CN109029363A - A kind of target ranging method based on deep learning - Google Patents
A kind of target ranging method based on deep learning Download PDFInfo
- Publication number
- CN109029363A CN109029363A CN201810562681.9A CN201810562681A CN109029363A CN 109029363 A CN109029363 A CN 109029363A CN 201810562681 A CN201810562681 A CN 201810562681A CN 109029363 A CN109029363 A CN 109029363A
- Authority
- CN
- China
- Prior art keywords
- target
- model
- training
- deep learning
- ranging
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
- G01C11/00—Photogrammetry or videogrammetry, e.g. stereogrammetry; Photographic surveying
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
- G01C11/00—Photogrammetry or videogrammetry, e.g. stereogrammetry; Photographic surveying
- G01C11/36—Videogrammetry, i.e. electronic processing of video signals from a single source or from different sources to give parallax or range information
Abstract
The invention discloses a kind of target ranging methods based on deep learning, specifically includes the following steps: the foundation of (1) based on the target database under different distance；(2) object ranging model is built；(3) loss function of design object ranging model；(4) training method of design object ranging model；(5) trained object ranging model is tested.Innovative point of the invention: based on the video image target detection algorithm in deep learning, regression problem is converted by object ranging problem and incorporates algorithm of target detection model, to realize target detection and object ranging in an algorithm model.The advantages of this method is: by this algorithm, only needing common monocular-camera that can obtain target category information and target range information, to may replace traditional binocular camera or laser radar to obtain the depth information of target, reduces hardware cost.
Description
Technical field
The present invention relates to a kind of target ranging methods based on deep learning, belong to artificial intelligence field.
Background technique
With the generation of big data and the promotion of computer computation ability, artificial intelligence technology has obtained further hair
Exhibition.How emerging technology to be applied on traditional problem is our urgent problems to be solved.It is traditional based on binocular camera
Stereoscopy passive ranging method higher cost, the target information of acquisition is more single, therefore cannot be promoted well.
Deep learning (deep learning) is the branch of machine learning, be one kind attempt using comprising labyrinth or
The multiple process layers being made of multiple nonlinear transformation carry out the algorithm of higher level of abstraction to data；Deep learning is in machine learning
A kind of algorithm based on to data progress representative learning；Various ways can be used to indicate in observation (such as piece image),
Such as vector of each pixel intensity value, or be more abstractively expressed as a series of sides, the region of specific shape etc.；And it uses certain
Specific representation method is easier learning tasks (for example, recognition of face or human facial expression recognition) from example；Deep learning
Benefit is that the feature learning and layered characteristic with non-supervisory formula or Semi-supervised extract highly effective algorithm to substitute acquisition feature by hand.
Target ranging method based on deep learning algorithm need to only be embedded in object ranging algorithm on common monocular-camera
Target category information and target range information can be obtained；This method reduces hardware cost instead of traditional binocular camera,
Enrich target signature information.
Summary of the invention
The invention discloses a kind of target ranging methods based on deep learning to be surveyed target by deep learning method
It is converted into regression problem away from problem, to reach the measurement of target range.
A goal of the invention according to the present invention, the present invention provides a kind of target ranging method based on deep learning,
Include the following:
First step: the foundation based on the target database under different distance；
Second step: object ranging model is built；
Third step: the loss function of design object ranging model；
Four steps: the training method of design object ranging model；
5th step: trained object ranging model is tested.
Further, in the first step, the foundation of the target database specifically includes that (a) data acquire；(b)
Data scaling；Wherein the source of data acquisition is the video information of specified video camera shooting；Data acquisition target includes 8 classes, point
Not are as follows: pedestrian, bicycle, motorcycle, automobile, bus, bird, cat, dog；Data scaling includes: the bounding box of detected target
Coordinate information, the bounding box classification information for being detected target, the depth information for being detected target.
Further, in the first step, the establishment step of the target database is specifically included that
First step: video data acquiring is carried out using binocular vision video camera；
Second step: video sub-frame processing is carried out to collected every section of video data, to obtain every frame image and figure
The depth information of object as in；
Third step: manually demarcating the single-frame images of acquisition using calibration tool, and calibration content includes: detected
The bounding box center point coordinate (x, y) of target, the bounding box width w and height h that are detected target, the bounding box for being detected target
The range information L of classification information c and detected target range video camera；
Four steps: training set, test set are divided into according to a certain percentage to the data demarcated.
