CN113869239A - Traffic signal lamp countdown identification system and construction method and application method thereof - Google Patents

Traffic signal lamp countdown identification system and construction method and application method thereof Download PDF

Info

Publication number
CN113869239A
CN113869239A CN202111160244.2A CN202111160244A CN113869239A CN 113869239 A CN113869239 A CN 113869239A CN 202111160244 A CN202111160244 A CN 202111160244A CN 113869239 A CN113869239 A CN 113869239A
Authority
CN
China
Prior art keywords
traffic signal
loss
signal lamp
countdown
model
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
CN202111160244.2A
Other languages
Chinese (zh)
Inventor
曹秋阳
周昆阳
杨启硕
宋锦伟
邵叶秦
施佺
高瞻
许长勇
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.)
Nantong University
Original Assignee
Nantong 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 Nantong University filed Critical Nantong University
Priority to CN202111160244.2A priority Critical patent/CN113869239A/en
Publication of CN113869239A publication Critical patent/CN113869239A/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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Traffic Control Systems (AREA)

Abstract

The invention provides a traffic signal lamp countdown identification system and a construction method and an application method thereof, which relate to the technical field of computer detection system design, wherein the construction method comprises the following steps: making a traffic signal lamp data set; a feature extraction network of the traffic signal lamp countdown identification model; constructing a multi-scale fusion module based on attention; and completing construction of an end-to-end traffic signal lamp countdown recognition model, preprocessing a training data set and a label, and training the end-to-end traffic signal lamp countdown recognition model. And inputting the newly acquired picture on the mobile terminal into the trained end-to-end traffic signal lamp countdown identification model to obtain the identification result of the traffic signal lamp countdown. The application provides a traffic signal lamp identification system that counts down, the rate of accuracy of model is higher, and the rate of missing examining is low, and is fast. The traffic light recognition method has the advantages that the recognition effect on the traffic light can be better under the complex environment, and the generalization and the practicability of the model are higher.

