CN107194343B - Traffic lights detection method based on the relevant convolution in position Yu Fire model - Google Patents
Traffic lights detection method based on the relevant convolution in position Yu Fire model Download PDFInfo
- Publication number
- CN107194343B CN107194343B CN201710342500.7A CN201710342500A CN107194343B CN 107194343 B CN107194343 B CN 107194343B CN 201710342500 A CN201710342500 A CN 201710342500A CN 107194343 B CN107194343 B CN 107194343B
- Authority
- CN
- China
- Prior art keywords
- traffic lights
- network
- firenet
- relevant
- convolutional
- 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.)
- Active
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
- G06V20/584—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of vehicle lights or traffic lights
-
- 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
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)
- Traffic Control Systems (AREA)
- Image Analysis (AREA)
Abstract
The traffic lights detection method based on the relevant convolution in position Yu Fire model that the present invention provides a kind of.This method uses multiple Fire models of connecting as trunk convolutional network FireNet, the relevant convolutional layer of point of addition behind FireNet the last layer characteristic pattern, and is trained to network, and trained network is utilized to carry out traffic lights detection.Since the parameter of Fire model is few, running space demand is low, it is more conducive to running on embedded device；FireNet has more powerful feature representation ability as convolutional network, can sufficiently excavate the information of traffic lights in different scenes, obtain more accurate feature representation；The relevant convolutional layer of point of addition, can be improved the accuracy of positioning behind FireNet the last layer characteristic pattern, and lesser and plurality of classes traffic lights are effectively detected.
Description
Technical field
The invention belongs to computer visions, object detection technical field, and in particular to one kind based on the relevant convolution in position with
The traffic lights detection method of Fire model.
Background technique
Can be accurate in vehicle travel process, the position and state for quickly determining the traffic lights in front are that advanced auxiliary is driven
Sail the technology important with unmanned middle one kind.However illumination, day due to the complexity of true traffic scene, such as acute variation
The problems such as differences in resolution of gas, the influence for the factors such as blocking and imaging device, so that the inspection of the traffic lights in real scene
It surveys relatively difficult.With the prevalence of deep learning, in target detection, image recognition, multiple computer visions such as image segmentation
Field all obtains thrilling effect.By the method for deep learning apply traffic lights detection on can preferably handle it is above-mentioned
There are the problem of.Currently, the detection method for traffic lights has three categories:
The first is the method based on image procossing.This method is operated by Threshold segmentation, morphological transformation etc. to image
It is handled, obtains interested object area in picture, then by specific priori knowledge, such as the connectivity of region, length and width
Than, shape, relative position etc., these regions are handled, are screened layer by layer, what is finally obtained is exactly the region where traffic lights, is passed through
Setpoint color threshold value or the color that traffic lights are judged using special color space.R.de Charette et al. is in document
“R.de Charette and F.Nashashibi,Real time visual traffic lights recognition
based on Spot Light Detection and adaptive traffic lights templates,IEEE
A kind of method of traffic lights detection is proposed in Intelligent Vehicles Symposium, pp.358-363,2009 ",
This method is converted by morphological image, and Threshold segmentation obtains object candidate region, then screens candidate regions using appearance ratio
Domain obtains the state of traffic lights finally by template matching.Its shortcoming is can not to adapt to changeable scene, and threshold value is excessively
Sensitivity, inadequate robust.
Second is the method based on Orientation on map, and the traffic lights information for being measured by accurate GPS and manually being marked obtains
To more accurately traffic lights priori knowledge, when close to traffic lights, the candidate region of object is obtained using geometric transformation,
Then classify on candidate region.V.John et al. document " V.John, K.Yoneda, Z.Liu, and S.Mita,
Saliency Map Generation by the Convolutional Neural Network for Real-Time
Traffic Light Detection Using Template Matching.IEEE Trans.Computational
It proposes to generate offline conspicuousness map by GPS in Imaging, vol.1, no.3, pp.159-173, Sept.2015 ", benefit
With in-vehicle camera parameter when close to traffic lights, the region that traffic lights occur is obtained by trigonometry, then uses convolution
Neural network and template matching detect traffic lights classification.Its deficiency is excessively to rely on sensor device, under same effect,
Cost is excessively high.