Further, in the second step, object ranging model builds the convolution mind referred to based on deep learning
Building through network model, a total of 33 layers of the model, wherein first 25 layers are characterized extract layer, latter 8 layers are characterized the pre- of information
Survey layer.
Further, in the third step, the loss function of design object ranging model is specific as follows:
Wherein, λcoordIndicate the coefficient of the loss function based on coordinate points；λconfidenceIndicate the loss based on confidence level
The coefficient of function；λdistanceIndicate the coefficient of the loss function based on range information；λprobabilityIt indicates based on classification information
The coefficient of loss function.
Further, in the four steps, the training method of design object ranging model is specific as follows: (1) pre-training
Sorter network: one sorter network of pre-training on 1000 class data set of ImageNet, this sorter network is in second step
Preceding 14 layer network；(2) training objective ranging model: trained model in (1) is saved, after adding again on its basis
19 layer networks, random initializtion weight, and model training is carried out under the training set in the Multi-target Data library under different distance；
(3) training method under the training set in the Multi-target Data library under different distance is as follows: every batch of 256 samples of training；Setting
Training maximum number of iterations is 100,000 times；Learning rate is set as 0.0001；Trained model is saved.
Further, in the 5th step, the actual effect of test target ranging model；To damage by training pattern
It loses function and obtains minimum value, to test trained model.
Innovative point of the invention: based on the video image target detection algorithm in deep learning, object ranging problem is turned
It turns to regression problem and incorporates algorithm of target detection model, to realize target detection and object ranging in an algorithm model.
The advantages of this method is: by this algorithm, only needing common monocular-camera that can obtain target category information and target range letter
Breath, to may replace traditional binocular camera or laser radar to obtain the depth information of target, reduces hardware cost.
Detailed description of the invention
Fig. 1 is a kind of flow chart of distance measuring method based on deep learning of the invention；
Fig. 2 is the flow chart of a kind of training of distance measuring method based on deep learning of the invention, test phase；
Fig. 3 is a kind of flow chart of the practical stage of distance measuring method based on deep learning of the invention；
Fig. 4 is data acquisition schematic diagram of the present invention；
Fig. 5 is database structure schematic diagram of the present invention；
Fig. 6 is inventive network model structure；
Fig. 7 is inventive network model parameter table；
Fig. 8 is mapping relations figure of the loss function of the present invention on network model fc_33 layer.
Specific embodiment
Come that the present invention will be described in detail hereinafter with reference to attached drawing and in conjunction with the embodiments.It should be noted that in the feelings not conflicted
Under condition, the features in the embodiments and the embodiments of the present application be can be combined with each other.
Embodiment
This implementation provides a kind of distance measuring method based on deep learning, referring to figs. 1 to Fig. 3, according to embodiments of the present invention
A kind of distance measuring method based on deep learning, include the following steps.
Step S101, the foundation based on the target database under different distance；
Step S102, object ranging model are built；
Step S103, the loss function of design object ranging model；
Step S104, the training method of design object ranging model；
Step S105 tests trained object ranging model.
The distance measuring method based on deep learning will be discussed in detail by following more excellent implementation in we.
Data acquisition platform is built first.The focal length of i.e. fixed binocular vision camera position and its camera, acquires same
The maximum distance of a target set distance video camera in the video information apart from video camera different location, this method is 50 meters,
As shown in Figure 4.Acquire the pretreatment of data.
(1) video sub-frame processing is carried out to collected video information, to acquire the depth of every frame and objects in images
Spend information.
(2) collected single-frame images is demarcated using calibration tool, calibration content includes: the side of detected target
Boundary's frame center point coordinate (x, y), the width w of bounding box and height h, the bounding box classification information c and the frame figure for being detected target
The distance L of target range video camera is detected as in.
(3) it completes according to above step to 8 class targets, the acquisition of every 10,000 section of video data of class target.Database
Structure is as shown in Figure 5.
Convolutional neural networks model based on deep learning is built.Specifically:
(1) the pre- of network model input data is built on the deep learning frame TensorFlow based on Python
Processing module.
(2) network model module is built on the deep learning frame TensorFlow based on Python.Specific packet
It includes: the loss function of network architecture and model.
