CN114548376A - Intelligent transportation system-oriented vehicle rapid detection network and method - Google Patents

Intelligent transportation system-oriented vehicle rapid detection network and method Download PDF

Info

Publication number
CN114548376A
CN114548376A CN202210170714.1A CN202210170714A CN114548376A CN 114548376 A CN114548376 A CN 114548376A CN 202210170714 A CN202210170714 A CN 202210170714A CN 114548376 A CN114548376 A CN 114548376A
Authority
CN
China
Prior art keywords
data set
tiny
convolution
network structure
yolo network
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
CN202210170714.1A
Other languages
Chinese (zh)
Inventor
陈超
贺巧云
刘海波
张松华
成利香
邹艳葵
吴薇
宋小凤
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.)
Hunan Institute of Technology
Original Assignee
Hunan Institute of Technology
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 Hunan Institute of Technology filed Critical Hunan Institute of Technology
Priority to CN202210170714.1A priority Critical patent/CN114548376A/en
Publication of CN114548376A publication Critical patent/CN114548376A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (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)
  • Traffic Control Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention belongs to the technical field of convolutional neural networks, and discloses a vehicle rapid detection network and a method for an intelligent traffic system. The invention has the beneficial effects that: the detection accuracy of the road vehicle detection method is obviously improved in the environment with uneven illumination, images with different scales and small targets, the number of parameters of the model is obviously reduced, the operation speed is faster, and the road vehicle detection method is light in weight and high in identification accuracy.

