CN110276247A - A kind of driving detection method based on YOLOv3-Tiny - Google Patents
A kind of driving detection method based on YOLOv3-Tiny Download PDFInfo
- Publication number
- CN110276247A CN110276247A CN201910387559.7A CN201910387559A CN110276247A CN 110276247 A CN110276247 A CN 110276247A CN 201910387559 A CN201910387559 A CN 201910387559A CN 110276247 A CN110276247 A CN 110276247A
- Authority
- CN
- China
- Prior art keywords
- yolov3
- tiny
- algorithm
- driving
- detection method
- 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
- G06F18/23—Clustering techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
-
- 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
-
- 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
Abstract
The driving detection method based on YOLOv3-Tiny that the invention discloses a kind of includes the following steps: that S1, training sample are obtained；S2, YOLOv3-Tiny algorithm is improved；S3, model training and output: modified hydrothermal process network in step S2 is trained using the training sample in step S1, iterative learning exports new deep learning model；S4, driving image obtain: utilizing unmanned aerial vehicle onboard camera shooting driving image；S5, driving image detection: the image in S4 is carried out using innovatory algorithm model in S3 to detect figure of obtaining a result.Compared with the algorithm before improvement, the accuracy of former algorithm target detection can be improved in the case where guaranteeing real-time and occupying the lesser situation in video memory space, improve driving detection accuracy.
Description
Technical field
The present invention relates to driving detection method, in particular to a kind of driving detection methods based on YOLOv3-Tiny.
Background technique
Target detection (object detection) finds the specific of target in given image, video or scene
Position, length and width information, while classification belonging to target is judged, target detection will generally solve two problems: object
Body where, the classification of target object (i.e. what target object is).Target detection is widely used in practical problem, especially
With the fast development of deep learning, while target detection technique is also rapidly developing, especially compared to passing through conventional method
Target detection, the target detection based on deep learning can improve the Efficiency and accuracy of detection, and answer in every field
With.It and is also the foundation stone of the tasks such as target identification, target following, the quality of target detection performance influences target following, target
The performance of the tasks such as identification.But it is more demanding to hardware device based on the target detection of deep learning, in real life
Difficulty is larger, still not mature enough in technical research, therefore there is an urgent need to more effective, practical object detection methods.
Summary of the invention
Goal of the invention: it is an object of the present invention to provide one kind to guarantee that target detection real-time and occupancy video memory are smaller
In the case where have compared with high measurement accuracy the driving detection method based on YOLOv3-Tiny.
Technical solution: the present invention provides a kind of driving detection method based on YOLOv3-Tiny, includes the following steps:
S1, training sample obtain；
S2, YOLOv3-Tiny algorithm is improved；
S3, model training and output: modified hydrothermal process network in step S2 is carried out using the training sample in step S1
Training, iterative learning export new deep learning model；
S4, driving image obtain: utilizing unmanned aerial vehicle onboard camera shooting driving image；
S5, driving image detection: the image in S4 is carried out using innovatory algorithm model in S3 to detect figure of obtaining a result.
Further, in the step S1, KITTI data set is chosen as training sample.
Further, model training uses and is uniformly distributed strategy in the step S3.
Further, include the following steps: in the step S2
S2.1, change network inputs image pixel size；
S2.2, default candidate frame clustering；
S2.3, YOLOv3-Tiny algorithm network structure is improved.
Further, image pixel by 416 × 416 is changed to 672 × 224 in the step S2.1.In the step S2.2
Clustering is carried out to default candidate frame using kmeans algorithm.In the step S2.3, network increases the convolution that 3 sizes are
The convolutional layer that layer and 3 sizes are.
The utility model has the advantages that the present invention compares R-CNN series and YOLO serial algorithm, the present invention requires not hardware device
Height, video memory space hold is smaller, specifically about 1GB, can be widely used in simple object detection system；Comparison
YOLO serial algorithm, modified hydrothermal process complexity of the present invention are lower；The improved YOLOv3-Tiny algorithm of the present invention, with improvement
Preceding algorithm compares, and can improve former algorithm target detection in the case where guaranteeing real-time and occupying the lesser situation in video memory space
Accuracy, improve driving detection accuracy.
