CN109344684A - Battery-operated motor cycle image detecting method based on units of variance model - Google Patents
Battery-operated motor cycle image detecting method based on units of variance model Download PDFInfo
- Publication number
- CN109344684A CN109344684A CN201810877373.5A CN201810877373A CN109344684A CN 109344684 A CN109344684 A CN 109344684A CN 201810877373 A CN201810877373 A CN 201810877373A CN 109344684 A CN109344684 A CN 109344684A
- Authority
- CN
- China
- Prior art keywords
- feature
- image
- battery
- motor cycle
- operated motor
- 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
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Abstract
The invention discloses a kind of battery-operated motor cycle image detecting methods based on units of variance model, initially set up sample set, and carry out feature extraction, root template is obtained by feature training, initial root template is obtained in a manner of spatial clustering, feature pyramid, the training initial root template are established to sample；Initialisation unit model and be loaded into initial root template by training obtain final root template and final partial model；For image to be detected, feature extraction is carried out to image to be detected, is then matched with the final root template and final partial model, obtains the position of battery-operated motor cycle in image to be detected.Deformable part model of the present invention constructs characteristics of image pyramid, and feature can be extracted on different scale, to adapt to the image detection of different scale；Deformable part model of the present invention is to use multicomponent mixed model, is suitable for different perspectives variation, illumination etc..
Description
Technical field
The present invention relates to a kind of image detecting methods, and in particular to a kind of battery-operated motor cycle figure based on units of variance model
As detection method, to detect electric motorcycle truck position in traffic video monitoring image.
Background technique
Current traffic video monitoring image tends to complicated multiplicity, and there are visual angle change, illumination variation, background complexity, target
The problems such as form of diverse and noise jamming, so that carrying out detection to specific objective in the picture becomes a difficult point.It is existing
Many of technology is directed to the algorithm of target detection of pedestrian, vehicle, bicycle etc., but in terms of the detection identification of battery-operated motor cycle
Research it is seldom, but the accidents such as the traffic accident as caused by battery-operated motor cycle, pilferage, fire are many.
Current vehicle identification algorithm constitutes one by extracting Partial Feature in each sample using being SIFT feature
Feature space carries out sparse coding to it and obtains sparse coding feature basic matrix, utilizes Linear SVM and sparse coding combined training
Then template out carries out SIFT feature to target image and template and matches to obtain Optimum Matching position, to detect vehicle.
But the algorithm has the disadvantage that (1) mainly in the large-scale target such as car, pick up, truck, bus, to electricity
The recognition effect of dynamic motorcycle is bad.(2) vulnerable to external condition influence: for blocking, different angle shooting and other effects it is undesirable,
Simultaneously vulnerable to illumination effect；(3) by target itself affect: when deformation occurs for target, easily causing missing inspection or false retrieval.
Summary of the invention
The problem of for currently existing technology, the object of the present invention is to provide a kind of based on units of variance model
Battery-operated motor cycle image detecting method detects units of variance model applied to battery-operated motor cycle, can be with using multicomponent model
Different perspectives is adapted to, target detection accuracy rate is improved.
In order to realize above-mentioned task, the invention adopts the following technical scheme:
Battery-operated motor cycle image detecting method based on units of variance model, comprising the following steps:
Step 1, positive sample collection and negative sample collection are established；
Step 2, classify to positive sample collection, then the grouping of negative sample collection carries out feature to each positive sample, negative sample and mentions
It takes, obtains the feature of positive sample and the feature of negative sample；
The feature of the training positive sample and the feature of negative sample, obtain root template；
Step 3, the root template is obtained into initial root template by spatial clustering, to all positive samples, negative sample
Feature pyramid is established respectively, then the training initial root template；
Step 4, the components number and shape for setting battery-operated motor cycle utilize the initial root after the training of component scanning step 3
Template, when component shape and the energy maximum of initial root template overlay area, which is the position of component, thus
Obtain initial part model；
It is initial in template after initial part model to be loaded into training, and in the positive sample collection and negative sample collection
On be trained, obtain final root template and final partial model；
Step 5, for image to be detected, feature extraction is carried out to image to be detected, then with the final root template
And final partial model is matched, and the position of battery-operated motor cycle in image to be detected is obtained.
