
Corresponding position PrAs area-of-interest, it is denoted as R, ROI region 9 in Fig. 2.In froFeature corresponding to upper extraction R carries out behavior
Classification and position adjustment.Behavioral value network 11, is denoted as DetectNet.Wherein, two full articulamentums 10 respectively include 1024
Implicit unit.Detecting network output is behavior classification score 12 and the position adjustment amount 13 to R.Score of classifying includes to all rows
For the score of classification and background classes, position adjustment amount includes in log space to the center of R and wide, high adjusted value；
5) single frames behavioral value is calculated as a result, constructing behavioral chain using dynamic growth algorithm.For video V to be detected,
Construct a behavioral chain candidate pool Pk, k=1,2 ..., Nk.With a binary group (τk, bk) indicate behavioral chain Pk, wherein τkFor
The accumulation score of behavioral chain, by constituting PkEach operation frame score cumulative form；bkFor PkThe last one current operation frame.Fig. 3
For behavioral chain dynamic growth algorithm flow chart, concrete operations are described below, and following algorithm carries out each behavior classification respectively:
A) it calculates testing result and accumulates score 14.If currently processed t frame image, T is that video V includes picture frame number,
If t < T, handles t frame testing result.If t frame includes N number of testing result altogether, it is denoted asWhereinI-th of operation frame is represented,For its DetectNet classification score.It calculates
Accumulation scoreAs shown in formula (1):
WhereinIndicate j-th of operation frame of t-1 frame image.
B) regeneration behavior chain candidate pool 15.
(i) (τk, bk) be updated toWhereinWith bkIt is connected, andWith maximum score.With
bkBe defined as being connected to satisfy two conditions:
1. friendship and ratio between them:2.WhereinWith
fro(bk) be respectivelyAnd bkIn froThe corresponding feature vector in upper region；
(ii) ifThenIt is updated to
C) recall behavioral chain testing result 16.If video V is disposed, byIt is counter to solve behavior
Each motion frame on chainWherein tsAnd teRespectively behavior starting and ending position.
6) behavioral chain P is calculatedkScore: s (Pk)=τk/(te-ts), threshold xi=0.7 is set, s (P is retainedk) >=ξ institute is right
The P answeredkFor final testing result.
It is above proposed by the present invention a kind of based on shape-movement double fluid information fusion video behavior detection method
Specific embodiment.This embodiment is carried out on actual video data collection UCF-101, and with evaluation criterion generally acknowledged at present
MAP (mean Average Precision) assesses experimental result.In IoU (Intersection over
Union) in the range of 0.1 to 0.6, method proposed by the present invention has all reached current leading detection accuracy, at present other
The comparison of method is as shown in table 1
Table 1
1. testing result comparison sheet of table, '-' expression do not refer to that the higher the better for result value
The literature-recitation that compares occurred in table 1 is as follows:
[1]Weinzaepfel P.,Harchaoui Z.,and Schmid C.,Learning to track for
spatio-temporal action localization.IEEE International Conference on Computer
Vision,2015,(pp.3164-3172).
[2]Saha S.,Singh G.,Sapienza M.,Torr P.H.,and Cuzzolin F.,Deep
learning for detecting multiple space-time action tubes in videos.arXiv:
1608.01529,2016.
[3]Peng X.,and Schmid C.,Multi-region two-stream R-CNN for action
detection.European Conference on Computer Vision,2016,October,(pp.744-759).
It should be noted that the purpose for publicizing and implementing example is to help to further understand the present invention, but the skill of this field
Art personnel, which are understood that, not to be departed from the present invention and spirit and scope of the appended claims, and various substitutions and modifications are all
It is possible.Therefore, the present invention should not be limited to embodiment disclosure of that, and the scope of protection of present invention is with claim
Subject to the range that book defines.
Claims (4)
1. a kind of based on shape-movement double fluid information fusion video behavior detection method, comprising the following steps:
Step 1: light stream image being calculated to present frame, extracts the depth expressing feature of RGB image and light stream image, specially to working as
Preceding video frame extraction light stream image, building depth convolutional network calculate separately the expressing feature of RGB image and light stream image；
Step 2: RGB and light stream depth characteristic being merged, specific is using convolutional network that obtained RGB and light stream is deep
Degree feature is merged；
Step 3: carrying out single frames behavioral value, obtain behavior classification score and motion frame position, be that selection behavior is proposed, gone
It is returned for classification and behavior position, obtains single frames behavioral value result；
Step 4: the link algorithm building increased using dynamic corresponds to the behavior path of same moving target, specially using dynamic
The link algorithm building that state increases belongs to the behavior path of same moving target, and final detection knot is determined by behavioral chain score
Fruit.
