CN109685118A - A kind of Weak Classifier Adaboost vehicle checking method based on convolutional neural networks feature - Google Patents
A kind of Weak Classifier Adaboost vehicle checking method based on convolutional neural networks feature Download PDFInfo
- Publication number
- CN109685118A CN109685118A CN201811500514.8A CN201811500514A CN109685118A CN 109685118 A CN109685118 A CN 109685118A CN 201811500514 A CN201811500514 A CN 201811500514A CN 109685118 A CN109685118 A CN 109685118A
- Authority
- CN
- China
- Prior art keywords
- weak classifier
- convolutional neural
- neural networks
- vehicle
- feature
- 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
Classifications
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
Abstract
A kind of Weak Classifier Adaboost vehicle checking method based on convolutional neural networks feature, it is high for the vehicle detection discrimination based on convolutional neural networks, but due to the bad problem of complicated network structure bring real-time, design a kind of vehicle checking method of efficient robust, the convolutional neural networks compared with shallow-layer are initially set up to extract vehicle characteristics, then classification is completed in the feature space that convolutional neural networks extract using Weak Classifier Adaboost, be finally completed vehicle detection.The method of the present invention efficiently uses the Fast Classification ability of the powerful ability in feature extraction and Weak Classifier Adaboost of CNN, while the trained defect big with operand in use of the CNN in turn avoiding depth.By on self-built data set with have traditional CNN algorithm, Gabor+SVM algorithm, HOG+SVM algorithm and carried out control experiment, algorithm proposed in this paper behaves oneself best, and real-time is more preferable, and accuracy rate is also up to 98.1%, preferably solves the problems, such as vehicle detection.
Description
Technical field
The invention belongs to computer vision fields, are related to field of vehicle detection under video traffic environment, in particular to a kind of
Weak Classifier Adaboost vehicle checking method based on convolutional neural networks feature.
Background technique
Currently, actively auxiliary drives and Unmanned Systems have become research hotspot, vehicle detection is being total to for these technologies
With basis.After decades of development, vehicle testing techniques make great progress, and detection accuracy is improved.Vehicle detection
General there are two types of modes: nonvisual sensor detection and vision-based inspection.Vehicle checking method based on computer vision
It is cheap, easy to install and use, thus more application prospect, while the also concern by more researchers.But due to multiple
Heterocycle border, imaging angle, illumination variation, vehicle diversity etc. influence, and vehicle detection based on computer vision is still one
Extremely challenging problem.
Vehicle checking method based on computer vision mainly has based on feature, based on optical flow method, based on template and be based on
Four kinds of methods of machine learning.With the continuous development of machine learning Domain Theory and technology, the vehicle detection based on machine learning
Method has become main stream approach, and wherein deep learning method is even more to start a burst of research boom.
Convolutional neural networks as one of deep learning method avoid tional identification using image as the input of network
Complicated feature extraction and data reconstruction processes in algorithm.Meanwhile this network structure to translation, scaling, inclination or
The deformation of its form has height invariance altogether, and vehicle has many variations on scale, position, direction in driving process, because
CNN is applied to vehicle detection by this, and the precision of detection can be improved.But to obtain preferable recognition effect, it is necessary to establish deep
Layer model.Such as in 2014ImageNet LSVRC contest, the GoogLeNet based on convolutional neural networks model is obtained most
Good achievement, the network share 22 layers, if including down-sampling layer, up to 27 layers.In such identifying system, CNN is not only
For doing feature extraction, and it is used to carry out identification classification.Generally for higher discrimination is reached, network architecture is very
Complexity at least needs three to four layers of network structure (only calculating the number of convolutional layer) for this task of vehicle detection, this
Complicated model structure can seriously affect real-time, be unfavorable for the raising of detected vehicular velocity.
Summary of the invention
The problem to be solved in the present invention is: it is high for the vehicle detection discrimination based on convolutional neural networks, but due to net
Network structure is complicated the bad problem of bring real-time, designs a kind of vehicle checking method of efficient robust.
The technical solution of the present invention is as follows: being completed using the excellent ability in feature extraction of convolutional neural networks to vehicle characteristics
It extracts, constructs the convolutional neural networks compared with shallow-layer and the high dimensional feature of vehicle is extracted, reduce the complexity of network, make
The training process of CNN and the operand of use process are greatly lowered, and improve real-time under the premise of guaranteeing characteristic validity;
On this basis, it using the good classification capacity of Weak Classifier Adaboost classification device and rapidity, realizes to vehicle
Detection.
