CN109508682A - A kind of detection method on panorama parking stall - Google Patents
A kind of detection method on panorama parking stall Download PDFInfo
- Publication number
- CN109508682A CN109508682A CN201811385268.6A CN201811385268A CN109508682A CN 109508682 A CN109508682 A CN 109508682A CN 201811385268 A CN201811385268 A CN 201811385268A CN 109508682 A CN109508682 A CN 109508682A
- Authority
- CN
- China
- Prior art keywords
- parking stall
- information
- parking
- convolutional neural
- neural networks
- 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
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
- G06V20/586—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of parking space
Abstract
A kind of detection method on panorama parking stall, belongs to parking stall detection field.This method is a kind of parking space detection method based on panoramic picture.This method marks the image on panorama parking stall first, establishes the probabilistic model on parking stall, and generate label information to the parking stall of mark.Then the contour images for establishing the convolutional neural networks extraction parking stall of multiple features fusion separate the location information on parking stall finally according to the profile information on parking stall.The method does not need shooting background image, insensitive to the variation of external environment, greatly improves the satisfaction of user.The algorithm acquires parking lot image information at any time, under the busy state status in parking stall, can quickly accurately detect parking stall, hence it is evident that better than other method for detecting parking stalls.
Description
Technical field
The invention belongs to the technical fields that Digital Image Processing and parking stall are detected, and in particular to a kind of panorama parking stall
Detection method.
Background technique
The wire frame of the image information on the parking stall acquired at present, parking stall can break or smudgy, this gives parking stall
Detection bring difficulty；Common parking stall detection algorithm needs shooting background image (parking stall figure when without vehicle first
Picture), this performance difficulty, if camera needs to re-shoot, this can be to user since external force (wind etc.) causes to shift
Bring undesirable experience.
Summary of the invention
Existing parking stall detection technique needs shooting background image (parking field picture when without vehicle) first, in fact
Difficulty is applied, the interference by external environment is easy, causes the bad experience of user.In view of the above problems, the invention discloses one kind
Panorama parking space detection method.This method is based on panorama parking bit image, and the convolutional neural networks for establishing multiple features fusion extract
Then the profile information on parking stall detects parking stall using the parking stall detection algorithm based on profile.
Technical scheme is as follows:
A kind of detection method on panorama parking stall includes:
Obtain the panoramic picture on parking stall；
The panoramic picture on the parking stall is labeled, obtains markup information, and stop according to markup information foundation
The probabilistic model of parking stall；
According to the probabilistic model on the parking stall, the label information on parking stall is generated；The wherein probability mould on each parking stall
Type and label information are one-to-one；
Establish the convolutional neural networks of multiple features fusion；The convolutional neural networks include multiple stages, and each stage is negative
Duty extracts the characteristic information in the stage；The convolutional neural networks have merged the characteristic information in multiple stages；The convolutional Neural
Network is trained using the label information on the parking stall, obtains the parameter of the convolutional neural networks；The convolutional Neural
Network utilizes the parameter of the convolutional neural networks, classifies to the parking bit image of input, exports the probability graph on parking stall,
The probability graph is handled by image normalization, obtains the profile information of parking image；
According to the profile information on parking stall, the location information on parking stall is isolated.
Further, the panoramic picture to the parking stall is labeled, and obtains markup information, and according to the mark
Note information establishes the probabilistic model on parking stall, comprises the following steps:
The panoramic picture on the parking stall is repeatedly marked, the markup information of parking bit image is obtained；
According to the markup information of the parking bit image, the probabilistic model on parking stall is established；
Wherein, the probability of each pixel on parking stall are as follows: marking this pixel is the labeled times on parking stall divided by mark
The total degree of this pixel；
Further, the panoramic picture to the parking stall is labeled, and obtains markup information, and according to the mark
The probabilistic model that note information establishes parking stall also includes before:
Distortion correction is carried out using affine transformation to the panoramic picture on the parking stall；
Further, the markup information according to the parking bit image, is established before the probabilistic model on parking stall also
Include:
To every parking bit image for being repeated as many times mark, merged within the scope of error σ；
Further, the stage described in the convolutional neural networks of the multiple features fusion further include, in first stage not
There are warp laminations, there is a warp lamination in remaining stage.
Further, the stage described in the convolutional neural networks of the multiple features fusion further includes wrapping in each stage
Convolution kernel containing 1*1.
Further, the stage described in the convolutional neural networks of the multiple features fusion further includes each of each stage
The output that the pixel value of the corresponding position of convolutional layer was summed as the current generation；
Further, the convolutional neural networks of the multiple features fusion using the loss function with sample tilt correction item into
Row training；
Further, the profile information according to parking stall, isolates the location information on parking stall, comprising following
Step:
According to the profile information on parking stall, the framework information on this parking stall is calculated；
According to the framework information on parking stall, straight line is detected；
The attribute for extracting parking stall filters the straight line, obtains first straight line information；
Clustering is carried out to the first straight line information according to the slope information of the first straight line information, deletes tuftlet
Class obtains second straight line information；
Straight line is carried out to second straight line information and extends operation, obtains first area；
Using area threshold T, the first area is deleted, the region that area is less than threshold value T is deleted, obtains the secondth area
Domain；
According to the feature of adjacent area, the second area is deleted, deletes the lesser area of adjacent area similarity
Domain obtains third region；
The third region is parking stall image information.