It is further, described to establish positive sample collection and negative sample collection, comprising:
Using multiple images comprising battery-operated motor cycle as positive sample collection, positive sample is concentrated into the electricity in each positive sample
Dynamic motorcycle carries out rectangle frame mark；What negative sample collection then chose the offer of PASCAL VOC data set does not include battery-operated motor cycle
Image.
Further, the step 2 includes:
Step 2.1, the rectangle callout box for the positive sample that positive sample is concentrated is divided into m class according to length-width ratio, negative sample collection with
Machine is divided into m group；
Step 2.2, feature extraction is carried out to each positive sample, negative sample, comprising:
For positive sample or negative sample image, multiple cells are divided the image into, to each pixel in cell point 18
A have a symbol gradient direction and 9 are sought gradient without symbol gradient direction, then with gradient direction in the histogram of cell plus
Power projection obtains the gradient orientation histogram of the cell, the as feature vector of corresponding 27 dimension of the cell；It then will be every
2*2 adjacent cells lattice form a block, then each block corresponds to the feature vector of one 108 dimension, by carrying out to it
108 original dimensional features are carried out the cumulative 4+27=31 that is down to each row and column respectively and tieed up by principal component analysis and parsing dimensionality reduction, will
All Block Characteristic vectors, which are together in series, can be obtained the feature of whole image；Thus the feature and negative sample of positive sample are obtained
Feature；
Step 2.3, according to the feature of the feature of positive sample and negative sample, m root template F is obtained with SVM training1,F2,…,
Fm。
It is further, described that feature pyramid is established respectively to all positive samples, negative sample, comprising:
For positive sample or negative sample image, using image as the pyramidal first layer of feature, then feature is pyramidal
Other layers are one layer 2 above1/intervalMultiple sampled, each layer pyramidal for feature, according to the side of step 2.2
Method carries out feature extraction, to obtain each layer of feature pyramid of feature.
Further, carry out target detection, it is described that feature extraction is carried out to image to be detected, then with it is described final
Root template and final partial model are matched, and the position of battery-operated motor cycle in image to be detected is obtained, comprising:
The feature pyramid for establishing image to be detected distinguishes the pyramidal each layer of this feature and the final root template
Convolution operation is carried out, selects maximum convolution value as root template score；
From the feature pyramid of image to be detected, select twice of resolution characteristics of image to be detected final with each
Partial model carries out convolution respectively, obtains the corresponding score of each final partial model；
The deformation for calculating final partial model is spent；
The total algebraical sum for asking described template score, the corresponding score of final partial model and deformation to spend is as detection
The position of total score, highest scoring is the position of battery-operated motor cycle.
The present invention has following technical characterstic compared with prior art:
1. deformable part model of the present invention constructs characteristics of image pyramid, can be extracted on different scale
Feature, to adapt to the image detection of different scale；
2. deformable part model of the present invention is to use multicomponent mixed model, it is suitable for different perspectives change
Change, illumination etc.；
3. deformable part model of the present invention uses the partial model of twice of resolution ratio, small range can be solved
Occlusion issue.
Detailed description of the invention
Fig. 1 is the flow chart of the method for the present invention；
(a) of Fig. 2 is the schematic diagram of final root template, (b) is the schematic diagram of final partial model, (c) spends for deformation
Schematic diagram.
Fig. 3 is the result schematic diagram that battery-operated motor cycle detection is carried out using the method for the present invention.