2. according to claim 1 based on shape-movement double fluid information fusion video behavior detection method, feature exists
In: fusion is carried out to RGB and light stream depth characteristic described in step 2 and refers to building convolution converged network, specifically:
Setting convolution kernel is 1*1, and convolutional layer number is the depth convolutional network for inputting RGB or Optical-flow Feature number of layers, and being used for will
Resemblance and motion feature are set ground depth convolution fusion by turn, obtain shape-motion information fusion feature.
3. according to claim 1 based on shape-movement double fluid information fusion video behavior detection method, feature exists
In, the module of single frames behavioral value described in step 3 specifically includes behavior sorter network and motion frame position Recurrent networks,
In:
Behavior sorter network refers to and proposes behavior to carry out more classification marking, including to all behavior classifications and background classes；
Motion frame position Recurrent networks refer to that the center propose behavior and wide, height are adjusted, and the operation of adjustment amount
It is to be carried out in log space.
4. according to claim 1 based on shape-movement double fluid information fusion video behavior detection method, feature exists
In, generation behavior path described in step 4 is to increase link algorithm using motion frame dynamic, it specifically includes:
Behavioral chain candidate pool is constructed, the behavioral chain specified number for maintaining currently most possibly to occur for giving video；
Construct behavioral chain and state binary group, including behavioral chain accumulation score and the last one motion frame position of behavioral chain, for pair
Behavioral chain is ranked up and calculates the associated relation between behavioral chain and current frame motion detection block, and specific is to use binary group
(τk, bk) indicate behavioral chain Pk, wherein τkFor the accumulation score of behavioral chain, by constituting PkEach operation frame score it is cumulative and
At；bkFor PkThe last one current operation frame.
Motion frame is accumulated score and is calculated, and is specifically behavior for evaluating and testing designated movement frame chaining in a possibility that existing behavioral chain
FrameAccumulation scoreIt is calculated using mode shown in formula (1):
WhereinIndicate j-th of operation frame of t-1 frame image；For t frame i-th of operation frame of image,ForBehavior point
Class score；
Behavioral chain candidate pool updates rule, low for realizing the growth, the new judgement for behavioral chain occur, score of existing behavioral chain
The deletion of behavioral chain, it is specific to be are as follows:
(i) (τk, bk) be updated toWhereinWith bkIt is connected, andWith maximum score, realize existing
The growth of behavioral chain；
(ii) ifThenIt is updated toIt realizes and new behavioral chain is added.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811298485.1A CN109376677B (en) | 2018-11-02 | 2018-11-02 | Video behavior detection method based on appearance-motion double-flow information fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811298485.1A CN109376677B (en) | 2018-11-02 | 2018-11-02 | Video behavior detection method based on appearance-motion double-flow information fusion |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109376677A true CN109376677A (en) | 2019-02-22 |
| CN109376677B CN109376677B (en) | 2022-04-01 |
Family
ID=65396749
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811298485.1A Active CN109376677B (en) | 2018-11-02 | 2018-11-02 | Video behavior detection method based on appearance-motion double-flow information fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109376677B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111091582A (en) * | 2019-12-31 | 2020-05-01 | 北京理工大学重庆创新中心 | Single-vision target tracking algorithm and system based on deep neural network |
| CN111125405A (en) * | 2019-12-19 | 2020-05-08 | 国网冀北电力有限公司信息通信分公司 | Power monitoring image anomaly detection method and device, electronic equipment and storage medium |
| CN111325253A (en) * | 2020-02-12 | 2020-06-23 | 杭州涂鸦信息技术有限公司 | Double-flow-method behavior identification method and system based on deep learning |
| CN111709410A (en) * | 2020-08-20 | 2020-09-25 | 深兰人工智能芯片研究院（江苏）有限公司 | Behavior identification method for strong dynamic video |
| CN115285885A (en) * | 2022-06-22 | 2022-11-04 | 广州先进技术研究所 | Unmanned forklift path and task joint generation method and system based on warehousing environment |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106650655A (en) * | 2016-12-16 | 2017-05-10 | 北京工业大学 | Action detection model based on convolutional neural network |
| US20170262995A1 (en) * | 2016-03-11 | 2017-09-14 | Qualcomm Incorporated | Video analysis with convolutional attention recurrent neural networks |
