CN111738300A - Optimization algorithm for detecting and identifying traffic signs and signal lamps - Google Patents

Optimization algorithm for detecting and identifying traffic signs and signal lamps Download PDF

Info

Publication number
CN111738300A
CN111738300A CN202010463642.0A CN202010463642A CN111738300A CN 111738300 A CN111738300 A CN 111738300A CN 202010463642 A CN202010463642 A CN 202010463642A CN 111738300 A CN111738300 A CN 111738300A
Authority
CN
China
Prior art keywords
classification
feature
stage
candidate
region
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
Application number
CN202010463642.0A
Other languages
Chinese (zh)
Inventor
王卓曜
金城
刀坤
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Fudan University
Original Assignee
Fudan University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Fudan University filed Critical Fudan University
Priority to CN202010463642.0A priority Critical patent/CN111738300A/en
Publication of CN111738300A publication Critical patent/CN111738300A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133—Distances to prototypes
    • G06F18/24137—Distances to cluster centroïds
    • G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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
    • 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
    • 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
    • G06V20/584—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of vehicle lights or traffic lights

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 invention discloses an optimization algorithm for detecting and identifying traffic signs and signal lamps; the algorithm is divided into a feature extraction stage, a region candidate stage and a hierarchical classification stage. In the feature extraction stage, a Ghost bottleeck module is introduced to construct a feature extraction network, big data training is carried out on ImageNet to obtain a pre-training model, image features are extracted through the pre-trained feature extraction network, and the feature graph is subjected to pooling treatment; in the region candidate stage, an RPN subnet is adopted to obtain a candidate region, and the feature graphs corresponding to the candidate region are cut and scaled, so that the feature subgraphs to be classified have the same size; and in the hierarchical classification stage, the images are classified by a hierarchical classification method, and traffic lights and traffic signs are identified. Compared with a baseline algorithm, the algorithm disclosed by the invention has the advantages that the performance in all aspects is greatly improved, and the requirements on the real-time performance and the reliability of an automatic driving system can be met.

