CN110309729A - Tracking and re-detection method based on anomaly peak detection and twin network - Google Patents
Tracking and re-detection method based on anomaly peak detection and twin network Download PDFInfo
- Publication number
- CN110309729A CN110309729A CN201910506402.1A CN201910506402A CN110309729A CN 110309729 A CN110309729 A CN 110309729A CN 201910506402 A CN201910506402 A CN 201910506402A CN 110309729 A CN110309729 A CN 110309729A
- Authority
- CN
- China
- Prior art keywords
- network
- target
- ship
- target ship
- detection
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- 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
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
- G06V20/42—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items of sport video content
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of tracking and re-detection method based on anomaly peak detection and twin network, comprising steps of S1, for video first frame, the size and center position of initialized target ship, and obtain the characteristic pattern of target ship;S2, judge whether target ship is lost according to the variation of every frame peak value of response: if target ship is lost, entering step S3, otherwise enter step S7;S3, the highest image of peak value of response is saved, as template image;S4, object candidate area is extracted by RPN network;S5, using template image and object candidate area as twin network inputs, select the highest region of similarity to export as target area;S6, the information for re-entering target ship;S7, continue to track target ship.This method tracks accuracy rate 82.6% or more in tracking target ship, while re-detection network gives particular boat for change and works well, and ship gives rate for change 86.5% or more, and it is fast to give speed for change.
Description
Technical field
The present invention relates to technical field of ships, and in particular to it is a kind of based on anomaly peak detection and twin network tracking and
Re-detection method.
Background technique
Target following technology is one of the hot spot in computer vision research field.During the motion, target will not be always
In perfect condition, it may occur that some cosmetic variations are blocked or deformation etc. such as posture, scale or the variation of illumination, these
All certain difficulty is caused to target following research.How to solve the problems, such as target occlusion, realizes long term object tracking, be current mesh
Mark the research emphasis in tracking technique.In the prior art, after target ship is blocked, the tracking accuracy rate and ship of ship are given for change
Rate is relatively low, so that being unable to complete the long-term follow to target ship.
Summary of the invention
To overcome the above deficiencies, the invention provides it is a kind of based on anomaly peak detection and twin network with
Track and re-detection method.
The present invention overcomes the technical solution used by its technical problem to be:
A kind of tracking and re-detection method based on anomaly peak detection and twin network, includes the following steps:
S1, it is directed to video first frame, the size and center position of initialized target ship, and obtains the spy of target ship
Sign figure;
S2, judge whether target ship is lost according to the variation of every frame peak value of response: if target ship is lost, entering step
Rapid S3, on the contrary enter step S7;
S3, the highest image of peak value of response is saved, as template image;
S4, object candidate area is extracted by RPN network;
S5, using template image and object candidate area as twin network inputs, select the highest region of similarity as
Target area output;
S6, the information for re-entering target ship;
S7, continue to track target ship.
Further, in the step S1, the characteristic pattern of target ship is obtained using Faster R-CNN network, it is specific
Method are as follows:
S1.1, the convolution characteristic pattern for calculating target ship image;
S1.2, convolution characteristic pattern is handled using RPN network, obtains target Suggestion box;
S1.3, characteristic pattern is extracted to target Suggestion box using RoI Pooling.
Further, in the step S2, the tool whether target ship is lost is judged according to the variation of every frame peak value of response
Body method are as follows:
If VpreIt indicates the peak value of response of present frame, gathers { Vi| i=1,2 ..., n indicate present frame period the last period
Interior peak set;
In above-mentioned formula, u indicates that the average value of peak value in period the last period of present frame, σ indicate its standard deviation;
If the peak distance mean value u of present frame has been more than λ σ, which is flagged as exceptional value, can sentence at this time
Disconnected ship has been lost.
Further, in the step S4, object candidate area is extracted by RPN network method particularly includes:
A part of ship image that the ship image data of acquisition is concentrated is as training dataset, and another part is as survey
Examination collection, RPN network carry out end-to-end training in the training stage;
S4.1, RPN network parameter is initialized with pre-training network model, passes through stochastic gradient descent algorithm and backpropagation
Algorithm finely tunes RPN network parameter;
S4.2, Faster R-CNN target detection network parameter is initialized with pre-training network model, and with step S4.1
In RPN network extract candidate region, training objective detect network;
S4.3, it is reinitialized with target detection network in step S4.2 and finely tunes RPN network parameter;
S4.4, object candidate area is extracted with RPN network in step S4.3.
Further, in the step S5, using template image and object candidate area as the specific of twin network inputs
Method are as follows:
The size of template image and object candidate area is adjusted to 107 pixels × 107 pixels, and it is inputted respectively
Two twin sub-networks, the Euclidean distance between feature that two twin sub-networks export are used to characterize between input picture pair
Similarity, twin network finally select the highest region of similarity as target area between object candidate area set
Output.
