CN110008848A - A kind of travelable area recognizing method of the road based on binocular stereo vision - Google Patents
A kind of travelable area recognizing method of the road based on binocular stereo vision Download PDFInfo
- Publication number
- CN110008848A CN110008848A CN201910187216.6A CN201910187216A CN110008848A CN 110008848 A CN110008848 A CN 110008848A CN 201910187216 A CN201910187216 A CN 201910187216A CN 110008848 A CN110008848 A CN 110008848A
- Authority
- CN
- China
- Prior art keywords
- image
- layer
- road
- stereo vision
- binocular stereo
- 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
- 230000004438 eyesight Effects 0.000 title claims abstract description 53
- 238000000034 method Methods 0.000 title claims abstract description 40
- 238000012549 training Methods 0.000 claims abstract description 35
- 230000011218 segmentation Effects 0.000 claims abstract description 32
- 238000012360 testing method Methods 0.000 claims abstract description 11
- 238000013480 data collection Methods 0.000 claims abstract description 7
- 238000012795 verification Methods 0.000 claims abstract description 4
- 239000010410 layer Substances 0.000 claims description 93
- 238000013527 convolutional neural network Methods 0.000 claims description 14
- 238000005070 sampling Methods 0.000 claims description 10
- 230000000694 effects Effects 0.000 claims description 8
- 238000000605 extraction Methods 0.000 claims description 8
- 238000011156 evaluation Methods 0.000 claims description 6
- 239000000463 material Substances 0.000 claims description 6
- 230000006872 improvement Effects 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 5
- 238000003475 lamination Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 239000002344 surface layer Substances 0.000 claims description 4
- 230000004888 barrier function Effects 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 3
- 238000010276 construction Methods 0.000 claims description 3
- 230000002708 enhancing effect Effects 0.000 claims description 3
- 230000006870 function Effects 0.000 claims description 3
- 230000000717 retained effect Effects 0.000 claims description 3
- 238000013519 translation Methods 0.000 claims description 3
- 230000015572 biosynthetic process Effects 0.000 claims 1
- 235000013399 edible fruits Nutrition 0.000 claims 1
- 238000003786 synthesis reaction Methods 0.000 claims 1
- 230000008447 perception Effects 0.000 abstract description 5
- 238000013528 artificial neural network Methods 0.000 description 7
- 238000005516 engineering process Methods 0.000 description 5
- 230000008901 benefit Effects 0.000 description 4
- 238000011161 development Methods 0.000 description 4
- 230000018109 developmental process Effects 0.000 description 4
- 230000000007 visual effect Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012216 screening Methods 0.000 description 2
- 206010039203 Road traffic accident Diseases 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000002955 isolation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000004218 nerve net Anatomy 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
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
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of roads based on binocular stereo vision can travel area recognizing method, comprising steps of the authentic data collection generally acknowledged to automatic Pilot field is acquired by (1) left and right monocular cam is input to pretreated image acquires spatial depth figure, the spatial depth figure is synthesized to the RGB-D image of binocular stereo vision, composing training, test and verification data set with original left camera RGB image;(2) training set is input to and is trained final semantic segmentation model based on improved Alexnet shallow-layer network, then test data is inputted into parted pattern, thus image of the output comprising can travel region;(3) start and completed to install on intelligent vehicle, demarcated complete binocular stereo vision camera group acquirement binocular stereo vision image, being entered into trained model realization road can travel the identification in region.The present invention provides more accurate, comprehensive perception data source for Vehicular intelligent driving.
Description
Technical field
The advanced auxiliary of automobile of the present invention drives and automatic Pilot field, and in particular to a kind of road based on binocular stereo vision
Road can travel area recognizing method.
Background technique
The fast development of auto industry, computer industry and internet is just promoting entire automobile industry to occur one jointly
Secondary innovation revolution.It is loaded with the vehicle of advanced DAS (Driver Assistant System), various technology senses full new automobile and automatic Pilot technology
Development, all teaching that distance " not needing driver " it is following within sight.But from the angle of technology, we are not
Hardly possible discovery will realize that the intelligent driving of vehicle, the intelligent driving of especially LV3 and the above rank still have many problems demand solutions
Certainly, such as sensing layer data extrapolating problem, decision-making level's logical problem and control layer actual effect problem.Wherein, accurate judgement vehicle
Travelable region should be addressed first, to travelable region have accurate anticipation and to vehicle carry out rationally, effective control
System, could further advance to vehicle and make right instructions, could preferably guarantee driving safety, the hair to avoid traffic accident
It is raw.
For intelligent driving vehicle, environment sensing is the basis that vehicle realizes decision and control, due to intelligent vehicle
The road traffic environment faced is extremely complex changeable, therefore, how to realize " the perfection perception " to road environment as far as possible, quasi-
True identifies that road can travel region and just become the field for having important research to be worth.
The common sensor of intelligent driving includes visual sensor, ultrasonic sensor, laser radar, GPS and IMU at present
Deng wherein visual sensor obtains the favor of numerous researchers with its cheap cost and preferable reliability, is current intelligence
One of most widely used sensor in driving.Traditional travelable area recognizing method is carried out based on monocular image mostly
, without considering space geometry feature, and the timeliness of network model processing is poor, thus this patent is by binocular tri-dimensional
Feel technology combination shallow semantic divides network, and being introduced into road can travel in region recognition, and method focuses more on travelable region
Geometric depath information.
Semantic segmentation is a problem of field of machine vision all the time, is also difficult to find a kind of generalization and Shandong so far
The good parted pattern of stick.But its importance is self-evident, we weigh image, semantic segmentation from following standard
Advantage and disadvantage:
(1) inside of different cut zone is smooth, and feature is consistent, and meets similitude and principle of whole;
(2) it is obvious to divide relied on property difference for different picture regions;
(3) boundary gradient in adjacent segmentation region changes greatly, so that boundary is obvious.Although many computer visions
The expert in field proposes many solutions, so that semantic segmentation has pretty good development, but since it has complexity
Principle, engineering feature, so there are still many problems to need to solve.