Description

Optimization algorithm for detecting and identifying traffic signs and signal lamps
Technical Field
The invention belongs to the technical field of statistical pattern recognition and image processing, and particularly relates to an optimization algorithm for traffic sign and signal lamp detection and recognition.
Background
Traffic sign and signal light detection are important components of the perception of road scenes for autonomous vehicles. The key problem of traffic sign and signal lamp detection is that the target is positioned and identified under the condition of meeting the real-time requirement, in particular to a technology for checking whether a target object exists in a complex fast moving image sequence and accurately and fast calculating the position of the target in the image, and the main problems to be solved are target identification and positioning under the conditions of complex illumination, complex background, multiple scales, multiple visual angles, shielding and the like.
For the identification of traffic lights and traffic signs, most studies currently view it as two distinct problems: namely traffic light identification and traffic sign identification. The traditional traffic signal lamp identification method mostly depends on visual characteristics such as color and shape of a signal lamp, so that the applicability is poor; the deep learning method is regarded as a special case of target detection, for example, a YOLO target detection algorithm is adopted to detect and identify traffic lights. The development of the traffic sign detection and identification method is similar, but the calculation cost is high by using a target detection method based on deep learning, the calculation amount cannot be reduced, a GPU with high power consumption and high calculation power is required to be used for calculation, and the real-time requirement is difficult to meet due to the power supply limitation of a vehicle-mounted power supply.
Disclosure of Invention
The invention aims to provide a traffic sign and traffic signal lamp detection and identification optimization algorithm based on deep learning; the algorithm is based on a hierarchical and fine-classified network structure of a lightweight backbone network, simultaneously positions and identifies traffic signal lamps and traffic signs in the image, can meet the requirements of real-time performance and accuracy, can be applied to a real-time automatic driving system, and can give early warning and limit to the road driving behavior of an automatic driving automobile so as to avoid traffic accidents. The technical scheme of the invention is specifically introduced as follows.
An optimization algorithm for detecting and identifying traffic signs and signal lamps is divided into three stages: a feature extraction stage, a region candidate stage and a hierarchical classification stage; in the feature extraction stage, a Ghost bottleeck module of GhostNet is introduced to construct a feature extraction network, big data training is carried out on a classical image classification data set ImageNet to obtain a pre-training model, rich semantic information contained in an image is extracted through the pre-trained feature extraction network, and a feature graph is subjected to pooling processing; in the Region candidate stage, a RPN subnet in the Faster-RCNN is adopted to obtain a candidate Region (Region Proposal), and feature graphs corresponding to the candidate Region are cut and scaled, so that all feature subgraphs to be classified have the same size; the hierarchical classification stage classification subnet classifies the images by a two-stage progressive hierarchical classification method so as to identify traffic lights and traffic signs in the images.
In the invention, the specific steps of the characteristic extraction stage are as follows:
(1) constructing a feature expression network, extracting the most basic original features of an image from an input image through a small common convolution kernel, inputting an output feature graph into a series of Ghost bottleeck modules, gradually increasing channels, dividing the Ghost bottleeck into different stages according to the size of the input feature graph, finally converting the feature graph into feature vectors by using a global average pool and a convolution layer for final classification, and performing big data training on a classical image classification data set ImageNet image classification training by using the feature expression network to obtain a pre-training model;
(2) separating the feature extraction network, taking 1-18 layers of the pre-trained feature expression network as feature extraction layers, fixing the parameters, gradients and Batchnormal coefficients of the first five layers, and finally performing average pooling treatment on the output feature map.
In the invention, the specific steps of the region candidate stage are as follows:
(1) acquiring a candidate region, acquiring the candidate region by adopting an RPN subnet in FasterR-CNN in a region candidate stage, taking a pooled feature map obtained in a feature extraction stage as input, performing convolution rich semantic information firstly, and then sending the information into an RPN network to acquire the coordinate of the candidate region;
(2) positive and negative candidate regions are distinguished
Figure BDA0002511905430000021
Indicating that the RPN network is locatedG represents all group Truth in the image, and the correct candidate area obtained by distinguishing RPN network
Figure BDA0002511905430000022
And wrong candidate region
Figure BDA0002511905430000023
Is defined as:
Figure BDA0002511905430000024
Figure BDA0002511905430000025
(3) screening positive candidate regions, sorting all candidate regions from large to small according to confidence degrees, and selecting the first 300 candidate regions to be sent to a subsequent classification stage;
(4) and obtaining a candidate region feature map, converting the coordinate expression of the candidate region into a feature map corresponding to the region, and scaling the feature map corresponding to the candidate region before the candidate region is sent into the subnet to 7 multiplied by 7, so that all feature subgraphs to be classified have the same size.
In the invention, the specific steps of the hierarchical classification stage are as follows:
(1) global classification phase
And in the global classification stage, the feature subgraphs are subjected to primary classification, namely, the feature subgraphs are classified into traffic signal lamps, traffic signs and background classes. If the global classification result is correct, performing the next subdivision;
(2) fine classification stage