The invention proposes a kind of detection method on panorama parking stall, this method is a kind of parking stall based on panoramic picture
Detection method.This method marks the image on panorama parking stall first, establishes the probabilistic model on parking stall, and to the parking stall of mark
Generate label information.Then the convolutional neural networks for establishing multiple features fusion extract the contour images on parking stall, finally use base
Parking space information is detected in the parking space detection method of profile.
The method does not need shooting background image, insensitive to the variation of external environment, greatly improves expiring for user
Meaning degree.The algorithm acquires parking lot image information at any time, can be quickly accurate under the busy state status in parking stall
Detect parking stall, hence it is evident that better than other method for detecting parking stalls.
Detailed description of the invention
Fig. 1 is a kind of detection method flow chart on panorama parking stall
Fig. 2 is the stage schematic diagram of the convolutional neural networks of multiple features fusion
Fig. 3 is the schematic diagram of the convolutional neural networks of multiple features fusion
Fig. 4 is a kind of flow chart of parking space detection method based on profile
Fig. 5 is a kind of flow chart of parking space detection method based on correction parking bit image
Specific embodiment
In order to enable the object, technical solutions and advantages of the present invention become apparent, root is described in detail below with reference to accompanying drawings
According to the example embodiment of the disclosure.Obviously, described embodiment is only a part of this disclosure embodiment, rather than this public affairs
The whole embodiments opened, it should be appreciated that the disclosure is not limited by example embodiment described herein.Based on described in the disclosure
The embodiment of the present disclosure, those skilled in the art's obtained all other embodiment in the case where not making the creative labor
It should all fall within the protection scope of the disclosure.
A kind of embodiment one: detection method on panorama parking stall
A kind of detection method on panorama parking stall, as shown in Figure 1, comprising:
S1: the panoramic picture on parking stall is obtained；
The panoramic picture on the parking stall can be obtained by the way of hardware and/or software.If using hardware mode
Panoramic picture is obtained, panorama camera can be used；If obtaining panoramic picture using software mode, can be obtained using fisheye camera
Image is taken, panoramic picture is then obtained using distortion correction；Or obtain the image of different zones respectively using more general cameras,
It is required that the visual field of adjacent camera at least N ° of coincidence, wherein N > 10, use panoramic mosaic technology for the multiple image of acquisition
Obtain panoramic picture.Wherein the distortion correction and panoramic mosaic technology have had mature algorithm.
The parking stall scene requirement diversification of acquisition.Distinguish from the position on parking stall: indoor parking position, high-speed service area stop
Parking stall, parking position, community parking position, market parking stall etc.；Distinguish from lighting angle: parking stall, rainy day under blazing sun are stopped
Parking stall, greasy weather parking stall, the parking stall in sand and dust, owl-light time-division parking stall, night parking stall；The type on parking stall: tiltedly parking
Position, upright parking stall etc..
The quantity on the parking stall of each scene obtained wants balanced.The image on the panorama parking stall obtained herein is for instructing
Practice the parameter of the convolutional neural networks of multiple features fusion.In order to avoid accuracy rate caused by data skew declines, each scene
It is too big that the quantity of parking bit image is unable to gap.
S3: being labeled the panoramic picture on the parking stall, obtains markup information, and establish according to the markup information
The probabilistic model on parking stall；
The present invention can be labeled the panoramic picture on the parking stall using automanual mode, it may be assumed that more people repeat
The parking bit image collected in step S1 is labeled using parking stall marking software, obtains its markup information.Parking stall mark
Note software is a tool for facilitating mark personnel's labeled data, and those skilled in the art can by purchase or voluntarily
Acquisition is write, the present invention will not do excessive elaboration.
Pay special attention to, when marking parking stall, the phenomenon that being obscured by an object if there is parking stall region, mark personnel should
Parking stall region is completely drawn out according to common sense.Wherein, it for the parking stall region blocked, to be labeled.
According to the markup information of the parking bit image, the probabilistic model on parking stall is established.
Assuming that image I is labelled with K times altogether, the pixel of picture position (x, y) is denoted as I (x, y), the i-th mark of I (x, y)
It is denoted as Ii(x, y), the i-th of I (x, y) are labeled as parking stall and are denoted asThe then probability of the position (x, y) of image I
Are as follows:
Wherein function f are as follows:
According to formula 1- formula 2, image I is calculated in the probability of the position position (x, y).And so on calculate it is every in image I
Probability at a position generates the probability graph P of image II。
S4: according to the probabilistic model on the parking stall, the label information on parking stall is generated；The wherein probability on each parking stall
Model and its label information are one-to-one；
The label information includes: positive sample and negative sample, is respectively used to indicate whether as parking stall.Present case uses
Threshold value η marks the position (x, y) of image I:
Wherein: LI(x, y)=1 is positive sample, indicates that this position is parking stall, LI(x, y)=0 is negative sample, indicates this position
Setting is not parking stall.
