CN117152513A - Vehicle boundary positioning method for night scene - Google Patents
Vehicle boundary positioning method for night scene Download PDFInfo
- Publication number
- CN117152513A CN117152513A CN202311105460.6A CN202311105460A CN117152513A CN 117152513 A CN117152513 A CN 117152513A CN 202311105460 A CN202311105460 A CN 202311105460A CN 117152513 A CN117152513 A CN 117152513A
- Authority
- CN
- China
- Prior art keywords
- model
- vehicle
- positioning
- image data
- boundary
- 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
- 238000000034 method Methods 0.000 title claims abstract description 26
- 238000012549 training Methods 0.000 claims abstract description 32
- 238000001514 detection method Methods 0.000 claims abstract description 31
- 238000004821 distillation Methods 0.000 claims abstract description 6
- 230000006870 function Effects 0.000 claims description 26
- 238000002372 labelling Methods 0.000 claims description 18
- 238000010606 normalization Methods 0.000 claims description 9
- 238000012795 verification Methods 0.000 claims description 9
- 238000012360 testing method Methods 0.000 claims description 8
- 230000004913 activation Effects 0.000 claims description 6
- 238000001914 filtration Methods 0.000 claims description 6
- 238000011176 pooling Methods 0.000 claims description 6
- 230000005764 inhibitory process Effects 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 4
- 230000008707 rearrangement Effects 0.000 claims description 4
- 238000013135 deep learning Methods 0.000 claims description 3
- 238000005286 illumination Methods 0.000 claims description 3
- 230000007246 mechanism Effects 0.000 claims description 3
- 238000012544 monitoring process Methods 0.000 claims description 3
- 206010039203 Road traffic accident Diseases 0.000 abstract description 3
- 238000004364 calculation method Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 5
- 230000008569 process Effects 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012216 screening Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000011897 real-time detection Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- 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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/22—Image preprocessing by selection of a specific region containing or referencing a pattern; Locating or processing of specific regions to guide the detection or recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural 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
Abstract
The invention belongs to the technical field of computer vision, and discloses a vehicle boundary positioning method for night scenes, which is used for collecting image data of night road scenes, carrying out frame marking on the preprocessed image data by using a detection model, and training the detection model by using a tag data training set to obtain a final teacher model; constructing a student model, distilling boundary positioning knowledge learned by a teacher model by using a positioning distillation method, and finally obtaining a lightweight network model; and finally obtaining the vehicle position of the night scene. The invention effectively solves the problem of difficult positioning of the fuzzy boundary by using positioning distillation, improves the positioning precision of the detection model, and can realize more accurate positioning by detecting surrounding vehicles in real time when the detection model is arranged in a vehicle-mounted 360-degree panoramic image system so as to avoid traffic accidents or provide better implementation conditions for automatic driving.
Description
Technical Field
The invention belongs to the technical field of computer vision, and particularly relates to a vehicle boundary positioning method for night scenes.
Background
Today, the development and popularization of artificial intelligence technology are increasing, and the computer vision technology is also becoming a hot research field in the field of artificial intelligence; among them, the image processing technology is widely and deeply applied in the intelligent traffic field, such as unmanned vehicles, unmanned logistics vehicles, road traffic monitoring, etc., which make the vehicle detection system based on the deep learning method more and more important. In the field of computer vision, vehicle bounding box position determination is an important problem, and the aim is to accurately position and determine the size of a vehicle on an image; however, the existing night image recognition system has the problem of poor recognition effect, and the main reason is the influence of light sources such as lamps of opposite vehicles or street lamps on roads on the system. The night illumination is unevenly distributed, no street lamp or vehicles far away can be darker, and vehicles near away can be too bright because of opposite vehicle lamps or street lamps, the change of the brightness can easily cause errors of a night recognition system of the automobile, the recognition accuracy is low, the boundary frame of the vehicle generates ambiguity, the boundary of the vehicle and the night environment background are easily integrated into a whole, the boundary of the vehicle cannot be well defined, the positioning accuracy of the boundary frame of the vehicle is reduced, and the like. In addition, the computational power of the hardware of the mobile terminal carried by the vehicle is generally low, and a lighter model is required for the network model. These seriously affect the detection accuracy and real-time performance of the vehicle boundary detection algorithm. Therefore, how to achieve accurate positioning of vehicle bounding boxes in night scenes and make network models as small as possible is an urgent need for current computer vision research.
