CN110210324A - A kind of road target quickly detects method for early warning and system - Google Patents
A kind of road target quickly detects method for early warning and system Download PDFInfo
- Publication number
- CN110210324A CN110210324A CN201910383340.XA CN201910383340A CN110210324A CN 110210324 A CN110210324 A CN 110210324A CN 201910383340 A CN201910383340 A CN 201910383340A CN 110210324 A CN110210324 A CN 110210324A
- Authority
- CN
- China
- Prior art keywords
- target
- area
- grades
- interest
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
- G06V20/50—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
- G06V20/588—Recognition of the road, e.g. of lane markings; Recognition of the vehicle driving pattern in relation to the road
-
- 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)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention discloses a kind of road targets quickly to detect method for early warning and system, belong to computer vision field, method includes: to acquire road scene image using image capture device, according to target range from closely extracting N grades of area-of-interests from road scene image to remote sequence;According to target range from closely to remote sequence, target detection successively is carried out to N grades of area-of-interests, after closer certain grade of area-of-interest of target range has detected target, carries out early warning, and no longer detect to subsequent level area-of-interest;Wherein, N >=2, the target range are the distance between target and image capture device.The detection warning algorithm of the targets such as the relatively existing pedestrian of the present invention, vehicle more rapidly accurately can carry out detection early warning to the target appeared in traffic route, have preferable practical value.
Description
Technical field
The invention belongs to computer vision field, more particularly, to a kind of road target quickly detect method for early warning and
System.
Background technique
Pedestrian, vehicle testing techniques are widely used in DAS (Driver Assistant System).In urban transportation scene, how to avoid handing over
Interpreter's event is always hot issue.Current auxiliary, which drives, mainly utilizes pedestrian, vehicle testing techniques, to traveling motor vehicle
The image in front is tested and analyzed, and is carried out initiative alarming for pedestrian, vehicle, driver can be assisted to evade in advance,
The generation to try to forestall traffic accidents.Current pedestrian, vehicle testing techniques mainly utilize the related algorithm based on deep learning, to defeated
Enter the targets such as pedestrian, the vehicle in image and carries out detection identification.One of DAS (Driver Assistant System) is important to be required to be exactly real-time.Vehicle
Quickly, if cracking cannot detect pedestrian and alarm, that will be unable to timely travel speed for people
Carry out hedging.In addition system needs accuracy with higher, reduces the occurrence of failing to report police, setting off false alarms to the greatest extent.
There are many outstanding deep learning pedestrians, vehicle detecting algorithm at present, compared with traditional algorithm, in precision and robust
Property aspect be higher by very much, also can quickly be run on high-performance server.If but these networks are grafted directly to embedded
The platforms such as equipment, mobile device, because of the difference of hardware platform calculated performance, the speed of service will have a greatly reduced quality.It is not able to satisfy
The requirement of rapidity.Therefore, the network run on embedded device and mobile device must be the neural network of lightweight.But
It is simple by pedestrian, vehicle detection network simplification, and is directly used in the detection and early warning of pedestrian, vehicle, although can obtains
Apparent acceleration effect, but also can inevitably cause detection accuracy decline more, it is unable to satisfy bicycle-mounted portable pedestrian inspection
The actual use demand of detection early warning system.
It can be seen that the prior art, which exists, fast and accurately to carry out detection early warning to the target being located in traffic route
The technical issues of.
Summary of the invention
Aiming at the above defects or improvement requirements of the prior art, the present invention provides a kind of road targets quickly to detect early warning
Thus it is pre- in the presence of that fast and accurately can not detect to the target being located in traffic route to solve the prior art for method and system
Alert technical problem.
To achieve the above object, according to one aspect of the present invention, it provides a kind of road target and quickly detects the pre- police
Method, comprising the following steps:
(1) road scene image is acquired using image capture device, according to target range from closely to remote sequence from road
N grades of area-of-interests are extracted in scene image;
(2) target detection successively is carried out to N grades of area-of-interests, in target from closely to remote sequence according to target range
After certain grade of area-of-interest being closer has detected target, carry out early warning, and no longer to subsequent level area-of-interest into
Row detection;
Wherein, N >=2, the target range are the distance between target and image capture device.