0<PI(x, y) < η (formula 4)
When the probability on the parking stall is less than the minimum probability of positive sample, this parking bit image does not generate label information.This
Sample in formula 4 is directly given up in invention, and different labelers recognizes difference to the sample in this section, if introducing such sample,
Meeting contamination data, influences disaggregated model.
S5: the convolutional neural networks of multiple features fusion are established；The convolutional neural networks include multiple stages, each stage
It is responsible for extracting the characteristic information in the stage；The convolutional neural networks have merged the characteristic information in multiple stages；The convolution mind
It is trained through network using the label information on the parking stall, obtains the parameter of the convolutional neural networks；The convolution mind
The parameter that the convolutional neural networks are utilized through network, classifies to the parking bit image of input, exports the probability on parking stall
Figure, the probability graph are handled by image normalization, obtain the profile information of parking image；
The present invention uses the convolutional neural networks of multiple features fusion.The convolutional neural networks fusion of the multiple features fusion is each
The characteristic information of layer convolutional layer can not only extract the texture information of image, but also can be with the semanteme letter of Simultaneous Extracting Image
Breath.The texture information of described image, marginal information, colouring information comprising image etc.；The semantic information of described image includes figure
Wheel information, headstock information of image of picture etc..In the panoramic picture on the parking stall, the present invention is directed to detector parking stall letters
Breath.The parking space information not only should include the marginal information of bottom, but also should include high-rise straight line information, right angle letter
Breath etc. can show that the authentication information of parking stall feature.The authentication information of the parking stall feature, is not limited to parking space information, can
To include information of vehicles.That is, vehicle, which exists, can be used as an auxiliary identification means existing for parking stall.
Further, there was only a small amount of object in the panoramic picture on the parking stall, characteristic information amount is few, therefore
When the present invention extracts feature using the convolutional neural networks of multiple features fusion, general convolutional layer is only considered.
The convolutional neural networks of the multiple features fusion include N number of stage.Analogy VGG model, the multiple features fusion
Convolutional neural networks include 5 stages, as shown in Figure 3.The convolutional neural networks of the multiple features fusion include stage 1 to the stage
5, i.e. stage1-stage5.Wherein each stage is approximately uniform, and maximum value pond layer is used between adjacent phases
maxpooling.Maximum value pond of the invention using common 2*2 core, step-length 2.It is conceivable that input feature vector passes through
The resolution ratio of maximum value Chi Huahou, feature can halve, and the calculation amount of network will be effectively reduced in this.In addition, the meeting pair of maximum value pondization
The characteristic information of input is summarized, and more advanced characteristic information is inputted.The network includes articulamentum concat, is used to
The characteristic information in each stage is summarized, the characteristic information on total parking stall is formed.This characteristic information not only includes bottom
Parking stall characteristic information, such as stage1, stage2；Also comprising high-rise characteristic information, such as stage4, stage5.It is worth
It is noted that each stage all can voluntarily optimize the parameter in the stage, this operation be in order to avoid gradient disperse or
Gradient explosion.Articulamentum concat connects the characteristic information before each perfecting by stage, i.e., the characteristic information in each stage passes through
Liang Ge branch, one branches into articulamentum concat, and a branch carries out the optimization layer of current generation.The feature in all stages
Information is summarized at articulamentum concat, then obtains the parking bitmap by the convolution kernel operation of the 1*1 in 1 channel
The probabilistic image of picture, the probabilistic image are sent into the cost loss layer LOSS of the convolutional neural networks of the multiple features fusion.Generation
Valence loss layer LOSS method optimizes the network parameter of the convolutional neural networks of the multiple features fusion using back-propagation algorithm.It is all
Network parameter generate the multiple features fusion multiple features fusion model.
The stage in the convolutional neural networks of the multiple features fusion is as shown in Figure 2.The stage includes two kinds of knots
Structure, as shown in Fig. 2-a and Fig. 2-b.Compared with Fig. 2-a, the more warp lamination DECONV of Fig. 2-b, i.e. Fig. 2-a is without warp lamination
There are a warp lamination DECONV by DECONV, Fig. 2-b.The convolutional neural networks of multiple features fusion of the invention include to include N
In a stage, wherein first stage corresponding diagram 2-a, does not include warp lamination DECONV；N-th stage corresponding diagram 2-b, packet
Containing a warp lamination DECONV, wherein N > 1.Other than above-mentioned distinguishing characteristics, the stage of the invention is identical, includes two-way
The 1*1 convolution in the N2 channel of 3*3 convolution sum in N1 channel, the two-way characteristic information summarize at EltWise.It is described
EltWise uses respective pixel sum operation, generates the characteristic information in the stage.If there are warp laminations in this stage
DECONV, then the characteristic information in the stage is up-sampled by warp lamination, exports the fusion feature in the stage；If there is no anti-
Convolutional layer DECONV, then the characteristic information in the stage is the fusion feature in the stage.It is worth noting that, in order to avoid gradient
Network caused by disperse or gradient are exploded is not restrained, and it is excellent using cost function LOSS progress that the present invention is directed to each stage respectively
Change, obtains the model parameter in this stage.