Patent application CN115171079a discloses a night scene based vehicle detection method, which utilizes dynamic filter networks to generate sample-specific convolution kernels, and uses different enhancement methods to constrain each enhancement subnetwork for different night image samples; however, the model realized by the method does not have a measure for light weight, the volume of the model is larger, the calculation force is required to be relatively strong, and when the model is actually deployed on a vehicle, the detection speed is possibly reduced and even the detection fails due to low calculation force of vehicle-mounted equipment, so that the model cannot be deployed on a similar low calculation force terminal; it also does not optimize for ambiguity of boundaries in a night-time environment, and in a real scenario, when a large vehicle such as a large truck is encountered, it is impossible to efficiently and accurately locate the boundaries of the vehicle, and thus a series of traffic accidents may occur due to inaccurate judgment of the boundaries of the vehicle.
Disclosure of Invention
In order to solve the technical problems, the invention provides a night scene-oriented vehicle boundary positioning method, which enhances the night vehicle boundary positioning effect and simultaneously lightens a network model.
The invention discloses a vehicle boundary positioning method facing a night scene, which comprises the following steps:
s1, acquiring image data of a night road scene;
s2, dividing the acquired image data into a training set, a verification set and a test set, and preprocessing the image data;
s3, performing frame labeling on the preprocessed image data by using a detection model to obtain a label image dataset; training the detection model by using the label image data set and the training set to obtain a final teacher model;
s4, constructing a student model, and distilling boundary positioning knowledge learned by a teacher model by using a positioning distillation method to finally obtain a lightweight network model;
s5, inputting the test set data into a lightweight network model, outputting position coordinate information of the vehicle, setting a threshold value filtering positioning frame, and deleting the repeated positioning frame by using non-maximum value inhibition to finally obtain the vehicle position of the night scene.
In the implementation process, the selected model can be replaced, and the teacher model can be any larger detection model with more parameters.
Further, S2 is specifically:
image data is acquired by using a vehicle-mounted camera or a road monitoring camera, an image shot by the camera is originally in a RAW format, and the format records original information of a camera sensor, including shutter speed aperture values and the like of the camera; the original format is used as input by using the deep learning network, which is equivalent to the introduction of the information, and the conventional RGB format picture with better effect after illumination enhancement can be output.
Further, S3 is specifically:
labeling the image data by using a detection model to obtain a labeling result, and correcting the labeling result; the obtained label is the coordinate information of the positioning frame, namely coordinate points of four corners of the rectangular frame, and the label can be visualized by using a tool; the error positioning frame can be seen by naked eyes that the boundary frame does not frame the target or the position deviation is larger, and the boundary frame is deleted manually;
and continuing to train the detection model by taking the corrected labeling data as training data until the detection model converges.
Further, the construction of the student model is specifically as follows:
the initial stage is normalized by Batch Normalization after convolution operation, then the feature map is reduced in size by a ReLU activation function and then a maximum pooling layer;
the model then falls into three phases: the first stage comprises a downsampling module and 4 basic modules which are sequentially connected; the second stage is sequentially connected with a downsampling module and 8 basic modules, and features are further extracted and processed; the third stage comprises a downsampling module and 4 basic modules, which are used for deepening the expression capacity of the features; in each stage, the base module consists of convolution, normalization, activation functions and a CBAM attention module; the connection between the basic modules is realized through a channel rearrangement mechanism;
then, the model is normalized by convolution and Batch Normalization again, and global average pooling is carried out, so that the feature map is converted into a fixed size;
finally, the features are mapped to the final output category through the full connection layer, and a prediction result is obtained.
Further, S4 is specifically: inputting the labeling result, the training set and the image data of the verification set into a teacher model and a student model simultaneously to respectively obtain vehicle positioning frame information;
taking the difference between the vehicle positioning frame information output by the teacher model and the vehicle positioning frame information output by the student model as a first part LOSS1 of the LOSS function; taking the difference between the result output by the student model and the real label as a second part LOSS2 of the LOSS function; taking the difference between the confidence coefficient of the positioning frame output by the student model and the confidence coefficient of the positioning frame output by the teacher model as a third part LOSS3 of the LOSS function; the LOSS1, LOSS2, LOSS3 all use a smooth L1 LOSS function; and training the student model by taking the weighted summation of LOSS1, LOSS2 and LOSS3 as a final LOSS function of the student model until convergence, so as to obtain a final lightweight network model. The LOSS1 and LOSS2 enable the student model to learn the ability of defining the boundary more accurately, and the LOSS3 enables the student model to learn the judgment ability of the teacher model on the confidence coefficient of the boundary frame.