In the fine classification stage, the feature map and the information of the global classification are combined, and more detailed classification is carried out under the category of the global classification, namely the global classification is of signal lamps and is further classified into red lamps, yellow lamps and green lamps; and for the global classification of the traffic signs, further classifying the global classification into various specific traffic signs, and adding a bounding box regression layer in the classification subnet, wherein the layer outputs fine bounding box coordinates of the target region.
Compared with the prior art, the invention has the following beneficial effects:
(1) in the feature extraction stage, a lightweight Ghost bottleeck module which is subjected to model pruning, compression optimization and backward shift and is suitable for edge calculation is adopted to construct network extraction features contained in a feature extraction image, and output feature graphs are subjected to pooling processing, so that feature redundancy is reduced, forward reasoning speed is high, and memory occupation is low;
(2) in the hierarchical classification stage, for each characteristic subgraph, a two-stage progressive hierarchical classification method is adopted, inter-class differences and intra-class differences before three classes of traffic signs, signal lamps and natural scenes are fully considered, and the classification accuracy is effectively improved. Compared with a baseline algorithm, the improved algorithm has greatly improved performance in all aspects, and can meet the requirements of real-time performance and reliability of an automatic driving system.
Drawings
FIG. 1: GhostNet network flow chart.
FIG. 2: the algorithm of the invention is a flow chart.
Detailed Description
The technical scheme of the invention is explained in detail in the following by combining the drawings and the embodiment.
The invention provides an optimization algorithm for detecting and identifying traffic signs and traffic signal lights, which is divided into 3 stages: feature extraction, region candidate stage and classification stage.
First, feature extraction stage
(1) Constructing a feature expression network
The method comprises the steps of constructing a feature expression network, firstly determining the size of an input image of the feature extraction network to be 512 multiplied by 3, extracting the most basic original features of the image by 16 smaller common convolution kernels with stride being 2, and inputting an output feature diagram into a series of Ghost bottleck channels which are gradually increased. These ghestbottleecks are divided into different stages according to the size of their input feature maps. All Ghost bottleecks were used with stride 1, except that the last Ghost bottleeck of each stage was stride 2. Finally, the feature map is converted into 1280-dimensional feature vectors by using the global average pool and the convolutional layer for final classification, and the specific structure of the feature expression network is shown in fig. 1. And (3) carrying out big data training on the classical image classification data set ImageNet image classification training by using the feature expression network to obtain a pre-training model.
(2) Separation feature extraction network
The method comprises the steps of extracting rich basic semantic features contained in an image well through a pre-trained feature expression network, taking the first 18 layers of the pre-trained feature expression network as feature extraction layers, fixing parameters, gradients and Batchnormal coefficients of the first five layers, closing dropout without being influenced by gradient back propagation, taking the dropout as an image basic feature extractor, carrying out 7 multiplied by 7 average pooling processing on an output feature map, and obtaining a feature map F1 to be sent to a subsequent regional candidate RPN.
Second, extracting candidate region stage
(1) Candidate region acquisition
The region candidate stage uses the RPN subnet in the Faster-RCNN to obtain the candidate region (RegionProposal). Taking the pooled feature map F1 obtained in the feature extraction stage as input, firstly performing common convolution with convolution kernel size of 1x1, stride of 1 and convolution kernel number of 960 to obtain a feature map F2 with rich semantics, and then sending the feature map F2 into an RPN network to obtain all interested candidate regions of the candidate regions obtained by the RPN network
Figure BDA0002511905430000041
The candidate region includes its coordinates and confidence.
(2) Positive and negative candidate region differentiation
Let G denote all group Truth in the image, in order to distinguish the correct candidate regions obtained by the RPN network
Figure BDA0002511905430000042
And wrong candidate region
Figure BDA0002511905430000043
The intersection ratio of the candidate area and the corresponding group Truth is carried outCalculating, wherein the intersection ratio of the candidate area and the corresponding group Truth is between 0.7 and 1, and the candidate area is listed as the positive sample
Figure BDA0002511905430000044
The intersection ratio of the candidate area and the corresponding group Truth is 0.01 to 0.3, and the candidate area is listed as a negative sample
Figure BDA0002511905430000045
And ignoring the candidate area with the intersection ratio of the candidate area and the corresponding group Truth between 0.3 and 0.7. The specific definition is as follows:
Figure BDA0002511905430000046
Figure BDA0002511905430000047
(3) positive candidate region screening
For all candidate regions obtained by RPN
Figure BDA0002511905430000048
Sorting according to the confidence degree from large to small, and selecting the first 300 candidate areas
Figure BDA0002511905430000049
And sending the obtained product to a subsequent classification stage.
(4) Candidate region feature map acquisition
Candidate region
Figure BDA00025119054300000410
And cutting and splicing the part F1 of the feature map corresponding to the coordinates. Extracting candidate regions
Figure BDA00025119054300000411
Corresponding characteristic diagram F1'. Before entering the sub-network, the candidate area is processed
Figure BDA00025119054300000412
Corresponding toThe feature map F1' is scaled to 7 × 7 to obtain F1 ", and all the feature maps F1" to be classified have the same size and are sent to the subsequent classification network.
Third, classification stage
For each characteristic sub-graph F1 ″, the classification sub-network will classify the images by a hierarchical classification method to identify traffic lights and traffic signs therein, specifically, the hierarchical classification method is divided into two stages: a global classification phase and a fine classification phase.
(1) Global classification phase