In each stage, non-linearization is carried out to the output of each convolutional layer.Such as the convolution of the 3*3 in N1 channel
It exports to nonlinear activation layer as inputting, non-linearization is carried out to convolutional layer, the output of nonlinear activation layer is sent into N2
The convolutional layer of the convolutional layer of the 1*1 in channel, the 1*1 in N2 channel passes through nonlinear activation function, is sent into EltWise layers.It is not difficult
Know, formally due to the presence of nonlinear activation function, convolution signal is subjected to non-linearization, the abundant of network could be extracted
Characteristic information.The present invention after each convolutional layer can connected nonlinearity activation primitive, for convolutional layer carry out non-linearization.
S6: according to the profile information on parking stall, the location information on parking stall is separated.
Its packet following steps, as shown in Figure 4:
Step S601: according to the profile information on parking stall, the framework information on this parking stall is calculated；
The panoramic picture on the parking stall is handled by the convolutional neural networks of the multiple features fusion, obtains the parking
The profile information of bit image.Compared to the marginal information that the edge detection algorithms such as canny extract, this profile information is believed better than edge
Breath has merged high-rise semantic information.For the bit image that stops, this profile information contains the profile information of vehicle, ground
Profile information, wheel aligner, rubber car lug, the pedestrian contour information on parking stall etc..The profile is a wheel in kind
Exterior feature has very strong anti-interference ability to noise.It is contemplated that the profile information on parking stall includes straight line information, in order in image
The information on parking stall is isolated in profile information, the present invention first extracts the skeleton line information of profile.Skeleton line extraction algorithm includes
Iterative algorithm and noniterative algorithm, noniterative algorithm generate a certain intermediate value or center line of lines by way of once traversing,
Without checking all single pixels, i.e., this algorithm is primary i.e. generation skeleton；Such as based on the method for range conversion.Run length
Coding refinement etc.；The pixel that iterative algorithm deduplication image border is met certain condition, finally obtains single pixel wideband backbone.
The present invention does not limit skeleton line extraction algorithm.
Step S602: according to the framework information on parking stall, straight line is detected；
The framework information includes some sparse marginal informations, and the present invention detects straight line information on this basis, is rejected
Remaining information.The marginal information one on parking stall is set to straight line information, rejects some non-directional edges here by straight-line detection
Information.Straight-line detection can be using Hough transformation (Hough Transform) or using LSD (Line Segment
Detector) algorithm etc., this belongs to algorithm appreciated by those skilled in the art, does not do excessive explanation herein.
Step S603: extracting the attribute on parking stall, filters the straight line, obtains first straight line information；
The straight line information isolated in the framework information, the straight line information on the parking stall being not all of may include one
A little noise informations, such as information, the information of people, the information in corner of license plate.The present invention extracts the attribute on parking stall, believes straight line
Breath is filtered.The attribute on parking stall herein can be length, the distance of adjacent straight line of straight-line segment.According to straight line line
Threshold value is arranged in the length of section, deletes very short straight-line segment；The distance of adjacent straight line is close, then deletes adjacent straight line.
Step S604: cluster point is carried out to the first straight line information according to the slope information of the first straight line information
Tuftlet class is deleted in analysis, obtains second straight line information；
The present invention also carries out clustering to the first straight line information using the method for cluster, this cluster is according to straight line
Slope is clustered, and is clustered as K cluster, and lesser cluster is deleted, and retains major class cluster.Clustering method herein can use
Kmeans or other clustering methods.
Step S605: straight line is carried out to second straight line information and extends operation, obtains first area；
For the second straight line information, the present invention extends operation using straight line, extends by straight line and operates, will acquire very
Multi-region domain information.
Step S606: using area threshold T, delete the first area, deletes the region that area is less than threshold value T, obtains
To second area；
For the first area, the method that the present invention first uses area threshold is deleted.For small area and big face
Long-pending region carries out deleting operation.Region at this time may have the case where overlapping.
Step S607: according to the feature of adjacent area, deleting the second area, deletes adjacent area similarity
Lesser region obtains third region；
For the second area, deleted according to the characteristic information of adjacent area.Characteristic information herein includes face
Product information, slope information etc..Even if there is affine transformation, area, slope of adjacent area etc. are not much different.It is deleted by this step
Subtract, guarantees that there is no the relationship for mutually including between region, i.e. region A is fully fallen in the B of region.
Step S608: the third region is parking stall location drawing picture.