Further, step (1) the following steps are included:
Road scene image is acquired, setting object height accounting threshold value is Thr, nearest target range is L0;
It is L according to the relationship f and nearest target range of the pixels tall of target and target range0, obtain the picture of target
Plain height H1, by H1As the length of first order area-of-interest, the length and road scene image of first order area-of-interest are utilized
The width of length-width ratio calculating first order area-of-interest;
For N grades of area-of-interests, by the long H of object height accounting threshold value Thr and N-1 grades of area-of-interestsN-1's
Product HNAs the length of N grades of area-of-interests, the length of N grades of area-of-interests and the length-width ratio meter of road scene image are utilized
Calculate the width of N grades of area-of-interests;
Is obtained according to the relationship f of the length of N grades of area-of-interests and wide and target pixels tall and target range
The maximum distance L of N-1 grades of area-of-interest detection targetsN-1, while LN-1It is also the nearest of N grades of area-of-interest detection targets
Distance;
Utilize the long H of object height accounting threshold value Thr and N grades of area-of-interestsNProduct and target pixel it is high
The relationship f of degree and target range obtains the maximum distance L of N grades of area-of-interest detection targetsN;N grades of area-of-interests
Object range detection range is (LN-1, LN)。
Further, step (1) further include:
Road scene image set of the different target under is acquired, the picture of target is obtained from road scene image set
Plain height and the discrete data with target range, the pixels tall of target and the relationship of target range are obtained using discrete data
f。
Further, the value range of object height accounting threshold value Thr is (0,1).
Further, the center of road scene image and N grades of area-of-interests is lane line joint.
Further, in step (2) target detection specific implementation are as follows:
Target detection, the training package of the target detection model are carried out to N grades of area-of-interests using target detection model
It includes:
N grades of area-of-interest samples are extracted for sample road scene image, and are marked in each area-of-interest sample
Object height accounting is greater than the target of object height accounting threshold value, obtains training sample set;
Building includes the target detection convolutional neural networks of feature extraction layer, concentrates N to training sample in feature extraction layer
Grade area-of-interest sample carries out feature extraction, obtains characteristic pattern;
Multiple pre-selection frames are set in characteristic pattern, calculate training sample concentrate be labeled in area-of-interest sample target with
The Duplication of each pre-selection frame is greater than the pre-selection frame and training sample set training objective detection convolution mind of preset value using Duplication
Through network, training result is obtained, calculates the penalty values of training result, backpropagation is carried out using penalty values and then more fresh target is examined
The parameter for surveying convolutional neural networks, obtains target detection model.
Further, the height accounting for preselecting frame is greater than object height accounting threshold value.
It is another aspect of this invention to provide that providing a kind of road target quickly detects early warning system, comprising:
Area-of-interest module is extracted, for acquiring road scene image using image capture device, according to target range
From N grades of area-of-interests are closely extracted from road scene image to remote sequence;
Module of target detection, for according to target range from closely to remote sequence, using target detection model to successively to N
Grade area-of-interest carries out target detection, after closer certain grade of area-of-interest of target range has detected target, carries out pre-
It is alert, and no longer subsequent level area-of-interest is detected;
Wherein, N >=2, the target range are the distance between target and image capture device.
Further, extracting area-of-interest module includes following submodule:
Initialization submodule, for acquiring road scene image, setting object height accounting threshold value is Thr, nearest target
Distance is L0;
First order area-of-interest length and width computational submodule, for according to the pixels tall of target and the relationship of target range
F and nearest target range are L0, obtain the pixels tall H of target1, by H1As the length of first order area-of-interest, is utilized
The length of level-one area-of-interest and the length-width ratio of road scene image calculate the width of first order area-of-interest;
N grades of area-of-interest length and width computational submodules are used for for N grades of area-of-interests, by object height accounting
The long H of threshold value Thr and N-1 grades of area-of-interestsN-1Product HNAs the length of N grades of area-of-interests, felt using N grades
The length in interest region and the length-width ratio of road scene image calculate the width of N grades of area-of-interests;
The object range detection range computation submodule of N grades of area-of-interests, for according to N grades of area-of-interests
The relationship f of long and wide and target pixels tall and target range obtains the farthest of N-1 grades of area-of-interest detection targets
Distance LN-1, while LN-1It is also the minimum distance of N grades of area-of-interest detection targets;Utilize object height accounting threshold value Thr
With the long H of N grades of area-of-interestsNProduct and target pixels tall and target range relationship f, obtain N grades sense
The maximum distance L of interest region detection targetN;The object range detection range of N grades of area-of-interests is (LN-1, LN)。
Further, the training of target detection model includes:
Training sample set acquisition submodule, for extracting N grades of area-of-interest samples for sample road scene image, and
Object height accounting in each area-of-interest sample is marked to obtain training sample greater than the target of object height accounting threshold value
Collection;
Characteristic pattern acquisition submodule, for constructing the target detection convolutional neural networks comprising feature extraction layer, in feature
Extract layer concentrates N grades of area-of-interest samples to carry out feature extraction training sample, obtains characteristic pattern;
Training submodule calculates training sample and concentrates area-of-interest sample for multiple pre-selection frames to be arranged in characteristic pattern
It is labeled the Duplication of target and each pre-selection frame in this, is greater than the pre-selection frame of preset value using Duplication and training sample is assembled for training
Practice target detection convolutional neural networks, obtains training result, calculate the penalty values of training result, reversely passed using penalty values
The parameter for broadcasting and then updating target detection convolutional neural networks, obtains target detection model.