In the global classification stage, the feature maps are subjected to primary classification through a full connection layer to obtain a global classification result p, wherein the global classification result p comprises traffic signal lamps, traffic signs and background classes. The global classification result is of the traffic light or the traffic sign, and the next sub-classification is carried out.
(2) Fine classification stage
In the fine classification stage, the information of the feature map F1' and the global classification p is combined, and the fine classification is carried out under the category of the global classification to obtain a fine classification result
Figure BDA0002511905430000051
For the global classification is traffic signal, further classifying it into red, yellow and green; and if the global classification is a traffic sign, the traffic sign is further classified into various specific traffic signs (such as no-pedestrian traffic and the like). And adding a bounding box regression layer in the classification subnet, wherein the output of the layer is the fine bounding box coordinate t of the target area.
(3) Expression of results
Finally, the whole network outputs the detection results in the input image, including the global classification result p and the fine classification result
Figure BDA0002511905430000052
And location information t. The output result is expressed as: global classification result p refinement classification result
Figure BDA0002511905430000053
And location information t (e.g. position information t): traffic signal lamp red lamp [0.1,0.2), (0.2,0.4)])。
Four, calculation of loss function
As can be seen from the above description of the network, the output of the network contains three items, namely the global classification result p and the fine classification result
Figure BDA0002511905430000054
And a bounding box location t. The loss function of this network therefore also measures the difference between the predicted and true values in three ways, respectively: global classification penalty LclsL 'is finely classified'clsAnd bounding box positioning loss Lloc. The specific definition is as follows:
Figure BDA0002511905430000055
global classification penalty LclsIn (c) piAnd uiGlobal classification results and their true values, N, respectively, for network predictiontotRepresenting a total number of candidate regions for normalizing global classification loss;
Figure BDA0002511905430000056
segment loss of L'clsIn
Figure BDA0002511905430000057
And
Figure BDA0002511905430000058
respectively the fine classification result of the network prediction and the true value thereof,
Figure BDA0002511905430000059
is 1 or 0 to indicate whether the result of the global classification is correct or not, by
Figure BDA00025119054300000510
Is set to 0, it is possible to avoid that the globally misclassified term is calculated twice, NgThen the correct global classification area number is indicated to normalize the fine classification loss;
Figure BDA00025119054300000511
bounding box positioning loss LlocMiddle tiAnd viRespectively, the predicted bounding box position of the network and its true value, N+Representing the total number of candidate regions with correct classification, and normalizing the bounding box positioning loss;
L(pi,ti)＝λLcls+L′cls+Lloc(6)
finally, the network loss function adopts weighted addition, a weight parameter lambda is used for balancing the positioning error of the bounding box and the global classification and fine classification losses, the lambda is a hyper-parameter, and the value of the lambda is determined to be 2 through experiments.
Network training
The training process of the network adopts the idea of transfer learning, namely the whole training process is divided into two stages: and a pre-training stage, namely transfer learning.
(1) Pre-training phase
In the pre-training stage, big data image classification training is carried out on the characteristic expression network on a classical image classification data set ImageNet to obtain a pre-training model. The network structure of the feature expression network is shown in fig. 1, and in the training phase, the hyper-parameters are set as follows: the batch size is 1024, the learning rate is 0.5, linear decay is adopted, weight decay is 4e-5, momentum is 0.9, label smooth is 0.1, and dropout is 0.1, and the total number of iterations is more than 30000.
(2) Transfer learning phase
In the transfer learning stage, the training of the invention adopts the MSCOCO target detection data set to carry out training, and a better initial model is obtained. In the second stage, based on the initial model, the traffic signal lamp and traffic sign data set Tsinghua-Tencent 100K used by the baseline author is used for fine-tuning to obtain a final detection model with strong generalization capability. Besides the two-stage training method, the following strategies are adopted in the training process: the larger batch size (32 for the experiment) was trained; setting a smaller Loss (such as 0.03) threshold or a larger iteration number (5 ten thousand iterations in the experiment) threshold; a method of doubling the learning rate every several iterations (e.g., 10 ten thousand) is employed.
Example 1
Multiple sets of comparative experiments were performed using the baseline algorithm author's test set Tsinghua-Tencent 100K, which was 1500 images in total, for a total of 3120 traffic signs and traffic lights. Using a reference pair area used by the author of less than 322(small) area is 322To 962(middle) area is more than 962The (large) detection results are respectively counted, the threshold value of the cross-over ratio (IoU) is set to be 0.5, the precision rate, the recall rate and the mAP are respectively calculated, and in addition, in order to compare and research the calculated amount and the memory consumption of the invention, the base line algorithm and the average forward reasoning speed, the memory occupation and the model size of the invention are counted.
Table 1 comparative experiment performance reference table
Figure BDA0002511905430000061
Figure BDA0002511905430000071
In contrast to the baseline method: the invention adopts the lightweight characteristic extraction network after model compression and pruning, and the table shows that the invention ensures that the difference between the precision rate and the recall rate of the large, medium and small targets and the baseline method is not more than two percentage points, and the recall rate and the precision rate on the medium target detection are improved by 1.2 to 1.46 percentage points compared with the baseline method, however, the memory occupation is only 906MB, which is one third of the baseline method, the operation inference speed can reach 15 frames per second, which is 4 times of the baseline method, and the invention can meet the requirement of high real-time performance of an automatic driving system.

