CN110378239A - A kind of real-time traffic marker detection method based on deep learning - Google Patents
A kind of real-time traffic marker detection method based on deep learning Download PDFInfo
- Publication number
- CN110378239A CN110378239A CN201910552184.5A CN201910552184A CN110378239A CN 110378239 A CN110378239 A CN 110378239A CN 201910552184 A CN201910552184 A CN 201910552184A CN 110378239 A CN110378239 A CN 110378239A
- Authority
- CN
- China
- Prior art keywords
- real
- convolution
- deep learning
- detection method
- method based
- 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
- 238000001514 detection method Methods 0.000 title claims abstract description 27
- 238000013135 deep learning Methods 0.000 title claims abstract description 16
- 239000003550 marker Substances 0.000 title claims abstract description 12
- 238000003062 neural network model Methods 0.000 claims abstract description 25
- 230000006870 function Effects 0.000 claims abstract description 12
- 238000010606 normalization Methods 0.000 claims description 12
- 230000004927 fusion Effects 0.000 claims description 10
- 238000013461 design Methods 0.000 claims description 7
- 238000005070 sampling Methods 0.000 claims description 6
- 238000001914 filtration Methods 0.000 claims description 3
- 238000004321 preservation Methods 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 2
- 238000012545 processing Methods 0.000 claims description 2
- 238000000034 method Methods 0.000 abstract description 4
- 238000004880 explosion Methods 0.000 abstract description 3
- 230000001629 suppression Effects 0.000 abstract description 3
- 238000012549 training Methods 0.000 abstract description 3
- 238000000605 extraction Methods 0.000 abstract description 2
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 8
- 238000013527 convolutional neural network Methods 0.000 description 4
- 230000004913 activation Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000005286 illumination Methods 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 206010039203 Road traffic accident Diseases 0.000 description 1
- 238000007792 addition Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000000151 deposition Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000006073 displacement reaction Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005562 fading Methods 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000011897 real-time detection Methods 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Classifications
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
- G06V20/582—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of traffic signs
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of real-time traffic marker detection method based on deep learning, this method devises a kind of improved neural network model, PW convolution, DW convolution are merged to reduce parameter amount in the front end portion of model basis convolution module, promote detection speed, rear end is using global average pondization and full articulamentum, make model and obtain automatically by way of study the significance level in each feature channel, improves the ability of model extraction feature；It is disappeared and gradient explosion issues using the gradient that residual error structure solves deep layer network simultaneously；Model-aided network uses FPN structure, robustness of the lift scheme for different size road traffic sign detections；Model replaces batch dimension to be normalized using channel dimension, reduces batch size to trained influence；Simultaneously using softening non-maxima suppression, traffic sign omission factor is reduced；And intersection entropy loss is improved, easy classification samples are inhibited in loss function, model is made more to focus on difficult classification samples in the training process.
Description
Technical field
The invention belongs to automatic Pilot environment sensing fields, and in particular to a kind of real-time traffic mark based on deep learning
Detection method.
Background technique
With constantly improve for municipal intelligent traffic, to the quick detection of traffic sign become in automatic Pilot one it is non-
Often important work.Road traffic sign detection system (TSRS) is the important component of advanced driving assistance system (ADAS),
There is highly important effect in terms of reducing traffic accident.Due to the complexity of road traffic scene, such as dimensional variation, movement
It obscures, block, illumination, colour fading etc., real-time traffic sign robust detection is still a huge challenge.
Traffic sign has significant color characteristic and shape feature, can first specific color space (such as HIS,
HSV, RGB etc.) it is middle using color threshold extraction traffic sign candidate region, it is then further according to contour feature and area features
It is sent into after screening in the classifier of pre-training and realizes fine grit classification；But the road traffic sign detection based on color and shape feature
Computation complexity is higher, takes a long time, and robustness is poor.
Due to the fast development of deep learning, convolutional neural networks achieve huge success in computer vision field,
There are many outstanding algorithm of target detection.Current target detection frame is generally divided into two kinds: 1. based on the single-order of recurrence
Section algorithm (SSD, YOLO etc.) can be trained end to end by positioning object to the intensive detection of original image；②
Based on the dual-stage algorithm (R-CNN, Fast R-CNN, Faster R-CNN etc.) that candidate region is extracted, with Area generation network
(RPN) sparse object candidate area is extracted, then candidate region is sent into network and carries out predicted position and classification.Due to depositing
The limitation in space and power consumption is stored up, the storage and calculating of single phase and dual-stage algorithm are all larger, and detection speed is slower.