The beneficial effects of the invention are as follows: aiming at the problem of fuzzy vehicle boundary frames in night scenes, the method adopts a teacher-student model to process vehicle data acquired at night; firstly, marking by using a detection model, carrying out manual auditing and correction after a marking result comes out, taking corrected marking data as training data to train the detection model, using the trained detection model to mark again, repeating the processes to realize iterative marking and training, and finally obtaining that the model convergence is obtained when the performance and marking quality of the model are close to a stable state; the strategy can finally achieve the purposes of data labeling and model training at the same time, and the trained model can be used as a teacher model; the constructed student model has compact and efficient structure, greatly reduces the calculation and memory expenditure through the technologies of channel rearrangement, grouping convolution and the like, and is suitable for equipment with limited resources; and meanwhile, the CBAM attention module is adopted, so that the characteristic representation capability of the model is further enhanced, and the performance and accuracy are improved. The invention introduces a positioning distillation technology based on a low calculation force platform with higher algorithm deployment requirement, so as to effectively solve the problem of difficult positioning of a fuzzy boundary, and prevent vehicle collision or even collision caused by inaccurate boundary judgment because of the compatibility of the vehicle and the background; the positioning accuracy of the detection model is improved, and the vehicle violation conditions in the intelligent traffic scene can be accurately judged; the vehicle-mounted 360-degree panoramic image system can realize more accurate positioning through real-time detection of surrounding vehicles, so as to avoid traffic accidents or provide better implementation conditions for automatic driving.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a schematic diagram of a student model training architecture;
fig. 3 is a schematic diagram of a student model structure.
Detailed Description
In order that the invention may be more readily understood, a more particular description of the invention will be rendered by reference to specific embodiments that are illustrated in the appended drawings.
As shown in fig. 1 and 2, the vehicle boundary positioning method facing the night scene of the present invention includes the following steps:
s1: the method comprises the steps of collecting data sets at 7 intersections of urban roads, wherein each intersection is provided with two cameras, the cameras are set to estimate the distance between vehicles when the approaching of the vehicles is recognized, the vehicles are shot when the distance between the vehicles and the cameras is 3 meters, 6 meters and 10 meters respectively, the picture information of the vehicles passing through the time period from 8 nights to 2 nights in the morning is recorded, and the image data of the vehicles in the dark environment are collected for 10 ten thousand total days after one week of collection and screening.
S2, dividing the acquired image data into a training set, a verification set and a test set, and preprocessing the image data;
s3, performing frame labeling on the preprocessed image data by using a detection model to obtain a label image dataset; training the detection model by using the label image data set and the training set to obtain a final teacher model;
s4, constructing a student model, wherein the student model is an efficient neural network model which receives RGB images as input as shown in FIG. 3. The initial stage is normalized by Batch Normalization after convolution operation, then the initial stage is subjected to a ReLU activation function and then a maximum pooling layer, so that the size of the feature map is effectively reduced. Subsequently, the model is divided into three phases. The first stage consists of one downsampling module (DownsampleUnit) and 4 basic modules (basic unit), which are connected to each other in order. The second stage consists of a downsampling module and 8 basic modules, which are connected in turn to further extract and process features. The third stage comprises a downsampling module and 4 basic modules for deepening the expressive power of the features. In each stage, the basic module (basicmit) is the core of the model, consisting of convolution, normalization, activation functions, and CBAM attention modules, etc. The connection between the basic modules is realized through a channel rearrangement mechanism, so that the exchange of the features and the information flow are effectively promoted, and the capability of feature representation is improved.
After three phases, the model is again convolved and Batch Normalization normalized, followed by global averaging pooling to convert the feature map to a fixed size, ready for the fully connected layer. Finally, the features are mapped to the final output category through the full connection layer, and a prediction result is obtained.
As shown in fig. 2, the boundary positioning knowledge learned by the teacher model is distilled by using a positioning distillation method, the labeling result and the image data of the training set and the verification set are simultaneously input into the teacher model and the student model to respectively obtain the vehicle positioning frame information, and finally the lightweight network model is obtained.
Taking the difference between the vehicle positioning frame information output by the teacher model and the vehicle positioning frame information output by the student model as a first part LOSS1 of the LOSS function; taking the difference between the result output by the student model and the real label as a second part LOSS2 of the LOSS function; taking the difference between the confidence coefficient of the positioning frame output by the student model and the confidence coefficient of the positioning frame output by the teacher model as a third part LOSS3 of the LOSS function; the LOSS1, LOSS2, LOSS3 all use a smooth L1 LOSS function; and training the student model by taking the weighted summation of LOSS1, LOSS2 and LOSS3 as a final LOSS function of the student model until convergence, so as to obtain a final lightweight network model. The LOSS1 and LOSS2 enable the student model to learn the ability of defining the boundary more accurately, and the LOSS3 enables the student model to learn the judgment ability of the teacher model on the confidence coefficient of the boundary frame.
