CN107985189A - Towards driver's lane change Deep Early Warning method under scorch environment - Google Patents
Towards driver's lane change Deep Early Warning method under scorch environment Download PDFInfo
- Publication number
- CN107985189A CN107985189A CN201711012609.0A CN201711012609A CN107985189A CN 107985189 A CN107985189 A CN 107985189A CN 201711012609 A CN201711012609 A CN 201711012609A CN 107985189 A CN107985189 A CN 107985189A
- Authority
- CN
- China
- Prior art keywords
- vehicle
- lane change
- driver
- early warning
- type
- 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
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B60—VEHICLES IN GENERAL
- B60Q—ARRANGEMENT OF SIGNALLING OR LIGHTING DEVICES, THE MOUNTING OR SUPPORTING THEREOF OR CIRCUITS THEREFOR, FOR VEHICLES IN GENERAL
- B60Q9/00—Arrangement or adaptation of signal devices not provided for in one of main groups B60Q1/00 - B60Q7/00, e.g. haptic signalling
- B60Q9/008—Arrangement or adaptation of signal devices not provided for in one of main groups B60Q1/00 - B60Q7/00, e.g. haptic signalling for anti-collision purposes
Landscapes
- Engineering & Computer Science (AREA)
- Human Computer Interaction (AREA)
- Mechanical Engineering (AREA)
- Image Analysis (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention discloses a kind of driver's lane change Deep Early Warning method under environment towards scorch, this vehicle rear side Image Acquisition is carried out using camera, this vehicle rear side image collected is passed in computer, by the deep learning network for establishing driver's lane change Deep Early Warning, complete to map end to end between camera inputs and export type of vehicle to this vehicle rear side vehicle on highway shooting image, complete identification and the mark identification frame of type of vehicle at the same time using the network, and identify the image of type of vehicle into the measuring and calculating of row distance output, then the operation conditions of vehicle is judged by vehicle ECU, if 0 100m detection ranges have identification type of vehicle, driver is given to give warning in advance, reach driver's lane change Deep Early Warning safety under scorch environment.Present invention reduces the error rate of identification, improves the accuracy of identification and realizes real-time monitoring.Realize to the real-time automatic identification of this vehicle rear side region vehicle, distance exam, early warning.
Description
Technical field
The invention belongs to field of automobile safety, the driver's lane change Deep Early Warning being related under a kind of environment towards scorch
Security system.
Background technology
Reduce traffic accident caused by loss into people increasingly pay close attention to the problem of.Automotive safety aids in driving technology
As the effective means for reducing traffic accident, reducing causality loss, oneself is subjected to the favor of national governments, and automotive safety auxiliary is driven
Sail the hot spot that technology is studied as scholars and Automobile Enterprises.It has been proposed that such as：Vehicle lane change video detection based on ARM
System, vehicle lane change monitoring system based on vehicle-mounted millimeter wave etc..But traditional system is mainly at a high speed based on elimination
The vision dead zone of driver under driving environment.Mainly it experienced following four developing stage：Install mirror on mirror (blind area roundlet mirror),
Ultrasonic Range Finder for Parking, 360 degree of full-view cameras of installation, installation higher-frequency radar etc. are installed.Traditional measure is paid attention to based on change
The acquisition of road dead zone information, does not carry out the analysis and processing of depth to the information of acquisition.The present invention uses deep learning net
Network substitutes driver and intelligent depth analysis and identification is carried out to the vehicle lane change information of blind area, to avoid driver fatigue and point
Erroneous judgement under refreshing state.
The content of the invention
The object of the present invention is to provide a kind of driver's lane change Deep Early Warning method under environment towards scorch, effectively
Reduce after driver attention under scorch environment do not concentrate, cannot find oversize vehicle, offside that side rear follows in time
The erroneous judgement of square vehicle following distance and caused by accident it is dead, severely injured, ensure safety of life and property.
The present invention is directed to and the moving object image in this vehicle rear side car door overlay area is answered based on image processing techniques
The problem of target detection discrimination is low caused by miscellaneous background.It is proposed driver's lane change depth under a kind of environment towards scorch
Early warning security system, it is characterized in that this vehicle rear side Image Acquisition is carried out using camera, this vehicle rear side figure that will be collected
As being passed in computer, by establish truck, car, dilly type identification driver's lane change Deep Early Warning depth
Learning network, completes camera and is arrived to being held between the input of this vehicle rear side vehicle on highway shooting image and output type of vehicle
The mapping at end, identification and the mark identification frame of type of vehicle are completed using the network at the same time, and identify type of vehicle to output
Image into row distance measuring and calculating, then by vehicle ECU judge vehicle operation conditions, if 0-100m detection ranges have identification
During type of vehicle, give driver and give warning in advance, reach driver's lane change Deep Early Warning safety under scorch environment.Such as figure
Shown in 1.