(3) network model training module is built on the deep learning frame TensorFlow based on Python.
(4) network model memory module is built on the deep learning frame TensorFlow based on Python.
The object ranging model that training is put up.It is specific as follows:
(1) pre-training sorter network: the pre-training one on ImageNet 1000-class competition dataset
A sorter network, this network are first 14 layers (network inputs are 452*452 at this time) in Fig. 7.
(2) training objective ranging model: trained model in (1) is saved, 19 after adding again on its basis
Layer network, random initializtion weight, and Multi-target Data library (the Multi-target database under different distance
Different distances (the Multi-target Data library under different distance)) under be trained.
(3) trained model is saved.
The actual effect of test target ranging model.Loss function is made to obtain minimum value by training pattern, thus right
Trained model is tested.
As a preferably embodiment, the present invention provides a kind of target ranging method based on deep learning,
By deep learning method, regression problem is converted by object ranging problem, to reach the measurement of target range.Method is specific
Steps are as follows:
Step 1: data acquisition platform is built.The focal length of i.e. fixed binocular vision camera position and its camera, acquisition are same
One target is in the video information apart from video camera different location.The sets of video data finally acquired should meet following requirement:
(1) single hop video length preferably must be held between 10 to 50 seconds.
(2) every section of video must be related to this method target (pedestrian, bicycle, motorcycle, automobile, bus,
Bird, cat, dog) at least one entity it is associated.
(3) by 80,000 sections of durations, equal video data does not form total video data.
(4) video total duration is 694 hours.
(5) video total size is 1.2T.
Step 2: video sub-frame processing is carried out to collected every section of video data, to obtain object in every frame and image
The depth information of body.
Step 3: demarcating the single-frame images of acquisition using calibration tool, and calibration content includes:
The bounding box center point coordinate (x, y) of detected target,
The bounding box width w and height h of detected target,
The bounding box classification information c of detected target,
It is detected the distance L of target range video camera.
Step 4: training set, test set are divided into according to a certain percentage to the data demarcated.
By Step 1: Step 2: step 3 processing after the available Multi-target Data library based under different distance
(Multi-target database at different distances (the Multi-target Data library under different distance)).
Step 5: the convolutional neural networks model based on deep learning is built.A total of 33 layers of the model, wherein preceding 25
Layer is characterized extract layer, the latter 8 layers prediction interval for being characterized information.
Step 6: the design of the loss function based on coordinate points.The loss function passes through Euclidean distance coordinates computed point
Penalty values are optimized to minimum in the training process by the penalty values between predicted value and true value.It is specific as follows:
Wherein:
I indicates i-th of grid, a total of S*S grid；
J indicates that j-th of bounding box in i-th of grid, a total of B are a；
Indicate the parameter value of the bounding box arrived by neural network forecast；
Indicate the true value of bounding-box perimeter；
Belong to the four-tuple being made of (x, y, w, h).
Step 7: the design of the loss function based on confidence level.The loss function calculates confidence level by error sum of squares
Predicted value and true value between penalty values, penalty values are optimized to minimum in the training process.It is specific as follows:
Wherein:
I indicates i-th of grid, a total of S*S grid；
J indicates that j-th of bounding box in i-th of grid, a total of B are a；
Indicate that j-th of predicted boundary frame is reliable in grid i；
Indicate the predicted value of confidence level；
CiIndicate the true value of confidence level.
Herein, the calculation formula of confidence level are as follows:
Wherein, Pr (Classi) indicate that the target in predicted boundary frame belongs to the probability of the i-th class；
Indicate the matching degree of predicted boundary frame and real border frame.
HereCalculation formula are as follows:
Wherein, BoxtruthIndicate the region of real border frame,
BoxpredIndicate the region of predicted boundary frame.
Step 8: the design of the loss function based on range information.The loss function calculates confidence by error sum of squares
Penalty values are optimized to minimum in the training process by the penalty values between the predicted value and true value of degree.It is specific as follows:
Wherein:
I indicates i-th of grid, a total of S*S grid；
J indicates that j-th of bounding box in i-th of grid, a total of B are a；
Indicate there is predicted target category in grid i；
Indicate the predicted value of range information；
Li(c) true value of range information is indicated.