S5, inputting the test set data into a lightweight network model, outputting position coordinate information of the vehicle, setting a threshold value filtering positioning frame, and deleting the repeated positioning frame by using non-maximum value inhibition to finally obtain the vehicle position of the night scene.
Specifically, S2: for the acquired image information, firstly dividing the image data into a training set, a verification set and a test set according to the ratio of 8:1:1, then preprocessing the acquired image information by adopting a traditional image processing method, enabling human eyes to distinguish the boundary of a vehicle more easily by using Gaussian filtering, facilitating the follow-up labeling work, then carrying out first round detection on all data by using the weight of a vehicle identification model pre-trained by a Yolov5 model to obtain a detection result label file corresponding to the image, then using a picture labeling tool, importing the data set subjected to Gaussian filtering processing in the step S2 and the obtained label file into the tool, adding an undetected positioning frame by deleting a positioning frame with detection errors, and finely adjusting the boundary of the vehicle positioning frame to accurately define the boundary of the vehicle, thereby manually completing the second round labeling of the image.
Step S3: selecting a YoloV5 model as a teacher model, inputting a label generated by labeling and a training set and a verification set in unprocessed original pictures into the model to train the model, taking a smooth L1 loss function of a model detection result of a vehicle position and an actual labeling result as a model loss function, using a pre-training model weight file Yolov5m. Pt as an initial weight to reduce the time and data quantity used for training, training 300 rounds on a GPU server with the model of Yingwei reaching 3090, and selecting a model weight file with the highest accuracy as a final teacher model.
Step S4: and selecting a SAD (ShufferAttentionDet) network model as a student model, inputting a label generated by labeling and a training set and a verification set in unprocessed original pictures into a teacher model and the student model, taking a weighted sum of a smooth L1 loss function of vehicle position information output by the teacher model and vehicle position information output by the student model and a smooth L1 loss function of confidence degrees respectively output by the two models as a training loss function, updating the student network only during counter propagation, training 200 rounds, and screening out the model weight with the highest accuracy rate from the training results as a final student model.
Step S5: inputting the data of the test set into the obtained student model, outputting the position coordinate information of the vehicle, setting the threshold value to 0.5, filtering the positioning frames with the confidence coefficient lower than 0.5, and deleting the redundant repeated positioning frames output in the network by using the non-maximum value inhibition, so as to finally obtain the vehicle position of the night scene.
The foregoing is merely a preferred embodiment of the present invention, and is not intended to limit the present invention, and all equivalent variations using the description and drawings of the present invention are within the scope of the present invention.
Claims (5)
1. A night scene-oriented vehicle boundary locating method, comprising the steps of:
s1, acquiring image data of a night road scene;
s2, dividing the acquired image data into a training set, a verification set and a test set, and preprocessing the image data;
s3, performing frame labeling on the preprocessed image data by using a detection model to obtain a label image dataset; training the detection model by using the label image data set and the training set to obtain a final teacher model;
s4, constructing a student model, and distilling boundary positioning knowledge learned by a teacher model by using a positioning distillation method to finally obtain a lightweight network model;
s5, inputting the test set data into a lightweight network model, outputting position coordinate information of the vehicle, setting a threshold value filtering positioning frame, and deleting the repeated positioning frame by using non-maximum value inhibition to finally obtain the vehicle position of the night scene.
2. The night scene-oriented vehicle boundary positioning method according to claim 1, wherein S2 specifically comprises:
image data is collected by using a vehicle-mounted camera or a road monitoring camera, and the original image data is in a RAW format; and using the deep learning network, taking the original format as input, and outputting the RGB format picture with enhanced illumination.
3. The night scene-oriented vehicle boundary positioning method according to claim 1, wherein S3 specifically comprises:
marking the image data by using a detection model to obtain a marking result, namely coordinate information of a vehicle positioning frame, correcting the marking result, and deleting the error positioning frame;
and continuing to train the detection model by taking the corrected labeling data as training data until the detection model converges, so as to obtain a final teacher model.
4. The night scene oriented vehicle boundary positioning method of claim 3, wherein constructing the student model is specifically:
the initial stage is normalized by Batch Normalization after convolution operation, then the feature map is reduced in size by a ReLU activation function and then a maximum pooling layer;
the model then falls into three phases: the first stage comprises a downsampling module and 4 basic modules which are sequentially connected; the second stage is sequentially connected with a downsampling module and 8 basic modules, and features are further extracted and processed; the third stage comprises a downsampling module and 4 basic modules, which are used for deepening the expression capacity of the features; in each stage, the base module consists of convolution, normalization, activation functions and a CBAM attention module; the connection between the basic modules is realized through a channel rearrangement mechanism;
then, the model is normalized by convolution and Batch Normalization again, and global average pooling is carried out, so that the feature map is converted into a fixed size;
finally, the features are mapped to the final output category through the full connection layer, and a prediction result is obtained.