Summary of the invention
Aiming at the problems existing in the prior art, the real-time traffic mark inspection based on deep learning that the present invention provides a kind of
Survey method solves the problems, such as that road traffic sign detection real-time is poor.
The present invention achieves the above technical objects by the following technical means.
A kind of real-time traffic marker detection method based on deep learning, after acquiring the picture comprising traffic sign and processing
Traffic sign data set is obtained, improved neural network model is created, using traffic sign data set to improved neural network
Model is trained, and the model parameter of preservation is imported in improved neural network model, is carried out to the traffic sign in picture
Real-time detection.
Further, the creation of the improved neural network model includes design basis convolution module, auxiliary convolutional layer, returns
One changes mode and improves intersection entropy loss；It is described basis convolution module front end portion use DW convolution sum PW convolution, DW volumes
For product for filtering, PW convolution is used for ALT-CH alternate channel, and rear end part passes through global average pondization and two full articulamentums is combined to obtain often
The weight in a feature channel, is normalized weight using sigmoid function, and the Weight after normalization is led to each
In the feature in road；The design of the auxiliary convolutional layer specifically: by 26 × 26 characteristic patterns up-sampling and 52 × 52 of basic network
Characteristic pattern carries out Fusion Features, and 13 × 13 characteristic patterns up-sampling carries out Fusion Features with 26 × 26 characteristic patterns, exists respectively after fusion
It is detected on three various sizes of characteristic patterns；The normalization is changed to feature channel dimension by the normalization of batch dimension
On normalization；The improved cross entropy loss function expression formula are as follows:
In formula: γ is adjustment parameter, and y is true tag, and y' is that the bounding box of prediction is the probability of positive sample.
Further, when the traffic sign in the picture is measured in real time, there are multiple pre- for each target in picture
Bounding box is surveyed, rejects extra predicted boundary frame using softening non-maxima suppression.
The beneficial effects of the present invention are:
1, real-time is preferable: the present invention uses PW convolution sum DW convolution, reduces computation complexity, passes through the average pond of the overall situation
Change and the combination of two full articulamentums makes network and obtain automatically by way of study the significance level in each feature channel, mentions
It rises useful feature and inhibits useless feature, the gradient during neural network model is deepened is avoided by residual error structure and is disappeared
Gradient of becoming estranged explosion, reduces hardware requirement, real-time is effectively promoted, can detect the traffic mark on road in time
Will makes driver give warning in advance, and reduces the generation of road accident.
2, robustness is higher: present invention improves over normalization modes, are normalized in the level of channel, avoid batch
Dimension reduces influence of the batch size to neural network model；Simultaneously using FPN structure building auxiliary convolution module, losing
Easy classification samples are inhibited in function, so that neural network model is more focused on difficult classification samples, is enhanced by a variety of data
Mode increases scene diversity, and the road traffic sign detection relative to tradition based on color and shape, robustness of the present invention is more preferable, can
With detect different weather, it is different block, the traffic sign of different scale.
Detailed description of the invention
Fig. 1 is a kind of real-time traffic Mark Detection flow chart based on deep learning of the invention；
Fig. 2 is the structure chart of improved neural network model；
Fig. 3 is improved basic convolution module figure；
Fig. 4 is test picture visualization figure.
Specific embodiment
The solution of the present invention is described in detail with reference to the accompanying drawings and detailed description, but protection of the invention
Range is not limited to this.
As shown in Figure 1, a kind of real-time traffic marker detection method based on deep learning, comprising the following steps:
Step (1), acquisition comprising traffic sign picture (picture of acquisition need to comprising the picture under various scenes, such as
Different roads, different weather, shooting of different angle etc.), and the traffic sign in picture is labeled using annotation tool,
Obtain traffic sign data set；Traffic sign data set is organized into VOC data set format, VOC data set format includes 3 sons
File, respectively JPEGImages file, Annotations file, ImageSets file, JPEGImages store picture,
Annotations stores the mark file of xml type, and ImageSets file stores txt text, every a line of txt text corresponding one
The title of a picture, improved neural network model read filename according to txt text, then to JPEGImages and
Corresponding picture and mark are found in Annotations file, and the mark of traffic sign is extracted in the picture mark searched out
Information obtains the frame parameter of markup information.