In general, through the invention it is contemplated above technical scheme is compared with the prior art, can obtain down and show
Beneficial effect:
(1) present invention employs hierarchical objectives to detect prediction policy, so that can use most fast speed for target nearby
Degree detects, shortens pre-warning time as far as possible for this most dangerous situation, and can detecte the target to distant place very little.Cause
This, the detection warning algorithm of the targets such as the relatively existing pedestrian of the present invention, vehicle, inspection that can to greatest extent to nearby target
It surveys early warning to be accelerated, while being also able to detect the target of early warning distant place, it can be more rapidly accurately to appearing in traffic route
On target carry out detection early warning, have preferable practical value.
(2) existing algorithm of target detection, for Small object Detection accuracy generally than big goal discrepancy.The present invention is logical
It crosses and simplifies the design of target detection network structure, promote network operations speed.Using the strategy of this hierarchical detection, can allow each
The detection of grade only needs to detect biggish target.The speed of network can be promoted in this way.
(3) present invention is by utilizing area-of-interests at different levels, and label target height accounting is greater than object height accounting threshold
The area-of-interest of value is used to train, and improves target detection network to the detection accuracy of target in real screen, avoids inspection
It measures compared with Small object.
Detailed description of the invention
Fig. 1 is the flow diagram that a kind of road target provided in an embodiment of the present invention quickly detects method for early warning;
Fig. 2 is preferred embodiment flow chart provided in an embodiment of the present invention;
Fig. 3 is a kind of three-level area-of-interest exacting method provided in an embodiment of the present invention;
Fig. 4 is the flow diagram of target detection model training data set mark provided in an embodiment of the present invention;
Fig. 5 is characteristic pattern provided in an embodiment of the present invention and pre-selection frame design diagram;
Fig. 6 is target detection convolutional neural networks structural schematic diagram provided in an embodiment of the present invention.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.As long as in addition, technical characteristic involved in the various embodiments of the present invention described below
Not constituting a conflict with each other can be combined with each other.
As shown in Figure 1, a kind of road target quickly detects method for early warning, comprising the following steps:
(1) road scene image is acquired using image capture device, according to target range from closely to remote sequence from road
N grades of area-of-interests are extracted in scene image;
(2) target detection successively is carried out to N grades of area-of-interests, in target from closely to remote sequence according to target range
After certain grade of area-of-interest being closer has detected target, carry out early warning, and no longer to subsequent level area-of-interest into
Row detection;
Wherein, N >=2, the target range are the distance between target and image capture device.
Further, step (1) the following steps are included:
Road scene image is acquired, setting object height accounting threshold value is Thr, nearest target range is L0;Acquisition is different
Road scene image set under target range obtains the pixels tall and and target of target from road scene image set
The discrete data of distance obtains the pixels tall of target and the relationship f of target range using discrete data;
It is L according to the relationship f and nearest target range of the pixels tall of target and target range0, obtain the picture of target
Plain height H1, by H1As the length of first order area-of-interest, the length and road scene image of first order area-of-interest are utilized
The width of length-width ratio calculating first order area-of-interest;
For N grades of area-of-interests, by the long H of object height accounting threshold value Thr and N-1 grades of area-of-interestsN-1's
Product HNAs the length of N grades of area-of-interests, the length of N grades of area-of-interests and the length-width ratio meter of road scene image are utilized
Calculate the width of N grades of area-of-interests;
Is obtained according to the relationship f of the length of N grades of area-of-interests and wide and target pixels tall and target range
The maximum distance L of N-1 grades of area-of-interest detection targetsN-1, while LN-1It is also the nearest of N grades of area-of-interest detection targets
Distance;
Utilize the long H of object height accounting threshold value Thr and N grades of area-of-interestsNProduct and target pixel it is high
The relationship f of degree and target range obtains the maximum distance L of N grades of area-of-interest detection targetsN;N grades of area-of-interests
Object range detection range is (LN-1, LN)。
Further, the value range of object height accounting threshold value Thr is (O, 1), it is preferable that object height accounting threshold value
The value of Thr is 0.5.