Description

Intelligent transportation system-oriented vehicle rapid detection network and method
[ technical field ] A
The invention relates to the technical field of convolutional neural networks, in particular to a vehicle rapid detection network and a method for an intelligent traffic system.
[ background of the invention ]
In recent years, with the acceleration of urbanization construction, the progress of the times and the rapid development of economy, the living standard of people is higher and higher, and the number of various vehicles is increased. Traffic is indispensable in our daily life, provides great convenience and comfort for people's trip, also brings a series of social problems to people simultaneously. Serious traffic problem drawbacks appear in the front of people, which leads to frequent traffic congestion, especially the occurrence of early peak and late peak congestion. In addition, the occurrence frequency of traffic safety accidents is very high, potential safety hazards are brought to the traveling of pedestrians, and the daily life of people is influenced.
Nowadays, society pays more and more attention to road safety and traffic order, and intelligent traffic becomes a hot content of key discussion by means of computer vision. The intelligent traffic system greatly solves the problems existing in the current road traffic, uses a small amount of personnel to participate as much as possible, uses the existing high-tech computer technology to ensure the safety of the road traffic, helps to improve the efficiency of traffic operation, and controls the traffic transportation to dredge. In the application of intelligent transportation systems, pedestrians and vehicles are important key components in road traffic. Therefore, the detection of pedestrians and vehicles by using artificial intelligence technology is a key technology of intelligent transportation and smart cities. The technology automatically identifies pedestrians and vehicles in the input traffic image or video, detects their real positions and marks them with a prediction box.
With the rapid development of deep learning, machine vision and deep learning are combined through a deep convolutional neural network and a large amount of training data, and a large number of methods based on deep learning are used for target detection, namely methods based on Region Proposal (Region Proposal) represented by R-CNN, SPP-Net, Fast R-CNN and Fast R-CNN; the R-CNN detection model proposed by Girshick et al adopts a convolutional neural network to extract image features, and improves the expression capability of samples from an experience-driven artificial feature paradigm to a data-driven learning paradigm; the SPP-Net model proposed by Kaiming He et al improves the R-CNN feature extraction step by adopting spatial pyramid pooling; although Fast R-CNN integrates the advantages of R-CNN and SPP-Net, the candidate box is extracted by adopting a selective search algorithm, and the real-time requirement cannot be met. The Faster R-CNN proposes a Region Proposal Network (RPN). Compared with the method, the method based on the area proposal has greatly improved detection precision, but still faces huge challenges to successfully realize real-time target detection.
In order to solve the problem of detection accuracy and speed balance, another type of algorithm based on a regression idea is proposed, which is mainly represented by YOLO, SSD, YOLOv2, and YOLOv 3. Among them, Redmon et al first proposed a regression-based YOLO network to increase the recognition speed to 45 frames/sec, but sacrifice some detection accuracy, and is prone to false detection and missed detection of small targets. Liu Wei et al propose an SSD model, which predicts on feature maps of different scales, so that the detection accuracy of an image can be ensured even under the condition of small resolution. And then Redmon et al propose a YOLOv2 network through a series of improved methods such as batch standardization, dimension clustering, multi-scale training and the like, further improve the detection speed, and simultaneously keep the detection precision equal to that of the SSD. The YOLOv3 introduces a deep residual error network ResNet on the basis of YOLOv2, so that the detection precision and speed are well balanced, but the network computing complexity is too high due to too many network layers, and the requirement on hardware is high.
Therefore, it is necessary to provide a vehicle rapid detection network and a vehicle rapid detection method for an intelligent transportation system, which are used for rapid detection of road vehicles in the intelligent transportation system, and which significantly improve detection accuracy for uneven-illumination environments, images of different scales and small targets, significantly reduce the number of parameters of models, and achieve faster operation speed, and are provided with a road vehicle detection method which is light in weight and has higher identification accuracy.
[ summary of the invention ]
The invention discloses a vehicle rapid detection network and a method for an intelligent traffic system, which can effectively solve the technical problems involved in the background technology.
In order to achieve the purpose, the technical scheme of the invention is as follows:
a vehicle rapid detection network facing an intelligent transportation system comprises an improved Tiny-YOLO network structure, wherein the improved Tiny-YOLO network structure is improved on the basis of the Tiny-YOLO network structure, the Tiny-YOLO network structure comprises 7 layers, the first 1-6 layers comprise 3 x 3 convolution and 2 x2 pooling, and the 7 th layer comprises 3 x 3 convolution, 3 x 3 convolution and 1 x 1 convolution;
the improved Tiny-YOLO network structure has the improvement points that:
11. selecting at least one layer from the first 1-6 layers of the Tiny-YOLO network structure to add 3 x 3 convolution, wherein the adding position is between 3 x 3 convolution and 2 x2 pooling;