5. The night scene oriented vehicle boundary positioning method according to claim 4, wherein S4 specifically is:
inputting the labeling result, the training set and the image data of the verification set into a teacher model and a student model simultaneously to respectively obtain vehicle positioning frame information;
taking the difference between the vehicle positioning frame information output by the teacher model and the vehicle positioning frame information output by the student model as a first part LOSS1 of the LOSS function; taking the difference between the result output by the student model and the real label as a second part LOSS2 of the LOSS function; taking the difference between the confidence coefficient of the positioning frame output by the student model and the confidence coefficient of the positioning frame output by the teacher model as a third part LOSS3 of the LOSS function; the LOSS1, LOSS2, LOSS3 all use a smooth L1 LOSS function; and training the student model by taking the weighted summation of LOSS1, LOSS2 and LOSS3 as a final LOSS function of the student model until convergence, so as to obtain a final lightweight network model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311105460.6A CN117152513A (en) | 2023-08-30 | 2023-08-30 | Vehicle boundary positioning method for night scene |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311105460.6A CN117152513A (en) | 2023-08-30 | 2023-08-30 | Vehicle boundary positioning method for night scene |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117152513A true CN117152513A (en) | 2023-12-01 |
Family
ID=88911280
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311105460.6A Pending CN117152513A (en) | 2023-08-30 | 2023-08-30 | Vehicle boundary positioning method for night scene |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117152513A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117372819A (en) * | 2023-12-07 | 2024-01-09 | 神思电子技术股份有限公司 | Target detection increment learning method, device and medium for limited model space |
-
2023
- 2023-08-30 CN CN202311105460.6A patent/CN117152513A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117372819A (en) * | 2023-12-07 | 2024-01-09 | 神思电子技术股份有限公司 | Target detection increment learning method, device and medium for limited model space |
| CN117372819B (en) * | 2023-12-07 | 2024-02-20 | 神思电子技术股份有限公司 | Target detection increment learning method, device and medium for limited model space |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111368687B (en) | Sidewalk vehicle illegal parking detection method based on target detection and semantic segmentation | |
| CN109977812B (en) | Vehicle-mounted video target detection method based on deep learning | |
| CN110163187B (en) | F-RCNN-based remote traffic sign detection and identification method | |
| CN111310862B (en) | Image enhancement-based deep neural network license plate positioning method in complex environment | |
| CN110399856B (en) | Feature extraction network training method, image processing method, device and equipment | |
| CN107563372B (en) | License plate positioning method based on deep learning SSD frame | |
| CN112464910A (en) | Traffic sign identification method based on YOLO v4-tiny | |
| CN111814621A (en) | Multi-scale vehicle and pedestrian detection method and device based on attention mechanism | |
| CN104700099A (en) | Method and device for recognizing traffic signs | |
| DE102012218390A1 (en) | Optimizing the detection of objects in images | |
| CN110659601B (en) | Depth full convolution network remote sensing image dense vehicle detection method based on central point | |
| CN112287896A (en) | Unmanned aerial vehicle aerial image target detection method and system based on deep learning | |
| CN117152513A (en) | Vehicle boundary positioning method for night scene | |
| CN113052159A (en) | Image identification method, device, equipment and computer storage medium | |
| CN114913498A (en) | Parallel multi-scale feature aggregation lane line detection method based on key point estimation | |
| CN110852358A (en) | Vehicle type distinguishing method based on deep learning | |
| CN112785610B (en) | Lane line semantic segmentation method integrating low-level features | |
| CN112509321A (en) | Unmanned aerial vehicle-based driving control method and system for urban complex traffic situation and readable storage medium | |
| CN112308066A (en) | License plate recognition system | |
| CN112364864A (en) | License plate recognition method and device, electronic equipment and storage medium | |
| CN112288701A (en) | Intelligent traffic image detection method | |
| CN116597270A (en) | Road damage target detection method based on attention mechanism integrated learning network | |
| CN116434056A (en) | Target identification method and system based on radar fusion and electronic equipment | |
| CN114550016B (en) | Unmanned aerial vehicle positioning method and system based on context information perception | |
| CN114882469A (en) | Traffic sign detection method and system based on DL-SSD model |
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 |