The concrete technical scheme of the present invention is as follows：A kind of driver's lane change Deep Early Warning peace under the environment towards scorch
Total system, it is characterized in that comprising the following steps：
The first, camera is installed：Camera is installed on the lower edge of rearview mirror, camera optical axis direction facing tailstock, with car
Angle between body longitudinal centre line is 3 °, and left and right rearview mirror respectively installs the folder of a camera, camera optical axis and horizontal plane
Angle is 4 °；
2nd, this vehicle rear side region is continuously shot using camera, obtains this vehicle rear side area image；
3rd, driver's lane change Deep Early Warning learning network model is established, to moving vehicle type identification under high velocity environment；Should
Process is divided into four parts：Establish vehicle type recognition under high velocity environment and train required external image storehouse；Build driver
The network model of lane change Deep Early Warning security system vehicle type recognition；The inhomogeneity collected with camera to vehicle side rear
The vehicle image of type optimizes network model parameter as training object；Vehicle type recognition network model parameter is instructed
Practice, achieve the purpose that different type vehicle identification；
4th, this vehicle rear side area image is inputted into driver's lane change Deep Early Warning learning network model, exports identification
Type of vehicle：Driver's lane change Deep Early Warning learning network model obtains optimal network model parameter by training, inputs again
This vehicle rear side area image, then by detecting the computing of bounding box+confidence level and class probability figure, exports the vehicle of identification
Type；
5th, to identifying the image of type of vehicle into the measuring and calculating of row distance；Pass through taking the photograph installed in vehicle mirrors lower edge
As head realizes the measuring and calculating of side front vehicle distance；
6th, the type of vehicle to output and the distance of measuring and calculating are analyzed, handled to realize driver's lane change Deep Early Warning：
The operation conditions of vehicle is judged by vehicle ECU, if it is determined that vehicle, which has lane change intention and recognizes this vehicle rear side, truck, visitor
When car, dilly enter 0-100m detection ranges, voice is prompted automatically：Lane change please notes safety；Conversely, then voice does not carry
Show.
Advantage of the present invention：The present invention is using driver's lane change Deep Early Warning learning network model to being transported under scorch environment
The identification of motor-car type, and the distance laterally with car is calculated, improve the accuracy of identification and realize real-time prison
Survey.Identification of driver's lane change Deep Early Warning learning network model to vehicle on highway type, improves traditional treatment method
Deficiency, can effectively overcome that traditional recognition methods discrimination is low, whole identification process spends the time long, image background more
Sample, complex background for the interference of target present situation.Towards driver's lane change Deep Early Warning safety under scorch environment
The exploitation of system, will realize to the real-time automatic identification of this vehicle rear side region vehicle, distance exam, early warning.The invention can
The error rate of identification is reduced, improves the speed of identification, helps to reduce driver attention under scorch environment and does not concentrate, no
Can find the erroneous judgement of oversize vehicle, offside front vehicle following distance that side rear follows in time and caused by accident it is dead, again
Wound, ensures safety of life and property.
Brief description of the drawings
Fig. 1 is work flow diagram of the present invention.
Fig. 2 is camera installation diagram.
Fig. 3 is driver's lane change Deep Early Warning learning network identification schematic diagram.
Fig. 4 is rear area target vehicle measuring and calculating distance model figure.
Fig. 5 is driver's lane change Deep Early Warning learning network structural model figure of the present invention.
Drawing reference numeral：Rearview mirror 1；Camera 2.
Embodiment
Below in conjunction with attached drawing to present invention further illustrate, but the present invention practical methods be not limited in it is following
Embodiment.Based on the embodiments of the present invention, those of ordinary skill in the art institute without making creative work
The every other embodiment obtained, belongs to the scope of protection of the invention.
A kind of driver's lane change Deep Early Warning method under environment towards scorch, comprises the following steps：
First, as shown in Fig. 2, installation camera.Camera 2 is installed on the lower edge of rearview mirror 1, camera optical axis direction court
It it is 3 ° to the angle between the tailstock, with vehicle body longitudinal centre line, left and right rearview mirror respectively installs a camera, installs such as Fig. 2 institutes
Show.The angle of camera optical axis and horizontal plane is 4 ° (pitch angle), as shown in Figure 4.