The third is the method based on machine learning, as Shi et al. document " Z.Shi, Z.Zhou, and C.Zhang,
Real-Time Traffic Light Detection With Adaptive Background Suppression
Filter.IEEE Trans.Intelligent Transportation Systems,vol.17,no.3,pp.690-700,
It proposes in Oct.2015 " by learning to the sample in training set, carry out filtering background that can be adaptive, to obtain
Then interested target area again classifies to obtained result.It can effectively avoid manually setting based on machine learning
A variety of threshold values are set, by the model for learning to obtain, there is stronger generalization ability.One point as machine learning of deep learning
Branch has more powerful learning ability, is increasingly becoming object detection field at present compared to traditional machine learning model
Mainstream algorithm.
Summary of the invention
A plurality of types of traffic lights are detected while in order to improve accuracy in detection, the invention proposes one kind to be based on position
The traffic lights detection method of relevant convolution and Fire model.The main thought of this method is: using multiple Fire models of connecting
As trunk convolutional network FireNet, and the relevant convolutional layer of point of addition behind FireNet the last layer characteristic pattern, In
The accuracy that positioning is improved while reducing network parameter can be effectively detected lesser and plurality of classes using the network
Traffic lights, while the complex scenes such as different illumination and weather can be successfully managed, obtain preferable detection effect.
A kind of traffic lights detection method based on the relevant convolution in position Yu Fire model, feature the following steps are included:
Step 1: acquire it is various in true traffic scene under the conditions of traffic lights sample, to collected each frame image into
Rower note, marks out the bounding box and specific category of traffic lights, the pictures after mark is divided into training set and test set, respectively
For the training and verifying of network model, training set picture number is greater than test set picture number；The classification includes level
Turn left with green light, red light, the green light straight trip of vertical both direction, green light, red light straight trip, red light left-hand rotation；
Step 2: N number of Fire model being together in series and is combined into trunk convolutional network FireNet, to be used to extract feature；
Wherein, the Fire model is to be combined by the convolutional layer that M1 convolution kernel size is 1 × 1 and M2 convolution kernel size is 3 × 3
It forms, N > 5, M1 > 2, M2 > 2；
Step 3: the training set obtained using step 1 is carried out classification pre-training to trunk convolutional network FireNet, obtained
Netinit weight；
Step 4: a light weight based on Anchor is added after the last layer characteristic pattern of trunk convolutional network FireNet
Grade detection layers；The relevant convolutional layer of point of addition after the characteristic pattern of trunk convolutional network FireNet the last layer；One is added again
A pond layer combines the relevant convolutional layer of detection layers and position based on Anchor, obtains final network；
Step 5: the initialization weight obtained using step 3 initializes the final network that step 4 obtains, then benefit
It is trained with the training set that step 1 obtains, and test image is detected with trained network, can be obtained red
The position of green light and specific category.
The beneficial effects of the present invention are: using the multiple Fire models of connecting as trunk convolutional network FireNet, compared to
General convolutional network is more conducive to transporting on embedded device since the parameter of Fire model is few, running space demand is low
Row；FireNet has more powerful feature representation ability as convolutional network, can sufficiently excavate traffic lights in different scenes
Information obtains more accurate feature representation；The relevant convolutional layer of point of addition behind FireNet the last layer characteristic pattern,
The accuracy that positioning can be improved, is effectively detected lesser traffic lights, and the relevant convolutional layer in position can also be to the volume of coupling
Product feature is separated, and the different spatial information of traffic lights is obtained, and is thus allowed for more accurate position and is returned；By
In using full convolution, it is possible to reduce the accumulation of process error further promotes the precision of detection, while difference can be uniformly processed
Scale, different classes of classification problem.
Detailed description of the invention
Fig. 1 is a kind of traffic lights detection method flow chart based on the relevant convolution in position Yu Fire model of the invention
Fig. 2 is the schematic diagram of the method for the present invention Fire model
Fig. 3 is the relevant convolutional layer model schematic in the method for the present invention position
Fig. 4 is the result figure that traffic lights detection is carried out using the method for the present invention
Specific embodiment
Present invention will be further explained below with reference to the attached drawings and examples, and the present invention includes but are not limited to following implementations
Example.