Step 9: the design of the loss function based on classification information.The loss function calculates the target class by cross entropy
Penalty values are optimized to minimum in the training process by the penalty values between the predicted value and true value of other probability.It is specific as follows:
Wherein:
I indicates i-th of grid, a total of S*S grid；
J indicates that j-th of bounding box in i-th of grid, a total of B are a；
Indicate there is predicted target category in grid i；
Indicate the predicted value of target category probability；
Pi(c) true value of target category probability is indicated.
Step 10: the design of the total losses function of object ranging model.It is specific as follows:
Wherein,
λcoordIndicate the coefficient of the loss function based on coordinate points；
λconfidenceIndicate the coefficient of the loss function based on confidence level；
λdistanceIndicate the coefficient of the loss function based on range information；
λprobabilityIndicate the coefficient of the loss function based on classification information.
Mapping relations of the loss function on network model fc_33 layer are as shown in Figure 8.
Step 11: the training method of design object ranging model.
(1) pre-training sorter network: the pre-training one on ImageNet 1000-class competition dataset
A sorter network, this network are first 14 layers (network inputs are 452*452 at this time) in Fig. 7.
(2) training objective ranging model: trained model in (1) is saved, 19 after adding again on its basis
Layer network, random initializtion weight, and Multi-target Data library (the Multi-target database under different distance
Different distances (the Multi-target Data library under different distance)) training set under be trained.It is specific as follows:
(a) 256 samples of every batch of training, every just to randomly choose new dimension of picture by 10 batches of training, adjustment network arrives
Continue training pattern after respective dimensions.
(b) training maximum number of iterations is set as 100,000 times.
(c) learning rate is set as 0.0001.
(3) trained model is saved.
Step 12: the actual effect of test target ranging model.
Loss function is made to obtain minimum value by training pattern, to test trained model.
Experimental example
Technical progress in order to further illustrate the present invention is now further illustrated using comparative experiments.
For the distance measuring method based on deep learning, We conducted following groups experiments.
Experiment 1: the test comparison result when test object is pedestrian, under the different distance based on binocular vision camera
(as shown in the table):
Experiment 2: when test object is pedestrian, the test based on the distance measuring method of deep learning under different distance is compared
As a result (as shown in the table):
| Test serial number | Target actual depth (mm) | Target survey depth (mm) | Relative error |
| 1 | 1000 | 1025 | 2.5% |
| 2 | 2000 | 2034 | 1.7% |
| 3 | 3000 | 2976 | - 0.8% |
| 4 | 4000 | 4092 | 2.3% |
| 5 | 5000 | 5160 | 3.2% |
Pass through the above comparative experiments, it has been found that the ranging accuracy rate of the distance measuring method of the invention based on deep learning
Basic and binocular vision camera ranging accuracy rate is consistent.In particular, being based on deep learning with the growth of measuring distance
The accuracy rate of distance measuring method be higher than the accuracy rate based on traditional binocular vision camera.
Although describing the illustrative embodiments for specific application in the disclosure, it should be understood that, implement
Example is not limited to this.Other embodiments are possible, and the spirit and scope and embodiment instructed herein will have it is great
In the other field of function, embodiment can be modified.Further, when describe in conjunction with the embodiments special characteristic, structure or
When characteristic, it is considered that, regardless of whether being expressly recited, feature, structure or characteristic in conjunction with as other embodiments realization are in correlation
In the knowledge of field technical staff.
Claims (7)
1. a kind of target ranging method based on deep learning, it is characterised in that: include the following:
First step: the foundation based on the target database under different distance；
Second step: object ranging model is built；
Third step: the loss function of design object ranging model；
Four steps: the training method of design object ranging model；
5th step: trained object ranging model is tested.
2. the target ranging method according to claim 1 based on deep learning, it is characterised in that: the first step
In, the foundation of the target database specifically includes that (a) data acquire；(b) data scaling；Wherein the source of data acquisition is
The video information of specified video camera shooting；Data acquisition target include 8 classes, be respectively as follows: pedestrian, bicycle, motorcycle, automobile,
Bus, bird, cat, dog；Data scaling includes: the bounding box coordinates information of detected target, the bounding box class for being detected target
Other information, the depth information for being detected target.