Further, the center of road scene image and N grades of area-of-interests is lane line joint.
Further, in step (2) target detection specific implementation are as follows:
Target detection, the training package of the target detection model are carried out to N grades of area-of-interests using target detection model
It includes:
N grades of area-of-interest samples are extracted for sample road scene image, and are marked in each area-of-interest sample
Object height accounting is greater than the target of object height accounting threshold value, obtains training sample set;
Building includes the target detection convolutional neural networks of feature extraction layer, concentrates N to training sample in feature extraction layer
Grade area-of-interest sample carries out feature extraction, obtains characteristic pattern;
Multiple pre-selection frames are set in characteristic pattern, calculate training sample concentrate be labeled in area-of-interest sample target with
The Duplication of each pre-selection frame is greater than the pre-selection frame and training sample set training objective detection convolution mind of preset value using Duplication
Through network, training result is obtained, calculates the penalty values of training result, backpropagation is carried out using penalty values and then more fresh target is examined
The parameter for surveying convolutional neural networks, obtains target detection model.The height accounting for preselecting frame is greater than object height accounting threshold value, in advance
If value is 0.5.
As shown in Fig. 2, road scene image is acquired for the case where N is 3, according to target range from closely to remote sequence
Three-level area-of-interest is extracted from road scene image;It is successively emerging to three-level sense according to target range from closely to remote sequence
Interesting region carries out target detection, carries out first order detection to first order area-of-interest first, if detecting target, carries out the
Otherwise level-one alarm carries out second level detection to second level area-of-interest and carries out second level alarm if detecting target,
Otherwise third level detection is carried out to third level area-of-interest and carries out third level alarm if detecting target, otherwise stop inspection
It surveys.
Region in wheelpath is divided into three-level area-of-interest, be respectively used to detection it is close, in, the target of distant place.Energy
It is enough that the detection early warning of nearby target is accelerated to greatest extent, while being also able to detect the target of early warning distant place.
As shown in figure 3, the case where N is 3, is arranged three-level area-of-interest, is specifically included following described in step (1)
Step:
(21) road scene image of acquisition is analyzed, extracts lane line track in the picture and joint;
Set object height accounting threshold value Thr;
(22) the relationship f of pixels tall and target range of the analysis target in acquisition image;
(23) setting detects target range a recently, obtains the pixels tall H of target at this time according to relationship f1, by H1As
The length of level-one area-of-interest calculates first order sense using the length of first order area-of-interest and the length-width ratio of road scene image
The width in interest region.
(24) according to threshold value Thr, the minimum target pixels tall H of first order area-of-interest detection is obtained2=H1*Thr。
By H2As the length of second level area-of-interest, the length of second level area-of-interest and the length-width ratio meter of road scene image are utilized
Calculate the width of second level area-of-interest.
(25) by H2, the maximum distance b of first order area-of-interest detection target is obtained according to relationship f, while being also the
The minimum distance of second level area-of-interest detection target.
(26) according to threshold value Thr, the minimum target pixels tall H of second level area-of-interest detection is obtained3=H2*Thr。
By H3As the length of third level area-of-interest, the length of third level area-of-interest and the length-width ratio meter of road scene image are utilized
Calculate the width of third level area-of-interest.
(27) by H3, the maximum distance c of second level area-of-interest detection target is obtained according to relationship f, while being also the
The minimum distance of three-level area-of-interest detection target.
(28) according to threshold value Thr, the minimum target pixels tall H of third level area-of-interest detection is obtained4=H3*Thr。
(29) by H4, the maximum distance d of third level area-of-interest detection target is obtained according to relationship f.