Specific embodiment
The invention discloses a kind of battery-operated motor cycle image detecting methods based on units of variance model, including following step
It is rapid:
Step 1, positive sample collection and negative sample collection are established
Using multiple images comprising battery-operated motor cycle as positive sample collection P, positive sample is concentrated in each positive sample
Battery-operated motor cycle carry out rectangle frame mark, this programme using weak mark mode, i.e., only to target position (i.e. electric motorcycle
Vehicle position) it is labeled；In the present embodiment, positive sample collection uses the figure comprising battery-operated motor cycle of 1000 shootings
Picture；
Negative sample collection N then chooses 1000 images not comprising battery-operated motor cycle of PASCAL VOC data set offer.
Step 2, classify to positive sample collection, then the grouping of negative sample collection carries out feature to each positive sample, negative sample and mentions
It takes, obtains the feature of positive sample and the feature of negative sample；The feature of the training positive sample and the feature of negative sample, obtain root
Template；Specific step is as follows:
Step 2.1, the rectangle callout box for the positive sample that positive sample is concentrated is divided into m class according to length-width ratio, negative sample collection with
Machine is divided into m group；
Step 2.2, feature extraction is carried out to each positive sample, negative sample, comprising:
For positive sample or negative sample image (image criticizes sample, negative sample), multiple cells are divided the image into,
Each cell is 8*8 pixel size in the present embodiment；To each pixel in cell point, 18 have symbol gradient direction
(0 ° -360 °) and 9 seek gradient without symbol gradient direction (0 ° -180 °), then with gradient direction in the histogram of cell
Weighted projection obtains the gradient orientation histogram of the cell, the as feature vector of the corresponding 18+9=27 dimension of the cell；
Then it is (such as adjacent less than four up and down every 2*2 adjacent cells lattice (four cells up and down) to be formed into a block
0) a cell is then mended in the part of missing, then each block corresponds to the feature vector of 4*27=108 dimension, by right
It carries out principal component analysis and parsing dimensionality reduction, and 108 original dimensional features add up to each row and column respectively and are down to 4+27=
31 dimensions, all Block Characteristic vectors, which are together in series, can be obtained the feature of whole image.
The image of the image of each positive sample, negative sample is handled all in accordance with the above method to get positive sample has been arrived
The feature of feature originally and negative sample.
Step 2.3, according to the feature of the feature of all positive samples and negative sample, with the training of standard support vector machines
Obtain m root template F1,F2,…,Fm。
Step 3, the root template is obtained into initial root template by spatial clustering, to all positive samples, negative sample
Feature pyramid is established respectively, then the training initial root template；It specifically includes:
Step 3.1, by the m root template F1,F2,…,FmInitial root template F is obtained by spatial clustering0；
Step 3.2, feature pyramid is established respectively to all positive samples, negative sample, positive sample or negative sample are established special
Levy that pyramidal steps are as follows:
For positive sample or negative sample image, using image as the pyramidal first layer of feature, then feature is pyramidal
Other layers are one layer 2 above1/intervalMultiple sampled, each layer pyramidal for feature, according to the side of step 2.2
Method carries out feature extraction, to obtain each layer of feature pyramid of feature.In the present embodiment, for positive sample interval=
5, for negative sample interval=2.
That is, using positive sample, negative sample original image as pyramidal first layer, the second layer is with first layer 21 /intervalMultiple sampled, third layer is with the second layer 21/intervalMultiple sampled, and so on；Each layer is adopted
Sample correspondence obtains piece image, which is extracted feature according to the feature extracting method of step 2.2, to obtain feature gold
The feature that each layer of word tower.
Step 3.3, the training initial root template
Mainly include two parts using the hidden support vector machines LSVM training initial root template in the present embodiment:
Step 3.3.1, sample optimization
If hidden variable zPFor the coordinate [x of rectangle frame in positive sample1,y1,x2,y2] (the upper left angle point of rectangle frame and bottom right
The coordinate of angle point), fixed initial root template calculates the position coordinates in rectangle frame with initial root mask convolution highest scoring,
Using the coordinate position as new positive sample callout box, positive sample is optimized with this；
For negative sample, fixed hidden variable zP, give up the negative sample far from interface, successively optimize negative sample.