Case two, a kind of parking space detection method based on correction parking bit image
The image that video camera is generally shot will appear the phenomenon that " near big and far smaller ", this phenomenon is known as distorting.For this distortion
Image, the present invention are corrected using affine transformation.A kind of parking space detection method such as Fig. 5 institute based on correction parking bit image
Show.
The present invention increases step S2 between the step S1 and step S3 of case one:
S2: distortion correction is carried out using affine transformation to parking bit image；
The phenomenon that parking bit image usually obtained will appear " near big and far smaller ", this phenomenon present invention, which is referred to as, to distort.For
Such distortion, the present invention are corrected using the method for affine transformation.Affine transformation is scaling, translation, rotation, reflection, mistake cuts
Deng a multiple any combination.Method known to its technical staff for ability, the present invention do not do excessive explanation.
Scheme three, a kind of mask method for the bit image that stops
In the image labeling of parking stall, for some parking stall lines blocked by vehicle etc., need manually according to common sense
It is drawn.Since understanding of the different people to parking stall is different, different location of pixels may be plotted to.As two people draw
The line segment on two parking stalls be parallel lines, but this parallel lines is not overlapped, and the distance between this parallel lines are 5 pictures
Element.For this phenomenon, the notation methods on both parking stalls are correct, only there is some errors, and this programme passes through to difference
The same parking stall that people draws is merged, and above-mentioned error is eliminated.
S1: the panoramic picture on parking stall is obtained；
The panoramic picture on the parking stall can be obtained by the way of hardware and/or software.If using hardware mode
Panoramic picture is obtained, panorama camera can be used；If obtaining panoramic picture using software mode, can be obtained using fisheye camera
Image is taken, panoramic picture is then obtained using distortion correction；Or obtain the image of different zones respectively using more general cameras,
It is required that the visual field of adjacent camera at least N ° of coincidence, wherein N > 10, use panoramic mosaic technology for the multiple image of acquisition
Obtain panoramic picture.Wherein the distortion correction and panoramic mosaic technology have had mature algorithm.
The parking stall scene requirement diversification of acquisition.Distinguish from the position on parking stall: indoor parking position, high-speed service area stop
Parking stall, parking position, community parking position, market parking stall etc.；Distinguish from lighting angle: parking stall, rainy day under blazing sun are stopped
Parking stall, greasy weather parking stall, the parking stall in sand and dust, owl-light time-division parking stall, night parking stall；The type on parking stall: tiltedly parking
Position, upright parking stall etc..
The quantity on the parking stall of each scene obtained wants balanced.The image on the panorama parking stall obtained herein is for instructing
Practice the parameter of the convolutional neural networks of multiple features fusion.In order to avoid accuracy rate caused by data skew declines, each scene
It is too big that the quantity of parking bit image is unable to gap.
S3: being labeled the panoramic picture on the parking stall, obtains markup information, and establish according to the markup information
The probabilistic model on parking stall；
Wherein step S3 is comprised the following steps:
S301: repeatedly marking the panoramic picture on the parking stall, obtains the markup information of parking bit image；
The present invention can be labeled the panoramic picture on the parking stall using automanual mode, it may be assumed that more people repeat
The parking bit image collected in step S1 is labeled using parking stall marking software, obtains its markup information.Parking stall mark
Note software is a tool for facilitating mark personnel's labeled data, and those skilled in the art can by purchase or voluntarily
Acquisition is write, the present invention will not do excessive elaboration.
Pay special attention to, when marking parking stall, the phenomenon that being obscured by an object if there is parking stall region, mark personnel should
Parking stall region is completely drawn out according to common sense.Wherein, it for the parking stall region blocked, to be labeled.
S302: it to the markup information of the parking bit image, is merged within the scope of error σ；
When being labeled to the panoramic picture on the parking stall, for the parking stall region blocked by vehicle etc., root is needed
It is drawn according to common sense.Different people may draw different parking stalls in different positions.Such as two of two people drafting
One line segment on parking stall is parallel lines, but this parallel lines is not overlapped, and the distance between this parallel lines are 5 pixels.For
The notation methods of this phenomenon, both parking stalls are correct, only there is some errors.
Above-mentioned error will affect the convergence rate of machine learning algorithm.This error may not be more than 20pixel, Ren Leiji
It can't see, but for machine, this will be very big error.This error can obscure machine, allow machine to parking stall
Detect obscure, increase machine learning difficulty.Therefore the present invention eliminates this error by the way of straight line fusion.
Assuming thatFor i-th of people mark k-th of parking stall,For k-th of parking stall of j-th of people mark, lkFor straight line
Fused parking stall, then:
Wherein x is the abscissa when value straight line.