The beneficial effects of the present invention are:
The present invention is detected by anomaly peak and the tracking and re-detection method of twin network, in tracking target ship,
Accuracy rate is tracked 82.6% or more, while re-detection network gives particular boat for change and works well, ship gives rate for change 86.5%
More than, and it is fast to give speed for change.
Detailed description of the invention
Fig. 1 is video sequence part described in the embodiment of the present invention to target ship tracking procedure chart, wherein Fig. 1 (a) is mesh
The tracing figure that is not blocked of mark ship, Fig. 1 (b) are that target ship is blocked the tracing figure of a part, and Fig. 1 (c) is target ship
The tracing figure being blocked completely.
Fig. 2 is the response distribution map for respectively corresponding Fig. 1, wherein the response distribution map of Fig. 2 (a) corresponding diagram 1 (a), figure
The response distribution map of 2 (b) corresponding diagrams 1 (b), the response distribution map of Fig. 2 (c) corresponding diagram 1 (c).
Fig. 3 is the schematic diagram that RPN network exports result.
Fig. 4 is the result figure of twin network output.
In figure, 1, target ship, 2, shelter.
Specific embodiment
For a better understanding of the skilled in the art, being done in the following with reference to the drawings and specific embodiments to the present invention
It is further described, it is following to be merely exemplary that the scope of protection of the present invention is not limited.
Tracking and re-detection method of the present invention based on anomaly peak detection and twin network, including walk as follows
It is rapid:
Step S1, video first frame, the size and center position of initialized target ship, as shown in Figure 1, described are directed to
Central point is the central point of ship in video image, and the characteristic pattern of target ship is obtained using Faster R-CNN network.
Wherein, the characteristic pattern of target ship is obtained using Faster R-CNN network method particularly includes:
S1.1, the convolution characteristic pattern for calculating target ship image;
S1.2, convolution characteristic pattern is handled using RPN network, obtains target Suggestion box;
S1.3, characteristic pattern is extracted to target Suggestion box using RoI Pooling.
Step S2, judge whether target ship is lost according to the variation of every frame peak value of response: if target ship is lost, into
Enter step S3, otherwise enters step S7.
Wherein, judge what whether target ship was lost according to the variation of every frame peak value of response method particularly includes:
If VpreIt indicates the peak value of response of present frame, gathers { Vi| i=1,2 ..., n indicate present frame period the last period
Interior peak set;
In above-mentioned formula, u indicates that the average value of peak value in period the last period of present frame, σ indicate its standard deviation;
If the peak distance mean value u of present frame has been more than λ σ, which is flagged as exceptional value, can sentence at this time
Disconnected ship has been lost.
Step S3, the highest image of peak value of response is saved, as template image.
Step S4, object candidate area is extracted by RPN network, method particularly includes:
A part of ship image that the ship image data of acquisition is concentrated is as training dataset, and another part is as survey
Examination collection, for the pre-training model that the present embodiment uses for ResNet50, RPN network carries out end-to-end training in the training stage;
S4.1, RPN network parameter is initialized with pre-training network model, passes through stochastic gradient descent algorithm and backpropagation
Algorithm finely tunes RPN network parameter;
S4.2, Faster R-CNN target detection network parameter is initialized with pre-training network model, and with step S4.1
In RPN network extract candidate region, training objective detect network;
S4.3, it is reinitialized with target detection network in step S4.2 and finely tunes RPN network parameter;
S4.4, object candidate area is extracted with RPN network in step S4.3.
By operating above, RPN network (region is also made to suggest network) can export more accurate object candidate area,
The object candidate area that the network area RPN generates is as shown in figure 3, Regional Representative's object candidate area in box.
Step S5, using template image and object candidate area as twin network inputs, the highest region of similarity is selected
It is exported as target area.
Wherein, using template image and object candidate area as twin network inputs method particularly includes:
The size of template image and object candidate area is adjusted to 107 pixels × 107 pixels, and it is inputted respectively
Two twin sub-networks, the Euclidean distance between feature that two twin sub-networks export are used to characterize between input picture pair
Similarity, twin network finally select the highest region of similarity as target area between object candidate area set
Output, the result of twin network output is as shown in figure 4, the part i.e. in box circle is that target area exports result.
Step S6, the information of target ship is re-entered.
Step S7, continue to track target ship.
In above-mentioned steps, target ship is navigation of turning left from the right side.Response distribution map in the above process as shown in Fig. 2,
The response distribution map for the tracing figure that wherein Fig. 2 (a) corresponding diagram 1 (a) target ship is not blocked, Fig. 2 (b) corresponding diagram 1 (b) mesh
Mark ship be blocked a part tracing figure response distribution map, Fig. 2 (c) corresponding diagram 1 (c) target ship is blocked completely
The response distribution map of tracing figure.