12. adding a 1 × 1 convolution in the layer to which the 3 × 3 convolution is added, at a position between the 3 × 3 convolution and the added 3 × 3 convolution;
13. a 1 × 1 convolution is added between the 7 th layer 3 × 3 convolution and the 3 × 3 convolution.
As a preferred improvement of the present invention: optionally adding a 3 x 3 convolution to two of the first 1-6 layers of the Tiny-YOLO network structure.
A vehicle rapid detection method facing an intelligent transportation system is realized based on the detection network, and comprises the following steps:
s1, constructing a comprehensive data set, wherein the comprehensive data set comprises a self-constructed data set and an open source data set;
s2, preprocessing the comprehensive data set, wherein the preprocessed comprehensive data set is divided into a training data set and a testing data set;
s3, constructing an improved Tiny-YOLO network structure;
s4, training the improved Tiny-YOLO network structure by using a training data set;
s5, testing the trained improved Tiny-YOLO network structure by using a test data set, observing the detection effect, and executing the step S6 if the detection effect is qualified;
s6, detecting the road photo by using the modified Tiny-YOLO network structure.
As a preferred improvement of the present invention: in step S1, the starting data set is the vehicle target image in the KITTI, and the self-established data set is the image captured under the video monitoring of different road scenes.
As a preferred improvement of the present invention: in step S2, the integrated data set is preprocessed, including image normalization, image enhancement, and image manual annotation.
As a preferred improvement of the present invention: in the step S3, after the modified Tiny-YOLO network structure is constructed, the method further includes analyzing the labeling boxes of the training data set and the test data set by using a K-Means ii algorithm, and calculating the optimal number and size of anchors.
As a preferred improvement of the present invention: and the manual image annotation comprises the step of manually annotating the vehicles in the images in the comprehensive data set by adopting Label Img software.
As a preferred improvement of the present invention: in step S4, the improved Tiny-YOLO network structure is trained by using a pralu activation function and using a training data set.
As a preferred improvement of the present invention: in the step S6, the new road scene image is captured to perform a secondary test on the improved Tiny-YOLO network structure, the detection effect is observed, and if the detection effect is qualified, the improved Tiny-YOLO network structure is used to detect the road photo.
As a preferred improvement of the present invention: in step S1, the ratio of the self-created data set to the developed source data set is 1: 9;
in step S2, the ratio of the training data set to the test data set is 4: 1.
The invention has the following beneficial effects:
1. the invention not only reduces the model parameter and the operand, further lightens the network structure on the original Tiny-YOLO, improves the system operation processing speed, but also greatly improves the vehicle detection accuracy of the intelligent traffic system, and meets the requirement of high real-time performance of the intelligent traffic system;
2. the method is suitable for vehicle detection in complex environments such as uneven illumination and images with different scales (including small targets) by optimizing the network model, the algorithm is high in real-time performance and precision, and hardware configuration is not increased while the detection precision is increased;
3. the improved Tiny-YOLO network can be successfully converted to an embedded development platform Jetson TX2 and successfully operates, and the model can be applied to an embedded system, so that the model can be really applied to real-time detection of road vehicles in an intelligent traffic system.
[ description of the drawings ]
In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without inventive efforts, wherein:
FIG. 1 is a logic diagram of a vehicle rapid detection method for an intelligent transportation system according to the present invention;
FIG. 2 is a schematic flow chart of a Tiny-YOLO network structure;
FIG. 3 is a schematic flow chart of a modified Tiny-YOLO network structure.
[ detailed description ] embodiments
The technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the following embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
It should be noted that all the directional indicators (such as up, down, left, right, front, and rear … …) in the embodiment of the present invention are only used to explain the relative position relationship between the components, the movement situation, etc. in a specific posture (as shown in the drawing), and if the specific posture is changed, the directional indicator is changed accordingly.
In addition, the descriptions related to "first", "second", etc. in the present invention are only for descriptive purposes and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality" means at least two, e.g., two, three, etc., unless specifically limited otherwise.
In the present invention, unless otherwise expressly stated or limited, the terms "connected," "secured," and the like are to be construed broadly, and for example, "secured" may be a fixed connection, a removable connection, or an integral part; can be mechanically or electrically connected; they may be directly connected or indirectly connected through intervening media, or they may be connected internally or in any other suitable relationship, unless expressly stated otherwise. The specific meanings of the above terms in the present invention can be understood by those skilled in the art according to specific situations.
In addition, the technical solutions in the embodiments of the present invention may be combined with each other, but it must be based on the realization of those skilled in the art, and when the technical solutions are contradictory or cannot be realized, such a combination of technical solutions should not be considered to exist, and is not within the protection scope of the present invention.