Description

Traffic signal lamp countdown identification system and construction method and application method thereof
Technical Field
The invention relates to the technical field of computer detection system design, in particular to a traffic signal lamp countdown identification system and a construction method and an application method thereof.
Background
Traffic lights are important information in vehicle assisted driving and autonomous driving. Under the influence of factors such as environment, the current target detection model has low identification accuracy and high omission factor of the traffic signal lamp, and has larger potential safety hazard. An accurate and efficient traffic signal lamp detection and identification algorithm is an important research direction for auxiliary driving and automatic driving. With the continuous development of the neural network technology, some scholars introduce a lightweight network MobileNet by using an NVIDIA Jetson Tegra X2 embedded platform on the basis of improving a feature extraction network of YOLOv3, so that the detection speed is effectively improved, but the detection effect on a remote traffic signal lamp is still to be improved.
In the prior art, although the method based on the neural network is greatly improved in the prediction speed, the method is used for complex scenes such as: the traffic signal lamp detection and identification accuracy rate in foggy days and rainy days is low.
Disclosure of Invention
The invention aims to solve the technical problem that the detection and identification accuracy rate is low in a complex scene in the prior art.
In order to achieve the purpose, the invention adopts the following technical scheme:
a construction method of a traffic signal lamp countdown identification system comprises the following steps:
s1: making a traffic signal lamp data set: collecting traffic signal lamp pictures, classifying the pictures and making the pictures into a data set in a VOC2007 format;
s2: the feature extraction network of the end-to-end traffic signal lamp countdown identification model comprises the following steps: selecting ResNet50vd as a main network for feature extraction, constructing an inclusion-CSP module, and replacing a 3 × 3 standard convolution of a network stage4 with a 3 × 3 deformable convolution DCN; removing the overlapped frames by using a matrix non-maximum value inhibition method, and using logistic regression as a traffic signal lamp countdown classifier;
s3: constructing a multi-scale fusion module based on attention to realize the fusion of multi-scale features;
s4: and completing network construction, preprocessing the training data set and the label, and then using the preprocessed training data set as an input training end-to-end traffic signal lamp countdown recognition model.
Preferably, the number of collected traffic light pictures in S1 is 8000, and in the classification process, the pictures are subjected to size normalization, random saturation adjustment, random contrast adjustment, random brightness adjustment, and Mosaic data enhancement.
Preferably, in S3: the final output characteristics are obtained by using different weights w1, w2 of Feature1 and Feature2, and the CSPNet is added to the PANET.
Preferably, the specific step of S4 is:
s4.1: calculating a Loss function, wherein the Loss function is obtained by weighting three Loss functions, and the three Loss functions are respectively class LossclassLoss of confidence LossconfLoss of coordinate deviation LosscoordTherein, LossclassUsing binary cross entropy Loss, LossconfLoss using binary cross entropyconf-objAnd Lossconf_noobj，LosscoordUsing CIOULoss, the weight ratio is 1: 2, and the concrete formula is as follows:
Loss＝0.25×Lossclass+0.25×Lossconf+0.5×Losscoord；
s4.2: selecting a cosine learning rate and an Exponential Moving Average (EMA);
s4.3: the prior box is re-determined by the K-means algorithm.
Preferably, the relationship between the cosine learning rate in S4.2 and the number of training rounds (epochs) of the model is as follows:
Figure BDA0003289746290000031
in the above formula, begin _ rate is the initial learning rate, epoch is the current round number;
the calculation formula of the exponential moving average EMA is as follows:
Wt＝α×Wt-1+(1-α)×W(t≥1)
in the above formula, α is the attenuation coefficient, WtIs an exponential moving average with an initial value of 0.
Preferably, the distance function calculated in S4.3 is:
d(box，centroid)＝1-IOU(box，centroid)
in the above formula, a rectangle frame in the center of the centroid cluster, a box represents a labeled rectangle frame, and an IOU represents the intersection ratio of the box and the centroid.
The application also provides a traffic signal lamp countdown identification system which is constructed and formed by using the construction method of the traffic signal lamp countdown identification system.
The application also provides an application method of the traffic signal lamp countdown identification system, and by using the traffic signal lamp countdown identification system, the traffic signal lamp countdown identification system is deployed in mobile terminal equipment to finish identification of the traffic signal lamp countdown in an actual scene.
Compared with the original YOLOv4, the construction method of the traffic signal lamp countdown identification system has the advantages that through improvement of YOLOv4, the accuracy of an end-to-end traffic signal lamp countdown identification model is higher, the omission factor is lower, and the speed is higher. The traffic light recognition method has the advantages that the recognition effect on the traffic light can be better under the complex environment, and the generalization and the practicability of the model are higher.
Drawings
Fig. 1 is a flowchart of a traffic signal light countdown identification method according to the present invention;
FIG. 2 is a ResNet50vd network structure according to an embodiment of the present invention;
fig. 3 is a modified inclusion-CSP structure in accordance with an embodiment of the present invention;
FIG. 4 is a block diagram of an attention-based multi-scale fusion module according to an embodiment of the present invention;
FIG. 5 is a modified CSPNet structure in accordance with one embodiment of the present invention;
fig. 6 is an expanded view of an improved PANet according to an embodiment of the present invention.
Fig. 7 is a flow chart of an application of the method herein.
Detailed Description
In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention, however, the present invention may be practiced in other ways than those specifically described herein, and thus the present invention is not limited to the specific embodiments of the present disclosure.
Part of English explanation:
LabelImg: visual image calibration tool
Feature: feature(s)
And (3) PANET: path Aggregation Network (Path Aggregation Network)
CSPNet: cross-phase local Network (Cross Stage Partial Network)
DCN: deep crossing Network (Deep Cross Network)
Matrix NMS: matrix Non-Maximum Suppression (Matrix Non-Maximum Suppression)
AMF: attention-based multiscale Fusion (Attention-based Multi-Scale Fusion)
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments.
Referring to fig. 1, a method for constructing a traffic signal light countdown identification system includes the following steps:
s1: making a traffic signal lamp data set under a complex scene:
specifically, firstly, pictures of traffic lights in various environments, scenes and weathers are collected, in one embodiment, the picture sources are Google gallery, 360 galleries, hundred-degree network galleries and the like, and 8000 countdown pictures are obtained from the galleries; then preprocessing the picture, in one embodiment, labeling the category and the position of a traffic signal lamp countdown target in the collected picture by using LabelImg, and manufacturing the traffic signal lamp countdown target into a VOC2007 format, wherein 20 types of data sets are provided, specifically, 10 types of green lamp countdown 0-9 and 10 types of red lamp countdown 0-9 are provided, image enhancement is performed on the data sets while classification is performed, and in one embodiment, size normalization, random saturation adjustment, random contrast adjustment, random brightness adjustment and Mosaic data enhancement are performed on the picture, so that a traffic signal lamp countdown data set is finally obtained.