A kind of Weak Classifier Adaboost vehicle checking method based on convolutional neural networks feature, for intelligent vehicle or
Automatic driving vehicle initially sets up the convolutional neural networks compared with shallow-layer to extract vehicle characteristics, then utilizes Weak Classifier
Adaboost completes classification in the feature space that convolutional neural networks extract, and is finally completed vehicle detection, comprising the following steps:
1) a convolutional neural networks feature extractor is established using training sample, completion mentions vehicle high dimensional feature
It takes, specifically:
1. establishing a convolutional neural networks model, whole network mainly includes input layer, two convolutional layers, is adopted under two
Sample layer, full articulamentum and output layer, as shown in Figure 2.Using the sample training model, preferable recognition capability is made it have;
2. utilizing the high dimensional feature of trained model extraction vehicle.Training sample is rejoined into trained model
In, convolutional layer C1, down-sampling layer S2, convolutional layer C3 and down-sampling layer S4 shown in Fig. 2 are stepped into, later by the output of S4
As the feature vector that CNN is extracted, completes vehicle characteristics and extract.
2) a Weak Classifier Adaboost decision is established using the feature vector that convolutional neural networks extract training sample
Classifier, specifically:
1. being looked for after the feature vector for extracting training sample using convolutional neural networks according to Weak Classifier Adaboost algorithm
Optimal threshold corresponding to each feature vector out obtains a Weak Classifier;
2. calculating each Weak Classifier to the classification error rate of sample set, and then obtain the optimal weak typing of epicycle training
Device;
3. updating sample weights, iteration is taken turns by T, obtains T Weak Classifier;
4. Weak Classifier is merged to obtain a final Decision Classfication device.
3) test sample is added in step 1 in the CNN* model of training, obtain feature corresponding to test sample to
Amount, enters step in 2 trained Decision Classfication devices, finally obtains testing result.
The present invention has the following advantages compared with prior art: CNN is used as feature extractor, Weak Classifier by the present invention
Adaboost algorithm is as feature selecting and classifier.Design in this way can extract effective feature to characterize vehicle
And background, and the terseness and high efficiency having due to Weak Classifier Adaboost algorithm improve the speed and essence of vehicle detection
Degree.The Fast Classification ability of this method effective use CNN powerful ability in feature extraction and Weak Classifier Adaboost, simultaneously
In turn avoid the trained defect big with operand in use of CNN of depth.Algorithm proposed by the present invention performance is most in comparative experiments
Good, real-time is more preferable, and accuracy rate is also up to 98.1%, preferably solves the problems, such as vehicle detection.
Detailed description of the invention
Fig. 1 is implementing procedure of the invention.
Fig. 2 is the convolutional neural networks model that the present invention establishes.
Fig. 3 is the vehicle characteristics that the convolutional neural networks feature extractor that the present invention establishes proposes.
Fig. 4 is the training process of Weak Classifier Adaboost Decision Classfication device of the invention.
Fig. 5 is vehicle detection effect of the invention.
Fig. 6 be the present invention on self-built data set with the comparison result on other methods recognition performance.
Specific embodiment
The invention proposes a kind of Weak Classifier Adaboost vehicle checking method based on convolutional neural networks feature, such as
Shown in Fig. 1, comprising the following steps:
1) diversity (such as different scenes, different weather) for considering sample makes that network is trained to have stronger adaptation
Ability extracts picture from the video obtained on running recording instrument of automobile (ADR), a part as sample;Another part takes
From MIT Car Database, by its sample level mirror image, enlarged sample collection constitutes the self-built data set of sample.
2) it establishes succinct shallow-layer CNN network and extracts the high dimensional feature of vehicle using the network.It compared two kinds of networks
Structure-CNN-1 and CNN-2, to verify influence of the heterogeneous networks structure to feature extraction performance.The only setting one in CNN-1
Convolutional layer and a down-sampling layer, and 2 convolutional layers and down-sampling layer are then provided in CNN-2.Specific implementation step are as follows:
1. keeping two network structures optimal by adjusting parameter, best convergence rate and discrimination are obtained;
2. test sample is separately input in CNN-1 model and CNN-2 model, the S1 in CNN-1 model is taken out respectively
The output of S2 layer in the output of layer and CNN-2 model is as test sample feature vector;
3. its each layer feature is shown using visualization method for single sample.By the feature after visualizing
Figure is as it can be seen that CNN-2 can effectively extract the further feature of vehicle and background, as shown in Figure 3 with preferable feature representation ability.