The invention provides a vehicle rapid detection network facing an intelligent transportation system, which comprises an improved Tiny-YOLO network structure, wherein the improved Tiny-YOLO network structure is improved on the basis of the Tiny-YOLO network structure, the Tiny-YOLO network structure comprises 7 layers, the first 1-6 layers comprise 3 x 3 convolution and 2 x2 pooling, and the 7 th layer comprises 3 x 3 convolution, 3 x 3 convolution and 1 x 1 convolution;
the improved Tiny-YOLO network structure has the improvement points that:
11. selecting at least one layer from the first 1-6 layers of the Tiny-YOLO network structure to add 3 x 3 convolution, wherein the adding position is between 3 x 3 convolution and 2 x2 pooling; preferably, 3 x 3 convolutions are added at two optional layers in the Tiny-YOLO network structure.
12. Adding a 1 × 1 convolution in the layer to which the 3 × 3 convolution is added, at a position between the 3 × 3 convolution and the added 3 × 3 convolution;
13. a 1 × 1 convolution is added between the 7 th layer 3 × 3 convolution and the 3 × 3 convolution.
The convolutional neural network has the advantages that the detection accuracy of the environment with uneven illumination, images with different scales and small targets is remarkably improved, the number of parameters of the model is remarkably reduced, and the operation speed is faster. The road vehicle detection method is light in weight and high in identification accuracy, and the dependence on equipment hardware configuration in a traffic detection system is not increased, so that a network model meets the requirements of an intelligent traffic system, particularly the real-time rapid vehicle detection on complex roads. The model is converted into a model on an embedded hardware platform Jetson TX2 of NVIDIA for deployment and successful operation, and a detection result with high precision and strong real-time performance is obtained.
The improved Tiny-YOLO network structure increases the number of network layers, compared with the original Tiny-YOLO model, the optimized structure can extract more features under the same condition after the number of convolution layers is increased, and the identification accuracy is also improved. In addition, the parameter quantity and the model size can be reduced by combining an NIN (network in network) network on the basis of adding the convolution layer number, and the features at the same position of the multichannel layer can be aggregated by the 1 x 1 convolution kernel.
In the application of the actual intelligent transportation system, the detection is required to have high precision and real-time performance no matter the intelligent transportation system is used for detecting traffic illegal behaviors or processing and penalizing traffic accidents. Although the original Tiny-YOLO detection algorithm is a YOLO simplified fast model, in real-world applications, the detection accuracy of the Tiny-YOLO network model is limited due to the fact that the Tiny-YOLO model has fewer network layers because of the cost reduction and no increase of hardware resources. Therefore, the invention optimizes the method to improve the detection precision and maintain smaller calculation amount to meet the detection real-time property.
Referring to fig. 1, the present invention further provides a vehicle rapid detection method for an intelligent transportation system, including the following steps:
s1, constructing a comprehensive data set, wherein the comprehensive data set comprises a self-constructed data set and an open source data set;
s2, preprocessing the comprehensive data set, wherein the preprocessed comprehensive data set is divided into a training data set and a testing data set;
s3, constructing an improved Tiny-YOLO network structure;
s4, training the improved Tiny-YOLO network structure by using a training data set;
s5, testing the trained improved Tiny-YOLO network structure by using a test data set, observing the detection effect, and executing the step S6 if the detection effect is qualified;
s6, detecting the road photo by using the modified Tiny-YOLO network structure.
The method comprises the steps of constructing a comprehensive data set through a self-constructed data set and an opening source data set, training an improved Tiny-YOLO network structure after preprocessing images in the comprehensive data set, detecting an average precision mean mAP and an actual recognition effect of actual road vehicle detection, and realizing rapid and accurate detection of vehicles on complex roads in the intelligent traffic system.
Preferably, in step S1, the opening source data set is a vehicle target image in the KITTI, and the self-established data set is derived from images captured under video surveillance of different road scenes. The self-built data set is mainly used for improving the pertinence and generalization capability of an optimized network structure in the field of intelligent transportation. The data set is augmented by utilizing an open source data set KITTI and combining a self-established data set to form a comprehensive data set, wherein the quantity ratio of the open source data set to the self-established data set in the comprehensive data set is 9: 1.
in step S2, the integrated data set is preprocessed, including image normalization, image enhancement, and image manual annotation. Normalization: the objective is to normalize the original images in the integrated dataset to the same and appropriate size, as required by the input of the Tiny-YOLO model. Image enhancement: the method aims to increase the number of data sets and scenes of various road environments, such as image illumination, brightness and the like. Manual image annotation: the method aims to improve the quality of a training data set, vehicles in an image of a comprehensive data set are manually marked by using Label Img software, and the software generates a corresponding xml file after marking.