3. the target ranging method according to claim 1 based on deep learning, it is characterised in that: the first step
In, the establishment step of the target database specifically includes that (1) carries out video data acquiring using binocular vision video camera；(2)
Video sub-frame processing is carried out to collected every section of video data, to obtain the depth letter of every frame image and objects in images
Breath；(3) single-frame images of acquisition is manually demarcated using calibration tool, calibration content includes: the boundary of detected target
Frame center point coordinate (x, y), the bounding box width w and height h that are detected target, the bounding box classification information c for being detected target
With the range information L of detected target range video camera；(4) data demarcated are divided into training according to a certain percentage
Collection, test set.
4. the target ranging method according to claim 1 based on deep learning, it is characterised in that: the second step
In, building for object ranging model refers to building for the model of the convolutional neural networks based on deep learning, and the model is a total of
33 layers, wherein first 25 layers are characterized extract layer, the latter 8 layers prediction interval for being characterized information.
5. the target ranging method according to claim 1 based on deep learning, it is characterised in that: the third step
In, the loss function of design object ranging model is specific as follows:
Wherein, λcoordIndicate the coefficient of the loss function based on coordinate points；λconfidanceIndicate the loss function based on confidence level
Coefficient；λdistanceIndicate the coefficient of the loss function based on range information；λprobabilityIndicate the loss based on classification information
The coefficient of function.
6. the target ranging method according to claim 1 based on deep learning, it is characterised in that: the four steps
In, the training method of design object ranging model is specific as follows: (1) pre-training sorter network: in ImageNet1000 class data
One sorter network of pre-training on collection, this sorter network are preceding 14 layer networks in second step；(2) training objective ranging mould
Type: trained model in (1) is saved, 19 layer network after adding again on its basis, random initializtion weight, and
Model training is carried out under the training set in the Multi-target Data library under different distance；(3) the Multi-target Data library under different distance
Training set under training method it is as follows: every batch of training 256 samples；Training maximum number of iterations is set as 100,000 times；Setting
Learning rate is 0.0001；Trained model is saved.
7. the target ranging method according to claim 1 based on deep learning, it is characterised in that: the 5th step
In, the actual effect of test target ranging model；Loss function is made to obtain minimum value by training pattern, thus to training
Model tested.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810562681.9A CN109029363A (en) | 2018-06-04 | 2018-06-04 | A kind of target ranging method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810562681.9A CN109029363A (en) | 2018-06-04 | 2018-06-04 | A kind of target ranging method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109029363A true CN109029363A (en) | 2018-12-18 |
Family
ID=64612033
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810562681.9A Pending CN109029363A (en) | 2018-06-04 | 2018-06-04 | A kind of target ranging method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109029363A (en) |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109859216A (en) * | 2019-02-16 | 2019-06-07 | 深圳市未来感知科技有限公司 | Distance measuring method, device, equipment and storage medium based on deep learning |
| CN110068308A (en) * | 2019-05-16 | 2019-07-30 | 合刃科技（深圳）有限公司 | A kind of distance measuring method and range-measurement system based on more mesh cameras |
| CN110136186A (en) * | 2019-05-10 | 2019-08-16 | 安徽工程大学 | A kind of detection target matching method for mobile robot object ranging |
| CN110807360A (en) * | 2019-09-19 | 2020-02-18 | 潍柴动力股份有限公司 | Method, device and system for detecting car stopping wall based on camera |
| CN111387932A (en) * | 2019-01-02 | 2020-07-10 | 中国移动通信有限公司研究院 | Vision detection method, device and equipment |