As shown in Figure 1, a kind of traffic lights detection method based on the relevant convolution in position Yu Fire model of the invention, real
It is existing that steps are as follows:
1. preparing data set
Image is labeled, and generates training and tests required data set.Specifically, pass through the equipment such as in-vehicle camera
Traffic lights picture under the conditions of different weather, different location, different time etc. are various during acquisition real driving, and to acquisition
To image marked frame by frame, mark out the bounding box and specific category of traffic lights, the pictures after mark be divided into training
Collection and test set, are respectively used to the training and verifying of network model, and training set picture number is greater than test set picture number；It is described
Classification include horizontal and vertical directions green light, red light, green light straight trip, green light turn left, red light straight trip, red light turn left
Deng.
2. planned network Fire model, and combine and obtain FireNet
Fire model is composed of the convolutional layer that several convolution kernel sizes are 1 × 1 and 3 × 3, in the present embodiment
Fire model is that the convolutional layer that 1 × 1 and 3 convolution kernel size is 3 × 3 is constituted by 3 convolution kernel sizes, as shown in Figure 2.Using
The convolutional layer that convolution kernel size is 1 × 1 greatly reduces number of parameters, available lesser model.3 × 3 convolution kernel
It is the minimum convoluted core scale that can learn to Local Structure of Image information.It, will be several according to the specific requirements to model size
A Fire model is cascaded, and obtains trunk convolutional network FireNet.9 Fire models are connected on one in the present embodiment
It rises and constitutes trunk convolutional network FireNet.
3. pair FireNet model carries out pre-training
The training set obtained using step 1 carries out classification pre-training to trunk convolutional network FireNet, and the present embodiment is pre-
Training uses stochastic gradient descent method, to the loss function L of target classificationcls=-logpuIt optimizes, wherein
U is true class label, and N is class number, and e is natural constant, puIt is softmax function, LclsFor classification loss function.It is logical
It crosses and minimizes the loss function, obtain netinit weight.
4. adding the lightweight detection layers based on Anchor
The lightweight detection based on Anchor is added after the last layer characteristic pattern of trunk convolutional network FireNet
Layer, i.e., using Ren et al. in document " S.Ren, K.He, R.Girshick, and J.Sun, Faster R-CNN:Towards
Real-Time Object Detection with Region Proposal Networks,in Proc.Neural
The thought of the anchor proposed in Information Processing Systems, pp.91-99,2015. " designs one light
The sliding window of magnitude detects network, and after being added to the last layer characteristic pattern of trunk convolutional network FireNet, for generating
Candidate region.In the present embodiment, the sliding window that lightweight is made of the convolutional layer that convolution kernel size is 3 × 3 detects network, packet
4 scales (25,50,80,120), 3 length-width ratios (1:2,1:1,2:1) totally 12 anchor are contained.
5. the relevant convolutional layer of point of addition
The relevant convolutional layer of point of addition after the output of FireNet the last layer, for extracting the space bit of traffic lights
Confidence breath.The relevant convolutional layer in the position used in the present embodiment is as shown in figure 3, be divided into upper left, the right side for the location information of target
Upper, lower-left, bottom right and in totally five parts, respectively correspond the feature of the different position of traffic lights.
6. adding pond layer
It adds a pond layer to combine the relevant convolutional layer of detection layers and position based on Anchor, obtain final
Network.
7. the modified network of training
The initialization weight obtained using step 3 initializes the final network that step 6 obtains, recycle step
1 obtained training set is trained it.The present embodiment carries out multitask loss function using stochastic gradient descent method excellent
Change, multitask loss function are as follows:
Wherein, LclsFor Classification Loss function, p, u are the classification and true classification of prediction, p respectivelyuIt is softmax letter
Number.LlocIt is positioning loss function, tuFor the position of prediction, v is the result manually marked；[u >=1] indicates that value is when u >=1
1, it is otherwise 0；I is callout box information, wherein (x, y) is the top left co-ordinate of the bounding box of mark, and w is the width of bounding box, h
For the height of bounding box；λ is that a hyper parameter is used to balance two loss functions, λ=2 in the present embodiment.