Above only describes basic principle of the invention and preferred embodiment, those skilled in the art can be according to foregoing description
Many changes and improvements are made, these changes and improvements should be within the scope of protection of the invention.
Claims (5)
1. a kind of tracking and re-detection method based on anomaly peak detection and twin network, which is characterized in that including walking as follows
It is rapid:
S1, it is directed to video first frame, the size and center position of initialized target ship, and obtains the feature of target ship
Figure;
S2, judge whether target ship is lost according to the variation of every frame peak value of response: if target ship is lost, entering step
S3, on the contrary enter step S7;
S3, the highest image of peak value of response is saved, as template image;
S4, object candidate area is extracted by RPN network;
S5, using template image and object candidate area as twin network inputs, select the highest region of similarity as target
Region output;
S6, the information for re-entering target ship;
S7, continue to track target ship.
2. the tracking and re-detection method according to claim 1 based on anomaly peak detection and twin network, feature
It is, in the step S1, the characteristic pattern of target ship is obtained using Faster R-CNN network, method particularly includes:
S1.1, the convolution characteristic pattern for calculating target ship image;
S1.2, convolution characteristic pattern is handled using RPN network, obtains target Suggestion box;
S1.3, characteristic pattern is extracted to target Suggestion box using RoI Pooling.
3. the tracking and re-detection method according to claim 1 based on anomaly peak detection and twin network, feature
It is, in the step S2, judges what whether target ship was lost according to the variation of every frame peak value of response method particularly includes:
If VpreIt indicates the peak value of response of present frame, gathers { Vi| i=1,2 ..., n } it indicates in period the last period of present frame
Peak set;
In above-mentioned formula, u indicates that the average value of peak value in period the last period of present frame, σ indicate its standard deviation;
If the peak distance mean value u of present frame has been more than λ σ, which is flagged as exceptional value, may determine that ship at this time
Oceangoing ship has been lost.
4. the tracking and re-detection method according to claim 1 based on anomaly peak detection and twin network, feature
It is, in the step S4, object candidate area is extracted by RPN network method particularly includes:
A part of ship image that the ship image data of acquisition is concentrated is as training dataset, and another part is as test
Collection, RPN network carry out end-to-end training in the training stage;
S4.1, RPN network parameter is initialized with pre-training network model, passes through stochastic gradient descent algorithm and back-propagation algorithm
Finely tune RPN network parameter;
S4.2, Faster R-CNN target detection network parameter is initialized with pre-training network model, and in step S4.1
RPN network extracts candidate region, and training objective detects network;
S4.3, it is reinitialized with target detection network in step S4.2 and finely tunes RPN network parameter;
S4.4, object candidate area is extracted with RPN network in step S4.3.
5. the tracking and re-detection method according to claim 1 based on anomaly peak detection and twin network, feature
It is, in the step S5, using template image and object candidate area as twin network inputs method particularly includes:
The size of template image and object candidate area is adjusted to 107 pixels × 107 pixels, and it is inputted two respectively
Twin sub-network, the Euclidean distance between feature that two twin sub-networks export are used to characterize similar between input picture pair
Degree, twin network finally select the highest region of similarity defeated as target area between object candidate area set
Out.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910506402.1A CN110309729A (en) | 2019-06-12 | 2019-06-12 | Tracking and re-detection method based on anomaly peak detection and twin network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910506402.1A CN110309729A (en) | 2019-06-12 | 2019-06-12 | Tracking and re-detection method based on anomaly peak detection and twin network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110309729A true CN110309729A (en) | 2019-10-08 |
Family
ID=68077391
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910506402.1A Pending CN110309729A (en) | 2019-06-12 | 2019-06-12 | Tracking and re-detection method based on anomaly peak detection and twin network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110309729A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111524162A (en) * | 2020-04-15 | 2020-08-11 | 上海摩象网络科技有限公司 | Method and device for retrieving tracking target and handheld camera |
CN111724409A (en) * | 2020-05-18 | 2020-09-29 | 浙江工业大学 | Target tracking method based on densely connected twin neural network |