Detailed description of the invention
Fig. 1 is algorithm network structure of the invention；
Fig. 2 is the detection comparison diagram of algorithm and former algorithm accuracy of the invention, and wherein Our Network is that the present invention changes
Algorithm testing result after, YOLOv3-Tiny are existing algorithm testing results；
Fig. 3 is KITTI data set test chart of the present invention；
Fig. 4 is unobstructed situation detection effect figure；
Fig. 5 is Small object object detection effect picture；
Fig. 6 is to have circumstance of occlusion detection effect figure.
Specific embodiment
As shown in figures 1 to 6, the driving detection method based on YOLOv3-Tiny of the present embodiment, comprising the following steps:
S1: choose training sample: travelling data collection needed for choosing network model training, the data set of selection is KITTI number
According to collection, training sample 5984 is opened in the data set, test sample 748 is opened.KITTI travelling data collection background is also complex, deposits
The occlusion situations such as.In order to make KITTI data set be suitable for YOLOv3-Tiny algorithm, format conversion is carried out to data set,
It first converts the travelling data collection of KITTI to the data set of VOC format, is converted into is suitable for improving network on this basis
Data set format.
S2: the improvement based on YOLOv3-Tiny algorithm: change input image pixels size, default candidate frame clustering,
Network model improves, specifically includes the following steps:
S2.1: changing input image pixels size, will specifically input size from 416 × 416 and be changed to 672 × 224；
S2.2: default candidate frame clustering is split it using grid when input picture in handling S2.1,
K are arranged in each grid with reference to anchor, training is classified using GroundTruth (true frame) as benchmark and returns damage
It loses.K anchor boxes corresponds to k different scale, all has independent sorting as a result, optimizing the accuracy rate of network, finds
The high-dimensional and optimal k value of optimal anchor wide.The objective function of cluster are as follows:
In formula (1), i indicates that the classification number of cluster, j indicate the quantity of data set, and Box [i] is the pre- of each cluster centre
If the size of frame, Truth [j] indicates each pedestrian's frame size in data set, and objective function makes prediction block and true frame to the greatest extent may be used
It can be overlapped.
S2.3: network structure is improved, there are three types of types for common convolution kernel size: 3 × 3,5 × 5,7 × 7, two sizes
5 × 5 convolution kernel can be substituted for 3 × 3 convolution kernel stacking, the convolution kernel stacking that three sizes are 3 × 3 can substitute 7 × 7
Convolution kernel.The number of plies of network is too deep to have that gradient explosion or gradient disappear, and the network number of plies is excessively shallow then cannot be abundant
The feature of learning objective, improved network structure are the convolutional layers for increasing by 33 × 3 on the basis of original core network, are reached
Deepen the effect of network.It is improved since the increase of the network number of plies will lead to the operand of network and the increase of model parameter
The convolutional layer that 3 sizes are 1 × 1 is introduced in network, it is therefore an objective to reduce model parameter and improve the learning ability of network.
S3: training, sample model training and model output: are iterated to data set with improved YOLOv3-Tiny algorithm
This picture passes through training iterative learning and exports deep learning model.The model training of the present embodiment uses and is uniformly distributed strategy, if
The initial learning rate of cover half type is 0.001, and after 20000 times and 25000 iteration, learning rate value is multiplied by 0.1, momentum coefficient
It is set as 0.9, weight attenuation coefficient is 0.001.
S4: driving image obtains: obtaining driving image data using unmanned aerial vehicle onboard camera.The unmanned plane used is four
Rotor wing unmanned aerial vehicle, model use predator 680, and airborne monocular cam model is to fly jade-like stone 8s, pixel size for be 1280 ×
720.Unmanned machine equipment includes flight controller, remote control, GPS, holder etc., the present embodiment using wireless image transmission technology by nobody
The image transmitting of the airborne camera captured in real-time of machine is into computer.
S5: driving image detection: detecting the driving image in S4 using the training pattern in S3, obtains detection knot
Fruit figure.
The present embodiment modified hydrothermal process and former algorithm performance detect comparing result such as table 1:
1 the present embodiment modified hydrothermal process of table and former algorithm performance detect comparing result
| Detection method | MAP/% | Recall/% | Averagely hands over and compare IOU/% | Detect speed/(frame/second) |
| YOLOv3-Tiny | 84.78 | 83 | 70.88 | 29.7 |
| Improved YOLOv3-Tiny | 87.97 | 85 | 74.88 | 27.5 |
Comparison improves the performance of both front and back algorithm, it can be found that improved YOLOv3-Tiny algorithm is average in mean value
3.19% is improved on precision mAP, recall rate Recall improves 2.00%, and it is average to hand over and improve 4.00% than IOU, it is examining
There is decline slightly in terms of degree of testing the speed, detects speed under normal conditions in the 25 frames/more than second can reach target identification real-time
It is required that the experimental results showed that improved algorithm can improve the precision of target identification in the case where meeting real-time.