The power that training obtains is saved until convergence to multitask loss function training using stochastic gradient descent method
It weighs to get trained network is arrived.
8. traffic lights detect
Test image is detected using network trained in step 7, the position of traffic lights and specific can be obtained
Classification.
The present embodiment is run in K40, CentOS operating system of video memory 12G, carries out emulation experiment using Python.It is real
The real roads scene video image that trained and test data is acquired both from automobile data recorder used in testing, resolution ratio are
1280 × 720, and each frame video image is marked, 9000 pictures are finally chosen altogether sets up training set, 2000 picture groups
Build test set.The picture of selection contains various complicated weather conditions, such as rainy day, dense fog, also contains other mal-conditions,
Such as Qiang Guang, it backlight, blocks.It contains the two major classes traffic lights of horizontal and vertical type, contains green light, red in each major class again
Lamp, green light straight trip, green light turns left, red light is kept straight on, red light turns left.
Fig. 4 is wherein piece image testing result schematic diagram, due to data set relative difficult used in the present invention, and
The scale of traffic lights target is generally smaller, by test of many times, for the average detected accuracy rate of various types of other traffic lights
About 68.26%, average recall rate is about 89.74%.Show that this method can effectively detect the traffic lights of plurality of classes, together
When keep lower omission factor.By the design of Fire model, network weight parameter is saved with double precision data type, it is last
Network weight model size is 4.4MB, is far below VGG16, and the weight parameter of the network model of ResNet101 can be used for vehicle
It carries on embedded platform.By the design of the relevant convolutional layer in position, so that only can obtain more accurately determining by detection
Position effect, has certain facilitation simultaneously for the detection of distance small target traffic lights.
Claims (1)
1. a kind of traffic lights detection method based on the relevant convolution in position Yu Fire model, feature the following steps are included:
Step 1: acquire it is various in true traffic scene under the conditions of traffic lights sample, collected each frame image is marked
Note, marks out the bounding box and specific category of traffic lights, the pictures after mark is divided into training set and test set, are respectively used to
The training and verifying of network model, training set picture number are greater than test set picture number；The classification includes horizontal and vertical
Green light, red light, green light straight trip, the green light of straight both direction turn left, red light is kept straight on, red light turns left；
Step 2: N number of Fire model being together in series and is combined into trunk convolutional network FireNet, to be used to extract feature；Wherein,
The Fire model is to be composed of the convolutional layer that M1 convolution kernel size is 1 × 1 and M2 convolution kernel size is 3 × 3,
N > 5, M1 > 2, M2 > 2；
Step 3: the training set obtained using step 1 is carried out classification pre-training to trunk convolutional network FireNet, obtains network
Initialize weight；
Step 4: the lightweight inspection based on Anchor is added after the last layer characteristic pattern of trunk convolutional network FireNet
Survey layer；The relevant convolutional layer of point of addition after the characteristic pattern of trunk convolutional network FireNet the last layer；A pond is added again
Change layer to combine the relevant convolutional layer of detection layers and position based on Anchor, obtains final network；
Step 5: the initialization weight obtained using step 3 initializes the final network that step 4 obtains, and recycles step
Rapid 1 obtained training set is trained it, and is detected with trained network to test image, and traffic lights can be obtained
Position and specific category.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710342500.7A CN107194343B (en) | 2017-05-16 | 2017-05-16 | Traffic lights detection method based on the relevant convolution in position Yu Fire model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710342500.7A CN107194343B (en) | 2017-05-16 | 2017-05-16 | Traffic lights detection method based on the relevant convolution in position Yu Fire model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107194343A CN107194343A (en) | 2017-09-22 |
| CN107194343B true CN107194343B (en) | 2019-11-22 |
Family
ID=59872740