And dividing the preprocessed comprehensive data set into two parts according to the ratio of 4:1, wherein the two parts comprise a training data set and a testing data set, the training data set is used for inputting an improved Tiny-YOLO network model for optimization training, and the testing data set is used for testing the trained model. In order to reduce the complexity of model training, a labeling box of a data set is analyzed by using a K-Means II algorithm, and finally the optimal Anchor size and number are determined.
The improved Tiny-YOLO network structure is trained by using a training data set by adopting a PReLU (parametric reconstructed Linear Unit) activation function, and the matching capability of the model can be improved by the PReLU on the premise of increasing a few parameters.
Preferably, in the step S6, the new road scene image is intercepted to perform a secondary test on the improved Tiny-YOLO network structure, the detection effect is observed, and if the detection effect is qualified, the improved Tiny-YOLO network structure is used to detect the road photo. Specifically, the pictures of the test data set are input into the trained improved Tiny-YOLO network structure, and the accuracy of the output vehicle detection is verified. And then, forming a verification data set by using the new monitoring video frame pictures, directly inputting the preprocessed monitoring video frame pictures into a trained improved Tiny-YOLO network model for secondary verification, and detecting the average precision mean mAP and the actual recognition effect of the actual road vehicle detection, thereby realizing the rapid and accurate detection of the vehicles on the complex road in the intelligent traffic system.
When the improved structure test result of the Tiny-YOLO network is not good, the number of pictures of the training data set is not enough or not wide enough, more sample pictures can be added to the comprehensive data set, and the fitting degree of model training is higher. In addition, it may be that the modified Tiny-YOLO network structure is problematic, and the modified Tiny-YOLO network structure may be adjusted to increase or decrease the number of layers of convolution addition, or adjust the position of convolution addition (e.g., layer 3 addition is changed to layer 4 addition), or adjust the training parameters, and make appropriate adjustments according to the results of the test.
As an implementation mode, a vehicle rapid detection method facing an intelligent transportation system is realized by using a modified Tiny-YOLO network structure.
1. Preparing a data set: the data set needs to be subjected to data amplification, and the comprehensive data set after the data amplification is divided into two parts, including an open source data set KITTI and a self-constructed data set. 900 vehicle images are selected from the KITTI data set, 100 video monitors for intercepting different road scenes are selected from the self-established data set, and the total number of the data sets is 1000.
2. Preprocessing a comprehensive data set: the preprocessing is carried out around three aspects of image normalization, image enhancement and image manual labeling. Image size normalization: the normalized image size is 416 x 416. Manually labeling the data set: preferably, the open source labeling tool Label Img labels the road scene image of the integrated data set, then the software generates an xml file, the generated xml file stores the information such as the size, the position and the type of the target in the image, and the road scene image is combined with the xml file to form the data set. Dividing the preprocessed integrated data set into two parts, namely a training set and a testing set, wherein the proportion of the training set to the testing set is 4:1, namely 800 training sets and 200 testing sets, wherein the self-built data set is distributed into 80 training sets and 20 testing sets.
3. The optimized network structure of the Tiny-type optical network is shown in fig. 2, the optimized network structure of the embodiment is divided into the following three aspects, and the improved Tiny-type optical network is shown in fig. 3.
1) Increasing the number of network layers: 3 multiplied by 3 convolutional layers are respectively added between the convolution and the pooling of the fifth layer and the sixth layer, and more network layers mean that the detection precision, particularly the accuracy of small target detection can be improved;
2) insert 1 × 1 convolutional layer: according to the 1 × 1 convolution layer proposed in NIN, a 1 × 1 convolution kernel is applied to the present optimization model using the advantage of the convolution kernel, and a 1 × 1 convolution is inserted between the original 3 × 3 convolution and the added 3 × 3 convolution. Because the number of network layers increased in the previous step 1) can increase the model parameter number and the operation complexity, the 1 × 1 convolution can realize the aggregation of information at the same position of different channels of the characteristic diagram, so that the convolution parameter number is reduced, and the capability of extracting nonlinear characteristics of the improved Tiny-YOLO network structure is improved;
3) data set clustering analysis: the average can be calculated quickly and the optimum anchor size and number can be determined using the K-Means II algorithm. The invention can obtain the optimal Anchor number of 4 and the optimal size of (8.21, 24.89), (17.58, 60.62), (45.32, 144.87), (137.31, 430.38) through experiments.
4. Network training: the training of the improved Tiny-YOLO network structure using the training data set can be significantly less complex since the optimal number and size of anchors have been calculated prior to training. The improved Tiny-YOLO network structure adopts a PReLU activation function for training optimization, and compared with a ReLU function of the original Tiny-YOLO network, the PReLU only increases a few parameter quantity, but can improve the fitting capability of the network.