S303: according to the markup information of the parking bit image, the probabilistic model on parking stall is established；
According to formula 1- formula 2, image I is calculated in the probability of the position position (x, y).And so on calculate it is every in image I
Probability at a position generates the probability graph P of image II。
Case four, using the multiple features fusion with sample correction term convolutional neural networks optimization method
In order to optimize the convolutional neural networks of the multiple features fusion, present invention introduces the loss functions with sample correction term
LOSS.Loss function uses basic logistic regression, and loss function is as follows:
Wherein: η is the threshold value of parking probabilistic model, is greater than this threshold value, shows there is parking stall；yiFor the parking really marked
The probabilistic model of position；P(Xi；It W is) the image X at model parameter WiBelong to the probability on parking stall；α is the penalty coefficient of negative sample；
The penalty coefficient of β sample as evidence；|Y+| it is the number of positive sample；|Y-| it is the number of samples of negative sample；λ is to balance positive negative sample
Coefficient.
A kind of case five: detection method on residue parking stall
The remaining parking stall for counting parking lot, is the important function in one, present parking lot.Stopped based on panorama of the invention
Driving skills art can very easily count the number of free parking space.
The present invention increases step S7 after case one: according to the parking stall location drawing picture, counting parking stall sum；For
Each parking stall location drawing picture detects information of vehicles, calculating vehicle sum；It is total according to the parking stall sum and vehicle
Number, obtains remaining parking stall number.
Using the panorama parking stall detection technique of scheme one to scheme five, each parking stall is detected, and count parking stall
Number be Ntotal.Using vehicle detecting algorithm, current vehicle is detected, and the number of calculating vehicle is Ncar.Then current residual is stopped
The number of parking stall is (Ntotal-Ncar)。
Case six, parking stall recommended technology
The parking stall management urgent need in parking lot wants a kind of parking stall recommended technology at present.Many people have such experience, open
Vehicle in parking lot around for a long time, just in order to find a parking stall.It altogether include the parking of four layers of underground especially for megastore
, user needs to look for for a long time available parking places.Therefore, it is badly in need of wanting a kind of parking stall recommended technology, meets the needs of users.
The present invention increases step S8 after case one: according to the parking stall location drawing picture, establishing parking stall schematic diagram；Needle
To each parking stall location drawing picture, information of vehicles is detected, and establishes idle parking stall list；According to parking stall schematic diagram and sky
Not busy parking stall list guides vehicle parking.
Using the panorama parking stall detection technique of scheme one to scheme five, each parking stall is detected, and to each parking
Position successively label, such as KiIndicate i-th of parking stall.Current parking stall is differentiated using vehicle testing techniques for each parking stall
Whether there is vehicle, if current parking is no vehicle, is added into idle parking stall list Q.By this idle parking stall list hair
It send and is shown in garage large-size screen monitors, user is facilitated to search suitable parking stall.
Case seven, a kind of detection method of parking stall thievery
Car has become a part of people's life, and some valuables are put in car by many people's habits, this will lead
Vehicle theft event is caused to occur again and again.The present invention proposes a kind of parking stall thievery detection method, for vehicle theft event
The behavior of generation and alarm, in the case where reducing for losing, to thief with watchful.
The present invention increases step S9 after case one: detecting pedestrian for the panoramic picture on the parking stall, and establishes row
The trace information of people；For each parking stall location drawing picture, information of vehicles is detected, and establishes idle parking stall list；Analysis
Pedestrian track information is alarmed if pedestrian track is hovered near the idle parking stall of expense.
Using the panorama parking stall detection technique of scheme one to scheme five, each parking stall is detected.Using pedestrian detection
The pedestrian in algorithm detection parking lot and the trace information for establishing pedestrian, if the trace information of pedestrian continues on some parking stall
On hover or the trace information of pedestrian loses on some parking stall and this parking stall is in busy before the deadline
State is then alarmed, and parking stall thievery will occur for prompt.
Claims (9)
1. a kind of detection method on panorama parking stall, includes:
Obtain the panoramic picture on parking stall；
The panoramic picture on the parking stall is labeled, obtains markup information, and parking stall is established according to the markup information
Probabilistic model；
According to the probabilistic model on the parking stall, the label information on parking stall is generated；Wherein the probabilistic model on each parking stall and
Its label information is one-to-one；
Establish the convolutional neural networks of multiple features fusion；The convolutional neural networks include multiple stages, and each stage extracts should
The characteristic information in stage；The convolutional neural networks have merged the characteristic information in multiple stages；The convolutional neural networks use
The label information on the parking stall is trained, and obtains the parameter of the convolutional neural networks；The convolutional neural networks utilize
The parameter of the convolutional neural networks classifies to the panoramic picture on the parking stall of input, exports the probability graph on parking stall, institute
The probability graph for stating parking stall is handled by image normalization, obtains the profile information of parking image；
According to the profile information on parking stall, the location information on parking stall is separated.
2. the method as described in claim 1, it is characterised in that: the panoramic picture to the parking stall is labeled, and is obtained
Markup information is taken, and establishes the probabilistic model on parking stall according to the markup information, is comprised the following steps:
The panoramic picture on the parking stall is repeatedly marked, the markup information of parking bit image is obtained；
According to the markup information of the parking bit image, the probabilistic model on parking stall is established；
Wherein, the probability of each pixel on parking stall are as follows: marking this pixel is the labeled times on parking stall divided by this picture of mark
The total degree of element.