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710342500.7A Active CN107194343B (en) | 2017-05-16 | 2017-05-16 | Traffic lights detection method based on the relevant convolution in position Yu Fire model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107194343B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108108761B (en) * | 2017-12-21 | 2020-05-01 | 西北工业大学 | Rapid traffic signal lamp detection method based on deep feature learning |
| CN110659540A (en) * | 2018-06-29 | 2020-01-07 | 北京京东尚科信息技术有限公司 | Traffic light detection method and device |
| CN110663971B (en) * | 2018-07-02 | 2022-03-29 | 天津工业大学 | Red date quality classification method based on double-branch deep fusion convolutional neural network |
| CN109035808A (en) * | 2018-07-20 | 2018-12-18 | 上海斐讯数据通信技术有限公司 | A kind of traffic lights switching method and system based on deep learning |
| CN111160282B (en) * | 2019-12-31 | 2023-03-24 | 合肥湛达智能科技有限公司 | Traffic light detection method based on binary Yolov3 network |
| CN113077630B (en) * | 2021-04-30 | 2022-06-28 | 安徽江淮汽车集团股份有限公司 | Traffic light detection method, device, equipment and storage medium based on deep learning |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106651765A (en) * | 2016-12-30 | 2017-05-10 | 深圳市唯特视科技有限公司 | Method for automatically generating thumbnail by use of deep neutral network |
-
2017
- 2017-05-16 CN CN201710342500.7A patent/CN107194343B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106651765A (en) * | 2016-12-30 | 2017-05-10 | 深圳市唯特视科技有限公司 | Method for automatically generating thumbnail by use of deep neutral network |
Non-Patent Citations (2)
| Title |
|---|
| Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks;S.Ren er.al;《in Proc.Neural Information Processing Systems》;20151231(第1期);第91-99页 * |
| SQUEEZENET: ALEXNET-LEVEL ACCURACY WITH 50X FEWER PARAMETERS AND <0.5MB MODEL SIZE;Forrest N. Iandola er.al;《arXiv:1602.07360v4 [cs.CV]》;20161104(第1期);第1-13页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107194343A (en) | 2017-09-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107194343B (en) | Traffic lights detection method based on the relevant convolution in position Yu Fire model | |
| CN108596101B (en) | Remote sensing image multi-target detection method based on convolutional neural network | |
| CN107576960B (en) | Target detection method and system for visual radar space-time information fusion | |
| CN105930819B (en) | Real-time city traffic lamp identifying system based on monocular vision and GPS integrated navigation system | |
| US10817731B2 (en) | Image-based pedestrian detection | |
| Liu et al. | Multiscale U-shaped CNN building instance extraction framework with edge constraint for high-spatial-resolution remote sensing imagery | |
| CN108229366B (en) | Deep learning vehicle-mounted obstacle detection method based on radar and image data fusion | |
| CN108694386B (en) | Lane line detection method based on parallel convolution neural network | |
| CN110163187B (en) | F-RCNN-based remote traffic sign detection and identification method | |
| CN105160309B (en) | Three lanes detection method based on morphological image segmentation and region growing | |
| CN104700414B (en) | A kind of road ahead pedestrian's fast ranging method based on vehicle-mounted binocular camera | |
| CN103679674B (en) | Method and system for splicing images of unmanned aircrafts in real time | |
| CN110942000A (en) | Unmanned vehicle target detection method based on deep learning | |
| CN108710875A (en) | A kind of take photo by plane road vehicle method of counting and device based on deep learning | |
| CN109543600A (en) | A kind of realization drivable region detection method and system and application | |
| CN109726627A (en) | A kind of detection method of neural network model training and common ground line | |
| CN110009010A (en) | Wide area optical remote sensing target detection method based on the re-detection of interest region | |
| CN106295605A (en) | Traffic lights detection and recognition methods | |
| CN113111727B (en) | Feature alignment-based method for detecting rotating target in remote sensing scene | |
| CN113255589B (en) | Target detection method and system based on multi-convolution fusion network | |
| Crommelinck et al. | Interactive cadastral boundary delineation from UAV data | |
| CN115035251A (en) | Bridge deck vehicle real-time tracking method based on domain-enhanced synthetic data set | |
| Bourja et al. | Real time vehicle detection, tracking, and inter-vehicle distance estimation based on stereovision and deep learning using YOLOv3 | |
| Shan et al. | Unmanned aerial vehicle (UAV)-Based pavement image stitching without occlusion, crack semantic segmentation, and quantification | |
| CN110909656A (en) | Pedestrian detection method and system with integration of radar and camera |
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 |