Step 3.3.2 optimizes initial root mould using stochastic gradient descent algorithm using the positive sample and negative sample after optimization
Plate F0。
Step 4, the components number and shape for setting battery-operated motor cycle utilize the initial root after the training of component scanning step 3
Template, when component shape and the energy maximum of root template overlay area, which is the position of component, to obtain
Initial part model；Specific step is as follows:
Step 4.1, the components number of battery-operated motor cycle is set first as 8, this 8 components correspond respectively to battery-operated motor cycle
8 parts；The shape of set parts is rectangle, and the area of each rectangle is identical, and the gross area of this 8 rectangles is step 3.3
The 80% of initial root template size after optimization.Component (i.e. the rectangle) scanning step 3.3 is utilized respectively to obtain after training
Overlay area when initial root template energy maximum after component and the optimization, is determined as the component by initial root template
Position to obtain the corresponding initial part model of the component, and sets 0 for the energy of the overlay area.Using identical
Method, initialize all 8 components, obtain the corresponding initial part model of each component.
Step 4.2, initial part model is loaded into the initial root template after the training, using LSVM in positive sample
It is trained on collection and negative sample collection, to obtain optimal electric motorcycle vehicle model, which is final battery-operated motor cycle mould
Type, (a) including final root template (initial root template obtains after training) such as Fig. 2 is shown, final partial model (initial portion
Part model obtains after training) as shown in (b) of Fig. 2.
Step 5, for image to be detected, feature extraction is carried out to image to be detected, then with the final root template
And final partial model is matched, and the position of battery-operated motor cycle in image to be detected is obtained.The matching algorithm can be adopted
With existing matching algorithm in the prior art.A kind of plain algorithm of matching is provided in the present embodiment, is specifically included:
Step 5.1, the feature of image to be detected is established first with step 3.2 identical method for image to be detected
Pyramid specifies interval=10 to carry out feature extraction to image to be detected herein.
Step 5.2, it is matched, is obtained electric in image to be detected with the final root template and final partial model
The position of dynamic motorcycle, the specific steps are as follows:
Step 5.2.1 carries out the pyramidal each layer of the feature of image to be detected with the final root template respectively
Convolution operation selects maximum convolution value as root template score in the convolution value that each layer is calculated.
Step 5.2.2 rolls up the final partial model of twice of resolution characteristics and each of image to be detected respectively
Product obtains the corresponding score of each final partial model；Twice of resolution characteristics of described image to be detected refer to be checked
Twice of resolution characteristics in altimetric image feature pyramid.
Step 5.2.3, the deformation for calculating final partial model are spent
As shown in (c) of Fig. 2, the deformation cost of i-th of final partial model is exactly the partial model in image to be detected
In physical location and anchor point position between departure degree；The anchor point position refers to the most terminal part after step 4.2 training
The position of central point of the part model in the final root template.
The algebraical sum of step 5.2.1 to step 5.2.3 acquired results is finally acquired, that is, seeks described template score, final
For total algebraical sum that the corresponding score of partial model and deformation are spent as detection total score, the position of highest scoring is target
Position.
Fig. 3 is the schematic diagram that battery-operated motor cycle testing result is carried out using the method for the present invention, from the figure, it can be seen that the mould
Type be suitable for different scale, different illumination, different deformation image battery-operated motor cycle detection, meanwhile, also can solve small range
Occlusion issue can be applied to the monitoring of battery-operated motor cycle in traffic.