Convolutional neural networks are a kind of neural networks, but are that convolutional neural networks have day with the difference of general neural network
The advantage of feature is so extracted, convolutional neural networks contain the feature extractor being made of convolutional layer and sub-sampling layer, Neng Gougeng
Increase effect, intelligently extract characteristics of image, this point is that the tional identification algorithm based on pixel and boundary cannot compare.It
Before, it is higher that the neural network due to comparing deep layer (referring generally to three layers or more) calculates force request to hardware, but now with depth
The rapid development of learning algorithm and hardware is spent, problems have been obtained overcoming well, also provided for the Project Realization of algorithm
It may.
Summary of the invention
To solve the above problems existing in the prior art, the present invention to design a kind of road based on binocular stereo vision can
Running region recognition methods, it is intended to which judgement road that can be more acurrate, more reliable can travel region, preferably perception vehicle driving
Environment, to preferably guide the traveling of vehicle in intelligently auxiliary drives.
In order to achieve the above object, technical solution of the present invention and method are as described below:
A kind of travelable area recognizing method of the road based on binocular stereo vision, comprising steps of
(1) by the authentic data collection that automatic Pilot field is generally acknowledged be acquired with pretreated image be input to it is left,
Right monocular cam simultaneously acquires spatial depth figure by binocular stereo vision, by the spatial depth figure and original left camera
RGB image synthesizes the RGB-D image of binocular stereo vision, composing training, test and verification data set;
(2) training set is input to and semantic segmentation is carried out based on improved Alexnet shallow-layer network and is constantly changed
Generation training, Optimal Parameters, reach scheduled model convergence state, and final curing model parameter completes training as final
Semantic segmentation model, then test data is inputted into parted pattern, thus image of the output comprising can travel region;
(3) start and completed to install on intelligent vehicle, demarcated complete binocular stereo vision camera group, to traffic environment
Information carries out test image acquisition, obtains binocular stereo vision image, it is feasible to be entered into trained model realization road
Sail the identification in region.
Further, the step (1) specifically includes:
Stereoscopic vision matching is carried out by left and right monocular image and acquires disparity map, later according to system binocular vision model
Spatial depth figure is acquired, is finally superimposed upon spatial depth figure (representing space geometry information and feature) as fourth lane
On RGB image, RGB-D image is formed.Increase space geometry feature, i.e. depth information i.e. on RGB image feature, so that net
Network has more features " can remember ".
Left and right monocular image obtains the depth information of image pixel by technique of binocular stereoscopic vision, restores to a certain degree
The three-dimensional geometry feature of traffic route, and the advantage of visual sensor quickly, stable, inexpensive can be given full play to.
Further, the encoder in the step (2), by construction convolutional neural networks as feature extraction, wherein
The improvement of Alexnet shallow-layer network is specifically included that
(1) full articulamentum is removed, the convolutional layer and pond layer of network are only retained;
(2) all pond layers of original Alexnet are changed to 2x2, get rid of all pond 3x3 layers;
(3) size of first layer convolutional layer is changed to 3x3 by 11x11, its step-length is changed to 1, and make padding=1;
(4) identical convolutional layer is added with second layer convolutional layer between second layer convolutional layer and second layer pond layer, increases
The depth of screening network, so that the convolution number of plies of new network is 6, and the pond number of plies then remains unchanged.
Further, in step (2), the training process of the improved Alexnet shallow-layer network includes:
(1) the spatial depth figure and original left camera RGB image are synthesized into RGB-D image, is arranged as instruction
Practice data set;
(2) with the training dataset for having label, training data is input to designed semantic segmentation network, picture
By several convolutional layers and pond layer, feature extraction is carried out;
(3) pond layer is up-sampled using deconvolution, search up-sampling is restored by whole image by deconvolution,
It is mapped to picture pixels rank, probabilistic forecasting is carried out to the travelable area pixel of intelligent vehicle;
(4) realize that semantic segmentation indicates and exports the image in travelable region.
Further, the authentic data collection generally acknowledged to automatic Pilot field is acquired and pretreatment specifically includes:
Training image and evaluation image are obtained with open source KITTI and CITYSCAPES data set;
Data enhancing processing is carried out to the training image and evaluation image, is grasped by symmetrical, overturning, translation and rotation
Make, obtains the training dataset that quantity is bigger, covering surface is more complete.
Further, the semantic segmentation model of the improved Alexnet shallow-layer network includes: 12 convolutional layers, 3 ponds
Change layer, 3 warp laminations (up-sampling) and one layer Softmax layers are used to prediction pixel probability.
Further, it is specifically to utilize stochastic gradient that the travelable area pixel to intelligent vehicle, which carries out probabilistic forecasting,
The method for declining (mini-batch SGD) trains the Softmax of output layer to return layer, until the loss function returned is realized in advance
Fixed convergence reaches scheduled prediction segmentation recognition effect.
Further, the travelable region specifically refers to all roads for removing all barriers, allowing vehicle driving
Road region.
Further, it is described allow vehicle driving all road areas from physical structure level include structuring road surface,
Semi-structured road surface and unstructured road surface.
Further, its road edge of the road surface of structuring rule, the structure on road surface is single, neat, has obvious
Lane line and other handmarkings carriage way, structure sheaf executes according to certain standard, the color and material of surface layer
Keep unified;The semi-structured road surface refers to the road surface of general nonstandardized technique, and the color and material of top course exist
Difference, such as parking lot, square etc. further include some other distributor roads;The non-structured road surface refers to no structure
The natural road scene of layer.