2nd, as shown in Figure 1, driver's lane change Deep Early Warning learning network model is established, to moving vehicle under high velocity environment
Type identification.The process is broadly divided into four parts：Establish the required external image storehouse of vehicle recognition training under high velocity environment；
Build the network model of driver's lane change Deep Early Warning security system vehicle cab recognition；Vehicle side rear is collected with camera
Different types of vehicle image optimizes network model parameter as training object；To vehicle cab recognition network model parameter
Training, reaches the identification of different type vehicle.
Four part idiographic flows are as follows：
1st, the required external image storehouse of vehicle recognition training under high velocity environment is established
Under scorch environment, this vehicle rear side region is carried out by the camera that rearview mirror lower edge is installed continuous
The image set obtained is shot, in this, as the training set and test set in external image storehouse.In training set, truck is divided into light-duty
Truck (1.8 tons of ＜ gross mass≤6 ton), Medium Truck (6.0 tons of ＜ gross mass≤14 ton), heavy truck (gross mass ＞ 14
Ton), car is divided into middle bus (6m<Length of wagon≤9m), motorbus (length of wagon>9m), car, MPV, SUV are returned
Class is dilly, i.e., training set is always divided into 6 classes, that is to say, that has 6 type of vehicle.
2nd, the network model of driver's lane change Deep Early Warning security system vehicle cab recognition is built
Driver's lane change Deep Early Warning learning network is made of convolutional layer, pond layer and full articulamentum.Possess 24 volumes
Lamination and 2 full articulamentums, using 1 × 1 convolutional layer and 3 × 3 convolutional layers, wherein largely having used the cascade structure of convolution.It is logical
Cross camera and this vehicle rear side region different automobile types are carried out with Image Acquisition, it is then that vehicle image is down-sampled using size as 448
× 448 as input.The initial convolutional layer of network is completed to extract the function of feature from image, last full articulamentum prediction class
Probability and detection bounding box, using shown in leaky ReLU activation primitives such as formula (1) on activation primitive：
In above formula：φ is activation primitive；X be convolution kernel interior joint value and weight dot product and bias term parameter and
Computing.
In order to avoid over-fitting, Web vector graphic dropout structures.
As shown in figure 5, with reference to driver's lane change Deep Early Warning learning network structural model figure, the process of its work is carried out
Introduce in detail, process is as follows：
(1) the first hidden layer of driver's lane change Deep Early Warning learning network：The vehicle side rear area that camera collects
Image adds down-sampling layer using linear convolution layer.Using highway of the convolution kernel of 64 7 × 7 sizes after down-sampled to input
Side behind vehicle image does convolution operation, and step-length 2, after convolution operation, can generate 64 new reflection vehicle classes
The low-dimensional information of type, and size is reduced to the half of original image, i.e., 224 × 224.Followed by doing down-sampling operation, under adopt
Sample is dimensioned so as to 2 × 2, and step-length is also 2, and the low-dimensional information of the reflection type of vehicle exported after down-sampling operation remains as 64
It is a, but the size dimension of each reduces half again than the size of the vehicle image of last layer.Output be changed into 64 112 ×
The low-dimensional vehicle characteristics figure of 112 sizes.
(2) the second hidden layer of driver's lane change Deep Early Warning learning network：Upper strata is produced using the convolution kernels of 192 3 × 3
Low-dimensional vehicle characteristics figure carry out convolution, wherein step-length be 1, the size of the single unit vehicle feature of output does not change.Do
Down-sampling operation is sized to 2 × 2 size, step-length 2, and the size of vehicle characteristics figure is changed into original half, exports
It is changed into the feature of 192 56 × 56 sizes.
(3) the 3rd hidden layer of driver's lane change Deep Early Warning learning network：Upper strata is exported using the convolution kernels of 128 1 × 1
The information of vehicle carries out convolution, and wherein step-length is 1, so the size of output does not change.It is main using 1 × 1 convolution kernel
If for dimensionality reduction.Processing by step-length for 11 × 1 convolution kernel, exports the comparison higher-dimension of 128 56 × 56 sizes
Information of vehicles feature.Then it is identical with the operating principle of previous step, the layer using 256 3 × 3 sizes convolution kernel deconvolute on
The output of layer, when the information of vehicles inputs of 128 56 × 56 are come in, by convolution operation, generates the car of 128 56 × 56 sizes
Information.Then the processing by step-length for 11 × 1 convolution kernel again, output description vehicle characteristics are changed into 256 56 × 56
The feature of size, then the processing by step-length for 13 × 3 convolution kernel, output are changed into the spy of 512 56 × 56 sizes again
Sign, then by step-length be 2, size is 2 × 2 maximum down-sampling layer, and finally output is changed into the height of 512 28 × 28 sizes
In first two layers of feature.