(210) it is the center of each area-of-interest with joint, successively intercepts a length of H1, H2, H3Region, as first,
Two, the area-of-interest of three-level.
Distance a, b, c, d described in the step (23) (25) (27) (29) constitute first, second and third grade of region of interest
The target detection range section (a, b) in domain, (b, c), (c, d).
As shown in figure 4, for the case where N is 3, the training method of target detection model are as follows:
Three-level area-of-interest is extracted to sample road scene image, obtains input picture, and in each input picture
In, the complete object that height accounting is greater than THR is marked, training dataset T is obtained;Labeling method is the square that record frames target
The upper left corner of shape frame and bottom right angular coordinate are recorded in an xml formatted file.
Target detection convolutional neural networks are constructed, to training dataset on the characteristic layer of target detection convolutional neural networks
T carries out feature extraction;
As shown in figure 5, the design pre-selection frame on the characteristic layer of target detection convolutional neural networks, preselects the scale setting of frame
For a little higher than Thr, the target object position that height accounting in input picture is greater than Thr is tentatively obtained;Preselect frame aspect ratio
arIt is 2,3,4;Preselecting frame height isWidth is
The pre-selection frame substep training target detection convolutional Neural net being arranged using training dataset T and on characteristic layer
Network obtains target detection model.
As shown in fig. 6, when detecting, input picture input target's feature-extraction network being obtained characteristic pattern, is detected
Testing result is obtained, carries out whether subsequent processing carries out early warning using testing result.
Embodiment 1
A kind of road target quickly detects method for early warning, when N is 3, and target is pedestrian, comprising the following steps:
Step 1: acquisition road scene image analyzes the road scene image of acquisition, extracts lane line and scheming
Track and joint as in;Setting picture centre is lane line joint.Analyze pixels tall of the pedestrian in acquisition image
With the relationship f of pedestrian's distance;Detection pedestrian level accounting threshold value Thr is set as 0.5.
Detection pedestrian distance is 5 meters recently for setting.According to relationship f, the pixels tall for obtaining pedestrian at 5 meters is 480, if the
Level-one area-of-interest length and width are 480.(length-width ratio of road scene image is 1 at this time)
It is 480 by first order area-of-interest sprite length and width, and the minimum pedestrian level accounting threshold value that sets is 0.5,
Obtain the first order minimum row people pixels tall 240 to be detected.At this point, pedestrian's distance is 10 meters, if it is second level detection row
The minimum distance of people.And second level area-of-interest length and width are set as 240.
Be 240 by second level area-of-interest sprite length and width, and the minimum detection pedestrian level accounting threshold value that sets as
0.5, obtain the minimum row people pixels tall 120 to be detected the second level.At this point, pedestrian's distance is 20 meters, if it is third level inspection
Survey the minimum distance of pedestrian.And second level area-of-interest length and width are set as 120.
Third level area-of-interest sprite length and width are 120, and the minimum detection pedestrian level accounting threshold value that sets as
0.5, obtain the third level minimum row people pixels tall 60 to be detected.At this point, pedestrian's distance is 40 meters, target is detected for the third level
Maximum distance.
Step 2: at different levels area-of-interests of the interception containing pedestrian.Three-level area-of-interest is above-mentioned three-level region of interest
The region that domain extracting method obtains.According to the ratio of the pixels tall of the pedestrian in area-of-interest and sprite resolution height
Whether it is greater than Thr to decide whether to mark the pedestrian.Mark the pedestrian that accounting is greater than Thr.Mask method is that record frames pedestrian's
The upper left corner of rectangle frame and bottom right angular coordinate.
Design pre-selection frame, tentatively obtains pedestrian's mesh on the characteristic layer L of the single scale pedestrian detection convolutional neural networks
Mark position;
Partial Feature layer after selection network, characteristic layer resolution ratio are 3 × 3, and detection pedestrian's scale is 0.7;By input picture
It is divided into 3 × 3 grid.It is concentric with the midpoint of net region for preselecting the design principle of frame;Preselect frame aspect ratio arFor 2,3,
4;Preselecting frame height isWidth is
Using training dataset T and the pre-selection frame training single scale pedestrian detection convolutional neural networks, pedestrian's inspection is obtained
Survey model.
Step 3: pedestrian detection successively being carried out to three-level area-of-interest using pedestrian detection model.