Claims (4)

1. An optimization algorithm for detecting and identifying traffic signs and signal lamps is characterized by comprising three stages: a feature extraction stage, a region candidate stage and a hierarchical classification stage; in the characteristic extraction stage, a Ghostbottleneck module of GhostNet is introduced to construct a characteristic extraction network, big data training is carried out on a classical image classification data set ImageNet to obtain a pre-training model, rich semantic information contained in an image is extracted through the pre-trained characteristic extraction network, and a characteristic graph is subjected to pooling processing; in the region candidate stage, an RPN subnet in the Faster-RCNN is adopted to obtain a candidate region, and the feature graph corresponding to the candidate region is cut and scaled, so that all the feature subgraphs to be classified have the same size; the hierarchical classification stage classification subnet classifies the images by a two-stage progressive hierarchical classification method so as to identify traffic lights and traffic signs in the images.
2. The optimization algorithm according to claim 1, wherein the specific steps of the feature extraction stage are as follows:
(1) constructing a feature expression network, extracting the most basic original features of an image from an input image through a small common convolution kernel, inputting an output feature graph into a series of Ghost bottleeck modules, gradually increasing channels, dividing the Ghost bottleeck into different stages according to the size of the input feature graph, finally converting the feature graph into feature vectors by using a global average pool and a convolution layer for final classification, and performing big data training on a classical image classification data set ImageNet image classification training by using the feature expression network to obtain a pre-training model;
(2) separating the feature extraction network, taking 1-18 layers of the pre-trained feature expression network as feature extraction layers, fixing the parameters, gradients and Batchnormal coefficients of the first five layers, and finally performing average pooling treatment on the output feature map.
3. The optimization algorithm according to claim 1, wherein the specific steps of the region candidate stage are as follows:
(1) acquiring a candidate region, acquiring the candidate region by adopting an RPN subnet in FasterR-CNN in a region candidate stage, taking a pooled feature map obtained in a feature extraction stage as input, performing convolution rich semantic information firstly, and then sending the information into an RPN network to acquire the coordinate of the candidate region;
(2) positive and negative candidate regions are distinguished
Figure FDA0002511905420000011
All candidate areas positioned by the RPN network are shown, G shows all group Truth in the image, and the correct candidate areas obtained by the RPN network are distinguished
Figure FDA0002511905420000012
And wrong candidate region
Figure FDA0002511905420000013
Is defined as:
Figure FDA0002511905420000014
Figure FDA0002511905420000015
(3) screening positive candidate regions, sorting all candidate regions from large to small according to confidence degrees, and selecting the first 300 candidate regions to be sent to a subsequent classification stage;
(4) and obtaining a candidate region feature map, converting the coordinate expression of the candidate region into a feature map corresponding to the region, and scaling the feature map corresponding to the candidate region before the candidate region is sent into the subnet to 7 multiplied by 7, so that all feature subgraphs to be classified have the same size.
4. The optimization algorithm according to claim 1, wherein the specific steps of the hierarchical classification stage are as follows:
(1) global classification phase
In the global classification stage, the feature subgraphs are classified preliminarily, namely the feature subgraphs are classified into traffic signal lamps, traffic signs and background classes, the global classification result is correct, and the next step of fine classification is carried out;
(2) fine classification stage
In the fine classification stage, the feature map and the information of the global classification are combined, and more detailed classification is carried out under the category of the global classification, namely the global classification is of signal lamps and is further classified into red lamps, yellow lamps and green lamps; and for the global classification of the traffic signs, further classifying the global classification into various specific traffic signs, and adding a bounding box regression layer in the classification subnet, wherein the layer outputs fine bounding box coordinates of the target region.
CN202010463642.0A 2020-05-27 2020-05-27 Optimization algorithm for detecting and identifying traffic signs and signal lamps Pending CN111738300A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010463642.0A CN111738300A (en) 2020-05-27 2020-05-27 Optimization algorithm for detecting and identifying traffic signs and signal lamps