Compared with prior art, the invention proposes under a kind of intelligent driving running car environment, vehicle is for can travel area
The recognition methods in domain, this method are proposed for the key and necessity for identifying road in intelligent vehicle traveling based on binocular
The road of stereoscopic vision can travel the recognition methods in region: it is special to increase space geometry on the basis of monocular image R, G, B feature
Sign (depth) recycles convolutional neural networks to carry out image special to realize the application that binocular stereo vision is divided in image, semantic
Sign is extracted, and the feature extracted is mapped directly to whole picture by deconvolution neural network, realizes the classification of pixel one by one;
To based on improved Alexnet convolutional neural networks and deconvolution neural network composed by semantic segmentation model carry out it is continuous
Improvement and optimization, finally by above-mentioned semantic segmentation model be applied to vehicle can travel region identification in, and by generally acknowledge
Standard data set carries out recognition effect verifying, compares segmentation precision, advanced optimizes model, provides more for Vehicular intelligent driving
For accurate, comprehensive perception data source.
Detailed description of the invention
The present invention provides reference attached drawings in order to further understand to disclosure, and attached drawing constitutes the one of the application
Part, but it is only used for some invention non-limiting examples that diagram embodies concept of the invention, rather than for making any limit
System.
Fig. 1 is that the road based on binocular stereo vision of some example embodiments according to the present invention can travel the knowledge in region
The flow chart of other method.
Fig. 2 is the convolution sum up-sampling network structure of some example embodiments according to the present invention.
Fig. 3 is binocular stereo vision schematic diagram, is illustrated according to the principle that left and right monocular vision seeks geometric space depth map
Figure.
Fig. 4 is that obtained black region can travel area schematic according to the method for the present invention.
Specific embodiment
It elaborates with reference to the accompanying drawings and examples to the present invention.
As shown in Figure 1, a kind of road based on binocular stereo vision can travel area recognizing method, comprising steps of
(1) by the authentic data collection that automatic Pilot field is generally acknowledged be acquired with pretreated image be input to it is left,
Right monocular cam simultaneously acquires spatial depth figure by binocular stereo vision, by the spatial depth figure and original left camera
RGB image synthesizes the RGB-D image of binocular stereo vision, composing training, test and verification data set;
(2) training set is input to and semantic segmentation is carried out based on improved Alexnet shallow-layer network and is constantly changed
Generation training, Optimal Parameters, reach scheduled model convergence state, and final curing model parameter completes training as final
Semantic segmentation model, then test data is inputted into parted pattern, thus image of the output comprising can travel region;
(3) start and completed to install on intelligent vehicle, demarcated complete binocular stereo vision camera group, to traffic environment
Information carries out test image acquisition, obtains binocular stereo vision image, it is feasible to be entered into trained model realization road
Sail the identification in region.
Specifically, the step (1) specifically includes:
Stereoscopic vision matching is carried out by left and right monocular image and acquires disparity map, later according to system binocular vision model
Spatial depth figure is acquired, is finally superimposed upon spatial depth figure (representing space geometry information and feature) as fourth lane
On RGB image, RGB-D image is formed.Increase space geometry feature, i.e. depth information i.e. on RGB image feature, so that net
Network has more features " can remember ".
Wherein, as shown in figure 3, the principle of binocular stereo vision principle derivation space geometry depth map is as follows:
Wherein, f is camera focus, and b is binocular vision system parallax range, and z is Object Depth, that is, distance, uRIt is that object exists
Right camera pixel, uLIt is object in left camera pixel.
It arranges:
D=uL-uR
What d here was indicated is the difference of left and right pixel coordinate, and unit is pixel (pixel), is defined as parallax
(Disparity)。
As shown in Fig. 2, the present embodiment uses convolution, pond layer, up-samples this deep learning processing method, pass through improvement
Alexnet building shallow-layer semantic segmentation network, preferably the travelable region of road scene can be split, preferably
The identification in the travelable region of specific region under the conditions of ground realization intelligent driving, and there is relatively high accuracy of identification and efficiency,
Relatively good data are capable of providing to intelligent driving algorithm to support.
Specifically, the encoder in the step (2), by construction convolutional neural networks as feature extraction, wherein
The improvement of Alexnet shallow-layer network is specifically included that
(1) full articulamentum is removed, the convolutional layer and pond layer of network are only retained;
(2) all pond layers of original Alexnet are changed to 2x2, get rid of all pond 3x3 layers;
(3) size of first layer convolutional layer is changed to 3x3 by 11x11, its step-length is changed to 1, and make padding=1;
(4) identical convolutional layer is added with second layer convolutional layer between second layer convolutional layer and second layer pond layer, increases
The depth of screening network, so that the convolution number of plies of new network is 6, and the pond number of plies then remains unchanged.
Specifically, in step (2), the training process of the improved Alexnet shallow-layer network includes:
(1) the spatial depth figure and original left camera RGB image are synthesized into RGB-D image, is arranged as instruction
Practice data set;
(2) with the training dataset for having label, training data is input to designed semantic segmentation network, picture
By several convolutional layers and pond layer, feature extraction is carried out;
(3) pond layer is up-sampled using deconvolution, search up-sampling is restored by whole image by deconvolution,
It is mapped to picture pixels rank, probabilistic forecasting is carried out to the travelable area pixel of intelligent vehicle;
(4) realize that semantic segmentation indicates and exports the image in travelable region.
Specifically, the authentic data collection generally acknowledged to automatic Pilot field is acquired and pretreatment specifically includes:
Training image and evaluation image are obtained with open source KITTI and CITYSCAPES data set;
Data enhancing processing is carried out to the training image and evaluation image, is grasped by symmetrical, overturning, translation and rotation
Make, obtains the training dataset that quantity is bigger, covering surface is more complete, can also prevent the generation of over-fitting.