Figure BDA0003518026940000081
In the formula, xiIs an independent variable, aiFor learnable parameters, subscript i represents the different channels.
5. And (3) verifying the model: the method comprises the steps of using a test data set as input, testing an improved Tiny-YOLO network structure, testing the precision of output vehicle detection, then using a preprocessed new monitoring video frame picture for secondary verification, detecting the average precision mean mAP and the actual recognition effect of actual road vehicle detection, and realizing rapid and accurate detection of vehicles on complex roads in the intelligent traffic system. The image is enhanced when the data set is constructed, the data set is expanded, and meanwhile, the generalization capability and robustness of the network can be improved through a data enhancement technology, so that the detection capability of the improved Tiny-YOLO network structure on targets with different scales and under different environments (such as illumination, shielding and the like) in a complex road is improved.
An improved Tiny-YOLO network structure is used as a model, an open source deep learning frame Pythrch is adopted, and the problem of two classifications, namely a vehicle and a background, is solved. In the training process, each batch training set sample comprises 40 images, and the weight parameters are updated by adopting a batch gradient descent method. The hyper-parameters of the training are set as: setting the initial learning rate to be 0.001, setting the weight attenuation parameter to be 0.0005, setting the epoch to be 350, and obtaining a final network model until the training process converges by observing the loss value. And the test data set is used for testing the finally trained model, displaying the predicted labeling frames and categories on the output image, outputting all detection frames with the confidence degree larger than 0.6, eliminating redundant detection frames by adopting a non-maximum value inhibition method, and finally screening out the optimal target boundary frame.
The confidence C and the intersection ratio IoU use the following formula: c ═ p (o) x IoU
In the formula, C represents the confidence (confidence) of the predicted target frame; p (O) represents the probability that the prediction box contains the target; IoU denotes the Intersection over Union of the predicted box and the actual box, which represents the accuracy of the target location.
Figure BDA0003518026940000091
In the formula, AGTRepresenting a real area box; a. theDRA detection area frame is represented; the numerator represents the intersection of the detection region and the real region; the denominator represents the union of the detection region and the real region.
Figure BDA0003518026940000092
In the formula, Pr represents accuracy; TP represents the number of positive classes predicted as positive classes; FP represents the number of negative classes predicted as positive classes.
Figure BDA0003518026940000093
Wherein Rc represents recall; TP represents the number of positive classes predicted as positive classes; FN is the number of positive classes predicted as negative classes.
Figure BDA0003518026940000094
Figure BDA0003518026940000095
Wherein N (TotalImages) represents the number of images of a certain class of objects in the verification set; n (Classes) represents the number of all classes in the authentication set.
In this example, compared with other networks with the network structures of YOLO, Tiny-YOLO v2 and Tiny-YOLO v3, the performance evaluation indexes and the statistical data thereof are shown in the following table:
use model Accuracy (%) Recall (%) FPS (frame/second) mAP(％) mIoU(％)
YOLO 88.2 84.6 45 63.4 76.7
Tiny-YOLO v2 89.5 85.1 100 58.3 79.2
Tiny-YOLO v3 92.6 87.5 83 63.5 77.4
This example 94.2 91.7 75 64.8 83.2
The method mainly optimizes the structure of the original Tiny-YOLO network, increases the number of convolution layers to improve the precision measurement accuracy according to the characteristics of CNN (convolutional neural network), and reduces the network parameters to improve the operation speed of the model. On the premise of keeping light weight, the recognition accuracy rate of the road scene vehicles can reach 94.2%, the detection speed can reach 75 frames/second, and the road scene vehicles detection device has excellent detection capability and high detection speed.
As an embodiment, three convolution kernels of 3 × 3 and 1 × 1 are added to the fourth, fifth and sixth layers, and a convolution kernel of 1 × 1 is added to the seventh layer, and other steps are the same as the previous embodiment. Through tests, the identification accuracy rate of vehicles in a road scene can reach 94.9%.
The working principle is as follows: the optimization method of the Tiny-YOLO model for the rapid detection of the road vehicles of the intelligent traffic system is provided, the detection accuracy of the environment with uneven illumination, images with different scales and small targets is obviously improved, the number of parameters of the model is obviously reduced, the operation speed is faster, the method for detecting the road vehicles is light in weight and higher in identification accuracy, the dependence on equipment hardware configuration in the traffic detection system is not increased, and the network model meets the requirements of the intelligent traffic system, particularly the real-time rapid detection of the vehicles on complex roads.
While embodiments of the invention have been disclosed above, it is not limited to the applications set forth in the specification and the embodiments, which are fully applicable to various fields of endeavor for which the invention pertains, and further modifications may readily be made by those skilled in the art, it being understood that the invention is not limited to the details shown and described herein without departing from the general concept defined by the appended claims and their equivalents.