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010463642.0A CN111738300A (en) 2020-05-27 2020-05-27 Optimization algorithm for detecting and identifying traffic signs and signal lamps

Publications (1)

Publication Number Publication Date
CN111738300A true CN111738300A (en) 2020-10-02

Family

ID=72647915

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010463642.0A Pending CN111738300A (en) 2020-05-27 2020-05-27 Optimization algorithm for detecting and identifying traffic signs and signal lamps

Country Status (1)

Country Link
CN (1) CN111738300A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112926501A (en) * 2021-03-23 2021-06-08 哈尔滨理工大学 Traffic sign detection algorithm based on YOLOv5 network structure
CN113052254A (en) * 2021-04-06 2021-06-29 安徽理工大学 Multi-attention ghost residual fusion classification model and classification method thereof
CN113361643A (en) * 2021-07-02 2021-09-07 人民中科(济南)智能技术有限公司 Deep learning-based universal mark identification method, system, equipment and storage medium
CN113378890A (en) * 2021-05-17 2021-09-10 浙江工业大学 Lightweight pedestrian and vehicle detection method based on improved YOLO v4
CN113963350A (en) * 2021-11-08 2022-01-21 西安链科信息技术有限公司 Vehicle identification detection method, system, computer equipment, storage medium and terminal

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103903013A (en) * 2014-04-15 2014-07-02 复旦大学 Optimization algorithm of unmarked flat object recognition
US20150221116A1 (en) * 2013-05-29 2015-08-06 Capso Vision, Inc. Method of Overlap-Dependent Image Stitching for Images Captured Using a Capsule Camera
CN108009518A (en) * 2017-12-19 2018-05-08 大连理工大学 A kind of stratification traffic mark recognition methods based on quick two points of convolutional neural networks
CN108985169A (en) * 2018-06-15 2018-12-11 浙江工业大学 Across the door operation detection method in shop based on deep learning target detection and dynamic background modeling
CN110163187A (en) * 2019-06-02 2019-08-23 东北石油大学 Remote road traffic sign detection recognition methods based on F-RCNN
CN110929593A (en) * 2019-11-06 2020-03-27 哈尔滨工业大学（威海） Real-time significance pedestrian detection method based on detail distinguishing and distinguishing