The present embodiment improves mAP performance comparison figure such as Fig. 2 of front and back, and dotted line indicates before improving as a result, solid line expression changes
It is after as a result, comparison dotted line and solid line under area (the accuracy mAP that area represents algorithm), it can be clearly seen that under solid line
Area be greater than the area under dotted line, therefore improved algorithm accuracy of identification is higher.
The improved algorithm of the present embodiment is directed to KITTI data set testing result figure such as Fig. 3, it can be seen that detection block fitting
Car body can also obtain accurate testing result when driving a vehicle target farther out, and thus improved YOLOv3-Tiny algorithm exists
There is preferable performance on accuracy of identification.
The improved algorithm of the present embodiment is directed to image detection result such as Fig. 4, Fig. 5, Fig. 6 of unmanned plane shooting, respectively generation
The unobstructed situation detection effect figure of table, Small object object detection effect picture have circumstance of occlusion detection effect figure, according to three kinds of results
Figure is it can be found that the improved algorithm of this example has preferable detection performance.
Claims (7)
1. a kind of driving detection method based on YOLOv3-Tiny, characterized by the following steps:
S1, training sample obtain；
S2, YOLOv3-Tiny algorithm is improved；
S3, model training and output: being trained modified hydrothermal process network in step S2 using the training sample in step S1,
Iterative learning exports new deep learning model；
S4, driving image obtain: utilizing unmanned aerial vehicle onboard camera shooting driving image；
S5, driving image detection: the image in S4 is carried out using innovatory algorithm model in S3 to detect figure of obtaining a result.
2. the driving detection method according to claim 1 based on YOLOv3-Tiny, it is characterised in that: the step S1
In, KITTI data set is chosen as training sample.
3. the driving detection method according to claim 1 based on YOLOv3-Tiny, it is characterised in that: the step S3
Middle model training uses and is uniformly distributed strategy.
4. the driving detection method according to claim 1 based on YOLOv3-Tiny, it is characterised in that: the step S2
In include the following steps:
S2.1, change network inputs image pixel size；
S2.2, default candidate frame clustering；
S2.3, YOLOv3-Tiny algorithm network structure is improved.
5. the driving detection method according to claim 4 based on YOLOv3-Tiny, it is characterised in that: the step
Image pixel is changed to 672 × 224 by 416 × 416 in S2.1.
6. the driving detection method according to claim 4 based on YOLOv3-Tiny, it is characterised in that: the step
Clustering is carried out to default candidate frame using kmeans algorithm in S2.2.
7. the driving detection method according to claim 4 based on YOLOv3-Tiny, it is characterised in that: the step
In S2.3, network increases the convolutional layer that 3 sizes are and the convolutional layer that 3 sizes are.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910387559.7A CN110276247A (en) | 2019-05-09 | 2019-05-09 | A kind of driving detection method based on YOLOv3-Tiny |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910387559.7A CN110276247A (en) | 2019-05-09 | 2019-05-09 | A kind of driving detection method based on YOLOv3-Tiny |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110276247A true CN110276247A (en) | 2019-09-24 |
Family
ID=67959197
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910387559.7A Pending CN110276247A (en) | 2019-05-09 | 2019-05-09 | A kind of driving detection method based on YOLOv3-Tiny |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110276247A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110826520A (en) * | 2019-11-14 | 2020-02-21 | 燕山大学 | Port grab bucket detection method based on improved YOLOv3-tiny algorithm |
| CN110852358A (en) * | 2019-10-29 | 2020-02-28 | 中国科学院上海微系统与信息技术研究所 | Vehicle type distinguishing method based on deep learning |
| CN110942005A (en) * | 2019-11-21 | 2020-03-31 | 网易（杭州）网络有限公司 | Object recognition method and device |