In one embodiment, the sizes of the pictures are normalized to 416 × 416 because the collected sizes are not the same; the random saturation adjustment is to set a threshold value to be 0.5, then randomly extract a number x in an interval (0, 1), adjust the saturation to be x times of the original if x is larger than or equal to 0.5, randomly extract a number y in an interval (-x, x) if x is smaller than 0.5, and then adjust the saturation to be 1+ y times of the original.
S2: the feature extraction network of the end-to-end traffic signal lamp countdown identification model is used for improving the detection and identification capacity of the model:
specifically, in one embodiment, the method comprises the following steps:
s2.1: referring to fig. 2, originally, there are more CSPDarkNet53 parameters of yollov 4, and in order to save computational resources, ResNet50vd is selected as a main network for feature extraction, where ResNet50vd adopts a ResNet-D network, that is, an average pooling layer and a convolution with a step length of 2 × 3 are added to block1, so that the model accuracy is improved under the condition of constant speed. Meanwhile, in order to improve the feature extraction capability of the model on the remote traffic signal lamp numbers, stages 1, 2, 3 and 4 in the network are modified by CSP.
S2.2: referring to fig. 3, an inclusion-CSP module is constructed and added to the stages 2 and 3 in the network to improve the model detection capability. The CSPNet structure can enhance the learning capability of the convolutional neural network, and the inclusion network can enable different levels of receptive fields to be different. The inclusion-CSP module is added into the backbone network, so that the width and the depth of the network can be increased, the receptive field of the network is increased, and the detection capability of the network on objects is enhanced.
S2.3: replacing the 3 × 3 standard convolution of stage4 in the network with a 3 × 3 deformable convolution DCN to extract more efficient features;
s2.4: and removing overlapped frames by using a matrix non-maximum value inhibition method so as to effectively avoid the condition that prediction frames of the same type of objects conflict with each other.
For the detection of the same object, YOLOv4 may present multiple prediction boxes, and in one embodiment, Matrix NMS is used to remove overlapping boxes. And the Matrix NMS calculates the scores of the IOU and the prediction box through Matrix calculation, and obtains a penalty coefficient according to the IOU and the scores, so that parallel calculation is realized, the reasoning speed of the model is increased, and the detection accuracy is improved. In one embodiment, the Matrix NMS has a number of test boxes of 1, a score threshold of 0.2, and an IOU threshold of 0.45.
S2.5: and using logistic regression as a traffic light countdown classifier, and performing prediction by using the 13 × 13, 26 × 26 and 52 × 52 feature maps output by the end-to-end traffic light countdown recognition model.
S3: constructing a multi-scale fusion module based on attention to further improve the detection performance of the network:
specifically, referring to fig. 4, features of different scales are input as Feature1 and Feature2, where Feature1 is a small-scale Feature after resize and the large-scale Feature2 is the same size. Firstly, adding Feature1 and Feature2, wherein the spatial resolution of the obtained features is changed into 1 × 1 after global average pooling, and the number of channels of the features is changed into C/2 after 1 × 1 convolution and a ReLU activation function; then, the number of channels is adjusted to C through 1 × 1 convolution and bn (batch normalization), and finally, the weight w1 of the output is obtained through the full connection layer (FC) and the sigmoid function. Because Feature1 and Feature2 Feature information are different, the method that Feature1 and Feature2 share weight w1 is not adopted in the application, the obtained weight w1 is multiplied by Feature1, and the weight w2 is multiplied by Feature2 and then added to obtain the final output Feature, wherein w2 is 1-w 1. An attention-based multi-scale Feature fusion module is introduced, and final output features are obtained by adopting different weights w1 and w2 of Feature1 and Feature 2. For inputs of different scales, the feature multi-scale fusion module can calculate different weights w1, w2 so that inputs of different scales contribute differently to the output features.
Referring to fig. 5 and 6, in one embodiment, CSPNet is added to PANet to further enhance the feature fusion capability of the PANet module. And finally, fusing the features of two different scales by the multi-scale fusion module through the improved PANet, and extracting the features through CSPNet.
S4: and completing network construction, preprocessing the training data set and the label, and then using the preprocessed training data set as an input training end-to-end traffic signal lamp countdown recognition model.
Specifically, in one embodiment, the method comprises the following steps:
s4.1: calculating a loss function
In an embodiment, the Loss function is obtained by weighting three Loss functions, and each of the three Loss functions is a class LossclassLoss of confidence LossconfLoss of coordinate deviation LosscoordTherein, LossclassUsing binary cross entropy Loss, LossconfLoss using binary cross entropyconf_objAnd Lossconf_noobj，LosscoordUsing CIOULoss, the weight ratio is 1: 2, and the concrete formula is as follows:
Loss＝0.25×Lossclass+0.25×Lossconf+0.5×Losscoord
s4.2: selecting a cosine learning rate and an Exponential Moving Average (EMA);
in one embodiment, the relationship between the number of training rounds (epochs) and the learning rate of the model is shown as follows:
Figure BDA0003289746290000081
in the above formula, begin _ rate is the initial learning rate, and epoch is the current round number. In one embodiment, begin _ rate is 0.001 and epochs is 100000.
In one embodiment, the moving average value of the parameter updating process is calculated by the exponential moving average EMA through an exponential decay mode, and each parameter W has one corresponding to the exponential moving average value WtThe relationship between the two is shown as the formula:
Wt＝α×Wt-1+(1-α)×W(t≥1)
in the above formula, α is an attenuation coefficient, and in one embodiment, α is 0.998, and W istInitial value 0, using WtAnd updating the parameters.
S4.3: and re-determining the prior frame by a K-means algorithm, wherein the distance function is shown as a formula:
d(box，centroid)＝1-IOU(box，centroid)
in the above formula, a rectangle frame in the center of the centroid cluster, a box represents a labeled rectangle frame, and an IOU represents the intersection ratio of the box and the centroid.
Obtaining a prior box through the algorithm: (6, 19), (13,40), (7, 22), (4, 13), (5, 16), (8, 24), (8, 26), (8, 25), (9, 28), completing the construction of the detection system.
The application also provides a traffic signal lamp countdown identification system which is constructed by using the construction method.
Referring to fig. 7, the traffic signal lamp countdown identification system is deployed in mobile terminal equipment to complete identification of the countdown of the traffic signal lamp in an actual scene.
Specifically, in one embodiment, a trained detection system is deployed on a mobile device Jetson TX2 and is flushed using jetpack 4.3. The deep learning framework adopted by the invention is PaddlePaddle1.8.4, and the deployment software is PaddleLite.
Compared with the original YOLOv4, the construction method of the traffic signal lamp countdown identification system is higher in accuracy, lower in omission factor and higher in speed of an end-to-end traffic signal lamp countdown identification model by improving YOLOv 4. The traffic light recognition method has the advantages that the recognition effect on the traffic light can be better under the complex environment, and the generalization and the practicability of the model are higher.
The above description is only an example of the present invention, and is not intended to limit the present invention, and it is obvious to those skilled in the art that various modifications and variations can be made in the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims (8)