3) the corresponding strong classifier of Weak Classifier Adaboost algorithm training, specific implementation step are utilized are as follows:
1. the training sample set S of Weak Classifier Adaboost algorithm is established using the vehicle characteristics extracted in step 2, wherein
X and Y corresponds respectively to positive example sample and negative example sample;T is the maximum cycle of training;
2. initialization sample weight is 1/n, the as initial probability distribution of training sample;
3. carrying out first time iteration: seeking optimal threshold corresponding to each feature vector first, obtain a weak typing
Then device seeks each Weak Classifier to the classification error rate of sample set, and then obtain the optimal Weak Classifier of epicycle training;
4. obtaining T Weak Classifier after T iteration;
5. Weak Classifier is merged to obtain a final Decision Classfication device.
4) test sample is added in step 2) in the CNN-2 model of training, obtains feature corresponding to test sample
Vector enters step in 3) trained Decision Classfication device, finally obtains court verdict.
Implementation of the invention is on self-built data set and has traditional CNN algorithm, Gabor+SVM algorithm, HOG+
SVM algorithm is compared, and respective error rate is calculated, and specific experiment result is as shown in Figure 6.It can be seen that algorithm proposed in this paper, wrong
Accidentally rate is minimum, is 1.92%, it reduces 0.83% than two layers traditional of CNN model.If being realized with traditional CNN, at least
It needs three layers or four-layer network network just can achieve this error rate.This algorithm herein, for the sample of 96*96 pixel size, 1
It can detect 483 in second, and in traditional two layers of CNN identification, it can detect 158 samples in 1 second.According to more complicated net
Network structure, detection speed will be slower.Pass through this comparison, it is seen that the vehicle detection based on CNN and Weak Classifier Adaboost
Algorithm not only increases accuracy of identification, while improving recognition speed, this is for the very high vehicle detection of requirement of real-time
It is very important.
Claims (3)
1. a kind of vehicle checking method of the Weak Classifier Adaboost method based on convolutional neural networks feature, it is characterized in that with
It is driven in active auxiliary or automatic driving vehicle is to the vehicle detection in environment, initially set up the convolutional Neural net compared with shallow-layer
Network extracts vehicle characteristics, then complete in the feature space that convolutional neural networks extract using Weak Classifier Adaboost method
At vehicle detection, comprising the following steps:
1) a convolutional neural networks feature extractor is established using training sample, completes the extraction to vehicle high dimensional feature;
2) a Weak Classifier Adaboost Decision Classfication is established using the feature vector that convolutional neural networks extract training sample
Device;
3) test sample is added in step 1 in the CNN model of training, obtains feature vector corresponding to test sample, into
Enter in the trained Decision Classfication device of step 2, finally obtains testing result.
2. a kind of vehicle detection side Weak Classifier Adaboost based on convolutional neural networks feature according to claim 1
Method, it is characterized in that the convolutional neural networks feature extractor that step 1) is established, specifically:
1.1) a convolutional neural networks model is established, whole network mainly includes input layer, two convolutional layers, two down-samplings
Layer, full articulamentum and output layer make it have preferable recognition capability using the sample training model;
1.2) high dimensional feature of trained model extraction vehicle is utilized.Training sample is rejoined in trained model,
Step into convolutional layer C1, down-sampling layer S2, convolutional layer C3 and down-sampling layer S4 shown in Fig. 2, later using the output of S4 as
The feature vector that CNN is extracted is completed vehicle characteristics and is extracted.
3. a kind of vehicle detection side Weak Classifier Adaboost based on convolutional neural networks feature according to claim 1
Method, it is characterized in that step 2), establishes a Weak Classifier Adaboost Decision Classfication device using the feature vector that step 1) is extracted,
Specifically:
2.1) it after the feature vector for extracting training sample using convolutional neural networks, is found out according to Weak Classifier Adaboost algorithm
Optimal threshold corresponding to each feature vector obtains a Weak Classifier;
2.2) each Weak Classifier is calculated to the classification error rate of sample set, and then obtains the optimal Weak Classifier of epicycle training;
2.3) sample weights are updated, iteration is taken turns by T, obtains T Weak Classifier;
2.4) Weak Classifier is merged to obtain a final Decision Classfication device.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811500514.8A CN109685118A (en) | 2018-12-07 | 2018-12-07 | A kind of Weak Classifier Adaboost vehicle checking method based on convolutional neural networks feature |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811500514.8A CN109685118A (en) | 2018-12-07 | 2018-12-07 | A kind of Weak Classifier Adaboost vehicle checking method based on convolutional neural networks feature |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109685118A true CN109685118A (en) | 2019-04-26 |