| CN107609460A (en) * | 2017-05-24 | 2018-01-19 | 南京邮电大学 | A kind of Human bodys' response method for merging space-time dual-network stream and attention mechanism |
-
2018
- 2018-11-02 CN CN201811298485.1A patent/CN109376677B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170262995A1 (en) * | 2016-03-11 | 2017-09-14 | Qualcomm Incorporated | Video analysis with convolutional attention recurrent neural networks |
| CN106650655A (en) * | 2016-12-16 | 2017-05-10 | 北京工业大学 | Action detection model based on convolutional neural network |
| CN107609460A (en) * | 2017-05-24 | 2018-01-19 | 南京邮电大学 | A kind of Human bodys' response method for merging space-time dual-network stream and attention mechanism |
Non-Patent Citations (2)
| Title |
|---|
| BO HUANG ET AL: "《Convolutional Gated Recurrent Units Fusion for Video Action Recognition》", 《ICONIP 2017: NEURAL INFORMATION PROCESSING》 * |
| 郭会文: "《基于多载体的人群异常事件检测方法研究》", 《中国博士学位论文全文数据库 信息科技辑》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111125405A (en) * | 2019-12-19 | 2020-05-08 | 国网冀北电力有限公司信息通信分公司 | Power monitoring image anomaly detection method and device, electronic equipment and storage medium |
| CN111091582A (en) * | 2019-12-31 | 2020-05-01 | 北京理工大学重庆创新中心 | Single-vision target tracking algorithm and system based on deep neural network |
| CN111325253A (en) * | 2020-02-12 | 2020-06-23 | 杭州涂鸦信息技术有限公司 | Double-flow-method behavior identification method and system based on deep learning |
| CN111325253B (en) * | 2020-02-12 | 2023-05-05 | 杭州涂鸦信息技术有限公司 | Double-flow-method behavior recognition method and system based on deep learning |
| CN111709410A (en) * | 2020-08-20 | 2020-09-25 | 深兰人工智能芯片研究院（江苏）有限公司 | Behavior identification method for strong dynamic video |
| CN111709410B (en) * | 2020-08-20 | 2020-12-01 | 深兰人工智能芯片研究院（江苏）有限公司 | Behavior identification method for strong dynamic video |
| CN115285885A (en) * | 2022-06-22 | 2022-11-04 | 广州先进技术研究所 | Unmanned forklift path and task joint generation method and system based on warehousing environment |
| CN115285885B (en) * | 2022-06-22 | 2023-12-12 | 广州先进技术研究所 | Unmanned forklift path and task joint generation method and system based on warehouse environment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109376677B (en) | 2022-04-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109376677A (en) | A kind of video behavior detection method merged based on shape-movement double fluid information | |
| US20210342586A1 (en) | Systems and methods for extracting and vectorizing features of satellite imagery | |
| CN106127204B (en) | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks | |
| CN107016357B (en) | Video pedestrian detection method based on time domain convolutional neural network | |
| KR102319541B1 (en) | Learning method and testing method of object detector to be used for surveillance based on r-cnn capable of converting modes according to aspect ratios or scales of objects, and learning device and testing device using the same | |
| CN110059558A (en) | A kind of orchard barrier real-time detection method based on improvement SSD network | |
| CN112001339A (en) | Pedestrian social distance real-time monitoring method based on YOLO v4 | |
| CN108803617A (en) | Trajectory predictions method and device | |
| CN108399362A (en) | A kind of rapid pedestrian detection method and device | |
| CN109271990A (en) | A kind of semantic segmentation method and device for RGB-D image | |
| CN109446970A (en) | A kind of Intelligent Mobile Robot road scene recognition methods based on deep learning | |
| CN111680655A (en) | Video target detection method for aerial images of unmanned aerial vehicle | |
| CN104281853A (en) | Behavior identification method based on 3D convolution neural network | |
| CN103714181B (en) | A kind of hierarchical particular persons search method | |
| CN109241982A (en) | Object detection method based on depth layer convolutional neural networks | |
| CN107463892A (en) | Pedestrian detection method in a kind of image of combination contextual information and multi-stage characteristics | |
| CN108805900A (en) | A kind of determination method and device of tracking target | |
| CN111274921A (en) | Method for recognizing human body behaviors by utilizing attitude mask | |
| CN113160062B (en) | Infrared image target detection method, device, equipment and storage medium | |
| JP2020038660A (en) | Learning method and learning device for detecting lane by using cnn, and test method and test device using the same | |
| CN110705566B (en) | Multi-mode fusion significance detection method based on spatial pyramid pool | |
| CN106991370A (en) | Pedestrian retrieval method based on color and depth | |
| CN105404894A (en) | Target tracking method used for unmanned aerial vehicle and device thereof | |
| CN110210550A (en) | Image fine granularity recognition methods based on integrated study strategy | |
| CN109409307A (en) | A kind of Online Video behavioral value system and method based on space-time contextual analysis |
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 |