1. A construction method of a traffic signal lamp countdown identification system is characterized by comprising the following steps: comprises the following steps:
s1: making a traffic signal lamp data set: collecting traffic signal lamp pictures, classifying the pictures and making the pictures into a data set in a VOC2007 format;
s2: the feature extraction network of the end-to-end traffic signal lamp countdown identification model comprises the following steps: selecting ResNet50vd as a main network for feature extraction, constructing an inclusion-CSP module, and replacing a 3 × 3 standard convolution of a network stage4 with a 3 × 3 deformable convolution DCN; removing the overlapped frames by using a matrix non-maximum value inhibition method, and using logistic regression as a traffic signal lamp countdown classifier;
s3: constructing a multi-scale fusion module AMF based on attention to realize the fusion of multi-scale features;
s4: and completing network construction, preprocessing the training data set and the label, and then using the preprocessed training data set as an input training end-to-end traffic signal lamp countdown recognition model.
2. The method of claim 1, wherein the traffic signal light countdown identifying system comprises: the number of collected traffic light pictures in S1 is 8000, and in the classification process, size normalization, random saturation adjustment, random contrast adjustment, random brightness adjustment, and Mosaic data enhancement are performed on the pictures.
3. The method of claim 1, wherein the traffic signal light countdown identifying system comprises: in said S3: the final output characteristics are obtained by using different weights w1, w2 of Feature1 and Feature2, and the CSPNet is added to the PANET.
4. The method of claim 1, wherein the traffic signal light countdown identifying system comprises: the specific steps of S4 are as follows:
s4.1, calculating a Loss function, wherein the Loss function is obtained by weighting three Loss functions, and the three Loss functions are respectively class LossclassLoss of confidence LossconfLoss of coordinate deviation LosscoordTherein, LossclassUsing binary cross entropy Loss, LossconfLoss using binary cross entropyconf_objAnd Lossconf_noobj，LosscoordUsing CIOULoss, the weight ratio is 1: 2, and the concrete formula is as follows:
Loss＝0.25×Lossclass+0.25×Lossconf+0.5×Losscoord；
s4.2: selecting a cosine learning rate and an Exponential Moving Average (EMA);
s4.3: the prior box is re-determined by the K-means algorithm.
5. The method of claim 4, wherein the traffic signal light countdown identifying system comprises: the relation between the cosine learning rate in S4.2 and the number of training rounds (epochs) of the model is shown as follows:
Figure FDA0003289746280000021
in the above formula, begin _ rate is the initial learning rate, epoch is the current round number;
the calculation formula of the exponential moving average EMA is as follows:
Wt＝α×Wt-1+(1-α)×W(t≥1)
in the above formula, α is the attenuation coefficient, WtIs an exponential moving average with an initial value of 0.
6. The method of claim 4, wherein the traffic signal light countdown identifying system comprises: the distance function calculated in S4.3 is:
d(box，centroid)＝1-IOU(box，centroid)
in the above formula, a rectangle frame in the center of the centroid cluster, a box represents a labeled rectangle frame, and an IOU represents the intersection ratio of the box and the centroid.
7. A traffic signal lamp countdown identification system is characterized in that: the traffic signal light countdown identifying system construction method according to any one of claims 1 to 6.
8. An application method of a traffic signal lamp countdown identification system is characterized in that: the traffic signal light countdown identifying system according to claim 7, wherein the traffic signal light countdown identifying system is deployed in a mobile terminal device to complete identification of the traffic signal light countdown in an actual scene.
CN202111160244.2A 2021-09-30 2021-09-30 Traffic signal lamp countdown identification system and construction method and application method thereof Pending CN113869239A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111160244.2A CN113869239A (en) 2021-09-30 2021-09-30 Traffic signal lamp countdown identification system and construction method and application method thereof