Specifically, the semantic segmentation model of the improved Alexnet shallow-layer network includes: 12 convolutional layers, 3 ponds
Change layer, 3 warp laminations (up-sampling) and one layer Softmax layers are used to prediction pixel probability.
Generally, the classification for carrying out Pixel-level to image by convolution sum deconvolution, to solve the figure of semantic level
As segmentation problem.Initially classical CNN is after convolutional layer, the feature vector that usually can all full articulamentum is used to be defined,
Softmax classification is carried out again.And improved Alexnet has better feature extraction effect, and using warp lamination to selected
The characteristic pattern (feature map) for selecting convolutional layer is up-sampled, and so that it is restored to the identical size of input picture, or even will not
Up-sampling result with characteristic pattern is merged, and on the basis of retaining image space information, realizes the prediction to each pixel,
Finally classified pixel-by-pixel on the characteristic pattern of up-sampling.Last individual element calculates the loss (Loss) of Softmax classification,
That is, each pixel corresponds to a training sample.
Preferably, the inherent advantage of convolutional neural networks (CNN) is that it is multiple that its multilayered structure may learn image
The feature of level, i.e. the convolutional layer receptive field of shallow-layer compare smaller, and the main local feature for learning low-dimensional, the present invention passes through expansion
The width of convolution kernel number increase network;The convolutional layer receptive field of deep layer is more larger, main study to higher-dimension, more abstract
Feature.Its weight parameter of different features is different, different to the sensibility of feature, helps to improve network performance.
Specifically, it is specifically to utilize stochastic gradient that the travelable area pixel to intelligent vehicle, which carries out probabilistic forecasting,
The method for declining (mini-batch SGD) trains the Softmax of output layer to return layer, until the loss function returned is realized in advance
Fixed convergence reaches scheduled prediction segmentation recognition effect.
Specifically, the travelable region specifically refer to remove all barriers (such as: vehicle, pedestrian, isolation strip
Deng), allow vehicle driving all road areas.All road areas for allowing vehicle driving are from physical structure level packet
Include structuring road surface, semi-structured road surface and unstructured road surface.The road surface of the structuring its road edge rule, road
The structure in face is single, neat, has the carriage way of apparent lane line and other handmarkings, structure sheaf is according to certain mark
Standard executes, and the color and material of surface layer keep unified;The semi-structured road surface refers to the road surface of general nonstandardized technique, road
The color and material of face surface layer have differences, such as parking lot, square etc., further include some other distributor roads;It is described non-
The road surface of structuring refers to the natural road scene of not structure sheaf.In structuring, semi-structured road and unstructured road row
It sails under scene, the advanced auxiliary of automobile drives or automatic Pilot needs realizing route to plan, just has to realize to travelable region
Identification, as far as possible close to " perfection " perceive.
The weight parameter that the present invention is obtained by training network is loaded into the network structure saved, relatively clear clear
It identifies the travelable region of intelligent vehicle in motion, and carries out can travel indicating for region with clearly color is compared, such as
Shown in Fig. 4, black portions are two kinds and are based respectively on image, semantic cutting transformation schematic diagram of the invention, can also clearly obtain
The travelable region of intelligent vehicle.
The present invention is above-mentioned more in the prior art in order to solve the problems, such as, proposes a kind of intelligent driving running car environment
Under, the identification to travelable region is realized by binocular stereo vision.Binocular stereo vision figure is carried out by convolutional neural networks
As feature extraction, the feature extracted then is mapped directly into whole picture by deconvolution neural network, realizes picture one by one
The classification of element;Research is extended into binocular stereo vision from monocular vision image, to realize binocular stereo vision in image language
The application of justice segmentation, i.e., increase space geometry feature (depth) on the basis of monocular image R, G, B feature, so that semantic segmentation
Network has more fully feature that can learn;To convolutional neural networks and deconvolution nerve net based on improved Alexnet
Semantic segmentation model composed by network is constantly improved and is optimized, and the accuracy and efficiency of model is improved;By above-mentioned semanteme
Parted pattern is applied to vehicle and can travel in the identification in region, and carries out recognition effect verifying by the recognized standard data set,
Segmentation precision is compared, model is advanced optimized, drives more accurate, the comprehensive perception data source provided for Vehicular intelligent.
The present invention is directed to combine traditional algorithm, improve and construct it is suitable, have a wide range of application, the good depth convolution of adaptability
Neural network, and be applied to and divide specifically for the image, semantic in the advanced auxiliary driving of automobile or automatic Pilot field,
It realizes that vehicle can travel the identification and detection in region, strives for more improving Vehicular automatic driving technology to the letter of environment sensing level
Breath acquisition and analysis.The method being based on is the image realized under road scene in conjunction with binocular stereo vision and convolutional neural networks
Semantic segmentation, more accurately identifying road by visual sensor data and deep learning algorithm can travel region, for intelligence
The further decision of vehicle, control are of great significance.
Part Methods step of the invention and process may need to be executed by computer, thus with hardware, software, firmware and
The mode of any combination thereof is implemented.
The above is only a preferred embodiment of the present invention, is not intended to limit the implementation and interest field of invention, all according to this
Equivalence changes, modification, replacement that content described in patent application scope of patent protection is made etc. should all be included in the present patent application
In the scope of the patents.Those skilled in the art will appreciate that without departing from the scope and spirit of the present invention, it can be wider
It is changed and modified in wealthy various aspects.