Step (2), is clustered by the frame to the markup information in traffic sign data set, obtains 9 anchor frame rulers
It is very little, respectively (10 × 13), (16 × 30), (33 × 23), (30 × 61), (62 × 45), (59 × 119), (116 × 90),
(156×198)、(373×326)。
Picture, is randomly assigned to different batches by step (3), before being sent into improved neural network model, to picture
It into row stochastic rotation, cuts, adjustment illumination variation, to expand picture scene diversity, and dimension of picture is uniformly adjusted to
416×416。
Step (4), design basis convolution module, as shown in figure 3, the front end portion of basic convolution module uses PW convolution
(Depthwise Convolution) and DW convolution (Pointwise Convolution), DW convolution are responsible for filtering, PW convolution
It is responsible for ALT-CH alternate channel.PW convolution sum DW convolution is relative to Standard convolution parameter amount compression ratio are as follows:
In formula, DKIndicate convolution kernel size, M is input channel number, and N is output channel number.
Since the calculation amount of DW convolution is smaller, PW convolution is added before DW convolution to promote port number, obtain picture
More features after DW convolution, then with PW convolution restores port number.Then feature pressure is carried out to picture by global average pond
Contracting makes the feature in each feature channel become a real number, this real number has global receptive field to a certain extent, and
And the dimension of global average pondization output and the feature port number of input match, and then first full articulamentum is logical to feature
Road is compressed, then with a full articulamentum enlarging property port number, and doing so has more non-linear, can preferably be intended
The correlation of feature interchannel complexity is closed, the weight between 0~1 is then obtained by sigmoid function, then will be after normalization
On Weight to the feature in each channel, finally by fusion residual error network, is connected using jump, constitute residual block, solved
Gradient disappears and gradient explosion issues, and improved neural network model is made to guarantee good performance while deepening network.In addition,
When feature port number is lower, activation primitive Relu can make the tensor value 0 of the feature on some channel, and characteristic information is caused to damage
It loses, therefore replaces the Relu activation primitive of port number small layers with linear transformation layer after PW convolution.Improved neural network mould
The infrastructure network of type is as shown in table 1:
The infrastructure network table of the improved neural network model of table 1
| Input | Convolution | Output channel number | Number of repetition | Step-length |
| (416,416,3) | Conv2d (3,3) | 32 | 1 | 2 |
| (208,208,32) | Inverted Residual block | 16 | 1 | 1 |
| (208,208,16) | Inverted Residual block | 24 | 2 | 2 |
| (104,104,24) | Inverted Residual block | 32 | 3 | 2 |
| (52,52,32) | Inverted Residual block | 64 | 4 | 2 |
| (26,26,64) | Inverted Residual block | 96 | 3 | 1 |
| (26,26,96) | Inverted Residual block | 160 | 3 | 2 |
| (13,13,160) | Inverted Residual block | 320 | 1 | 1 |
Step (5), Design assistant convolutional layer
FPN module as shown in Figure 2, because basic network (is made of) low-level image feature with high-resolution basic convolution module
Rate, traffic sign in the picture location information in low-level image feature is more accurate, and receptive field is larger, so being suitble to the big object of prediction
Body, and basic network high-level characteristic semantic information is relatively abundant, and receptive field is smaller, so be suitble to prediction wisp, therefore by base
26 × 26 characteristic patterns up-sampling and 52 × 52 characteristic patterns of plinth network carry out Fusion Features, 13 × 13 characteristic patterns up-sampling with 26 ×
26 characteristic patterns carry out Fusion Features, are detected on three various sizes of characteristic patterns respectively after fusion.
Step (6) designs normalization mode
Common normalization is to carry out in the dimension of batch, but this dimension is variation, when picture training set and
When test set distribution is different, error generation, normalized general formulae will lead to are as follows:
X is the feature calculated by layer in formula, and μ and σ are average and standard deviations, when batch size is smaller, will lead to meter
Mean value and the variance inaccuracy of calculating, therefore the present invention is normalized on feature channel dimension, avoids batch dimension, meter
Each group of feature channel dimension of mean value and variance are calculated, influence of the batch size to neural network model is reduced, further increases
Neural network model performance.