The present invention is that a kind of road target suitable for DAS (Driver Assistant System) quickly detects method for early warning, uses divide first
Grade target detection prediction policy, is divided into three-level area-of-interest for the region in wheelpath.Be respectively used to detection it is close, in, it is remote
The target at place.Can the detection early warning to greatest extent to nearby target accelerate, while being also able to detect early warning distant place
Target.And according to the strategy of this hierarchical detection, devise fast target detection network.Use the pre-selection for meeting target signature
Frame and single scale characteristic pattern carry out target detection work, have obtained quick target detection network.The method achieve to mesh
Mark fast and accurately detects early warning.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should all include
Within protection scope of the present invention.
Claims (10)
1. a kind of road target quickly detects method for early warning, which comprises the following steps:
(1) road scene image is acquired using image capture device, according to target range from closely to remote sequence from road scene
N grades of area-of-interests are extracted in image;
(2) target detection successively is carried out to N grades of area-of-interests, in target range from closely to remote sequence according to target range
After closer certain grade of area-of-interest has detected target, early warning is carried out, and no longer examine to subsequent level area-of-interest
It surveys;
Wherein, N >=2, the target range are the distance between target and image capture device.
2. a kind of road target as described in claim 1 quickly detects method for early warning, which is characterized in that step (1) packet
Include following steps:
Road scene image is acquired, setting object height accounting threshold value is Thr, nearest target range is L0;
It is L according to the relationship f and nearest target range of the pixels tall of target and target range0, obtain the pixels tall of target
H1, by H1As the length of first order area-of-interest, the length of first order area-of-interest and the length-width ratio of road scene image are utilized
Calculate the width of first order area-of-interest;
For N grades of area-of-interests, by the long H of object height accounting threshold value Thr and N-1 grades of area-of-interestsN-1Product
HNAs the length of N grades of area-of-interests, the is calculated using the length of N grades of area-of-interests and the length-width ratio of road scene image
The width of N grades of area-of-interests;
N-1 grades are obtained according to the relationship f of the length of N grades of area-of-interests and wide and target pixels tall and target range
The maximum distance L of area-of-interest detection targetN-1, while LN-1It is also the minimum distance of N grades of area-of-interest detection targets;
Utilize the long H of object height accounting threshold value Thr and N grades of area-of-interestsNProduct and target pixels tall and mesh
The relationship f of subject distance obtains the maximum distance L of N grades of area-of-interest detection targetsN;The target of N grades of area-of-interests away from
It is (L from detection rangeN-1, LN)。
3. a kind of road target as claimed in claim 2 quickly detects method for early warning, which is characterized in that the step (1) is also
Include:
Road scene image set of the different target under is acquired, the pixel that target is obtained from road scene image set is high
Degree and the discrete data with target range, the pixels tall of target and the relationship f of target range are obtained using discrete data.
4. a kind of road target as claimed in claim 2 quickly detects method for early warning, which is characterized in that the object height accounts for
Value range than threshold value Thr is (0,1).
5. a kind of road target as described in claim 1-4 is any quickly detects method for early warning, which is characterized in that the road
The center of scene image and N grades of area-of-interests is lane line joint.
6. a kind of road target as claimed in claim 2 or 4 quickly detects method for early warning, which is characterized in that the step (2)
The specific implementation of middle target detection are as follows:
Target detection is carried out to N grades of area-of-interests using target detection model, the training of the target detection model includes:
N grades of area-of-interest samples are extracted for sample road scene image, and mark target in each area-of-interest sample
Height accounting is greater than the target of object height accounting threshold value, obtains training sample set;
Building includes the target detection convolutional neural networks of feature extraction layer, concentrates N grades of senses to training sample in feature extraction layer
Interest area sample carries out feature extraction, obtains characteristic pattern;
Multiple pre-selection frames are set in characteristic pattern, calculates training sample and concentrates and be labeled target and each in area-of-interest sample
The Duplication for preselecting frame is greater than the pre-selection frame of preset value using Duplication and training sample set training objective detects convolutional Neural net
Network obtains training result, calculates the penalty values of training result, carries out backpropagation using penalty values and then update target detection to roll up
The parameter of product neural network, obtains target detection model.
7. a kind of road target as claimed in claim 6 quickly detects method for early warning, which is characterized in that the height of the pre-selection frame
It spends accounting and is greater than object height accounting threshold value.