Claims (5)
1. the battery-operated motor cycle image detecting method based on units of variance model, which comprises the following steps:
Step 1, positive sample collection and negative sample collection are established；
Step 2, classify to positive sample collection, then the grouping of negative sample collection carries out feature extraction to each positive sample, negative sample, obtains
To the feature of positive sample and the feature of negative sample；
The feature of the training positive sample and the feature of negative sample, obtain root template；
Step 3, the root template is obtained into initial root template by spatial clustering, all positive samples, negative sample is distinguished
Feature pyramid is established, then the training initial root template；
Step 4, the components number and shape for setting battery-operated motor cycle, the initial root template after being trained using component scanning step 3,
When component shape and the energy maximum of initial root template overlay area, which is the position of component, to obtain
Initial part model；
It is initial in template and enterprising in the positive sample collection and negative sample collection after initial part model to be loaded into training
Row training, obtains final root template and final partial model；
Step 5, for image to be detected, feature extraction is carried out to image to be detected, then with the final root template and
Final partial model is matched, and the position of battery-operated motor cycle in image to be detected is obtained.
2. battery-operated motor cycle image detecting method of the as described in claim 1 based on units of variance model, which is characterized in that
Described establishes positive sample collection and negative sample collection, comprising:
Using multiple include battery-operated motor cycle images as positive sample collection, electronic rub what positive sample was concentrated in each positive sample
Motorcycle carries out rectangle frame mark；Negative sample collection then chooses the figure not comprising battery-operated motor cycle of PASCAL VOC data set offer
Picture.
3. battery-operated motor cycle image detecting method of the as described in claim 1 based on units of variance model, which is characterized in that
The step 2 includes:
Step 2.1, the rectangle callout box for the positive sample that positive sample is concentrated is divided into m class according to length-width ratio, negative sample collection divides at random
At m group；
Step 2.2, feature extraction is carried out to each positive sample, negative sample, comprising:
For positive sample or negative sample image, multiple cells are divided the image into, 18 have to each pixel in cell point
Symbol gradient direction and 9 seek gradient without symbol gradient direction, and throwing is then weighted in the histogram of cell with gradient direction
Shadow obtains the gradient orientation histogram of the cell, the as feature vector of corresponding 27 dimension of the cell；Then by every 2*2
A adjacent cells lattice form a block, then each block corresponds to the feature vector of one 108 dimension, by it is carried out it is main at
108 original dimensional features are carried out the cumulative 4+27=31 that is down to each row and column respectively and tieed up, will owned by analysis and parsing dimensionality reduction
Block Characteristic vector, which is together in series, can be obtained the feature of whole image；Thus the feature of positive sample and the spy of negative sample are obtained
Sign；
Step 2.3, according to the feature of the feature of positive sample and negative sample, m root template F is obtained with SVM training1,F2,…,Fm。
4. battery-operated motor cycle image detecting method of the as described in claim 1 based on units of variance model, which is characterized in that
Described establishes feature pyramid to all positive samples, negative sample respectively, comprising:
For positive sample or negative sample image, using image as the pyramidal first layer of feature, then feature it is pyramidal other
Layer is one layer 2 above1/intervalMultiple sampled, each layer pyramidal for feature, according to step 2.2 method into
Row feature extraction, to obtain each layer of feature pyramid of feature.