Claims (10)
1. a kind of road based on binocular stereo vision can travel area recognizing method, which is characterized in that comprising steps of
(1) the authentic data collection generally acknowledged to automatic Pilot field is acquired and is input to left and right list with pretreated image
Mesh camera simultaneously acquires spatial depth figure by binocular stereo vision, and the spatial depth figure and original left camera RGB are schemed
As the RGB-D image of synthesis binocular stereo vision, composing training, test and verification data set;
(2) training set is input to and semantic segmentation and continuous iteration instruction is carried out based on improved Alexnet shallow-layer network
Practice, Optimal Parameters, reach scheduled model convergence state, final curing model parameter completes training as final language
Adopted parted pattern, then test data is inputted into parted pattern, thus image of the output comprising can travel region;
(3) start and completed to install on intelligent vehicle, demarcated complete binocular stereo vision camera group, to traffic environment information
Test image acquisition is carried out, binocular stereo vision image is obtained, being entered into trained model realization road can travel area
The identification in domain.
2. the road described according to claim 1 based on binocular stereo vision can travel area recognizing method, which is characterized in that
The step (1) specifically includes:
Stereoscopic vision matching is carried out by left and right monocular image and acquires disparity map, is acquired later according to system binocular vision model
Spatial depth figure is finally superimposed upon on RGB image using spatial depth figure as fourth lane, forms RGB-D image.
3. the road described according to claim 1 based on binocular stereo vision can travel area recognizing method, which is characterized in that
Encoder in the step (2), by construction convolutional neural networks as feature extraction, wherein to Alexnet shallow-layer network
Improvement specifically include that
(1) full articulamentum is removed, the convolutional layer and pond layer of network are only retained;
(2) all pond layers of original Alexnet are changed to 2x2, get rid of all pond 3x3 layers;
(3) size of first layer convolutional layer is changed to 3x3 by 11x11, its step-length is changed to 1, and make padding=1;
(4) identical convolutional layer is added with second layer convolutional layer between second layer convolutional layer and second layer pond layer, increases net
The depth of network, so that the convolution number of plies of new network is 6, and the pond number of plies then remains unchanged.
4. the road described according to claim 1 based on binocular stereo vision can travel area recognizing method, which is characterized in that
In step (2), the training process of the improved Alexnet shallow-layer network includes:
(1) the spatial depth figure and original left camera RGB image are synthesized into RGB-D image, is arranged as training number
According to collection;
(2) with the training dataset for having label, training data is input to designed semantic segmentation network, picture passes through
Several convolutional layers and pond layer carry out feature extraction;
(3) pond layer is up-sampled using deconvolution, search up-sampling is restored by whole image by deconvolution, is mapped
To picture pixels rank, probabilistic forecasting is carried out to the travelable area pixel of intelligent vehicle;
(4) realize that semantic segmentation indicates and exports the image in travelable region.
5. the road described according to claim 1 based on binocular stereo vision can travel area recognizing method, which is characterized in that
The authentic data collection generally acknowledged to automatic Pilot field is acquired and pretreatment specifically includes:
Training image and evaluation image are obtained with KITTI and CITYSCAPES data set;
Data enhancing processing is carried out to the training image and evaluation image, by symmetrical, overturning, translation and rotation process, is obtained
The training dataset bigger to quantity, covering surface is more complete.
6. the road described according to claim 1 based on binocular stereo vision can travel area recognizing method, which is characterized in that
The semantic segmentation model of the improved Alexnet shallow-layer network includes: 12 convolutional layers, 3 pond layers, 3 warps
Lamination (up-sampling) and one layer Softmax layers are used to prediction pixel probability.
7. the road according to claim 4 based on binocular stereo vision can travel area recognizing method, which is characterized in that
It is specifically to utilize the method training of stochastic gradient descent defeated that the travelable area pixel to intelligent vehicle, which carries out probabilistic forecasting,
The Softmax of layer returns layer out, until the loss function returned realizes scheduled convergence, reaches scheduled prediction segmentation identification effect
Fruit.
8. the road described according to claim 1 based on binocular stereo vision can travel area recognizing method, which is characterized in that
The travelable region specifically refers to all road areas for removing all barriers, allowing vehicle driving.
9. the road according to claim 8 based on binocular stereo vision can travel area recognizing method, which is characterized in that
All road areas for allowing vehicle driving from physical structure level include structuring road surface, semi-structured road surface and non-
Structuring road surface.