3 fused characteristic patterns in step (5) are divided into grid, each neural network forecast 3 by pixel by step (7)
Bounding box, such as table 2, each characteristic pattern uses 3 anchor frames, improved neural network model output for prediction bounding box information,
Confidence level and class probability, bounding box information are the relative displacement of the grid top left co-ordinate of predicted boundary frame and prediction target
(tx,ty,tw,th), wherein tx、tyIndicate the central point cross of predicted boundary frame, the offset of ordinate, tw、thIndicate prediction side
The central point cross of prediction block, width, the height of ordinate and prediction block can be calculated according to formula (3) in boundary's frame width, high offset
(bx,by,bw,bh), in which:
Wherein: cx,cyOffset for current grid relative to current signature figure upper left corner grid, σ () function are
Logistic function is used to tx、tyIt is normalized between 0-1, pw,phIt is to hand over and with mark bounding box than maximum anchor frame
It is wide and high.
The matching of table 2 characteristic pattern and anchor frame
| Characteristic pattern | Receptive field | Anchor frame |
| 13×13 | Greatly | (116,90)(156,198)(373,326) |
| 26×26 | In | (30,61)(62,45)(59,119) |
| 52×52 | It is small | (10,13)(16,30)(33,23) |
All predicted boundary frames of improved neural network model are divided into positive sample (with mark bounding box area by step (8)
Hand over and compare>0.5 in domain) it (is handed over and with mark boxed area than<0.4) with negative sample, costing bio disturbance is carried out, in order to keep loss more next
Smaller, each batch, which is sent into improved neural network model, can all update Model Weight, until penalty values convergence, every iteration 100
Model parameter of secondary preservation.And under normal circumstances, the ratio that target accounts in picture is much smaller than the ratio that background accounts for, so just
Negative sample ratio difference is larger, and is all largely the negative sample easily classified, and loss function at this time is in a large amount of simple samples
Iterative process in relatively slowly and possibly can not be optimized to it is optimal.The present invention is directed to difficult, the easy unbalanced problem of classification samples number,
It is improved on the basis of standard intersects entropy loss, standard cross entropy loss function are as follows:
Y' is the probability that the bounding box of prediction is positive sample in formula, between 0-1；Y is true tag, it can be seen that is being handed over
The bigger loss of output probability for pitching positive sample in entropy loss is smaller, and it is smaller that the output probability of negative sample gets over small loss.It is improved
Shown in cross entropy loss function such as formula (5):
γ is adjustment parameter in formula, the rate that easy classified weight reduces is adjusted, so that improved neural network model is being instructed
More focus on difficult classification samples when practicing.
The model parameter saved in step (8) is imported in improved neural network model, is surveyed by step (9), deconditioning
Attempt piece and be sent into improved neural network model, exports the bounding box information and class probability of prediction.
Step (10), each target in picture has multiple predicted boundary frames at this time, and common non-maxima suppression will be pre-
It surveys bounding box to be ranked up by confidence level, retains the highest predicted boundary frame of confidence level, calculate other predicted boundary frames and set
The friendship of the highest predicted boundary frame of reliability and ratio are deleted the predicted boundary frame, are done so if handed over and than being greater than a certain threshold value
Be easy to cause the missing inspection of target, therefore the present invention is to reduce the omission factor of model, using softening non-maximum, will and confidence level
The confidence score of the friendship of highest predicted boundary frame and the predicted boundary frame than being greater than some threshold value reduces, rather than sets 0,
This step is recycled in remaining predicted boundary frame, finally obtains the corresponding predicted boundary frame of each traffic sign.
Step (11) visualizes detection picture with opencv, as shown in Figure 4.
Specific implementation in the present invention is merely illustrative for the purpose of the present invention, and not restrictive, belonging to the present invention
Those skilled in the art can make various modifications or additions to the described embodiments or using similar
Mode substitutes, but can't deviate spirit or beyond the scope defined by the appended claims of the invention.
Claims (7)
1. a kind of real-time traffic marker detection method based on deep learning, which is characterized in that acquisition includes the figure of traffic sign
Traffic sign data set is obtained after piece and processing, improved neural network model is created, using traffic sign data set to improvement
Neural network model be trained, the model parameter of preservation is imported in improved neural network model, to the friendship in picture
Logical mark is measured in real time.
2. the real-time traffic marker detection method based on deep learning as described in claim 1, which is characterized in that the improvement
Neural network model creation include design basis convolution module, auxiliary convolutional layer, normalization mode and improve cross entropy
Loss.