5. battery-operated motor cycle image detecting method of the as described in claim 1 based on units of variance model, which is characterized in that
It is described that feature extraction is carried out to image to be detected, it is then carried out with the final root template and final partial model
Match, obtain the position of battery-operated motor cycle in image to be detected, comprising:
The feature pyramid for establishing image to be detected carries out the pyramidal each layer of this feature and the final root template respectively
Convolution operation selects maximum convolution value as root template score；
From the feature pyramid of image to be detected, twice of resolution characteristics and each final component of image to be detected are selected
Model carries out convolution respectively, obtains the corresponding score of each final partial model；
The deformation for calculating final partial model is spent；
The total algebraical sum for asking described template score, the corresponding score of final partial model and deformation to spend must as detection
Point, the position of highest scoring is the position of battery-operated motor cycle.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810877373.5A CN109344684A (en) | 2018-08-03 | 2018-08-03 | Battery-operated motor cycle image detecting method based on units of variance model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810877373.5A CN109344684A (en) | 2018-08-03 | 2018-08-03 | Battery-operated motor cycle image detecting method based on units of variance model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109344684A true CN109344684A (en) | 2019-02-15 |
Family
ID=65296768
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810877373.5A Pending CN109344684A (en) | 2018-08-03 | 2018-08-03 | Battery-operated motor cycle image detecting method based on units of variance model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109344684A (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103295021A (en) * | 2012-02-24 | 2013-09-11 | 北京明日时尚信息技术有限公司 | Method and system for detecting and recognizing feature of vehicle in static image |
| CN104134071A (en) * | 2014-06-20 | 2014-11-05 | 北京工业大学 | Deformable part model object detection method based on color description |
| CN105225226A (en) * | 2015-09-02 | 2016-01-06 | 电子科技大学 | A kind of cascade deformable part model object detection method based on Iamge Segmentation |
-
2018
- 2018-08-03 CN CN201810877373.5A patent/CN109344684A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103295021A (en) * | 2012-02-24 | 2013-09-11 | 北京明日时尚信息技术有限公司 | Method and system for detecting and recognizing feature of vehicle in static image |
| CN104134071A (en) * | 2014-06-20 | 2014-11-05 | 北京工业大学 | Deformable part model object detection method based on color description |
| CN105225226A (en) * | 2015-09-02 | 2016-01-06 | 电子科技大学 | A kind of cascade deformable part model object detection method based on Iamge Segmentation |
Non-Patent Citations (1)
| Title |
|---|
| 马俊凯 等: "基于可变形模型的目标跟踪算法", 《红外与激光工程》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ahmad et al. | Object detection through modified YOLO neural network | |
| CN106874894B (en) | Human body target detection method based on regional full convolution neural network | |
| CN105046196B (en) | Front truck information of vehicles structuring output method based on concatenated convolutional neutral net | |
| CN107944450B (en) | License plate recognition method and device | |
| EP2648159A1 (en) | Object detecting method and object detecting device using same | |
| CN109376580B (en) | Electric power tower component identification method based on deep learning | |
| CN107480585B (en) | Target detection method based on DPM algorithm | |
| CN113920107A (en) | Insulator damage detection method based on improved yolov5 algorithm | |
| CN112295933B (en) | Method for robot to rapidly sort goods | |
| CN104881665A (en) | Chip character identification and verification method and apparatus | |
| CN111915583B (en) | Vehicle and pedestrian detection method based on vehicle-mounted thermal infrared imager in complex scene | |
| JP4946878B2 (en) | Image identification apparatus and program | |
| CN110728307A (en) | Method for realizing small sample character recognition of X-ray image by self-generating data set and label | |
| CN114049572A (en) | Detection method for identifying small target | |
| CN113076804B (en) | Target detection method, device and system based on YOLOv4 improved algorithm | |
| CN114332942A (en) | Night infrared pedestrian detection method and system based on improved YOLOv3 | |
| CN108564043B (en) | Human body behavior recognition method based on space-time distribution diagram | |
| CN112101113B (en) | Lightweight unmanned aerial vehicle image small target detection method | |
| CN116682101A (en) | License plate number recognition method and system | |
| CN109344684A (en) | Battery-operated motor cycle image detecting method based on units of variance model | |
| Ponsa et al. | Cascade of classifiers for vehicle detection | |
| Palmer et al. | Scale proportionate histograms of oriented gradients for object detection in co-registered visual and range data | |
| CN111178158B (en) | Rider detection method and system | |
| CN112700415A (en) | Pen electricity shell defect rapid detection algorithm | |
| CN113643544A (en) | Intelligent detection method and system for illegal parking in parking lot based on Internet of things |
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: 20190215 |