Claims (10)

1. A vehicle rapid detection network facing an intelligent transportation system is characterized by comprising an improved Tiny-YOLO network structure, wherein the improved Tiny-YOLO network structure is improved on the basis of the Tiny-YOLO network structure, the Tiny-YOLO network structure comprises 7 layers, the first 1-6 layers comprise 3 x 3 convolution and 2 x2 pooling, and the 7 th layer comprises 3 x 3 convolution, 3 x 3 convolution and 1 x 1 convolution;
the improved Tiny-YOLO network structure has the improvement points that:
11. selecting at least one layer from the first 1-6 layers of the Tiny-YOLO network structure to add 3 x 3 convolution, wherein the adding position is between 3 x 3 convolution and 2 x2 pooling;
12. adding a 1 × 1 convolution in the layer to which the 3 × 3 convolution is added, at a position between the 3 × 3 convolution and the added 3 × 3 convolution;
13. a 1 × 1 convolution is added between the 7 th layer 3 × 3 convolution and the 3 × 3 convolution.
2. The intelligent transportation system-oriented vehicle rapid detection network according to claim 1, wherein: optionally adding a 3 x 3 convolution to two of the first 1-6 layers of the Tiny-YOLO network structure.
3. A method for rapidly detecting vehicles facing an intelligent transportation system, characterized in that the method is implemented based on the network of any one of claims 1-2, and the method comprises the following steps:
s1, constructing a comprehensive data set, wherein the comprehensive data set comprises a self-constructed data set and an open source data set;
s2, preprocessing the comprehensive data set, wherein the preprocessed comprehensive data set is divided into a training data set and a testing data set;
s3, constructing an improved Tiny-YOLO network structure;
s4, training the improved Tiny-YOLO network structure by using a training data set;
s5, testing the trained improved Tiny-YOLO network structure by using a test data set, observing the detection effect, and executing the step S6 if the detection effect is qualified;
s6, detecting the road photo by using the modified Tiny-YOLO network structure.
4. The intelligent transportation system-oriented vehicle rapid detection method according to claim 3, characterized in that: in step S1, the starting data set is the vehicle target image in the KITTI, and the self-established data set is the image captured under the video monitoring of different road scenes.
5. The intelligent transportation system-oriented vehicle rapid detection method according to claim 3, characterized in that: in step S2, the integrated data set is preprocessed, including image normalization, image enhancement, and image manual annotation.
6. The intelligent transportation system-oriented vehicle rapid detection method according to claim 5, characterized in that: in the step S3, after the modified Tiny-YOLO network structure is constructed, the method further includes analyzing the labeling boxes of the training data set and the test data set by using a K-Means ii algorithm, and calculating the optimal number and size of anchors.
7. The intelligent transportation system-oriented vehicle rapid detection method according to claim 5, characterized in that: and the manual image annotation comprises the step of manually annotating the vehicles in the images in the comprehensive data set by adopting Label Img software.
8. The intelligent transportation system-oriented vehicle rapid detection method according to claim 3, characterized in that: in step S4, the improved Tiny-YOLO network structure is trained by using a pralu activation function and using a training data set.
9. The intelligent transportation system-oriented vehicle rapid detection method according to claim 3, characterized in that: in the step S6, the new road scene image is captured to perform a secondary test on the improved Tiny-YOLO network structure, the detection effect is observed, and if the detection effect is qualified, the improved Tiny-YOLO network structure is used to detect the road photo.
10. The intelligent transportation system-oriented vehicle rapid detection method according to claim 3, characterized in that: in step S1, the ratio of the self-created data set to the developed source data set is 1: 9;
in step S2, the ratio of the training data set to the test data set is 4: 1.
CN202210170714.1A 2022-02-24 2022-02-24 Intelligent transportation system-oriented vehicle rapid detection network and method Pending CN114548376A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210170714.1A CN114548376A (en) 2022-02-24 2022-02-24 Intelligent transportation system-oriented vehicle rapid detection network and method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210170714.1A CN114548376A (en) 2022-02-24 2022-02-24 Intelligent transportation system-oriented vehicle rapid detection network and method