| CN111738088A (en) * | 2020-05-25 | 2020-10-02 | 西安交通大学 | Pedestrian distance prediction method based on monocular camera |
| CN111935389A (en) * | 2019-12-30 | 2020-11-13 | 苏州臻迪智能科技有限公司 | Shot object switching method and device, shooting equipment and readable storage medium |
| CN112153362A (en) * | 2020-09-15 | 2020-12-29 | 清华大学深圳国际研究生院 | Method and system for measuring stereoscopic depth of naked eye 3D display system |
| CN112257566A (en) * | 2020-10-20 | 2021-01-22 | 哈尔滨工程大学 | Artificial intelligence target identification ranging method based on big data |
| CN112699748A (en) * | 2020-12-21 | 2021-04-23 | 南京理工大学 | Human-vehicle distance estimation method based on YOLO and RGB image |
| CN112781562A (en) * | 2020-12-29 | 2021-05-11 | 中国北方车辆研究所 | Multi-target reconnaissance and ranging method based on monocular camera |
| CN113269118A (en) * | 2021-06-07 | 2021-08-17 | 重庆大学 | Monocular vision forward vehicle distance detection method based on depth estimation |
| CN113421298A (en) * | 2021-06-17 | 2021-09-21 | 深圳市高格通讯技术有限公司 | Vehicle distance measuring method, vehicle control device, vehicle and readable storage medium |
| CN113963322A (en) * | 2021-10-29 | 2022-01-21 | 北京百度网讯科技有限公司 | Detection model training method and device and electronic equipment |
| CN114930798A (en) * | 2019-12-30 | 2022-08-19 | 苏州臻迪智能科技有限公司 | Shooting object switching method and device, and image processing method and device |
| WO2022183484A1 (en) * | 2021-03-05 | 2022-09-09 | 华为技术有限公司 | Method and apparatus for determining object detection model |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107220603A (en) * | 2017-05-18 | 2017-09-29 | 惠龙易通国际物流股份有限公司 | Vehicle checking method and device based on deep learning |
| CN107316010A (en) * | 2017-06-13 | 2017-11-03 | 武汉理工大学 | A kind of method for recognizing preceding vehicle tail lights and judging its state |
| CN107392965A (en) * | 2017-07-24 | 2017-11-24 | 王华锋 | A kind of distance-finding method being combined based on deep learning and binocular stereo vision |
| US20180136332A1 (en) * | 2016-11-15 | 2018-05-17 | Wheego Electric Cars, Inc. | Method and system to annotate objects and determine distances to objects in an image |
-
2018
- 2018-06-04 CN CN201810562681.9A patent/CN109029363A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180136332A1 (en) * | 2016-11-15 | 2018-05-17 | Wheego Electric Cars, Inc. | Method and system to annotate objects and determine distances to objects in an image |
| CN107220603A (en) * | 2017-05-18 | 2017-09-29 | 惠龙易通国际物流股份有限公司 | Vehicle checking method and device based on deep learning |
| CN107316010A (en) * | 2017-06-13 | 2017-11-03 | 武汉理工大学 | A kind of method for recognizing preceding vehicle tail lights and judging its state |
| CN107392965A (en) * | 2017-07-24 | 2017-11-24 | 王华锋 | A kind of distance-finding method being combined based on deep learning and binocular stereo vision |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111387932A (en) * | 2019-01-02 | 2020-07-10 | 中国移动通信有限公司研究院 | Vision detection method, device and equipment |
| CN109859216B (en) * | 2019-02-16 | 2021-06-25 | 深圳市未来感知科技有限公司 | Distance measurement method, device and equipment based on deep learning and storage medium |
| CN109859216A (en) * | 2019-02-16 | 2019-06-07 | 深圳市未来感知科技有限公司 | Distance measuring method, device, equipment and storage medium based on deep learning |
| CN110136186A (en) * | 2019-05-10 | 2019-08-16 | 安徽工程大学 | A kind of detection target matching method for mobile robot object ranging |
| CN110136186B (en) * | 2019-05-10 | 2022-09-16 | 安徽工程大学 | Detection target matching method for mobile robot target ranging |
| CN110068308A (en) * | 2019-05-16 | 2019-07-30 | 合刃科技（深圳）有限公司 | A kind of distance measuring method and range-measurement system based on more mesh cameras |
| CN110068308B (en) * | 2019-05-16 | 2021-03-23 | 合刃科技（深圳）有限公司 | Distance measurement method and distance measurement system based on multi-view camera |
| CN110807360A (en) * | 2019-09-19 | 2020-02-18 | 潍柴动力股份有限公司 | Method, device and system for detecting car stopping wall based on camera |
| CN111935389A (en) * | 2019-12-30 | 2020-11-13 | 苏州臻迪智能科技有限公司 | Shot object switching method and device, shooting equipment and readable storage medium |
| CN114930798A (en) * | 2019-12-30 | 2022-08-19 | 苏州臻迪智能科技有限公司 | Shooting object switching method and device, and image processing method and device |