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111160244.2A CN113869239A (en) 2021-09-30 2021-09-30 Traffic signal lamp countdown identification system and construction method and application method thereof

Publications (1)

Publication Number Publication Date
CN113869239A true CN113869239A (en) 2021-12-31

Family

ID=79001170

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111160244.2A Pending CN113869239A (en) 2021-09-30 2021-09-30 Traffic signal lamp countdown identification system and construction method and application method thereof

Country Status (1)

Country Link
CN (1) CN113869239A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115662132A (en) * 2022-10-27 2023-01-31 天津天瞳威势电子科技有限公司 Traffic light countdown time identification method and device

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115662132A (en) * 2022-10-27 2023-01-31 天津天瞳威势电子科技有限公司 Traffic light countdown time identification method and device

Similar Documents

Publication Publication Date Title
CN111814621B (en) Attention mechanism-based multi-scale vehicle pedestrian detection method and device
CN108647585B (en) Traffic identifier detection method based on multi-scale circulation attention network
CN112884064B (en) Target detection and identification method based on neural network
CN110956094B (en) RGB-D multi-mode fusion personnel detection method based on asymmetric double-flow network
CN111767882A (en) Multi-mode pedestrian detection method based on improved YOLO model
CN110263786B (en) Road multi-target identification system and method based on feature dimension fusion
CN107273832B (en) License plate recognition method and system based on integral channel characteristics and convolutional neural network
CN111401293B (en) Gesture recognition method based on Head lightweight Mask scanning R-CNN
CN114359851A (en) Unmanned target detection method, device, equipment and medium
EP3690744B1 (en) Method for integrating driving images acquired from vehicles performing cooperative driving and driving image integrating device using same
CN109919073B (en) Pedestrian re-identification method with illumination robustness
CN109657538B (en) Scene segmentation method and system based on context information guidance
CN110956158A (en) Pedestrian shielding re-identification method based on teacher and student learning frame
CN114202743A (en) Improved fast-RCNN-based small target detection method in automatic driving scene
CN114049572A (en) Detection method for identifying small target
Maggiolo et al. Improving maps from CNNs trained with sparse, scribbled ground truths using fully connected CRFs
CN114708566A (en) Improved YOLOv 4-based automatic driving target detection method
CN115019039A (en) Example segmentation method and system combining self-supervision and global information enhancement
CN116740516A (en) Target detection method and system based on multi-scale fusion feature extraction
CN114973199A (en) Rail transit train obstacle detection method based on convolutional neural network
CN114596548A (en) Target detection method, target detection device, computer equipment and computer-readable storage medium
CN114743126A (en) Lane line sign segmentation method based on graph attention machine mechanism network
CN112785610B (en) Lane line semantic segmentation method integrating low-level features
CN112347967B (en) Pedestrian detection method fusing motion information in complex scene
CN113869239A (en) Traffic signal lamp countdown identification system and construction method and application method thereof

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