Publications (1)

Publication Number Publication Date
CN114548376A true CN114548376A (en) 2022-05-27

Family

ID=81676944

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210170714.1A Pending CN114548376A (en) 2022-02-24 2022-02-24 Intelligent transportation system-oriented vehicle rapid detection network and method

Country Status (1)

Country Link
CN (1) CN114548376A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116958086A (en) * 2023-07-21 2023-10-27 盐城工学院 Metal surface defect detection method and system with enhanced feature fusion capability

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116958086A (en) * 2023-07-21 2023-10-27 盐城工学院 Metal surface defect detection method and system with enhanced feature fusion capability
CN116958086B (en) * 2023-07-21 2024-04-19 盐城工学院 Metal surface defect detection method and system with enhanced feature fusion capability

Similar Documents

Publication Publication Date Title
CN112380952B (en) Power equipment infrared image real-time detection and identification method based on artificial intelligence
Li et al. Application on traffic flow prediction of machine learning in intelligent transportation
CN110059554B (en) Multi-branch target detection method based on traffic scene
CN111626350B (en) Target detection model training method, target detection method and device
CN108537215B (en) Flame detection method based on image target detection
CN111079640B (en) Vehicle type identification method and system based on automatic amplification sample
CN112418117A (en) Small target detection method based on unmanned aerial vehicle image
CN105787472B (en) A kind of anomaly detection method based on the study of space-time laplacian eigenmaps
CN112183471A (en) Automatic detection method and system for standard wearing of epidemic prevention mask of field personnel
CN110188807A (en) Tunnel pedestrian target detection method based on cascade super-resolution network and improvement Faster R-CNN
Yang et al. A vehicle real-time detection algorithm based on YOLOv2 framework
CN104134364B (en) Real-time traffic sign identification method and system with self-learning capacity
CN110689043A (en) Vehicle fine granularity identification method and device based on multiple attention mechanism
CN112287896A (en) Unmanned aerial vehicle aerial image target detection method and system based on deep learning
Gündüz et al. A new YOLO-based method for real-time crowd detection from video and performance analysis of YOLO models
CN114332473B (en) Object detection method, device, computer apparatus, storage medium, and program product
Wang et al. Research on key technologies of intelligent transportation based on image recognition and anti-fatigue driving
CN112464701A (en) Method for detecting whether people wear masks or not based on light weight characteristic fusion SSD
CN116824335A (en) YOLOv5 improved algorithm-based fire disaster early warning method and system
CN115512387A (en) Construction site safety helmet wearing detection method based on improved YOLOV5 model
CN111339950B (en) Remote sensing image target detection method
CN112949510A (en) Human detection method based on fast R-CNN thermal infrared image
CN113781767A (en) Traffic data fusion method and system based on multi-source perception
CN116363532A (en) Unmanned aerial vehicle image traffic target detection method based on attention mechanism and re-parameterization
CN115272882A (en) Discrete building detection method and system based on remote sensing image

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