| CN111935389B (en) * | 2019-12-30 | 2021-11-19 | 苏州臻迪智能科技有限公司 | Shot object switching method and device, shooting equipment and readable storage medium |
| CN111738088A (en) * | 2020-05-25 | 2020-10-02 | 西安交通大学 | Pedestrian distance prediction method based on monocular camera |
| CN112153362A (en) * | 2020-09-15 | 2020-12-29 | 清华大学深圳国际研究生院 | Method and system for measuring stereoscopic depth of naked eye 3D display system |
| CN112257566A (en) * | 2020-10-20 | 2021-01-22 | 哈尔滨工程大学 | Artificial intelligence target identification ranging method based on big data |
| CN112257566B (en) * | 2020-10-20 | 2023-05-23 | 陕西合友网络科技有限公司 | Artificial intelligence target recognition ranging method based on big data |
| CN112699748A (en) * | 2020-12-21 | 2021-04-23 | 南京理工大学 | Human-vehicle distance estimation method based on YOLO and RGB image |
| CN112699748B (en) * | 2020-12-21 | 2022-09-27 | 南京理工大学 | Human-vehicle distance estimation method based on YOLO and RGB image |
| CN112781562A (en) * | 2020-12-29 | 2021-05-11 | 中国北方车辆研究所 | Multi-target reconnaissance and ranging method based on monocular camera |
| WO2022183484A1 (en) * | 2021-03-05 | 2022-09-09 | 华为技术有限公司 | Method and apparatus for determining object detection model |
| CN113269118A (en) * | 2021-06-07 | 2021-08-17 | 重庆大学 | Monocular vision forward vehicle distance detection method based on depth estimation |
| CN113421298A (en) * | 2021-06-17 | 2021-09-21 | 深圳市高格通讯技术有限公司 | Vehicle distance measuring method, vehicle control device, vehicle and readable storage medium |
| CN113963322A (en) * | 2021-10-29 | 2022-01-21 | 北京百度网讯科技有限公司 | Detection model training method and device and electronic equipment |
| CN113963322B (en) * | 2021-10-29 | 2023-08-25 | 北京百度网讯科技有限公司 | Detection model training method and device and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109029363A (en) | A kind of target ranging method based on deep learning | |
| CN110378381B (en) | Object detection method, device and computer storage medium | |
| CN111046880B (en) | Infrared target image segmentation method, system, electronic equipment and storage medium | |
| CN105809198B (en) | SAR image target recognition method based on depth confidence network | |
| EP3614308A1 (en) | Joint deep learning for land cover and land use classification | |
| CN108830188A (en) | Vehicle checking method based on deep learning | |
| CN109902646A (en) | A kind of gait recognition method based on long memory network in short-term | |
| US11694354B2 (en) | Geospatial object geometry extraction from imagery | |
| CN110135354B (en) | Change detection method based on live-action three-dimensional model | |
| CN110517311A (en) | Pest and disease monitoring method based on leaf spot lesion area | |
| US10943352B2 (en) | Object shape regression using wasserstein distance | |
| CN109766873B (en) | Pedestrian re-identification method based on hybrid deformable convolution | |
| Zhang et al. | A GANs-based deep learning framework for automatic subsurface object recognition from ground penetrating radar data | |
| CN113160062B (en) | Infrared image target detection method, device, equipment and storage medium | |
| CN112395987A (en) | SAR image target detection method based on unsupervised domain adaptive CNN | |
| CN110096982A (en) | A kind of video frequency vehicle big data searching method based on deep learning | |
| CN110992378B (en) | Dynamic updating vision tracking aerial photographing method and system based on rotor flying robot | |
| Digumarti et al. | Automatic segmentation of tree structure from point cloud data | |
| Wang et al. | Study on the method of transmission line foreign body detection based on deep learning | |
| Martins et al. | Deep learning applied to the identification of fruit fly in intelligent traps | |
| CN111444816A (en) | Multi-scale dense pedestrian detection method based on fast RCNN | |
| Alburshaid et al. | Palm trees detection using the integration between gis and deep learning | |
| CN111242134A (en) | Remote sensing image ground object segmentation method based on feature adaptive learning | |
| Maddileti et al. | Pseudo Trained YOLO R_CNN Model for Weapon Detection with a Real-Time Kaggle Dataset | |
| CN112668662B (en) | Outdoor mountain forest environment target detection method based on improved YOLOv3 network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181218 |
|
| RJ01 | Rejection of invention patent application after publication |