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150221116A1 (en) * 2013-05-29 2015-08-06 Capso Vision, Inc. Method of Overlap-Dependent Image Stitching for Images Captured Using a Capsule Camera
CN103903013A (en) * 2014-04-15 2014-07-02 复旦大学 Optimization algorithm of unmarked flat object recognition
CN108009518A (en) * 2017-12-19 2018-05-08 大连理工大学 A kind of stratification traffic mark recognition methods based on quick two points of convolutional neural networks
CN108985169A (en) * 2018-06-15 2018-12-11 浙江工业大学 Across the door operation detection method in shop based on deep learning target detection and dynamic background modeling
CN110163187A (en) * 2019-06-02 2019-08-23 东北石油大学 Remote road traffic sign detection recognition methods based on F-RCNN
CN110929593A (en) * 2019-11-06 2020-03-27 哈尔滨工业大学（威海） Real-time significance pedestrian detection method based on detail distinguishing and distinguishing

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
KAI HAN: "GhostNet: More Features from Cheap Operations", 《ARXIV》 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112926501A (en) * 2021-03-23 2021-06-08 哈尔滨理工大学 Traffic sign detection algorithm based on YOLOv5 network structure
CN113052254A (en) * 2021-04-06 2021-06-29 安徽理工大学 Multi-attention ghost residual fusion classification model and classification method thereof
CN113378890A (en) * 2021-05-17 2021-09-10 浙江工业大学 Lightweight pedestrian and vehicle detection method based on improved YOLO v4
CN113378890B (en) * 2021-05-17 2024-03-22 浙江工业大学 Lightweight pedestrian vehicle detection method based on improved YOLO v4
CN113361643A (en) * 2021-07-02 2021-09-07 人民中科(济南)智能技术有限公司 Deep learning-based universal mark identification method, system, equipment and storage medium
CN113963350A (en) * 2021-11-08 2022-01-21 西安链科信息技术有限公司 Vehicle identification detection method, system, computer equipment, storage medium and terminal

Similar Documents

Publication Publication Date Title
Hasegawa et al. Robust Japanese road sign detection and recognition in complex scenes using convolutional neural networks
CN108171136B (en) System and method for searching images by images for vehicles at multi-task gate
CN108830188B (en) Vehicle detection method based on deep learning
CN111738300A (en) Optimization algorithm for detecting and identifying traffic signs and signal lamps
Li et al. Traffic light recognition for complex scene with fusion detections
CN108921083B (en) Illegal mobile vendor identification method based on deep learning target detection
Zhang et al. Pedestrian detection method based on Faster R-CNN
CN110263786B (en) Road multi-target identification system and method based on feature dimension fusion
CN111368754B (en) Airport runway foreign matter detection method based on global context information
CN107038442A (en) A kind of car plate detection and global recognition method based on deep learning
CN109492596A (en) A kind of pedestrian detection method and system based on K-means cluster and region recommendation network
Pei et al. Localized traffic sign detection with multi-scale deconvolution networks
CN108985145A (en) The Opposite direction connection deep neural network model method of small size road traffic sign detection identification
CN112270286A (en) Shadow interference resistant monochrome video target tracking method
CN114882351B (en) Multi-target detection and tracking method based on improved YOLO-V5s
CN112949510A (en) Human detection method based on fast R-CNN thermal infrared image
Hasegawa et al. Robust detection and recognition of japanese traffic sign in the complex scenes based on deep learning
Yun et al. Part-level convolutional neural networks for pedestrian detection using saliency and boundary box alignment
CN112418207A (en) Weak supervision character detection method based on self-attention distillation
CN112307894A (en) Pedestrian age identification method based on wrinkle features and posture features in community monitoring scene
CN110555425A (en) Video stream real-time pedestrian detection method
CN115909276A (en) Improved YOLOv 5-based small traffic sign target detection method in complex weather
Chen et al. Research on vehicle detection and tracking algorithm for intelligent driving
Pazhoumand-Dar et al. DTBSVMs: A new approach for road sign recognition
Li et al. Chinese license plate recognition algorithm based on unet3+

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
WD01 Invention patent application deemed withdrawn after publication
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20201002