8. a kind of road target quickly detects early warning system characterized by comprising
Area-of-interest module is extracted, for acquiring road scene image using image capture device, according to target range from close
N grades of area-of-interests are extracted from road scene image to remote sequence;
Module of target detection, for according to target range from closely to remote sequence, using target detection model to successively feeling to N grades
Interest region carries out target detection, after closer certain grade of area-of-interest of target range has detected target, carries out early warning, and
No longer subsequent level area-of-interest is detected;
Wherein, N >=2, the target range are the distance between target and image capture device.
9. a kind of road target as claimed in claim 8 quickly detects early warning system, which is characterized in that the extraction is interested
Regions module includes following submodule:
Initialization submodule, for acquiring road scene image, setting object height accounting threshold value is Thr, nearest target range
For L0;
First order area-of-interest length and width computational submodule, for according to the relationship f of the pixels tall of target and target range with
And target range is L recently0, obtain the pixels tall H of target1, by H1As the length of first order area-of-interest, first is utilized
The length of grade area-of-interest and the length-width ratio of road scene image calculate the width of first order area-of-interest;
N grades of area-of-interest length and width computational submodules are used for for N grades of area-of-interests, by object height accounting threshold value
The long H of Thr and N-1 grades of area-of-interestsN-1Product HNIt is interested using N grades as the length of N grades of area-of-interests
The length in region and the length-width ratio of road scene image calculate the width of N grades of area-of-interests;
The object range detection range computation submodule of N grades of area-of-interests, for according to the length of N grades of area-of-interests and
The relationship f of wide and target pixels tall and target range obtains the maximum distance of N-1 grades of area-of-interest detection targets
LN-1, while LN-1It is also the minimum distance of N grades of area-of-interest detection targets;Utilize object height accounting threshold value Thr and N
The long H of grade area-of-interestNProduct and target pixels tall and target range relationship f, obtain N grades of region of interest
The maximum distance L of domain detection targetN;The object range detection range of N grades of area-of-interests is (LN-1, LN)。
10. a kind of road target quickly detects early warning system as claimed in claim 8 or 9, which is characterized in that the target inspection
Survey model training include:
Training sample set acquisition submodule for extracting N grades of area-of-interest samples for sample road scene image, and marks
Object height accounting is greater than the target of object height accounting threshold value in each area-of-interest sample, obtains training sample set;
Characteristic pattern acquisition submodule, for constructing the target detection convolutional neural networks comprising feature extraction layer, in feature extraction
Layer concentrates N grades of area-of-interest samples to carry out feature extraction training sample, obtains characteristic pattern;
Training submodule calculates training sample and concentrates in area-of-interest sample for multiple pre-selection frames to be arranged in characteristic pattern
The Duplication of labeled target and each pre-selection frame is greater than the pre-selection frame and training sample set training mesh of preset value using Duplication
Mark detection convolutional neural networks, obtain training result, calculate the penalty values of training result, using penalty values carry out backpropagation into
And the parameter of target detection convolutional neural networks is updated, obtain target detection model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910383340.XA CN110210324B (en) | 2019-05-08 | 2019-05-08 | Road target rapid detection early warning method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910383340.XA CN110210324B (en) | 2019-05-08 | 2019-05-08 | Road target rapid detection early warning method and system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110210324A true CN110210324A (en) | 2019-09-06 |
CN110210324B CN110210324B (en) | 2021-01-19 |
Family
ID=67785676
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910383340.XA Active CN110210324B (en) | 2019-05-08 | 2019-05-08 | Road target rapid detection early warning method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110210324B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111179533A (en) * | 2020-01-11 | 2020-05-19 | 刘惠敏 | Target identification system based on shape analysis |