(4) the 4th hidden layer of driver's lane change Deep Early Warning learning network：It is finally that 1024 step-lengths are that 1 size is 3 × 3 big
Small convolution kernel, the step-length of first 9 times is 1 in this layer, therefore the new feature size for exporting description vehicle is 28 × 28, last convolution
The number of core is still 1024.Operated by the maximum pond down-sampling that step-length is 2 × 2, this operation output is 1024 14
The hum pattern of × 14 sizes.
(5) the 5th hidden layer of driver's lane change Deep Early Warning learning network：It is similar with the 4th layer of operation.Output description vehicle
The high feature sizes of information are 7 × 7, and the number of convolution kernel is 1024.
(6) the 6th hidden layer of driver's lane change Deep Early Warning learning network：Using the convolution kernel of two groups of 1024 3 × 3 sizes
Convolution operation is carried out, has just obtained more high dimensional feature of the output for 1024 7 × 7 sizes.
(7) the full articulamentum of driver's lane change Deep Early Warning learning network：Namely driver's lane change Deep Early Warning study net
The layer 7 of network structure and the 8th layer.Full articulamentum is that the feature for the description vehicle diverse location for being extracted preceding networks is whole
It is arranged together.Driver's lane change Deep Early Warning learning network is a regression problem, by the different type vehicle pictures of identification
Small grid is divided into, and predicts frame, confidence level and class probability.The prediction of output is represented as shown in formula (2)：
T=S × S × (B × 5+C) (2)
In above formula, T is the number of tensor；S is the size of input picture grid；B is the number of detection bounding box；C is every
The species of a small grid prediction vehicle.For the parameter used for S=7, B=2, C=6, final network exports 784 tensors to 7 × 7
A grid is predicted, the probability and coordinate of each 6 class vehicle of grid forecasting.
(8) driver's lane change Deep Early Warning learning network output layer：Driver's lane change Deep Early Warning learning network is predicted
Different automobile types and the bounding box that has screened of network shown on picture.
3rd, using the different types of vehicle image that camera collects vehicle side rear as training object, network is joined
Number optimizes.
In driver's lane change Deep Early Warning learning network identifies schematic diagram, as shown in figure 3, input picture is divided into S
× S grid, it is assumed that the center of some vehicle falls in some grid, then corresponding grid is responsible for detecting the vehicle.It is each small
B detection bounding box (Detected Bounding Box) of grid forecasting and its corresponding confidence score.If in small grid not
Containing information of vehicles, then confidence level is 0, otherwise wishes that confidence score is equal to and predicts detection bounding box and normative reference frame
The IOU of (Ground Truth Box).Shown in confidence score such as formula (3):
In above formula, Score is confidence score, and Pr (object) is the probability that bounding box includes target vehicle；BBgtFor base
In the normative reference frame of training label；BBdtTo detect bounding box；Area () represents area.
Each detection bounding box includes 5 predicted values：X, y, w, h and confidence level.(x, y) represents the center of detection bounding box
With respect to the position of its female grid, w and h represent the wide and high of detection bounding box, confidence level represent the detection bounding box that predicts with
The IOU of any normative reference frame.Conditional probability Pr (the Class of each small grid prediction C kind vehiclesi| object), the probability tables
Show that the i-th class vehicle center falls into the probability of the grid.This probability rely in small grid whether the probability containing vehicle.No matter
How many is a by the number B of frame, and network only predicts the species of vehicle for each small grid.
The confidence score of each bounding box is calculated by formula (5), these scores represent that class of vehicle is appeared in frame
Probability, while also illustrate that detection bounding boxes of these predictions are adapted to the degree of vehicle.
In above formula, Pr (classi| object) it is the conditional probability that each small grid predicts C kind vehicles.
As shown in Figure 3, driver's lane change Deep Early Warning learning network identification schematic diagram is understood, 49 grid each grids
Predict the probability of 6 class type of vehicle, then piece image produces the probability of 294 predictions, and most of image-region is
There is no vehicle, this prediction probability that may result in most of region is 0, causes training process to dissipate, introducing a variable can
To solve the problems, such as this well：I.e. whether certain position has probability P r (object) existing for vehicle.There is Pr (object) to add again
On assume to have vehicle in certain region in the case of predict that the vehicle belongs to the probability of which kind of vehicle, formula is such as shown in (6).
Pr (class)=Pr (object) × Pr (class | object) (6)
In above formula, Pr (class) includes target carriage per a kind of unconditional probability, Pr (object) for certain position for bounding box
Probability, Pr (class | object) is conditional probability.
Pr (class) is updated in each position, and only in the presence of having vehicle just to Pr (Vehicle |
Object) update.
The design object of loss function is exactly to allow coordinate (x, y, w, h), confidence level, these three aspects of classifying reach good
Balance, the loss function of driver's lane change Deep Early Warning learning network use the wide high square root of frame, and formula is such as shown in (7)：
In above formula, λcoordFor the weight to position error bigger, λnoobjFor the confidence to the grid not comprising object center
Spend the weight of error smaller, S2For the size of grid,Represent that the jth vehicle frame No. i-th band of position is responsible for this
The identification mission in region,Represent to assume that target vehicle is appeared in No. i-th band of position.X and y difference representative images are worked as
Front position, the width and height of w and h representative pictures.C is total classification number of the vehicle of required identification, and p (c) represents the car
Belong to the probability of c class vehicles.
4th, the mesh of different type vehicle identification is reached by the training to driver's lane change Deep Early Warning learning network parameter
's
To reduce the gap between the output of vehicle cab recognition and type of vehicle, using small lot gradient descent method (mini-
Batch gradient descent scheme) and momentum (momentum) network parameter is trained, momentum has training process
The advantages of faster restraining.Using the derivative of driver's lane change Deep Early Warning learning network loss function, with back propagation
(Back-Propagation) parameter is constantly updated, reduces loss function value until convergence.Renewal rule is as follows：
ω(l)(t)=ω(l)(t)-Mω(l)(t) (9)
In above formula, Mω(l)(t) the parameter ω for being l layers in the t times iteration(l)Momentum；μ is momentum rate；α is study
Rate；λ is weight attenuation (weight decay term).In formula (8), the momentum of preceding an iteration be used to calculate current
The momentum of iteration.It can avoid being absorbed in local minimum and more rapid convergence by the skill.
3rd, this vehicle rear side area image is inputted into driver's lane change Deep Early Warning learning network model, exports identification
Vehicle.
Understand that driver's lane change Deep Early Warning learning network model obtains optimal net by training by above-mentioned formula (8), (9)
Network model parameter, inputs this vehicle rear side area image again, you can by by image using the grid of S × S as input, then
By detecting the computing of bounding box+confidence level and class probability figure, the final detection of different type vehicle under high velocity environment is exported
Classification.The process of its work is as shown in Figure 3.
4th, identify the image of type of vehicle into the measuring and calculating of row distance output.
By the measuring and calculating that side front vehicle distance is realized installed in the camera of vehicle mirrors lower edge.Calculate distance
Schematic diagram is as shown in Figure 4.By the similar property of projection relational expression and triangle, subpoint p can be calculated1',p'2To taking the photograph
As head vertical lens project to the horizontal distance of ground point.Calculation formula is such as shown in (10), (11)：
In above formula, θ is the pitch angle of video camera；F is the focal length of video camera；H is cam lens upright projection to ground
The vertical height of point；(x0,y0) be the plane of delineation origin, vehicle subpoint p on the ground1,p2With in image sit (x,
Y) point in marking is corresponding.
Shown in transformational relation such as formula (11), (12) between image coordinate and pixel coordinate:
In above formula, (u, v) be pixel coordinate system under coordinate, (u0,v0) be image pixel center coordinate.
Substitute the above in formula (10), (11).It can be obtained by the calibration of intrinsic parameters of the camera, under normal conditions
The pitch angle of video camera is smaller, and the intersection point p on camera optical axis and ground can be approximately thought beyond normal field range,
Shown in the measuring and calculating range formula such as formula (14) of actual use:
F in above formulai=f/dy is the effective focal length in vertical direction.
Based on above formula (14), the pixel coordinate point of target vehicle projection to be measured on the ground, this car side in image are obtained
The distance of vehicle distances this car camera of rear region identification can be calculated.
5th, the identification and distance that export sports model analyzed, handled to realize driver's lane change Deep Early Warning.
Based on it is above-mentioned in real time identify, measuring and calculating as a result, then passing through vehicle ECU (Electronic Control Unit)
The operation conditions of vehicle is judged, if it is determined that vehicle, which has lane change intention and recognizes this vehicle rear side, truck, car, dilly
During into 0-100m detection ranges, voice is prompted " lane change please notes safety " automatically；Conversely, then voice is not prompted.
Above-mentioned embodiment is used for explaining the present invention, rather than limits the invention, in spirit of the invention and
In scope of the claims, to any modifications and changes of the invention made, protection scope of the present invention is both fallen within.
Claims (3)
1. a kind of driver's lane change Deep Early Warning method under environment towards scorch, it is characterized in that carrying out this using camera
Vehicle rear side Image Acquisition, this vehicle rear side image collected is passed in computer, by establishing truck, car, small
The deep learning network of driver's lane change Deep Early Warning of type vehicle type recognition, it is public at a high speed to this vehicle rear side to complete camera
Road vehicles shooting image is inputted between output type of vehicle to be mapped end to end, and type of vehicle is completed at the same time using the network
Identification and mark identification frame, and then the image for identifying type of vehicle to output is sentenced into the measuring and calculating of row distance by vehicle ECU
Determine the operation conditions of vehicle, if 0-100m detection ranges have identification type of vehicle, give driver and give warning in advance, reach at a high speed
Driver's lane change Deep Early Warning safety under driving environment.
2. driver's lane change Deep Early Warning method under a kind of environment towards scorch as claimed in claim, it is characterized in that
Comprise the following steps：
Step 1: installation camera：Camera (2) is installed on the lower edge of rearview mirror (1), and camera optical axis direction is towards car
Angle between tail, with vehicle body longitudinal centre line is 3 °, and left and right rearview mirror respectively installs a camera, camera optical axis and water
The angle of plane is 4 °；
Step 2: being continuously shot using camera (2) to this vehicle rear side region, this vehicle rear side area image is obtained；
Step 3: driver's lane change Deep Early Warning learning network model is established, to moving vehicle type identification under high velocity environment；Should
Process is divided into four parts：Establish vehicle type recognition under high velocity environment and train required external image storehouse；Build driver
The network model of lane change Deep Early Warning security system vehicle type recognition；The inhomogeneity collected with camera to vehicle side rear
The vehicle image of type optimizes network model parameter as training object；Vehicle type recognition network model parameter is instructed
Practice, achieve the purpose that different type vehicle identification；
Step 4: this vehicle rear side area image is inputted driver's lane change Deep Early Warning learning network model, identification is exported
Type of vehicle：Driver's lane change Deep Early Warning learning network model obtains optimal network model parameter by training, inputs again
This vehicle rear side area image, then by detecting the computing of bounding box+confidence level and class probability figure, exports the vehicle of identification
Type；
Step 5: to identifying the image of type of vehicle into the measuring and calculating of row distance；Pass through taking the photograph installed in vehicle mirrors lower edge
As head realizes the measuring and calculating of side front vehicle distance；
Step 6: the distance of the type of vehicle and measuring and calculating to output is analyzed, handled to realize driver's lane change Deep Early Warning：
The operation conditions of vehicle is judged by vehicle ECU, if it is determined that vehicle, which has lane change intention and recognizes this vehicle rear side, truck, visitor
When car, dilly enter 0-100m detection ranges, voice is prompted automatically：Lane change please notes safety；Conversely, then voice does not carry
Show.
3. as claimed in claim 2 towards driver's lane change Deep Early Warning method under scorch environment, it is characterized in that step
Four part specific methods are as follows in rapid three：
(1), establish vehicle type recognition under high velocity environment and train required external image storehouse
Under scorch environment, this vehicle rear side region is continuously shot by the camera that rearview mirror lower edge is installed
The image set of acquisition, in this, as the training set and test set in external image storehouse, training set is always divided into 6 class type of vehicle：Gently
Type truck, Medium Truck, heavy truck, middle bus, motorbus, dilly；
(2), the network model of driver's lane change Deep Early Warning security system vehicle cab recognition is built
Driver's lane change Deep Early Warning learning network is made of convolutional layer, pond layer and full articulamentum, possesses 24 convolutional layers
With 2 full articulamentums, using 1 × 1 convolutional layer and 3 × 3 convolutional layers, by camera to this vehicle rear side region different automobile types into
Row Image Acquisition, then inputs down-sampled be used as using size as 448 × 448 of vehicle image, the initial convolutional layer of network is completed
The function of feature, last full articulamentum prediction class probability and detection bounding box are extracted from image, is made on activation primitive
With leaky ReLU activation primitives, Web vector graphic dropout structures；
(3), using the different types of vehicle image that camera collects vehicle side rear as training object, to network parameter
Optimize
Input picture is divided into S × S grid, it is assumed that the center of some vehicle falls in some grid, then corresponding grid
It is responsible for detecting the vehicle.Each small grid predicts B detection bounding box confidence score corresponding with its, if being free of in small grid
There is information of vehicles, then confidence level is 0, otherwise wishes that confidence score is equal to and predicts detection bounding box and normative reference frame
IOU；
(4), by the training to driver's lane change Deep Early Warning learning network parameter, different vehicle type identification is achieveed the purpose that
Network parameter is trained using small lot gradient descent method and momentum, is damaged using driver's lane change Deep Early Warning learning network
The derivative of function is lost, parameter is constantly updated with back propagation, reduces loss function value until convergence.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711012609.0A CN107985189B (en) | 2017-10-26 | 2017-10-26 | Early warning method for lane changing depth of driver in high-speed driving environment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711012609.0A CN107985189B (en) | 2017-10-26 | 2017-10-26 | Early warning method for lane changing depth of driver in high-speed driving environment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107985189A true CN107985189A (en) | 2018-05-04 |
| CN107985189B CN107985189B (en) | 2021-03-12 |
Family
ID=62031216
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711012609.0A Active CN107985189B (en) | 2017-10-26 | 2017-10-26 | Early warning method for lane changing depth of driver in high-speed driving environment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107985189B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108765826A (en) * | 2018-05-31 | 2018-11-06 | 四川物联亿达科技有限公司 | A kind of based on deep learning block fuzzy scene human figure identification method and system more |
| CN109934108A (en) * | 2019-01-31 | 2019-06-25 | 华南师范大学 | The vehicle detection and range-measurement system and implementation method of a kind of multiple target multiple types |
| CN109949611A (en) * | 2019-03-28 | 2019-06-28 | 百度在线网络技术（北京）有限公司 | The lane change method, apparatus and storage medium of unmanned vehicle |
| CN110281923A (en) * | 2019-06-28 | 2019-09-27 | 信利光电股份有限公司 | A kind of vehicle auxiliary lane change method, apparatus and system |
| CN110562258A (en) * | 2019-09-30 | 2019-12-13 | 驭势科技（北京）有限公司 | Method for vehicle automatic lane change decision, vehicle-mounted equipment and storage medium |
| CN111539243A (en) * | 2019-01-23 | 2020-08-14 | 通用汽车环球科技运作有限责任公司 | Automatic data collection for continuous improvement in the detection of an object of interest |
| CN111666876A (en) * | 2020-06-05 | 2020-09-15 | 北京百度网讯科技有限公司 | Method and device for detecting obstacle, electronic equipment and road side equipment |
| CN112785564A (en) * | 2021-01-15 | 2021-05-11 | 武汉纺织大学 | Pedestrian detection tracking system and method based on mechanical arm |
| CN114715038A (en) * | 2022-03-16 | 2022-07-08 | 深圳市德驰微视技术有限公司 | Domain controller-based blind area monitoring image post-processing method and system |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104442554A (en) * | 2014-10-24 | 2015-03-25 | 中国人民解放军理工大学 | Automobile dead zone detection and safety travelling method and system |
| US9418319B2 (en) * | 2014-11-21 | 2016-08-16 | Adobe Systems Incorporated | Object detection using cascaded convolutional neural networks |
| CN105975915A (en) * | 2016-04-28 | 2016-09-28 | 大连理工大学 | Front vehicle parameter identification method based on multitask convolution nerve network |
| CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
| CN107031661A (en) * | 2017-03-16 | 2017-08-11 | 浙江零跑科技有限公司 | A kind of lane change method for early warning and system based on blind area camera input |
| CN107229906A (en) * | 2017-05-08 | 2017-10-03 | 上海工程技术大学 | A kind of automobile overtaking's method for early warning based on units of variance model algorithm |
-
2017
- 2017-10-26 CN CN201711012609.0A patent/CN107985189B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104442554A (en) * | 2014-10-24 | 2015-03-25 | 中国人民解放军理工大学 | Automobile dead zone detection and safety travelling method and system |
| US9418319B2 (en) * | 2014-11-21 | 2016-08-16 | Adobe Systems Incorporated | Object detection using cascaded convolutional neural networks |
| CN105975915A (en) * | 2016-04-28 | 2016-09-28 | 大连理工大学 | Front vehicle parameter identification method based on multitask convolution nerve network |
| CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
| CN107031661A (en) * | 2017-03-16 | 2017-08-11 | 浙江零跑科技有限公司 | A kind of lane change method for early warning and system based on blind area camera input |
| CN107229906A (en) * | 2017-05-08 | 2017-10-03 | 上海工程技术大学 | A kind of automobile overtaking's method for early warning based on units of variance model algorithm |
Non-Patent Citations (3)
| Title |
|---|
| 王华莹: "《基于机器视觉的车辆检测和车距测量方法研究》", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
| 王宇宁等: "《基于YOLO算法的车辆实时检测》", 《武汉理工大学学报》 * |
| 邓柳: "《基于深度卷积神经网络的车型识别》", 《中国优秀硕士学位论文全文数据库工程科技II辑》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108765826A (en) * | 2018-05-31 | 2018-11-06 | 四川物联亿达科技有限公司 | A kind of based on deep learning block fuzzy scene human figure identification method and system more |
| CN111539243A (en) * | 2019-01-23 | 2020-08-14 | 通用汽车环球科技运作有限责任公司 | Automatic data collection for continuous improvement in the detection of an object of interest |
| CN111539243B (en) * | 2019-01-23 | 2023-12-22 | 通用汽车环球科技运作有限责任公司 | Continuous improved automatic data collection in detection of objects of interest |
| CN109934108A (en) * | 2019-01-31 | 2019-06-25 | 华南师范大学 | The vehicle detection and range-measurement system and implementation method of a kind of multiple target multiple types |
| CN109934108B (en) * | 2019-01-31 | 2023-01-31 | 华南师范大学 | Multi-target and multi-type vehicle detection and distance measurement system and implementation method |
| US11409288B2 (en) | 2019-03-28 | 2022-08-09 | Apollo Intelligent Driving Technology (Beijing) Co., Ltd. | Lane changing method, device for driverless vehicle and computer-readable storage medium |
| CN109949611A (en) * | 2019-03-28 | 2019-06-28 | 百度在线网络技术（北京）有限公司 | The lane change method, apparatus and storage medium of unmanned vehicle |
| CN109949611B (en) * | 2019-03-28 | 2021-11-30 | 阿波罗智能技术(北京)有限公司 | Lane changing method and device for unmanned vehicle and storage medium |
| CN110281923A (en) * | 2019-06-28 | 2019-09-27 | 信利光电股份有限公司 | A kind of vehicle auxiliary lane change method, apparatus and system |
| CN110562258A (en) * | 2019-09-30 | 2019-12-13 | 驭势科技（北京）有限公司 | Method for vehicle automatic lane change decision, vehicle-mounted equipment and storage medium |
| CN111666876A (en) * | 2020-06-05 | 2020-09-15 | 北京百度网讯科技有限公司 | Method and device for detecting obstacle, electronic equipment and road side equipment |
| CN111666876B (en) * | 2020-06-05 | 2023-06-09 | 阿波罗智联(北京)科技有限公司 | Method and device for detecting obstacle, electronic equipment and road side equipment |
| CN112785564A (en) * | 2021-01-15 | 2021-05-11 | 武汉纺织大学 | Pedestrian detection tracking system and method based on mechanical arm |
| CN114715038A (en) * | 2022-03-16 | 2022-07-08 | 深圳市德驰微视技术有限公司 | Domain controller-based blind area monitoring image post-processing method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107985189B (en) | 2021-03-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107985189A (en) | Towards driver's lane change Deep Early Warning method under scorch environment | |
| Ni et al. | An improved deep network-based scene classification method for self-driving cars | |
| CN108983219A (en) | A kind of image information of traffic scene and the fusion method and system of radar information | |
| CN111695448B (en) | Roadside vehicle identification method based on visual sensor | |
| CN112633176B (en) | Rail transit obstacle detection method based on deep learning | |
| CN108909624A (en) | A kind of real-time detection of obstacles and localization method based on monocular vision | |
| CN107031623A (en) | A kind of road method for early warning based on vehicle-mounted blind area camera | |
| CN106326893A (en) | Vehicle color recognition method based on area discrimination | |
| CN111553201A (en) | Traffic light detection method based on YOLOv3 optimization algorithm | |
| CN105574552A (en) | Vehicle ranging and collision early warning method based on monocular vision | |
| CN111738037A (en) | Automatic driving method and system and vehicle | |
| CN105260712A (en) | Method and system for detecting pedestrian in front of vehicle | |
| CN107031661A (en) | A kind of lane change method for early warning and system based on blind area camera input | |
| CN106446834A (en) | Vehicle type identification method and vehicle type identification device based on images | |
| CN106056102A (en) | Video-image-analysis-based road vehicle type classification method | |
| CN107796373A (en) | A kind of distance-finding method of the front vehicles monocular vision based on track plane geometry model-driven | |
| CN117975436A (en) | Three-dimensional target detection method based on multi-mode fusion and deformable attention | |
| CN111523415A (en) | Image-based two-passenger one-dangerous vehicle detection method and device | |
| CN106845359A (en) | Tunnel portal driving prompt apparatus and method based on infrared emission | |
| CN114926984B (en) | Real-time traffic conflict collection and road safety evaluation method | |
| CN110307995A (en) | A kind of test method for autonomous driving vehicle marginal test | |
| Tahir et al. | Object detection in autonomous vehicles under adverse weather: a review of traditional and deep learning approaches | |
| CN117334040A (en) | Cross-domain road side perception multi-vehicle association method and system | |
| Liu et al. | Research on security of key algorithms in intelligent driving system | |
| CN114120246B (en) | Front vehicle detection algorithm based on complex environment |
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 |