10. the road according to claim 9 based on binocular stereo vision can travel area recognizing method, feature exists
In, its road edge of the road surface of structuring rule, the structure on road surface is single, neat, there is apparent lane line and other
The carriage way of handmarking, structure sheaf are executed according to certain standard, and the color and material of surface layer keep unified;Described half
The road surface of structuring refers to the road surface of general nonstandardized technique, and the color and material of top course have differences;It is described non-structural
The road surface of change refers to the natural road scene of not structure sheaf.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910187216.6A CN110008848A (en) | 2019-03-13 | 2019-03-13 | A kind of travelable area recognizing method of the road based on binocular stereo vision |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910187216.6A CN110008848A (en) | 2019-03-13 | 2019-03-13 | A kind of travelable area recognizing method of the road based on binocular stereo vision |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110008848A true CN110008848A (en) | 2019-07-12 |
Family
ID=67166927
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910187216.6A Pending CN110008848A (en) | 2019-03-13 | 2019-03-13 | A kind of travelable area recognizing method of the road based on binocular stereo vision |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110008848A (en) |
Cited By (27)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110688937A (en) * | 2019-09-24 | 2020-01-14 | 重庆大学 | Binocular vision auxiliary driving parasitic system |
CN110992304A (en) * | 2019-10-30 | 2020-04-10 | 浙江力邦合信智能制动系统股份有限公司 | Two-dimensional image depth measuring method and application thereof in vehicle safety monitoring |
CN111079586A (en) * | 2019-12-03 | 2020-04-28 | 西安电子科技大学 | Automatic driving target detection system and method based on deep learning and binocular shooting |
CN111178584A (en) * | 2019-12-04 | 2020-05-19 | 常熟理工学院 | Unmanned behavior prediction method based on double-layer fusion model |
CN111292366A (en) * | 2020-02-17 | 2020-06-16 | 华侨大学 | Visual driving ranging algorithm based on deep learning and edge calculation |
CN111354028A (en) * | 2020-02-19 | 2020-06-30 | 山东大学 | Binocular vision-based power transmission channel hidden danger identification and tracking method |
CN111562788A (en) * | 2020-06-04 | 2020-08-21 | 哈尔滨理工大学 | Path planning and obstacle avoidance method for indoor intelligent cleaning robot |
CN111666833A (en) * | 2020-01-20 | 2020-09-15 | 深圳大学 | Road damage detection method based on vehicle formation in intelligent networking environment |
CN112150805A (en) * | 2020-09-01 | 2020-12-29 | 福瑞泰克智能系统有限公司 | Determination method, device and equipment of travelable area and storage medium |
CN112417976A (en) * | 2020-10-26 | 2021-02-26 | 深圳大学 | Pavement detection and identification method and device, intelligent terminal and storage medium |
CN112634341A (en) * | 2020-12-24 | 2021-04-09 | 湖北工业大学 | Method for constructing depth estimation model of multi-vision task cooperation |
CN112639821A (en) * | 2020-05-11 | 2021-04-09 | 华为技术有限公司 | Method and system for detecting vehicle travelable area and automatic driving vehicle adopting system |
CN112706764A (en) * | 2020-12-30 | 2021-04-27 | 潍柴动力股份有限公司 | Active anti-collision early warning method, device, equipment and storage medium |
CN113096176A (en) * | 2021-03-26 | 2021-07-09 | 西安交通大学 | Semantic segmentation assisted binocular vision unsupervised depth estimation method |
CN113240632A (en) * | 2021-04-22 | 2021-08-10 | 北京中科慧眼科技有限公司 | Road surface detection method and system based on semantic segmentation network and intelligent terminal |
CN113255520A (en) * | 2021-05-25 | 2021-08-13 | 华中科技大学 | Vehicle obstacle avoidance method based on binocular vision and deep learning and electronic equipment |
CN113269130A (en) * | 2021-06-11 | 2021-08-17 | 国电瑞源(西安)智能研究院有限公司 | Visual path searching method based on artificial neural network |
CN113313071A (en) * | 2021-06-28 | 2021-08-27 | 浙江同善人工智能技术有限公司 | Road area identification method and system |
CN113343875A (en) * | 2021-06-18 | 2021-09-03 | 深圳亿嘉和科技研发有限公司 | Driving region sensing method for robot |
WO2021175434A1 (en) * | 2020-03-05 | 2021-09-10 | Cambridge Enterprise Limited | System and method for predicting a map from an image |
CN113569620A (en) * | 2021-05-24 | 2021-10-29 | 惠州市德赛西威智能交通技术研究院有限公司 | Pavement marker instantiation identification method based on monocular vision |
CN113610970A (en) * | 2021-08-30 | 2021-11-05 | 上海智能网联汽车技术中心有限公司 | Automatic driving system, device and method |
CN113792707A (en) * | 2021-11-10 | 2021-12-14 | 北京中科慧眼科技有限公司 | Terrain environment detection method and system based on binocular stereo camera and intelligent terminal |
CN114154510A (en) * | 2021-11-30 | 2022-03-08 | 江苏智能网联汽车创新中心有限公司 | Control method and device for automatic driving vehicle, electronic equipment and storage medium |
CN115063594A (en) * | 2022-08-19 | 2022-09-16 | 清驰(济南)智能科技有限公司 | Feature extraction method and device based on automatic driving |
CN115187959A (en) * | 2022-07-14 | 2022-10-14 | 清华大学 | Method and system for landing flying vehicle in mountainous region based on binocular vision |
CN115179920A (en) * | 2022-09-07 | 2022-10-14 | 北京中科慧眼科技有限公司 | Vehicle running system adjusting method and system based on cross-country scene |
Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103871061A (en) * | 2014-03-17 | 2014-06-18 | 电子科技大学 | Method for processing fundus images based on binocular vision |
CN105760442A (en) * | 2016-02-01 | 2016-07-13 | 中国科学技术大学 | Image feature enhancing method based on database neighborhood relation |
CN105930830A (en) * | 2016-05-18 | 2016-09-07 | 大连理工大学 | Road surface traffic sign recognition method based on convolution neural network |