CN111986517A (en) * | 2020-08-26 | 2020-11-24 | 珠海大横琴科技发展有限公司 | Ship anomaly detection method and device and storage medium |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160019683A1 (en) * | 2014-07-17 | 2016-01-21 | Ricoh Company, Ltd. | Object detection method and device |
WO2018122459A1 (en) * | 2017-01-02 | 2018-07-05 | Tampereen Yliopisto | Calcium level analysis in cardiomyocyte |
CN108898620A (en) * | 2018-06-14 | 2018-11-27 | 厦门大学 | Method for tracking target based on multiple twin neural network and regional nerve network |
CN109508655A (en) * | 2018-10-28 | 2019-03-22 | 北京化工大学 | The SAR target identification method of incomplete training set based on twin network |
CN109766780A (en) * | 2018-12-20 | 2019-05-17 | 武汉理工大学 | A kind of ship smog emission on-line checking and method for tracing based on deep learning |
-
2019
- 2019-06-12 CN CN201910506402.1A patent/CN110309729A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160019683A1 (en) * | 2014-07-17 | 2016-01-21 | Ricoh Company, Ltd. | Object detection method and device |
WO2018122459A1 (en) * | 2017-01-02 | 2018-07-05 | Tampereen Yliopisto | Calcium level analysis in cardiomyocyte |
CN108898620A (en) * | 2018-06-14 | 2018-11-27 | 厦门大学 | Method for tracking target based on multiple twin neural network and regional nerve network |
CN109508655A (en) * | 2018-10-28 | 2019-03-22 | 北京化工大学 | The SAR target identification method of incomplete training set based on twin network |
CN109766780A (en) * | 2018-12-20 | 2019-05-17 | 武汉理工大学 | A kind of ship smog emission on-line checking and method for tracing based on deep learning |
Non-Patent Citations (3)
Title |
---|
BO LI等: "High performance visual tracking with siamese region proposal network", 《2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
刘延飞等: "基于异常值检测的KCF目标丢失预警方法研究", 《计算机工程与应用》 * |
方萍等: "《试验设计与统计》", 30 June 2003, 浙江大学出版社 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111524162A (en) * | 2020-04-15 | 2020-08-11 | 上海摩象网络科技有限公司 | Method and device for retrieving tracking target and handheld camera |
WO2021208261A1 (en) * | 2020-04-15 | 2021-10-21 | 上海摩象网络科技有限公司 | Tracking target retrieving method and device, and handheld camera |
CN111524162B (en) * | 2020-04-15 | 2022-04-01 | 上海摩象网络科技有限公司 | Method and device for retrieving tracking target and handheld camera |
CN111724409A (en) * | 2020-05-18 | 2020-09-29 | 浙江工业大学 | Target tracking method based on densely connected twin neural network |
CN111986517A (en) * | 2020-08-26 | 2020-11-24 | 珠海大横琴科技发展有限公司 | Ship anomaly detection method and device and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108154118B (en) | A kind of target detection system and method based on adaptive combined filter and multistage detection | |
CN107832672B (en) | Pedestrian re-identification method for designing multi-loss function by utilizing attitude information | |
CN108491835B (en) | Two-channel convolutional neural network for facial expression recognition | |
CN109657631B (en) | Human body posture recognition method and device | |
CN108052896B (en) | Human body behavior identification method based on convolutional neural network and support vector machine | |
Li et al. | Robust visual tracking based on convolutional features with illumination and occlusion handing | |
CN111814661B (en) | Human body behavior recognition method based on residual error-circulating neural network | |
Cheng et al. | Robust visual localization in dynamic environments based on sparse motion removal | |
CN104200495B (en) | A kind of multi-object tracking method in video monitoring | |
CN109685013B (en) | Method and device for detecting head key points in human body posture recognition | |
CN105069434B (en) | A kind of human action Activity recognition method in video | |
CN110674785A (en) | Multi-person posture analysis method based on human body key point tracking | |
CN111563452B (en) | Multi-human-body gesture detection and state discrimination method based on instance segmentation | |
CN107424161B (en) | Coarse-to-fine indoor scene image layout estimation method | |
CN114187665B (en) | Multi-person gait recognition method based on human skeleton heat map | |
CN110309729A (en) | Tracking and re-detection method based on anomaly peak detection and twin network | |
CN105718882A (en) | Resolution adaptive feature extracting and fusing for pedestrian re-identification method | |
CN104821010A (en) | Binocular-vision-based real-time extraction method and system for three-dimensional hand information | |
CN111126494B (en) | Image classification method and system based on anisotropic convolution | |
CN113608663B (en) | Fingertip tracking method based on deep learning and K-curvature method | |
CN113076891B (en) | Human body posture prediction method and system based on improved high-resolution network | |
CN112287906B (en) | Template matching tracking method and system based on depth feature fusion | |
CN106529441B (en) | Depth motion figure Human bodys' response method based on smeared out boundary fragment | |
CN106407978B (en) | Method for detecting salient object in unconstrained video by combining similarity degree | |
CN102663777A (en) | Target tracking method and system based on multi-view video |
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 |
Application publication date: 20191008 |
|
RJ01 | Rejection of invention patent application after publication |