3. the method as described in claim 1, it is characterised in that: the panoramic picture to the parking stall is labeled, and is obtained
Markup information is taken, and also includes before according to the probabilistic model that the markup information establishes parking stall:
Distortion correction is carried out using affine transformation to the panoramic picture on the parking stall.
4. method according to claim 2, it is characterised in that: the markup information according to the parking bit image is established
Also include before the probabilistic model on parking stall:
To the markup information of the parking bit image, merged within the scope of error σ.
5. the method as described in claim 1, it is characterised in that: the stage described in the convolutional neural networks of the multiple features fusion
Further include that warp lamination is not present in first stage, there is a warp lamination in remaining stage.
6. method as claimed in claim 5, it is characterised in that: the stage described in the convolutional neural networks of the multiple features fusion
Further including includes the convolution kernel of 1*1 in each stage.
7. method as claimed in claim 5, it is characterised in that: the stage described in the convolutional neural networks of the multiple features fusion
It further include the output that the pixel value of the corresponding position of each convolutional layer in each stage was summed as the current generation.
8. the method as described in claim 1, it is characterised in that: the convolutional neural networks of the multiple features fusion use band sample
The loss function of tilt correction item is trained.
9. the method as described in claim 1, it is characterised in that: the profile information according to parking stall, separation parking
The location information of position comprising the steps of:
According to the profile information on parking stall, the framework information on this parking stall is calculated；
According to the framework information on parking stall, straight line is detected；
The attribute for extracting parking stall filters the straight line, obtains first straight line information；
Clustering is carried out to the first straight line information according to the slope information of the first straight line information, deletes tuftlet class,
Obtain second straight line information；
Straight line is carried out to second straight line information and extends operation, obtains first area；
Using area threshold T, the first area is deleted, the region that area is less than threshold value T is deleted, obtains second area；
According to the feature of adjacent area, the second area is deleted, the lesser region of adjacent area similarity is deleted, obtains
To third region；
The third region is parking stall image information.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811385268.6A CN109508682A (en) | 2018-11-20 | 2018-11-20 | A kind of detection method on panorama parking stall |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811385268.6A CN109508682A (en) | 2018-11-20 | 2018-11-20 | A kind of detection method on panorama parking stall |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109508682A true CN109508682A (en) | 2019-03-22 |
Family
ID=65749345
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811385268.6A Pending CN109508682A (en) | 2018-11-20 | 2018-11-20 | A kind of detection method on panorama parking stall |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109508682A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110299028A (en) * | 2019-07-31 | 2019-10-01 | 深圳市捷顺科技实业股份有限公司 | Method, apparatus, equipment and the readable storage medium storing program for executing of line detection are got in a kind of parking |
| CN110458912A (en) * | 2019-08-08 | 2019-11-15 | 金瓜子科技发展（北京）有限公司 | A kind of vehicle icon treating method and apparatus |
| CN111428616A (en) * | 2020-03-20 | 2020-07-17 | 东软睿驰汽车技术（沈阳）有限公司 | Parking space detection method, device, equipment and storage medium |
| CN111862302A (en) * | 2019-04-12 | 2020-10-30 | 北京城市网邻信息技术有限公司 | Image processing method, image processing apparatus, object modeling method, object modeling apparatus, image processing apparatus, object modeling apparatus, and medium |
| CN117079359A (en) * | 2023-09-04 | 2023-11-17 | 厦门中卡科技股份有限公司 | Intelligent parking payment method and system based on neural network |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090085772A1 (en) * | 2007-09-28 | 2009-04-02 | Ching-Chun Huang | System And Method Of Image-Based Space Detection |
| JP2011215052A (en) * | 2010-03-31 | 2011-10-27 | Aisin Aw Co Ltd | Own-vehicle position detection system using scenic image recognition |
| CN107610506A (en) * | 2017-09-04 | 2018-01-19 | 浙江工商大学 | The detection method and system of parking position state |
| CN107738612A (en) * | 2017-09-22 | 2018-02-27 | 西安电子科技大学 | The detection of automatic parking parking stall and identifying system based on panoramic vision accessory system |
| CN108022431A (en) * | 2017-12-04 | 2018-05-11 | 珠海横琴小可乐信息技术有限公司 | A kind of method and system that parking stall is detected by video capture image |
| CN108766022A (en) * | 2018-06-11 | 2018-11-06 | 青岛串并联电子科技有限公司 | Parking position state identification method based on machine learning and system |