3. the real-time traffic marker detection method based on deep learning as claimed in claim 2, which is characterized in that the basis
The front end portion of convolution module uses DW convolution sum PW convolution, and for DW convolution for filtering, PW convolution is used for ALT-CH alternate channel, rear end
It is divided to the weight for combining two full articulamentums to obtain each feature channel by global average pondization, using sigmoid function to power
It is normalized again, it will be in the feature of the Weight after normalization to each channel.
4. the real-time traffic marker detection method based on deep learning as claimed in claim 2, which is characterized in that the auxiliary
The design of convolutional layer specifically: 26 × 26 characteristic patterns up-sampling and 52 × 52 characteristic patterns of basic network are subjected to Fusion Features,
13 × 13 characteristic patterns up-sampling carries out Fusion Features with 26 × 26 characteristic patterns, respectively in three various sizes of characteristic patterns after fusion
On detected.
5. the real-time traffic marker detection method based on deep learning as claimed in claim 2, which is characterized in that the normalizing
Change and the normalization on feature channel dimension is changed to by the normalization of batch dimension.
6. the real-time traffic marker detection method based on deep learning as claimed in claim 2, which is characterized in that the improvement
Cross entropy loss function expression formula are as follows:
In formula: γ is adjustment parameter, and y is true tag, and y' is that the bounding box of prediction is the probability of positive sample.
7. the real-time traffic marker detection method based on deep learning as described in claim 1, which is characterized in that the picture
In traffic sign when being measured in real time, each target in picture there are multiple predicted boundary frames, using soften it is non-greatly
Value inhibits to reject extra predicted boundary frame.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910552184.5A CN110378239A (en) | 2019-06-25 | 2019-06-25 | A kind of real-time traffic marker detection method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910552184.5A CN110378239A (en) | 2019-06-25 | 2019-06-25 | A kind of real-time traffic marker detection method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110378239A true CN110378239A (en) | 2019-10-25 |
Family
ID=68249230
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910552184.5A Pending CN110378239A (en) | 2019-06-25 | 2019-06-25 | A kind of real-time traffic marker detection method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110378239A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111160205A (en) * | 2019-12-24 | 2020-05-15 | 江苏大学 | Embedded multi-class target end-to-end unified detection method for traffic scene |
| CN111191608A (en) * | 2019-12-30 | 2020-05-22 | 浙江工业大学 | Improved traffic sign detection and identification method based on YOLOv3 |
| CN111243575A (en) * | 2020-01-15 | 2020-06-05 | 北京工业大学 | Dialect species identification method based on expanded convolutional neural network |
| CN111274970A (en) * | 2020-01-21 | 2020-06-12 | 南京航空航天大学 | Traffic sign detection method based on improved YOLO v3 algorithm |
| CN111340223A (en) * | 2020-02-27 | 2020-06-26 | 上海商汤临港智能科技有限公司 | Neural network compression method, target detection method, driving control method and device |
| CN111767878A (en) * | 2020-07-03 | 2020-10-13 | 中国科学院自动化研究所 | Deep learning-based traffic sign detection method and system in embedded device |
| CN111831956A (en) * | 2020-06-16 | 2020-10-27 | 五邑大学 | Method for adjusting high-degree-of-freedom imbalance-like loss function and storage medium |
| CN113506293A (en) * | 2021-09-08 | 2021-10-15 | 成都数联云算科技有限公司 | Image processing method, device, equipment and storage medium |
| CN113628759A (en) * | 2021-07-22 | 2021-11-09 | 中国科学院重庆绿色智能技术研究院 | Infectious disease epidemic situation safety region prediction method based on big data |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108229442A (en) * | 2018-02-07 | 2018-06-29 | 西南科技大学 | Face fast and stable detection method in image sequence based on MS-KCF |
| CN109685017A (en) * | 2018-12-26 | 2019-04-26 | 中山大学 | A kind of ultrahigh speed real-time target detection system and detection method based on light weight neural network |
| CN109801265A (en) * | 2018-12-25 | 2019-05-24 | 国网河北省电力有限公司电力科学研究院 | A kind of real-time transmission facility foreign matter detection system based on convolutional neural networks |
-
2019
- 2019-06-25 CN CN201910552184.5A patent/CN110378239A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108229442A (en) * | 2018-02-07 | 2018-06-29 | 西南科技大学 | Face fast and stable detection method in image sequence based on MS-KCF |