CN106709568A (en) * | 2016-12-16 | 2017-05-24 | 北京工业大学 | RGB-D image object detection and semantic segmentation method based on deep convolution network |
CN107657233A (en) * | 2017-09-28 | 2018-02-02 | 东华大学 | Static sign language real-time identification method based on modified single multi-target detection device |
CN107871098A (en) * | 2016-09-23 | 2018-04-03 | 北京眼神科技有限公司 | Method and device for acquiring human face characteristic points |
CN108234984A (en) * | 2018-03-15 | 2018-06-29 | 百度在线网络技术(北京)有限公司 | Binocular depth camera system and depth image generation method |
CN108520535A (en) * | 2018-03-26 | 2018-09-11 | 天津大学 | Object classification method based on depth recovery information |
CN108985194A (en) * | 2018-06-29 | 2018-12-11 | 华南理工大学 | A kind of intelligent vehicle based on image, semantic segmentation can travel the recognition methods in region |
CN109086799A (en) * | 2018-07-04 | 2018-12-25 | 江苏大学 | A kind of crop leaf disease recognition method based on improvement convolutional neural networks model AlexNet |
CN109117691A (en) * | 2017-06-23 | 2019-01-01 | 百度在线网络技术(北京)有限公司 | Drivable region detection method, device, equipment and storage medium |
-
2019
- 2019-03-13 CN CN201910187216.6A patent/CN110008848A/en active Pending
Patent Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103871061A (en) * | 2014-03-17 | 2014-06-18 | 电子科技大学 | Method for processing fundus images based on binocular vision |
CN105760442A (en) * | 2016-02-01 | 2016-07-13 | 中国科学技术大学 | Image feature enhancing method based on database neighborhood relation |
CN105930830A (en) * | 2016-05-18 | 2016-09-07 | 大连理工大学 | Road surface traffic sign recognition method based on convolution neural network |
CN107871098A (en) * | 2016-09-23 | 2018-04-03 | 北京眼神科技有限公司 | Method and device for acquiring human face characteristic points |
CN106709568A (en) * | 2016-12-16 | 2017-05-24 | 北京工业大学 | RGB-D image object detection and semantic segmentation method based on deep convolution network |
CN109117691A (en) * | 2017-06-23 | 2019-01-01 | 百度在线网络技术(北京)有限公司 | Drivable region detection method, device, equipment and storage medium |
CN107657233A (en) * | 2017-09-28 | 2018-02-02 | 东华大学 | Static sign language real-time identification method based on modified single multi-target detection device |
CN108234984A (en) * | 2018-03-15 | 2018-06-29 | 百度在线网络技术(北京)有限公司 | Binocular depth camera system and depth image generation method |
CN108520535A (en) * | 2018-03-26 | 2018-09-11 | 天津大学 | Object classification method based on depth recovery information |
CN108985194A (en) * | 2018-06-29 | 2018-12-11 | 华南理工大学 | A kind of intelligent vehicle based on image, semantic segmentation can travel the recognition methods in region |
CN109086799A (en) * | 2018-07-04 | 2018-12-25 | 江苏大学 | A kind of crop leaf disease recognition method based on improvement convolutional neural networks model AlexNet |
Non-Patent Citations (4)
Title |
---|
代具亭等: "基于彩色-深度图像和深度学习的场景语义分割网络", 《科学技术与工程》 * |
代具亭等: "基于彩色-深度图像和深度学习的场景语义分割网络", 《科学技术与工程》, no. 20, 18 July 2018 (2018-07-18) * |
赵远东等: "人脸识别中AlexNet网络设计和改进方法研究", 《通信技术》 * |
赵远东等: "人脸识别中AlexNet网络设计和改进方法研究", 《通信技术》, no. 03, 10 March 2019 (2019-03-10) * |
Cited By (40)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110688937A (en) * | 2019-09-24 | 2020-01-14 | 重庆大学 | Binocular vision auxiliary driving parasitic system |
CN110992304A (en) * | 2019-10-30 | 2020-04-10 | 浙江力邦合信智能制动系统股份有限公司 | Two-dimensional image depth measuring method and application thereof in vehicle safety monitoring |
CN110992304B (en) * | 2019-10-30 | 2023-07-07 | 浙江力邦合信智能制动系统股份有限公司 | Two-dimensional image depth measurement method and application thereof in vehicle safety monitoring |
CN111079586A (en) * | 2019-12-03 | 2020-04-28 | 西安电子科技大学 | Automatic driving target detection system and method based on deep learning and binocular shooting |
CN111079586B (en) * | 2019-12-03 | 2023-05-23 | 西安电子科技大学 | Automatic driving target detection system and method based on deep learning and binocular shooting |
CN111178584B (en) * | 2019-12-04 | 2021-12-07 | 常熟理工学院 | Unmanned behavior prediction method based on double-layer fusion model |
CN111178584A (en) * | 2019-12-04 | 2020-05-19 | 常熟理工学院 | Unmanned behavior prediction method based on double-layer fusion model |
CN111666833A (en) * | 2020-01-20 | 2020-09-15 | 深圳大学 | Road damage detection method based on vehicle formation in intelligent networking environment |
CN111292366A (en) * | 2020-02-17 | 2020-06-16 | 华侨大学 | Visual driving ranging algorithm based on deep learning and edge calculation |
CN111292366B (en) * | 2020-02-17 | 2023-03-10 | 华侨大学 | Visual driving ranging algorithm based on deep learning and edge calculation |
CN111354028B (en) * | 2020-02-19 | 2022-05-31 | 山东大学 | Binocular vision-based power transmission channel hidden danger identification and tracking method |
CN111354028A (en) * | 2020-02-19 | 2020-06-30 | 山东大学 | Binocular vision-based power transmission channel hidden danger identification and tracking method |
WO2021175434A1 (en) * | 2020-03-05 | 2021-09-10 | Cambridge Enterprise Limited | System and method for predicting a map from an image |
CN114282597B (en) * | 2020-05-11 | 2023-04-07 | 华为技术有限公司 | Method and system for detecting vehicle travelable area and automatic driving vehicle adopting system |
CN112639821A (en) * | 2020-05-11 | 2021-04-09 | 华为技术有限公司 | Method and system for detecting vehicle travelable area and automatic driving vehicle adopting system |
CN114282597A (en) * | 2020-05-11 | 2022-04-05 | 华为技术有限公司 | Method and system for detecting vehicle travelable area and automatic driving vehicle adopting system |
CN111562788A (en) * | 2020-06-04 | 2020-08-21 | 哈尔滨理工大学 | Path planning and obstacle avoidance method for indoor intelligent cleaning robot |
CN112150805B (en) * | 2020-09-01 | 2022-02-25 | 福瑞泰克智能系统有限公司 | Determination method, device and equipment of travelable area and storage medium |