| CN108830188A (en) * | 2018-05-30 | 2018-11-16 | 西安理工大学 | Vehicle checking method based on deep learning |
-
2018
- 2018-11-20 CN CN201811385268.6A patent/CN109508682A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090085772A1 (en) * | 2007-09-28 | 2009-04-02 | Ching-Chun Huang | System And Method Of Image-Based Space Detection |
| JP2011215052A (en) * | 2010-03-31 | 2011-10-27 | Aisin Aw Co Ltd | Own-vehicle position detection system using scenic image recognition |
| CN107610506A (en) * | 2017-09-04 | 2018-01-19 | 浙江工商大学 | The detection method and system of parking position state |
| CN107738612A (en) * | 2017-09-22 | 2018-02-27 | 西安电子科技大学 | The detection of automatic parking parking stall and identifying system based on panoramic vision accessory system |
| CN108022431A (en) * | 2017-12-04 | 2018-05-11 | 珠海横琴小可乐信息技术有限公司 | A kind of method and system that parking stall is detected by video capture image |
| CN108830188A (en) * | 2018-05-30 | 2018-11-16 | 西安理工大学 | Vehicle checking method based on deep learning |
| CN108766022A (en) * | 2018-06-11 | 2018-11-06 | 青岛串并联电子科技有限公司 | Parking position state identification method based on machine learning and system |
Non-Patent Citations (2)
| Title |
|---|
| 赵璐等: "基于先验形状信息和水平集方法的车辆检测", 《浙江大学学报(工学版)》 * |
| 金康等: "基于停车位可用概率的停车位发现算法", 《微型机与应用》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111862302A (en) * | 2019-04-12 | 2020-10-30 | 北京城市网邻信息技术有限公司 | Image processing method, image processing apparatus, object modeling method, object modeling apparatus, image processing apparatus, object modeling apparatus, and medium |
| CN110299028A (en) * | 2019-07-31 | 2019-10-01 | 深圳市捷顺科技实业股份有限公司 | Method, apparatus, equipment and the readable storage medium storing program for executing of line detection are got in a kind of parking |
| CN110299028B (en) * | 2019-07-31 | 2022-06-14 | 深圳市捷顺科技实业股份有限公司 | Parking line crossing detection method, device, equipment and readable storage medium |
| CN110458912A (en) * | 2019-08-08 | 2019-11-15 | 金瓜子科技发展（北京）有限公司 | A kind of vehicle icon treating method and apparatus |
| CN111428616A (en) * | 2020-03-20 | 2020-07-17 | 东软睿驰汽车技术（沈阳）有限公司 | Parking space detection method, device, equipment and storage medium |
| CN111428616B (en) * | 2020-03-20 | 2023-05-23 | 东软睿驰汽车技术（沈阳）有限公司 | Parking space detection method, device, equipment and storage medium |
| CN117079359A (en) * | 2023-09-04 | 2023-11-17 | 厦门中卡科技股份有限公司 | Intelligent parking payment method and system based on neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109508682A (en) | A kind of detection method on panorama parking stall | |
| CN108596277B (en) | Vehicle identity recognition method and device and storage medium | |
| CN108388888B (en) | Vehicle identification method and device and storage medium | |
| WO2022027873A1 (en) | Vehicle reidentification method and device based on multimodal information fusion | |
| WO2021142902A1 (en) | Danet-based unmanned aerial vehicle coastline floating garbage inspection system | |
| CN105488517B (en) | A kind of vehicle brand type identifier method based on deep learning | |
| CN110598654B (en) | Multi-granularity cross modal feature fusion pedestrian re-identification method and re-identification system | |
| CN102968637B (en) | Complicated background image and character division method | |
| Zhang et al. | An empirical study of multi-scale object detection in high resolution UAV images | |
| CN110569779B (en) | Pedestrian attribute identification method based on pedestrian local and overall attribute joint learning | |
| CN109711370A (en) | A kind of data anastomosing algorithm based on WIFI detection and face cluster | |
| CN108154110B (en) | Intensive people flow statistical method based on deep learning people head detection | |
| CN101848377B (en) | Device and method for intelligent linkage of multi-video recording device based on cloud computing and mass video searching | |
| CN101873414B (en) | Event video detection system based on hierarchical structure | |
| CN106331636A (en) | Intelligent video monitoring system and method of oil pipelines based on behavioral event triggering | |
| CN108846334A (en) | A kind of other automatic identifying method of the varieties of clouds and system | |
| CN109117745B (en) | Cloud face recognition and positioning method based on building information model | |
| CN101916383A (en) | Vehicle detecting, tracking and identifying system based on multi-camera | |
| CN110675421B (en) | Depth image collaborative segmentation method based on few labeling frames | |
| CN105787876A (en) | Panorama video automatic stitching method based on SURF feature tracking matching | |
| CN104463232A (en) | Density crowd counting method based on HOG characteristic and color histogram characteristic | |
| WO2020114116A1 (en) | Pedestrian detection method based on dense crowds, and storage medium and processor | |
| CN102693427A (en) | Method and device for forming detector for detecting images | |
| CN113723377A (en) | Traffic sign detection method based on LD-SSD network | |
| CN105335949A (en) | Video image rain removal method and system |
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: 20190322 |
|
| WD01 | Invention patent application deemed withdrawn after publication |