Family
ID=66187337
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811500514.8A Pending CN109685118A (en) | 2018-12-07 | 2018-12-07 | A kind of Weak Classifier Adaboost vehicle checking method based on convolutional neural networks feature |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109685118A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110307982A (en) * | 2019-06-19 | 2019-10-08 | 电子科技大学 | Bearing fault classification method based on CNN and Adaboost |
CN112183185A (en) * | 2020-08-13 | 2021-01-05 | 天津大学 | Liquid leakage detection method based on optical flow method and CNN-SVM |
CN116204784A (en) * | 2022-12-30 | 2023-06-02 | 成都天仁民防科技有限公司 | DAS-based subway tunnel external hazard operation intrusion recognition method |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106250838A (en) * | 2016-07-27 | 2016-12-21 | 乐视控股(北京)有限公司 | vehicle identification method and system |
CN107742132A (en) * | 2017-11-07 | 2018-02-27 | 江南大学 | Potato detection method of surface flaw based on convolutional neural networks |
EP3291146A1 (en) * | 2016-09-05 | 2018-03-07 | Fujitsu Limited | Knowledge extraction from a convolutional neural network |
CN108629279A (en) * | 2018-03-27 | 2018-10-09 | 哈尔滨理工大学 | A method of the vehicle target detection based on convolutional neural networks |
-
2018
- 2018-12-07 CN CN201811500514.8A patent/CN109685118A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106250838A (en) * | 2016-07-27 | 2016-12-21 | 乐视控股(北京)有限公司 | vehicle identification method and system |
EP3291146A1 (en) * | 2016-09-05 | 2018-03-07 | Fujitsu Limited | Knowledge extraction from a convolutional neural network |
CN107742132A (en) * | 2017-11-07 | 2018-02-27 | 江南大学 | Potato detection method of surface flaw based on convolutional neural networks |
CN108629279A (en) * | 2018-03-27 | 2018-10-09 | 哈尔滨理工大学 | A method of the vehicle target detection based on convolutional neural networks |
Non-Patent Citations (1)
Title |
---|
封晶: "基于区域卷积神经网络的车辆检测方法", 《科技广场》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110307982A (en) * | 2019-06-19 | 2019-10-08 | 电子科技大学 | Bearing fault classification method based on CNN and Adaboost |
CN110307982B (en) * | 2019-06-19 | 2021-01-26 | 电子科技大学 | Bearing fault classification method based on CNN and Adaboost |
CN112183185A (en) * | 2020-08-13 | 2021-01-05 | 天津大学 | Liquid leakage detection method based on optical flow method and CNN-SVM |
CN116204784A (en) * | 2022-12-30 | 2023-06-02 | 成都天仁民防科技有限公司 | DAS-based subway tunnel external hazard operation intrusion recognition method |
CN116204784B (en) * | 2022-12-30 | 2023-09-08 | 成都天仁民防科技有限公司 | DAS-based subway tunnel external hazard operation intrusion recognition method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111444821B (en) | Automatic identification method for urban road signs | |
CN111553387B (en) | Personnel target detection method based on Yolov3 | |
CN104050471B (en) | Natural scene character detection method and system | |
CN108090906B (en) | Cervical image processing method and device based on region nomination | |
CN108171112A (en) | Vehicle identification and tracking based on convolutional neural networks | |
CN108985217A (en) | A kind of traffic sign recognition method and system based on deep space network | |
CN106919902B (en) | Vehicle identification and track tracking method based on CNN | |
CN108960141A (en) | Pedestrian's recognition methods again based on enhanced depth convolutional neural networks | |
CN109543632A (en) | A kind of deep layer network pedestrian detection method based on the guidance of shallow-layer Fusion Features | |
CN109685118A (en) | A kind of Weak Classifier Adaboost vehicle checking method based on convolutional neural networks feature | |
CN113420607A (en) | Multi-scale target detection and identification method for unmanned aerial vehicle | |
CN111898432B (en) | Pedestrian detection system and method based on improved YOLOv3 algorithm | |
CN112418330A (en) | Improved SSD (solid State drive) -based high-precision detection method for small target object | |
CN110197494A (en) | A kind of pantograph contact point real time detection algorithm based on monocular infrared image | |
CN114972312A (en) | Improved insulator defect detection method based on YOLOv4-Tiny | |
CN104504381A (en) | Non-rigid target detection method and system thereof | |
CN114049572A (en) | Detection method for identifying small target | |
CN107330381A (en) | A kind of face identification method | |
CN111462140A (en) | Real-time image instance segmentation method based on block splicing | |
CN109919107B (en) | Traffic police gesture recognition method based on deep learning and unmanned vehicle | |
Liu et al. | CAFFNet: channel attention and feature fusion network for multi-target traffic sign detection | |
CN110751005B (en) | Pedestrian detection method integrating depth perception features and kernel extreme learning machine | |
CN109543519B (en) | Depth segmentation guide network for object detection | |
Wu et al. | Traffic sign detection based on SSD combined with receptive field module and path aggregation network | |
CN117079095A (en) | Deep learning-based high-altitude parabolic detection method, system, medium and equipment |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190426 |