CN112150805A (en) * | 2020-09-01 | 2020-12-29 | 福瑞泰克智能系统有限公司 | Determination method, device and equipment of travelable area and storage medium |
CN112417976B (en) * | 2020-10-26 | 2024-05-21 | 深圳大学 | Pavement detection and identification method and device, intelligent terminal and storage medium |
CN112417976A (en) * | 2020-10-26 | 2021-02-26 | 深圳大学 | Pavement detection and identification method and device, intelligent terminal and storage medium |
CN112634341A (en) * | 2020-12-24 | 2021-04-09 | 湖北工业大学 | Method for constructing depth estimation model of multi-vision task cooperation |
CN112706764A (en) * | 2020-12-30 | 2021-04-27 | 潍柴动力股份有限公司 | Active anti-collision early warning method, device, equipment and storage medium |
CN113096176B (en) * | 2021-03-26 | 2024-04-05 | 西安交通大学 | Semantic segmentation-assisted binocular vision unsupervised depth estimation method |
CN113096176A (en) * | 2021-03-26 | 2021-07-09 | 西安交通大学 | Semantic segmentation assisted binocular vision unsupervised depth estimation method |
CN113240632B (en) * | 2021-04-22 | 2024-04-16 | 北京中科慧眼科技有限公司 | Pavement detection method and system based on semantic segmentation network and intelligent terminal |
CN113240632A (en) * | 2021-04-22 | 2021-08-10 | 北京中科慧眼科技有限公司 | Road surface detection method and system based on semantic segmentation network and intelligent terminal |
CN113569620A (en) * | 2021-05-24 | 2021-10-29 | 惠州市德赛西威智能交通技术研究院有限公司 | Pavement marker instantiation identification method based on monocular vision |
CN113255520A (en) * | 2021-05-25 | 2021-08-13 | 华中科技大学 | Vehicle obstacle avoidance method based on binocular vision and deep learning and electronic equipment |
CN113269130A (en) * | 2021-06-11 | 2021-08-17 | 国电瑞源(西安)智能研究院有限公司 | Visual path searching method based on artificial neural network |
CN113343875A (en) * | 2021-06-18 | 2021-09-03 | 深圳亿嘉和科技研发有限公司 | Driving region sensing method for robot |
CN113313071A (en) * | 2021-06-28 | 2021-08-27 | 浙江同善人工智能技术有限公司 | Road area identification method and system |
CN113610970A (en) * | 2021-08-30 | 2021-11-05 | 上海智能网联汽车技术中心有限公司 | Automatic driving system, device and method |
CN113792707A (en) * | 2021-11-10 | 2021-12-14 | 北京中科慧眼科技有限公司 | Terrain environment detection method and system based on binocular stereo camera and intelligent terminal |
CN114154510A (en) * | 2021-11-30 | 2022-03-08 | 江苏智能网联汽车创新中心有限公司 | Control method and device for automatic driving vehicle, electronic equipment and storage medium |
CN114154510B (en) * | 2021-11-30 | 2024-09-20 | 江苏智能网联汽车创新中心有限公司 | Control method and device for automatic driving vehicle, electronic equipment and storage medium |
CN115187959A (en) * | 2022-07-14 | 2022-10-14 | 清华大学 | Method and system for landing flying vehicle in mountainous region based on binocular vision |
CN115063594A (en) * | 2022-08-19 | 2022-09-16 | 清驰(济南)智能科技有限公司 | Feature extraction method and device based on automatic driving |
CN115063594B (en) * | 2022-08-19 | 2022-12-13 | 清驰(济南)智能科技有限公司 | Feature extraction method and device based on automatic driving |
CN115179920A (en) * | 2022-09-07 | 2022-10-14 | 北京中科慧眼科技有限公司 | Vehicle running system adjusting method and system based on cross-country scene |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110008848A (en) | A kind of travelable area recognizing method of the road based on binocular stereo vision | |
CN108985194B (en) | Intelligent vehicle travelable area identification method based on image semantic segmentation | |
Bachute et al. | Autonomous driving architectures: insights of machine learning and deep learning algorithms | |
CN105160309B (en) | Three lanes detection method based on morphological image segmentation and region growing | |
CN107368787B (en) | Traffic sign identification method for deep intelligent driving application | |
CN110009765A (en) | A kind of automatic driving vehicle contextual data system and scene format method for transformation | |
CN102708356B (en) | Automatic license plate positioning and recognition method based on complex background | |
CN110531753A (en) | Control system, control method and the controller of autonomous vehicle | |
CN101900562B (en) | Clear path detection using divide approach | |
CN110062871A (en) | Method and system for video-based positioning and mapping | |
CN113095152B (en) | Regression-based lane line detection method and system | |
CN108428254A (en) | The construction method and device of three-dimensional map | |
CN113895464B (en) | Intelligent vehicle driving map generation method and system fusing personalized driving style | |
CN113887376A (en) | Target detection method, device, medium and equipment | |
CN106529417A (en) | Visual and laser data integrated road detection method | |
Wang et al. | Autonomous Driving System Driven by Artificial Intelligence Perception Fusion | |
CN111931683A (en) | Image recognition method, image recognition device and computer-readable storage medium | |
CN118155183A (en) | Unstructured scene automatic driving network architecture method for deep multi-mode perception | |
CN106529391B (en) | A kind of speed limit road traffic sign detection of robust and recognition methods | |
CN115273005A (en) | Visual navigation vehicle environment perception method based on improved YOLO algorithm | |
Schuegraf et al. | Deep learning for the automatic division of building constructions into sections on remote sensing images | |
CN109543568A (en) | A kind of vehicle-logo location method | |
CN116258820B (en) | Large-scale urban point cloud data set and building individuation construction method and related device | |
Chamola et al. | Overtaking mechanisms based on augmented intelligence for autonomous driving: Datasets, methods, and challenges | |
US11796338B2 (en) | Automated semantic mapping |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190712 |