| CN109801265A (en) * | 2018-12-25 | 2019-05-24 | 国网河北省电力有限公司电力科学研究院 | A kind of real-time transmission facility foreign matter detection system based on convolutional neural networks |
| CN109685017A (en) * | 2018-12-26 | 2019-04-26 | 中山大学 | A kind of ultrahigh speed real-time target detection system and detection method based on light weight neural network |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111160205A (en) * | 2019-12-24 | 2020-05-15 | 江苏大学 | Embedded multi-class target end-to-end unified detection method for traffic scene |
| CN111160205B (en) * | 2019-12-24 | 2023-09-05 | 江苏大学 | Method for uniformly detecting multiple embedded types of targets in traffic scene end-to-end |
| CN111191608A (en) * | 2019-12-30 | 2020-05-22 | 浙江工业大学 | Improved traffic sign detection and identification method based on YOLOv3 |
| CN111243575A (en) * | 2020-01-15 | 2020-06-05 | 北京工业大学 | Dialect species identification method based on expanded convolutional neural network |
| CN111274970A (en) * | 2020-01-21 | 2020-06-12 | 南京航空航天大学 | Traffic sign detection method based on improved YOLO v3 algorithm |
| CN111340223A (en) * | 2020-02-27 | 2020-06-26 | 上海商汤临港智能科技有限公司 | Neural network compression method, target detection method, driving control method and device |
| CN111831956A (en) * | 2020-06-16 | 2020-10-27 | 五邑大学 | Method for adjusting high-degree-of-freedom imbalance-like loss function and storage medium |
| CN111831956B (en) * | 2020-06-16 | 2024-05-28 | 五邑大学 | Method for adjusting high-degree-of-freedom class unbalance loss function and storage medium |
| CN111767878A (en) * | 2020-07-03 | 2020-10-13 | 中国科学院自动化研究所 | Deep learning-based traffic sign detection method and system in embedded device |
| CN113628759A (en) * | 2021-07-22 | 2021-11-09 | 中国科学院重庆绿色智能技术研究院 | Infectious disease epidemic situation safety region prediction method based on big data |
| CN113506293A (en) * | 2021-09-08 | 2021-10-15 | 成都数联云算科技有限公司 | Image processing method, device, equipment and storage medium |
| CN113506293B (en) * | 2021-09-08 | 2021-12-07 | 成都数联云算科技有限公司 | Image processing method, device, equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110378239A (en) | A kind of real-time traffic marker detection method based on deep learning | |
| CN109190636B (en) | Remote sensing image ship target information extraction method | |
| CN111160440B (en) | Deep learning-based safety helmet wearing detection method and device | |
| CN102096821B (en) | Number plate identification method under strong interference environment on basis of complex network theory | |
| CN111553201B (en) | Traffic light detection method based on YOLOv3 optimization algorithm | |
| CN109325418A (en) | Based on pedestrian recognition method under the road traffic environment for improving YOLOv3 | |
| CN109344821A (en) | Small target detecting method based on Fusion Features and deep learning | |
| CN106557579B (en) | Vehicle model retrieval system and method based on convolutional neural network | |
| CN108960198A (en) | A kind of road traffic sign detection and recognition methods based on residual error SSD model | |
| CN107167811A (en) | The road drivable region detection method merged based on monocular vision with laser radar | |
| CN105335702B (en) | A kind of bayonet model recognizing method based on statistical learning | |
| CN105354568A (en) | Convolutional neural network based vehicle logo identification method | |
| CN109118500A (en) | A kind of dividing method of the Point Cloud Data from Three Dimension Laser Scanning based on image | |
| CN109508675B (en) | Pedestrian detection method for complex scene | |
| CN106022232A (en) | License plate detection method based on deep learning | |
| CN106447658A (en) | Significant target detection method based on FCN (fully convolutional network) and CNN (convolutional neural network) | |
| CN110852347A (en) | Fire detection method using improved YOLO v3 | |
| CN110349260B (en) | Automatic pavement marking extraction method and device | |
| CN109101981B (en) | Loop detection method based on global image stripe code in streetscape scene | |
| CN104809481A (en) | Natural scene text detection method based on adaptive color clustering | |
| CN109583483A (en) | A kind of object detection method and system based on convolutional neural networks | |
| CN111967313B (en) | Unmanned aerial vehicle image annotation method assisted by deep learning target detection algorithm | |
| CN115205264A (en) | High-resolution remote sensing ship detection method based on improved YOLOv4 | |
| CN106897677B (en) | Vehicle feature classification retrieval system and method | |
| CN103093240A (en) | Calligraphy character identifying method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191025 |