WO2022012425A1 (en) * | 2020-07-16 | 2022-01-20 | 长沙智能驾驶研究院有限公司 | Target detection method and apparatus, and electronic device |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107066929A (en) * | 2017-01-06 | 2017-08-18 | 重庆大学 | The manifold freeway tunnel Parking hierarchical identification method of one kind fusion |
CN107911697A (en) * | 2017-10-30 | 2018-04-13 | 北京航空航天大学 | Unmanned plane image motion object detection method based on area-of-interest layering |
CN108399630A (en) * | 2018-01-22 | 2018-08-14 | 北京理工雷科电子信息技术有限公司 | Target fast ranging method in area-of-interest under a kind of complex scene |
-
2019
- 2019-05-08 CN CN201910383340.XA patent/CN110210324B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107066929A (en) * | 2017-01-06 | 2017-08-18 | 重庆大学 | The manifold freeway tunnel Parking hierarchical identification method of one kind fusion |
CN107911697A (en) * | 2017-10-30 | 2018-04-13 | 北京航空航天大学 | Unmanned plane image motion object detection method based on area-of-interest layering |
CN108399630A (en) * | 2018-01-22 | 2018-08-14 | 北京理工雷科电子信息技术有限公司 | Target fast ranging method in area-of-interest under a kind of complex scene |
Non-Patent Citations (6)
Title |
---|
HUIBAO LIN等: "Hierarchical region-of-interest detection", 《OPTICAL ENGINEERING 》 * |
HUIBAO LIN等: "KNOWLEDGE-BASED HIERARICAL REGION-OF-INEREST DETECTION", 《2002 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS,SPEECH,AND SIGNAL PROCESSING》 * |
KAI CHEN等: "Once for All: a Two-flow Convolutional Neural Network for Visual Tra cking", 《ARXIV:1604.07507V1 [CS.CV]》 * |
TONG LI等: "SDBD: A Hierarchical Region-of-Interest Detection Approach in Large-Scale Remote Sensing Image", 《IEEE GEOSCIENCE AND REMOTE SENSING LETTERS》 * |
向根: "基于DM8127的多目标远距离检测定位系统", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
张治国: "前方道路行人检测和距离估计研究", 《中国博士学位论文全文数据库 工程科技Ⅱ辑》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111179533A (en) * | 2020-01-11 | 2020-05-19 | 刘惠敏 | Target identification system based on shape analysis |
CN111179533B (en) * | 2020-01-11 | 2021-10-29 | 合肥源康信息科技有限公司 | Target identification system based on shape analysis |
WO2022012425A1 (en) * | 2020-07-16 | 2022-01-20 | 长沙智能驾驶研究院有限公司 | Target detection method and apparatus, and electronic device |
Also Published As
Publication number | Publication date |
---|---|
CN110210324B (en) | 2021-01-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110175576B (en) | Driving vehicle visual detection method combining laser point cloud data | |
CN103824452B (en) | A kind of peccancy parking detector based on panoramic vision of lightweight | |
CN102724482B (en) | Based on the intelligent vision sensing network moving target relay tracking system of GPS and GIS | |
CN105702048B (en) | Highway front truck illegal road occupation identifying system based on automobile data recorder and method | |
Yu et al. | Automated detection of urban road manhole covers using mobile laser scanning data | |
CN104616502B (en) | Car license recognition and alignment system based on combination type bus or train route video network | |
CN106845364B (en) | Rapid automatic target detection method | |
CN103366154B (en) | Reconfigurable clear path detection system | |
CN108389421B (en) | Parking lot accurate induction system and method based on image re-identification | |
Feng et al. | Mixed road user trajectory extraction from moving aerial videos based on convolution neural network detection | |
CN110532889A (en) | Track foreign matter detecting method based on rotor unmanned aircraft and YOLOv3 | |
CN104063882B (en) | Vehicle video speed measuring method based on binocular camera | |
CN111626170B (en) | Image recognition method for railway side slope falling stone intrusion detection | |
CN105608431A (en) | Vehicle number and traffic flow speed based highway congestion detection method | |
CN112749584B (en) | Vehicle positioning method based on image detection and vehicle-mounted terminal | |
CN112668375B (en) | Tourist distribution analysis system and method in scenic spot | |
CN106951829A (en) | A kind of notable method for checking object of video based on minimum spanning tree | |
CN114067384A (en) | Highway epidemic situation prevention and control personnel generate heat and detect and vehicle trajectory tracking roadside equipment | |
CN106845325A (en) | A kind of information detecting method and device | |
CN106845359A (en) | Tunnel portal driving prompt apparatus and method based on infrared emission | |
CN110210324A (en) | A kind of road target quickly detects method for early warning and system | |
CN114926984B (en) | Real-time traffic conflict collection and road safety evaluation method | |
Guerrieri et al. | Flexible and stone pavements distress detection and measurement by deep learning and low-cost detection devices | |
CN116311166A (en) | Traffic obstacle recognition method and device and electronic equipment | |
KR102562757B1 (en) | Prediction and recognition method of road marking information and road maintenance method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |