CN110263628A - Obstacle detection method, device, electronic equipment and storage medium - Google Patents
Obstacle detection method, device, electronic equipment and storage medium Download PDFInfo
- Publication number
- CN110263628A CN110263628A CN201910385611.5A CN201910385611A CN110263628A CN 110263628 A CN110263628 A CN 110263628A CN 201910385611 A CN201910385611 A CN 201910385611A CN 110263628 A CN110263628 A CN 110263628A
- Authority
- CN
- China
- Prior art keywords
- filter
- layer
- neural networks
- convolutional neural
- importance
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of obstacle detection method, device, electronic equipment and storage medium, and method includes: the image for obtaining the direction to be measured of vehicle;Image is inputted into convolutional neural networks model, and obtains the recognition result of convolutional neural networks model output, re -training was established after convolutional neural networks model deletes redundant filter, and redundant filter is the filter that whole importance is lower than threshold value;Recognition result is used to indicate the direction to be measured of vehicle, and whether there are obstacles.Obstacle detection method provided by the invention, amount of storage and calculation amount needed for electronic equipment when reducing detection barrier.
Description
Technical field
The present invention relates to unmanned field more particularly to a kind of obstacle detection method, device, electronic equipment and deposit
Storage media.
Background technique
Under unmanned field, road scene based on computer vision barrier (including pedestrian, non-motor vehicle,
Motor vehicle etc.) detection be a most basic task, it provides necessary information for the driving strategy of vehicle.Its core is logical
The detection of obstacles program sent an image in electronic equipment is crossed, so that the detection of obstacles program just can recognize that image
In whether there are obstacles.
In the prior art, the highest obstacle detection method of accuracy rate all uses depth convolutional neural networks model.
Depth convolutional neural networks model is made of multiple convolutional layers and full articulamentum, is the most important deep learning for image recognition
One of frame.
However, the existing obstacle detection method using depth convolutional neural networks, due to being deposited in convolutional neural networks
In bulk redundancy parameter and unnecessary calculating, cause during detection of obstacles, when detecting barrier needed for electronic equipment
Amount of storage and calculation amount it is larger.
Summary of the invention
The present invention provides a kind of obstacle detection method, device, electronic equipment and storage medium, to solve the prior art
Amount of storage needed for electronic equipment and calculation amount larger problem when middle detection barrier.
The first aspect of the invention provides a kind of obstacle detection method, comprising:
Obtain the image in the direction to be measured of vehicle;
Described image is inputted into convolutional neural networks model, and obtains the identification knot of the convolutional neural networks model output
Fruit, re -training was established after the convolutional neural networks model deletes redundant filter, and the redundant filter is whole
Body importance is lower than the filter of threshold value;The recognition result is used to indicate the direction to be measured of the vehicle with the presence or absence of obstacle
Object.
Optionally, before the image in the direction to be measured for obtaining vehicle, further includes:
By pruning algorithms model, the filter of redundancy in the convolutional neural networks model is deleted.
Optionally, after the recognition result for obtaining the convolutional neural networks model output, further includes:
According to the recognition result, the driving behavior of the vehicle is controlled.
Optionally, the pruning algorithms model, comprising: similarity determines that layer, mean value determine that layer, local importance determine
Layer, regular terms determine that layer, whole importance determine layer, judge layer and delete layer;
The similarity determines layer, for determining first filter and first mistake in the convolutional neural networks model
Linear similarity between the same layer filter of filter;
The mean value determines layer, for determining the maximum similarity of the first filter according to the linear similarity
Mean value;
The part importance determines layer, for the maximum similarity mean value according to the first filter, to described the
One filter carries out max normalized, determines the local importance of the first filter;
The regular terms determines layer, for the corresponding regular terms of layer where determining the first filter;
The entirety importance determines layer, for determining described first according to the regular terms and the local importance
The whole importance of filter;
The judgement layer, for determining the first filter according to the whole importance and preset deletion ratio
For redundant filter;
The deletion layer, for deleting the redundant filter in the convolutional neural networks model.
Optionally, if the first filter is located at the full articulamentum of the convolutional neural networks model, the similarity
It determines layer, is specifically used for using formula It calculates between the first filter and the same layer filter of the first filter
Linear similarity;
Wherein, sim is similarity function, and corr is linear dependence function, Xm1For the m1 input node, Xm2It is
M2 input node,For Xm1Corresponding weight,For Xm2Corresponding weight, μ are mean value, and σ is standard deviation, and E is scheduled to last
Prestige value.
Optionally, if the first filter is located at the convolutional layer of the convolutional neural networks model, the similarity is true
Given layer is specifically used for using formula
Calculate the first filter and institute
State the similarity between the filter of first filter same layer;
Wherein, sim is similarity function, and corr is linear dependence function, Xm1For the m1 input node, Xm2It is
M2 input node,For Xm1The weight being connect with each node,For Xm2The power being connect with each node
Weight, K are convolution kernel size.
Optionally, the local importance determines layer, is specifically used for using formula
Described in n ∈ [1, M], n ≠ m are calculated
The local importance of first filter;
Wherein, Imp is local importance, and sim is similarity function, and k is adjustable parameter, and TopK is to take maximum k number value
Function.
Optionally, the regular terms determines layer, is specifically used for using formula
U ∈ [1, L] calculates described first
The corresponding regular terms of layer where filter;
Wherein, ReglFor the regular terms of l layers of filter, β and γ are adjustable parameter, ClFor l layers of calculation amount, Cu
For u layers of calculation amount, SlFor l layer parameter amount, SuFor u layer parameter amount, L is the number of plies of neural network.
Optionally, the whole importance determines layer, is specifically used for using formula Calculate the whole importance of the first filter;
Wherein, ReImp is whole importance,For l layers of m-th of node, Imp is local importance, ReglIt is
The regular terms of l layers of filter.
The second aspect of the invention provides a kind of obstacle detector, comprising:
Obtain module, the image in the direction to be measured for obtaining vehicle;
Processing module for described image to be inputted convolutional neural networks model, and obtains the convolutional neural networks mould
The recognition result of type output, re -training was established after the convolutional neural networks model deletes redundant filter, described
Redundant filter is the filter that whole importance is lower than threshold value;The recognition result is used to indicate the direction to be measured of the vehicle
Whether there are obstacles.
Optionally, further includes:
Pruning module, for deleting the filter of redundancy in the convolutional neural networks model by pruning algorithms model.
Optionally, further includes:
Module is driven, for controlling the driving behavior of the vehicle according to the recognition result.
Optionally, the pruning algorithms model, comprising: similarity determines that layer, mean value determine that layer, local importance determine
Layer, regular terms determine that layer, whole importance determine layer, judge layer and delete layer;
The similarity determines layer, for determining first filter and first mistake in the convolutional neural networks model
Linear similarity between the same layer filter of filter;
The mean value determines layer, for determining the maximum similarity of the first filter according to the linear similarity
Mean value;
The part importance determines layer, for the maximum similarity mean value according to the first filter, to described the
One filter carries out max normalized, determines the local importance of the first filter;
The regular terms determines layer, for the corresponding regular terms of layer where determining the first filter;
The entirety importance determines layer, for determining described first according to the regular terms and the local importance
The whole importance of filter;
The judgement layer, for determining the first filter according to the whole importance and preset deletion ratio
For redundant filter;
The deletion layer, for deleting the redundant filter in the convolutional neural networks model.
Optionally, if the first filter is located at the full articulamentum of the convolutional neural networks model, the similarity
It determines layer, is specifically used for using formula It calculates between the first filter and the same layer filter of the first filter
Linear similarity;
Wherein, sim is similarity function, and corr is linear dependence function, Xm1For the m1 input node, Xm2It is
M2 input node,For Xm1Corresponding weight,For Xm2Corresponding weight, μ are mean value, and σ is standard deviation, and E is scheduled to last
Prestige value.
Optionally, if the first filter is located at the convolutional layer of the convolutional neural networks model, the similarity is true
Given layer is specifically used for using formula
Calculate the first filter and institute
State the similarity between the filter of first filter same layer;
Wherein, sim is similarity function, and corr is linear dependence function, Xm1For the m1 input node, Xm2It is
M2 input node,For Xm1The weight being connect with each node,For Xm2The weight being connect with each node,
K is convolution kernel size.
Optionally, the local importance determines layer, is specifically used for using formula
Described in n ∈ [1, M], n ≠ m are calculated
The local importance of first filter;
Wherein, Imp is local importance, and sim is similarity function, and k is adjustable parameter, and TopK is to take maximum k number value
Function.
Optionally, the regular terms determines layer, is specifically used for using formula
U ∈ [1, L] calculates described first
The corresponding regular terms of layer where filter;
Wherein, ReglFor the regular terms of l layers of filter, β and γ are adjustable parameter, ClFor l layers of calculation amount, Cu
For u layers of calculation amount, SlFor l layer parameter amount, SuFor u layer parameter amount, L is the number of plies of neural network.
Optionally, the whole importance determines layer, is specifically used for using formula Calculate the whole importance of the first filter;
Wherein, ReImp is whole importance,For l layers of m-th of node, Imp is local importance, ReglIt is
The regular terms of l layers of filter.
The third aspect of the invention provides a kind of electronic equipment, comprising:
Memory, for storing program instruction;
Processor executes the step of method described in first aspect for calling and executing the program instruction in the memory
Suddenly.
The fourth aspect of the invention provides a kind of storage medium, and computer journey is stored in the readable storage medium storing program for executing
Sequence, the computer program is for executing the described in any item methods of first aspect.
Obstacle detection method, device, electronic equipment and storage medium provided by the invention, by obtain vehicle to
Image is inputted convolutional neural networks model by the image for surveying direction, and obtains the identification of the convolutional neural networks model output
As a result.Wherein, re -training was established after convolutional neural networks model deletes redundant filter, and redundant filter is whole
Importance is lower than the filter of threshold value;Recognition result is used to indicate the direction to be measured of vehicle, and whether there are obstacles.Pass through deletion
The mode of redundant filter, while not losing neural network precision, it is possible to reduce the storage of convolutional neural networks model is empty
Between, also, reduce filter and reduce convolution algorithm accordingly, and then can reduce and be deposited needed for electronic equipment when detecting barrier
Reserves and calculation amount.
Detailed description of the invention
It, below will be to embodiment or the prior art in order to illustrate more clearly of the present invention or technical solution in the prior art
Attached drawing needed in description is briefly described, it should be apparent that, the accompanying drawings in the following description is of the invention one
A little embodiments for those of ordinary skill in the art without any creative labor, can also be according to this
A little attached drawings obtain other attached drawings.
Fig. 1 is a kind of application scenarios schematic diagram of obstacle detection method provided by the invention;
Fig. 2 is a kind of flow diagram of obstacle detection method provided by the invention;
Fig. 3 is the flow diagram of another obstacle detection method provided by the invention;
Fig. 4 is the flow diagram of another obstacle detection method provided by the invention;
Fig. 5 is a kind of structural schematic diagram of obstacle detector provided by the invention;
Fig. 6 is the structural schematic diagram of another obstacle detector provided by the invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
In the prior art, the highest obstacle detection method of accuracy rate all uses depth convolutional neural networks model.
Depth convolutional neural networks model is made of multiple convolutional layers and full articulamentum, is the most important deep learning for image recognition
One of frame.
However, the existing obstacle detection method using depth convolutional neural networks, due to being deposited in convolutional neural networks
In bulk redundancy parameter and unnecessary calculating, cause during detection of obstacles, terminal device detects required when barrier
Amount of storage and calculation amount it is larger.
In view of the above problem, the present invention provides a kind of obstacle detection methods, by the direction to be measured for obtaining vehicle
Image, by image input convolutional neural networks model, and obtain convolutional neural networks model output recognition result.Wherein,
Re -training was established after convolutional neural networks model deletes redundant filter, and redundant filter is that whole importance is lower than
The filter of threshold value;Recognition result is used to indicate the direction to be measured of vehicle, and whether there are obstacles.By deleting redundant filter
Mode, while not losing neural network precision, it is possible to reduce the memory space of convolutional neural networks model, also, more
Few filter means less convolution algorithm, so can reduce detection barrier when electronic equipment needed for amount of storage and
Calculation amount.
The executing subject of obstacle detection method provided by the present invention can be obstacle detector.The obstacle quality testing
Surveying device can be program, program code software, or the related medium for executing code is stored with, for example, USB flash disk etc..
In some embodiments, obstacle detector can also fill to integrate or being equipped with the related entity for executing code
It sets, for example, processor, chip, micro-control unit (Microcontroller Unit, abbreviation MCU), computer, computer etc..
Fig. 1 is a kind of application scenarios schematic diagram of obstacle detection method provided by the invention.Such as Fig. 1, vehicle 11 is in road
When the automatic running of road, the camera of 11 vehicle body of vehicle installation can obtain the image on one or several directions of vehicle 11 in real time.
When barrier 12 occurs in blocking the way road, the image that camera of the vehicle 11 in the barrier direction is shot will appear the barrier
12, at this point, can identify that whether there are obstacles in the image by convolutional neural networks model, and then control driving for vehicle
Sail behavior.Wherein, in a kind of embodiment, above-mentioned image can be sent in the terminal device in vehicle 11, so that terminal
Processor identifies the image by convolutional neural networks model in equipment.In another embodiment, above-mentioned image can be sent out
Server is given, so that processor by convolutional neural networks model identifies the image in server.
Below for integrated or be equipped with the processor of the related electronic equipment for executing code, with specifically embodiment pair
Technical solution of the present invention is described in detail.These specific embodiments can be combined with each other below, for identical or phase
As concept or process may be repeated no more in some embodiments.
Fig. 2 is a kind of flow diagram of obstacle detection method provided by the invention.What is involved is processing for the present embodiment
How device detects the process of the barrier on some direction of vehicle.As shown in Fig. 2, this method comprises:
The image in the direction to be measured of S101, acquisition vehicle.
In this step, the direction to be measured of vehicle can be one, or multiple.
The image for obtaining the direction to be measured of vehicle, can specifically be obtained in real time by the camera being arranged on vehicle body.Camera shooting
The setting of head can be specifically arranged in conjunction with the running environment of vehicle, to shoot the image in direction to be measured.
For example in conjunction with concrete condition, if the front end of automobile body, rear end, left and right side are each provided with a camera shooting
Head, then processor can obtain in real time vehicle front, rear, left and right side image.
S102, image is inputted into convolutional neural networks model, and obtains the recognition result of convolutional neural networks model output,
Re -training was established after convolutional neural networks model deletes redundant filter, and redundant filter is that whole importance is lower than
The filter of threshold value;Recognition result is used to indicate the direction to be measured of vehicle, and whether there are obstacles.
In this step, the image received is inputted convolutional neural networks model, the convolutional neural networks mould by processor
Type can identify image.A kind of enforceable identification method is only to identify that whether there are obstacles in image;It is another
Enforceable identification method is that specially which kind of barrier identified, so that processor can specifically be set according to the type of barrier
It sets and hides mode.
Usually having many filters in convolutional neural networks model includes similar or even identical information, i.e., redundancy compared with
Height, therefore, under the premise of guaranteeing not losing the precision of convolutional neural networks model, in order to compress and accelerate above-mentioned convolutional Neural
Network model, above-mentioned convolutional neural networks model need to delete redundant filter therein before use.Specifically, deleting superfluous
Remaining filter can calculate the whole importance between filter, and the filter for determining that entirety importance is lower than threshold value is redundancy
Filter, and then neural networks pruning is carried out to realize.Whole importance, it can be understood as filter is in entire convolutional Neural net
Significance level in network model.
Obstacle detection method provided in this embodiment, the image in the direction to be measured by obtaining vehicle, image is inputted
Convolutional neural networks model, and obtain the recognition result of the convolutional neural networks model output.Wherein, convolutional neural networks mould
Re -training was established after type deletes redundant filter, and redundant filter is the filter that whole importance is lower than threshold value;
Recognition result is used to indicate the direction to be measured of vehicle, and whether there are obstacles.By way of deleting redundant filter, do not damaging
While denervation neural network accuracy, it is possible to reduce the memory space of convolutional neural networks model, also, it is corresponding to reduce filter
Amount of storage and calculation amount needed for electronic equipment when reducing convolution algorithm, and then detection barrier can be reduced.
During deleting redundant filter, how to determine that the filter in convolutional neural networks model is redundant filter
It is key problem.The process for how deleting redundant filter in convolutional neural networks model to processor below is illustrated.
Fig. 3 is the flow diagram of another obstacle detection method provided by the invention.What is involved is places for the present embodiment
How reason device deletes the process of redundant filter in convolutional neural networks model.As shown in figure 3, this method comprises:
S201, pass through pruning algorithms model, delete the filter of redundancy in convolutional neural networks model.
To be illustrated to this step, a convolutional neural networks model can be set first.
If the neural network in convolutional neural networks model is L layers shared, wherein PlFor in above-mentioned convolutional neural networks model
L layers of neural network, PlInput be Xl, export as Yl, weight matrix Wl。
Work as PlWhen for full articulamentum, WlFor Ml×NlMatrix, wherein MlIt is the quantity of input node, NlFor output node
Quantity;For m-th of input node,For n-th of input node.It enablesFor WlIn m row, that is,
Work as PlWhen for convolutional layer, WlFor a Kl×Kl×Ml×NlTensor, wherein K is convolution kernel size, M and N difference
For the number of channels of input and output.XlIt is 1l×1l×NlTensor, YlFor Ol×Ol×Nl,For m-th of input feature vector figure,
On the basis of above-mentioned convolutional neural networks model, pruning algorithms model can be according in convolutional neural networks model
Filter between whole importance determine redundant filter, and carry out neural networks pruning.
Specifically, pruning algorithms model, may include: that similarity determines that layer, mean value determine that layer, local importance determine
Layer, regular terms determine that layer, whole importance determine layer, judge layer and delete layer.
Layer, which explains, to be determined to the similarity in pruning algorithms model below.
Similarity determines layer, for determining the same layer filter of first filter and first filter in convolutional neural networks
Between linear similarity.Wherein, first filter is any filter in convolutional neural networks model.
When first filter is located at the full articulamentum of above-mentioned convolutional neural networks model, Yn=∑mWM, nXm, if there is
M1, m2 ∈ [1, M], so thatIt sets up, then YnIt can be deformed into formula (1), formula (1) is specific as follows:
Wherein, ∈ isWithTo be unsatisfactory for losing caused by stringent linear relationship.
Therefore, it can be incited somebody to action by way of beta pruning and adjustment neural networkWithMerge.Due toThenWithAlways by similar feature activation, and then Xm1And Xm2Identical feature is always expressed, so
It can be measured with Pearson came relative coefficient linear similar between first filter and the same layer filter of first filter
Degree, specific as shown in formula (2):
Wherein, sim is similarity function, and corr is linear dependence function, Xm1For the m1 input node, Xm2It is
M2 input node,For Xm1Corresponding weight,For Xm2Corresponding weight, μ are mean value, and σ is standard deviation, and E is scheduled to last
Prestige value.
When first filter is located at the full articulamentum of convolutional neural networks model, since full articulamentum can regard convolution as
The convolutional layer that core size is 1, therefore, formula (2) can be used for convolutional layer by simple transformation.It is big for a convolution kernel
The small convolutional layer for K, the combination of filter are one 4 dimension tensors, and size is K × K × M × N.Each filter can be seen
Doing is therefore all filters are reassembled as the independent node of K × K group by K × K independent nodes, calculates this K × K group section
The correlation of point between any two, is then averaged the similarity of this K × K node, as the linear phase between filter
Like degree.Based on this, linear similarity between the same layer filter of first filter and first filter can use formula (3)
It calculates, formula (3) is specific as follows:
Wherein, sim is similarity function, and corr is linear dependence function, Xm1For the m1 input node, Xm2It is
M2 input node,For Xm1The weight being connect with each node,For Xm2The weight being connect with each node,
K is convolution kernel size.
Layer, which explains, to be determined to the mean value in pruning algorithms model below.
Mean value determines layer, for determining the maximum similarity mean value of first filter according to linear similarity.
After determining the linear similarity between first filter and the same layer filter of first filter, can choose with
The highest N number of filter of first filter similarity, and it is average similar to calculate the maximum between first filter and N number of filter
Degree.Maximum average similarity is higher, and the local importance of first filter is lower.Wherein, N can specifically be set according to the actual situation
It sets.
Layer, which explains, to be determined to the local importance in pruning algorithms model below.
Local importance determines layer, for the maximum similarity mean value according to first filter, carries out to first filter
Max normalized determines the local importance of first filter.
It is since the calculated similarity numberical range of different layers is different in convolutional neural networks model, above-mentioned similarity is true
Given layer and above-mentioned mean value determine that the maximum average similarity that layer is determined is used as local importance, can not carry out global comparison.Cause
This, needs for maximum average similarity to be all normalized, be compared after scaling it identical range.
In the prior art, mainstream algorithm normalizes the different degree of filter using L2, however L2 is used to normalize, beta pruning
Algorithm model can tend to think that the filter in the more layer of filter is unessential, and this tendency is wrong.
To solve the above problems, above-mentioned tendency can be avoided using max normalization.Phase after max normalization can be used
It determines the local importance of first filter like degree, can specifically be calculated with formula (4), formula (4) is specific as follows:
Wherein, Imp is local importance, and sim is similarity function, and k is adjustable parameter, and TopK is to take maximum k number value
Function.
Layer, which explains, to be determined to the regular terms in pruning algorithms model below.
Regular terms determines layer, for the corresponding regular terms of layer where determining first filter.
The compression and acceleration of convolutional neural networks model are two different optimization directions, in convolutional neural networks
While equivalent parameter is fallen in different location beta pruning, the calculation amount for the reduction that can not be equal.
In view of the above problem, after determining the local different degree of first filter, it is also necessary to determine first filter
Regular terms.By adjusting regular terms, different Pruning strategies can be taken, to carry out different optimization directions.
Specifically, can be using each layer of convolutional neural networks model of parameter amount and calculation amount as regular terms.Convolution mind
Quantity through l layers of filter in network model will affect the parameter amount and calculation amount of l and l+1 layers, therefore, with l
The relevant parameter amount (S of layerl) and calculation amount (Cl) can be determined by formula (5) and (6), formula (5) and (6) are as follows:
Sl=(KlKlMlNl)+(Kl+1Kl+1Ml+1Nl+1) (5)
Cl=(2IlIlKlKlMlNl)+(2Il+1Il+1Kl+1Kl+1Ml+1Nl+1) (6)
After determining parameter amount and calculation amount by formula (5) and (6), and then it can determine regular terms, it is specific such as formula
(7);
Wherein, ReglFor the regular terms of l layers of filter, β and γ are adjustable parameter, ClFor l layers of calculation amount, Cu
For u layers of calculation amount, SlFor l layer parameter amount, SuFor u layer parameter amount, L is the number of plies of neural network.
By formula (7), the regular terms of first filter can be determined.
Layer, which explains, to be determined to the whole importance in pruning algorithms model below.
Whole importance determines layer, for determining that the entirety of first filter is important according to regular terms and local importance
Property.
Determine that the local importance and regular terms that determine in layer determine the regular terms determined in layer in conjunction with local importance, it can
To determine the whole importance of first filter, specifically, such as formula (8):
Wherein, ReImp is whole importance,For l layers of m-th of node, Imp is local importance, ReglIt is
The regular terms of l layers of filter.
Layer is judged, for determining that first filter is redundant filter according to whole importance and preset deletion ratio.
It determines in convolutional neural networks model after the whole importance of first filter, it can be according to preset deletion ratio
Determine threshold value, the filter by whole different degree lower than the threshold value is determined as redundant filter.
Layer is deleted, for deleting the redundant filter in convolutional neural networks.
Optionally, by pruning algorithms model, the filter of redundancy in convolutional neural networks model is deleted, can also be made
Re -training is carried out to the convolutional neural networks model after beta pruning with sample data, thus to the parameter of convolutional neural networks model
Carry out fine tune.
In addition, separating convolution for the depth in convolutional neural networks model.One depth separates convolutional layer
One depth convolutional layer and convolutional layer, the size that point convolution can be counted as filter is that 1 × 1 × M × N is common
Convolution, thus we can as beta pruning common convolutional layer beta pruning point convolutional layer.For depth convolutional layer, this layer it is defeated
The port number for entering and exporting is always equal, and puts the output that convolution input is exactly depth convolution, therefore beta pruning point convolution is same
When depth convolutional layer also can be by beta pruning, without independent beta pruning depth convolutional layer again.
For the residual block in convolutional neural networks.One residual block all includes two or more convolutional layers, for
The residual block of one not no linear projection, its input channel quantity must be equal with output channel quantity.Therefore, for one
There is no the residual block of linear projection, it can be with the average different degree of input filter and output filter as respective entirety
Different degree, and wipe out the filter output and input simultaneously in beta pruning.
Finally it is to be appreciated that it will be appreciated by persons skilled in the art that convolutional neural networks involved in the present embodiment
Model and pruning algorithms model, can not only apply to the identification of barrier, can also apply to other and need to carry out image knowledge
In other scheme.
The image in the direction to be measured of S202, acquisition vehicle.
S203, image is inputted into convolutional neural networks model, and obtains the recognition result of convolutional neural networks model output,
Re -training was established after convolutional neural networks model deletes redundant filter, and redundant filter is that redundancy is more than threshold value
Filter;Recognition result is used to indicate the direction to be measured of vehicle, and whether there are obstacles.
Technical term, technical effect, technical characteristic and the optional embodiment of step S202-S203, can refer to Fig. 2
Shown step S101-S102 understands, for duplicate content, is not repeated herein.
Obstacle detection method provided in this embodiment cuts convolutional neural networks model by pruning algorithms model
Branch, to delete the redundant filter in convolutional neural networks model, so that convolutional neural networks model is compressed and accelerated,
And then amount of storage and calculation amount needed for electronic equipment when detecting barrier to reduce.
Fig. 4 is the flow diagram of another obstacle detection method provided by the invention.What is involved is places for the present embodiment
Manage the process how device carries out Driving control according to recognition result.As shown in figure 4, this method comprises:
The image in the direction to be measured of S301, acquisition vehicle.
S302, image is inputted into convolutional neural networks model, and obtains the recognition result of convolutional neural networks model output,
Re -training was established after convolutional neural networks model deletes redundant filter, and redundant filter is that whole importance is lower than
The filter of threshold value;Recognition result is used to indicate the direction to be measured of vehicle, and whether there are obstacles.
Technical term, technical effect, technical characteristic and the optional embodiment of step S301-S302, can refer to Fig. 2
Shown step S101-S102 understands, for duplicate content, is not repeated herein.
S303, according to recognition result, control the driving behavior of vehicle.
In this step, after convolutional neural networks model completes image recognition, processor can be tied directly according to identification
Fruit determines vehicle periphery, and whether there are obstacles and the position of barrier, and then controls the driving behavior of vehicle, reformulates
Drive route is with avoiding barrier.
If convolutional neural networks model identifies specific obstacle identity, can according to it is preset hide rule come into
One-step optimization drive route.Such as: if the barrier identified has pedestrian and vehicle, hide rule according to preset, preferentially
Hide pedestrian.
In a kind of optional embodiment, if processor judges that the barrier identified may cause driving safety
When hidden danger, processor can sound an alarm while controlling the driving behavior avoiding barrier in vehicle age to passengers inside the car, so that
Passengers inside the car terminate the automatic driving mode of vehicle and manually adapter tube vehicle.
Obstacle detection method provided by the invention, the image in the direction to be measured by obtaining vehicle, image is inputted and is rolled up
Product neural network model, and the recognition result of convolutional neural networks model output is obtained, and then control the driving behavior of vehicle.Its
In, re -training was established after convolutional neural networks model deletes redundant filter.By the above-mentioned means, in automatic Pilot
During, amount of storage and calculation amount needed for electronic equipment when can reduce detection barrier.
Those of ordinary skill in the art will appreciate that: realize that all or part of the steps of above method embodiment can pass through
The relevant hardware of program instruction is completed, and program above-mentioned can be stored in a computer readable storage medium, the program
When being executed, step including the steps of the foregoing method embodiments is executed;And storage medium above-mentioned includes: ROM, RAM, magnetic disk or light
The various media that can store program code such as disk.
Fig. 5 is a kind of structural schematic diagram of obstacle detector provided by the invention.The obstacle detector can be with
It can be aforementioned described processor by being implemented in combination with for software, hardware or both.
As shown in figure 5, the obstacle detector 40 includes: pruning module 41, obtains module 42, processing module 43, drives
Sail module 44.
Pruning module 41, for deleting the filter of redundancy in convolutional neural networks model by pruning algorithms model.
Optionally, pruning algorithms model, comprising: similarity determines that layer, mean value determine that layer, local importance determine layer, just
Then item determines that layer, whole importance determine layer, judge layer and delete layer;
Similarity determines layer, for determining the same layer filter of first filter and first filter in convolutional neural networks
Between linear similarity;
Mean value determines layer, for determining the maximum similarity mean value of first filter according to linear similarity;
Local importance determines layer, for the maximum similarity mean value according to first filter, carries out to first filter
Max normalized determines the local importance of first filter;
Regular terms determines layer, for the corresponding regular terms of layer where determining first filter;
Whole importance determines layer, for determining that the entirety of first filter is important according to regular terms and local importance
Property;
Layer is judged, for determining that first filter is redundant filter according to whole importance and preset deletion ratio;
Layer is deleted, for deleting the redundant filter in convolutional neural networks.
Optionally, if first filter is located at the full articulamentum of convolutional neural networks model, similarity determines layer, specific to use
In using formula
Calculate the
Linear similarity between one filter and the same layer filter of first filter;
Wherein, sim is similarity function, and corr is linear dependence function, Xm1For the m1 input node, Xm2It is
M2 input node,For Xm1Corresponding weight,For Xm2Corresponding weight, μ are mean value, and σ is standard deviation, and E is scheduled to last
Prestige value.
Optionally, if first filter is located at the convolutional layer of convolutional neural networks model, similarity determines layer, is specifically used for
Using formulaCalculate first filter and the first mistake
Similarity between the filter of filter same layer;
Wherein, sim is similarity function, and Corr is linear dependence function, Xm1For the m1 input node, Xm2It is
M2 input node,For Xm1The weight being connect with each node,For Xm2The weight being connect with each node,
K is convolution kernel size.
Optionally, local importance determines layer, is specifically used for using formula
N ∈ [1, M], n ≠ m calculate first
The local importance of filter;
Wherein, Imp is local importance, and sim is similarity function, and k is adjustable parameter, and TopK is to take maximum k number value
Function.
Optionally, regular terms determines layer, is specifically used for using formula
U ∈ [1, L] calculates the first filtering
The corresponding regular terms of layer where device;
Wherein, ReglFor the regular terms of l layers of filter, β and γ are adjustable parameter, ClFor l layers of calculation amount, Cu
For u layers of calculation amount, SlFor l layer parameter amount, SuFor u layer parameter amount, L is the number of plies of neural network.
Optionally, whole importance determines layer, is specifically used for using formula Calculate the whole importance of first filter;
Wherein, ReImp is whole importance,For l layers of m-th of node, Imp is local importance, ReglIt is
The regular terms of l layers of filter.
Obtain module 42, the image in the direction to be measured for obtaining vehicle;
Processing module 43 for image to be inputted convolutional neural networks model, and obtains the output of convolutional neural networks model
Recognition result, convolutional neural networks model deletes re -training after redundant filter and establishes, and redundant filter is whole
Body importance is lower than the filter of threshold value;Recognition result is used to indicate the direction to be measured of vehicle, and whether there are obstacles.
Module 44 is driven, for controlling the driving behavior of vehicle according to recognition result.
Obstacle detector provided by the invention can execute the movement of processor in above method embodiment, in fact
Existing principle is similar with technical effect, and details are not described herein.
Fig. 6 is the structural schematic diagram of another obstacle detector provided by the invention.As shown in fig. 6, the barrier
Detection device may include: at least one processor 51 and memory 52.Fig. 6 shows the electronics taken a processor as an example
Equipment.
Memory 52, for storing program.Specifically, program may include program code, and program code includes computer
Operational order.
Memory 52 may include high speed RAM memory, it is also possible to further include nonvolatile memory (non-volatile
Memory), a for example, at least magnetic disk storage.
Processor 51 is used to execute the computer executed instructions of the storage of memory 52, to realize obstacle detection method.
Wherein, processor 51 may be a central processing unit (Central Processing Unit, referred to as CPU),
Either specific integrated circuit (Application Specific IntegratedCircuit, referred to as ASIC), either
It is configured to implement one or more integrated circuits of the embodiment of the present application.
Optionally, in specific implementation, if communication interface, memory 52 and the independent realization of processor 51, communication connect
Mouth, memory 52 and processor 51 can be connected with each other by bus and complete mutual communication.Bus can be industrial mark
Quasi- architecture (Industry Standard Architecture, referred to as ISA) bus, external equipment interconnection
(Peripheral Component, referred to as PCI) bus or extended industry-standard architecture (Extended Industry
Standard Architecture, referred to as EISA) bus etc..Bus can be divided into address bus, data/address bus, control always
Line etc., it is not intended that an only bus or a type of bus.
Optionally, in specific implementation, if communication interface, memory 52 and processor 51 are integrated real on one chip
Existing, then communication interface, memory 52 and processor 51 can be completed to communicate by internal interface.
The present invention also provides a kind of computer readable storage medium, which may include: U
Disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access
Memory), the various media that can store program code such as disk or CD, specifically, the computer readable storage medium
In be stored with program instruction, program instruction is for the method in above-described embodiment.
Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention., rather than its limitations;To the greatest extent
Pipe present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that: its according to
So be possible to modify the technical solutions described in the foregoing embodiments, or to some or all of the technical features into
Row equivalent replacement;And these are modified or replaceed, various embodiments of the present invention technology that it does not separate the essence of the corresponding technical solution
The range of scheme.
Claims (12)
1. a kind of obstacle detection method characterized by comprising
Obtain the image in the direction to be measured of vehicle;
Described image is inputted into convolutional neural networks model, and obtains the recognition result of the convolutional neural networks model output,
Re -training was established after the convolutional neural networks model deletes redundant filter, and the redundant filter is whole weight
The property wanted is lower than the filter of threshold value;The recognition result is used to indicate the direction to be measured of the vehicle, and whether there are obstacles.
2. the method according to claim 1, wherein it is described obtain vehicle direction to be measured image before,
Further include:
By pruning algorithms model, the filter of redundancy in the convolutional neural networks model is deleted.
3. the method according to claim 1, wherein obtaining what the convolutional neural networks model exported described
After recognition result, further includes:
According to the recognition result, the driving behavior of the vehicle is controlled.
4. according to the method described in claim 2, it is characterized in that, the pruning algorithms model, comprising: similarity determine layer,
Mean value determines that layer, local importance determine that layer, regular terms determine that layer, whole importance determine layer, judge layer and delete layer;
The similarity determines layer, for determining first filter and the first filter in the convolutional neural networks model
Same layer filter between linear similarity;
The mean value determines layer, for determining the maximum similarity mean value of the first filter according to the linear similarity;
The part importance determines layer, for the maximum similarity mean value according to the first filter, to first mistake
Filter carries out max normalized, determines the local importance of the first filter;
The regular terms determines layer, for the corresponding regular terms of layer where determining the first filter;
The entirety importance determines layer, for determining first filtering according to the regular terms and the local importance
The whole importance of device;
The judgement layer, for determining that the first filter is superfluous according to the whole importance and preset deletion ratio
Remaining filter;
The deletion layer, for deleting the redundant filter in the convolutional neural networks model.
5. according to the method described in claim 4, it is characterized in that, if the first filter is located at the convolutional neural networks
The full articulamentum of model, the similarity determine layer, are specifically used for using formulaCalculate first mistake
Linear similarity between filter and the same layer filter of the first filter;
Wherein, sim is similarity function, and corr is linear dependence function, Xm1For the m1 input node, Xm2It is defeated for m2
Ingress,For Xm1Corresponding weight,For Xm2Corresponding weight, μ are mean value, and σ is standard deviation, and E is desired value.
6. according to the method described in claim 4, it is characterized in that, if the first filter is located at the convolutional neural networks
The convolutional layer of model, the similarity determine layer, are specifically used for using formulaCalculate the first filter and first mistake
Similarity between the filter of filter same layer;
Wherein, sim is similarity function, and corr is linear dependence function, Xm1For the m1 input node, Xm2It is defeated for m2
Ingress,For Xm1The weight being connect with each node,For Xm2The weight connecting with each node, K are volume
Product core size.
7. according to the method described in claim 4, it is characterized in that, the part importance determines layer, specifically for using public
Formula
Described in calculating
The local importance of first filter;
Wherein, Imp is local importance, and sim is similarity function, and k is adjustable parameter, and TopK is the letter for taking maximum k number value
Number.
8. according to the method described in claim 4, it is characterized in that, the regular terms determines layer, specifically for using formula
Calculate first mistake
The corresponding regular terms of layer where filter;
Wherein, ReglFor the regular terms of l layers of filter, β and γ are adjustable parameter, ClFor l layers of calculation amount, CuFor u
Layer calculation amount, SlFor l layer parameter amount, SuFor u layer parameter amount, L is the number of plies of neural network.
9. according to the method described in claim 4, it is characterized in that, the entirety importance determines layer, specifically for using public
FormulaCalculate the whole importance of the first filter;
Wherein, ReImp is whole importance,For l layers of m-th of node, Imp is local importance, ReglIt is l layers
Filter regular terms.
10. a kind of obstacle detector characterized by comprising
Obtain module, the image in the direction to be measured for obtaining vehicle;
Processing module, for described image to be inputted convolutional neural networks model, and it is defeated to obtain the convolutional neural networks model
Recognition result out, re -training was established after the convolutional neural networks model deletes redundant filter, the redundancy
Filter is the filter that whole importance is lower than threshold value;The recognition result be used to indicate the vehicle direction to be measured whether
There are barriers.
11. a kind of electronic equipment characterized by comprising memory and processor;
The memory, for storing the executable instruction of the processor;
The processor is configured to carry out any method of perform claim requirement 1-9 via the execution executable instruction.
12. a kind of storage medium, is stored thereon with computer program characterized by comprising when the program is executed by processor
Realize any method of claim 1-9.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910385611.5A CN110263628B (en) | 2019-05-09 | 2019-05-09 | Obstacle detection method, obstacle detection device, electronic apparatus, and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910385611.5A CN110263628B (en) | 2019-05-09 | 2019-05-09 | Obstacle detection method, obstacle detection device, electronic apparatus, and storage medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110263628A true CN110263628A (en) | 2019-09-20 |
CN110263628B CN110263628B (en) | 2021-11-23 |
Family
ID=67914546
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910385611.5A Active CN110263628B (en) | 2019-05-09 | 2019-05-09 | Obstacle detection method, obstacle detection device, electronic apparatus, and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110263628B (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110929805A (en) * | 2019-12-05 | 2020-03-27 | 上海肇观电子科技有限公司 | Neural network training method, target detection device, circuit and medium |
CN111126456A (en) * | 2019-12-05 | 2020-05-08 | 杭州飞步科技有限公司 | Processing method, device and equipment of neural network model and storage medium |
CN111397611A (en) * | 2020-03-05 | 2020-07-10 | 北京百度网讯科技有限公司 | Path planning method and device and electronic equipment |
JP7316566B1 (en) | 2022-05-11 | 2023-07-28 | ノタ、インコーポレイテッド | Neural network model weight reduction method and electronic device for performing the same |
US11775806B2 (en) | 2022-02-10 | 2023-10-03 | Nota, Inc. | Method of compressing neural network model and electronic apparatus for performing the same |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20040120330A1 (en) * | 2002-12-20 | 2004-06-24 | Rhine Scott Alan | Need-based filtering for rapid selection of devices in a tree topology network |
CN108399398A (en) * | 2018-03-22 | 2018-08-14 | 武汉云衡智能科技有限公司 | A kind of pilotless automobile obstacle recognition detection method based on deep learning |
CN108596009A (en) * | 2017-12-29 | 2018-09-28 | 西安智加科技有限公司 | A kind of obstacle detection method and system for agricultural machinery automatic Pilot |
CN108846445A (en) * | 2018-06-26 | 2018-11-20 | 清华大学 | A kind of convolutional neural networks filter technology of prunning branches based on similarity-based learning |
CN109447033A (en) * | 2018-11-14 | 2019-03-08 | 北京信息科技大学 | Vehicle front obstacle detection method based on YOLO |
CN109558808A (en) * | 2018-11-09 | 2019-04-02 | 同济大学 | A kind of road Edge Detection based on deep learning |
WO2019080484A1 (en) * | 2017-10-26 | 2019-05-02 | 北京深鉴智能科技有限公司 | Method of pruning convolutional neural network based on feature map variation |
-
2019
- 2019-05-09 CN CN201910385611.5A patent/CN110263628B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20040120330A1 (en) * | 2002-12-20 | 2004-06-24 | Rhine Scott Alan | Need-based filtering for rapid selection of devices in a tree topology network |
WO2019080484A1 (en) * | 2017-10-26 | 2019-05-02 | 北京深鉴智能科技有限公司 | Method of pruning convolutional neural network based on feature map variation |
CN108596009A (en) * | 2017-12-29 | 2018-09-28 | 西安智加科技有限公司 | A kind of obstacle detection method and system for agricultural machinery automatic Pilot |
CN108399398A (en) * | 2018-03-22 | 2018-08-14 | 武汉云衡智能科技有限公司 | A kind of pilotless automobile obstacle recognition detection method based on deep learning |
CN108846445A (en) * | 2018-06-26 | 2018-11-20 | 清华大学 | A kind of convolutional neural networks filter technology of prunning branches based on similarity-based learning |
CN109558808A (en) * | 2018-11-09 | 2019-04-02 | 同济大学 | A kind of road Edge Detection based on deep learning |
CN109447033A (en) * | 2018-11-14 | 2019-03-08 | 北京信息科技大学 | Vehicle front obstacle detection method based on YOLO |
Non-Patent Citations (2)
Title |
---|
ZHENGTAO WANG: "Towards thinner convolutional neural networks through gradually global pruning", 《2017 IEEE INTERNATIONAL CONFERENCE ON IMAGE PROCESSING (ICIP)》 * |
孙凯等: "基于改进卷积神经网络的动态障碍物检测方法", 《计算机应用与软件》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110929805A (en) * | 2019-12-05 | 2020-03-27 | 上海肇观电子科技有限公司 | Neural network training method, target detection device, circuit and medium |
CN111126456A (en) * | 2019-12-05 | 2020-05-08 | 杭州飞步科技有限公司 | Processing method, device and equipment of neural network model and storage medium |
CN111126456B (en) * | 2019-12-05 | 2023-06-23 | 杭州飞步科技有限公司 | Neural network model processing method, device, equipment and storage medium |
CN110929805B (en) * | 2019-12-05 | 2023-11-10 | 上海肇观电子科技有限公司 | Training method, target detection method and device for neural network, circuit and medium |
CN111397611A (en) * | 2020-03-05 | 2020-07-10 | 北京百度网讯科技有限公司 | Path planning method and device and electronic equipment |
US11775806B2 (en) | 2022-02-10 | 2023-10-03 | Nota, Inc. | Method of compressing neural network model and electronic apparatus for performing the same |
JP7316566B1 (en) | 2022-05-11 | 2023-07-28 | ノタ、インコーポレイテッド | Neural network model weight reduction method and electronic device for performing the same |
Also Published As
Publication number | Publication date |
---|---|
CN110263628B (en) | 2021-11-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110263628A (en) | Obstacle detection method, device, electronic equipment and storage medium | |
CN106485215B (en) | Face shielding detection method based on deep convolutional neural network | |
WO2018090706A1 (en) | Method and device of pruning neural network | |
CN109145860B (en) | lane line tracking method and device | |
CN108875648A (en) | A method of real-time vehicle damage and component detection based on mobile video stream | |
CN111667501A (en) | Target tracking method and device, computing equipment and storage medium | |
CN109558901B (en) | Semantic segmentation training method and device, electronic equipment and storage medium | |
EP3704558A1 (en) | Depth-aware object counting | |
CN107924568B (en) | Image processing apparatus, image processing method, and storage medium | |
CN111079533A (en) | Unmanned vehicle driving decision method, unmanned vehicle driving decision device and unmanned vehicle | |
CN108898057B (en) | Method, device, computer equipment and storage medium for tracking target detection | |
CN109308704A (en) | Background elimination method, device, computer equipment and storage medium | |
CN113052008A (en) | Vehicle weight recognition method and device | |
CN111798482A (en) | Target tracking method and device | |
KR20200095357A (en) | Learning method and learning device for heterogeneous sensor fusion by using merging network which learns non-maximum suppression | |
CN113239746B (en) | Electric vehicle detection method, device, terminal equipment and computer readable storage medium | |
CN113762220A (en) | Object recognition method, electronic device, and computer-readable storage medium | |
CN113139696A (en) | Trajectory prediction model construction method and trajectory prediction method and device | |
CN112580565A (en) | Lane line detection method, lane line detection device, computer device, and storage medium | |
CN116363628A (en) | Mark detection method and device, nonvolatile storage medium and computer equipment | |
CN115690732A (en) | Multi-target pedestrian tracking method based on fine-grained feature extraction | |
CN113112525A (en) | Target tracking method, network model, and training method, device, and medium thereof | |
CN110533158B (en) | Model construction method, system and non-volatile computer readable recording medium | |
CN110599542A (en) | Method and device for local mapping of adaptive VSLAM (virtual local area model) facing to geometric area | |
CN109949253A (en) | A kind of filter method of prediction block, electronic equipment and storage medium |
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 |