| CN111046787A (en) * | 2019-12-10 | 2020-04-21 | 华侨大学 | Pedestrian detection method based on improved YOLO v3 model |
| CN111444916A (en) * | 2020-03-26 | 2020-07-24 | 中科海微(北京)科技有限公司 | License plate positioning and identifying method and system under unconstrained condition |
| CN111563525A (en) * | 2020-03-25 | 2020-08-21 | 北京航空航天大学 | Moving target detection method based on YOLOv3-Tiny |
| CN112734794A (en) * | 2021-01-14 | 2021-04-30 | 北京航空航天大学 | Moving target tracking and positioning method based on deep learning |
| CN113536963A (en) * | 2021-06-25 | 2021-10-22 | 西安电子科技大学 | SAR image airplane target detection method based on lightweight YOLO network |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108647742A (en) * | 2018-05-19 | 2018-10-12 | 南京理工大学 | Fast target detection method based on lightweight neural network |
| CN108921875A (en) * | 2018-07-09 | 2018-11-30 | 哈尔滨工业大学（深圳） | A kind of real-time traffic flow detection and method for tracing based on data of taking photo by plane |
| CN109685152A (en) * | 2018-12-29 | 2019-04-26 | 北京化工大学 | A kind of image object detection method based on DC-SPP-YOLO |
-
2019
- 2019-05-09 CN CN201910387559.7A patent/CN110276247A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108647742A (en) * | 2018-05-19 | 2018-10-12 | 南京理工大学 | Fast target detection method based on lightweight neural network |
| CN108921875A (en) * | 2018-07-09 | 2018-11-30 | 哈尔滨工业大学（深圳） | A kind of real-time traffic flow detection and method for tracing based on data of taking photo by plane |
| CN109685152A (en) * | 2018-12-29 | 2019-04-26 | 北京化工大学 | A kind of image object detection method based on DC-SPP-YOLO |
Non-Patent Citations (1)
| Title |
|---|
| 刘军等: "基于增强Tiny YOLOV3算法的车辆实时检测与跟踪", 《农业工程学报》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110852358A (en) * | 2019-10-29 | 2020-02-28 | 中国科学院上海微系统与信息技术研究所 | Vehicle type distinguishing method based on deep learning |
| CN110826520A (en) * | 2019-11-14 | 2020-02-21 | 燕山大学 | Port grab bucket detection method based on improved YOLOv3-tiny algorithm |
| CN110942005A (en) * | 2019-11-21 | 2020-03-31 | 网易（杭州）网络有限公司 | Object recognition method and device |
| CN111046787A (en) * | 2019-12-10 | 2020-04-21 | 华侨大学 | Pedestrian detection method based on improved YOLO v3 model |
| CN111563525A (en) * | 2020-03-25 | 2020-08-21 | 北京航空航天大学 | Moving target detection method based on YOLOv3-Tiny |
| CN111444916A (en) * | 2020-03-26 | 2020-07-24 | 中科海微(北京)科技有限公司 | License plate positioning and identifying method and system under unconstrained condition |
| CN112734794A (en) * | 2021-01-14 | 2021-04-30 | 北京航空航天大学 | Moving target tracking and positioning method based on deep learning |
| CN112734794B (en) * | 2021-01-14 | 2022-12-23 | 北京航空航天大学 | Moving target tracking and positioning method based on deep learning |
| CN113536963A (en) * | 2021-06-25 | 2021-10-22 | 西安电子科技大学 | SAR image airplane target detection method based on lightweight YOLO network |
| CN113536963B (en) * | 2021-06-25 | 2023-08-15 | 西安电子科技大学 | SAR image airplane target detection method based on lightweight YOLO network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110276247A (en) | A kind of driving detection method based on YOLOv3-Tiny | |
| CN111274970B (en) | Traffic sign detection method based on improved YOLO v3 algorithm | |
| CN113807464B (en) | Unmanned aerial vehicle aerial image target detection method based on improved YOLO V5 | |
| CN110175615B (en) | Model training method, domain-adaptive visual position identification method and device | |
| CN113420643B (en) | Lightweight underwater target detection method based on depth separable cavity convolution | |
| CN108549866B (en) | Remote sensing airplane identification method based on dense convolutional neural network | |
| CN109376580A (en) | A kind of electric tower component identification method based on deep learning | |
| CN111126278A (en) | Target detection model optimization and acceleration method for few-category scene | |
| CN116385958A (en) | Edge intelligent detection method for power grid inspection and monitoring | |
| CN111540203B (en) | Method for adjusting green light passing time based on fast-RCNN | |
| CN111340019A (en) | Grain bin pest detection method based on Faster R-CNN | |
| CN114419413A (en) | Method for constructing sensing field self-adaptive transformer substation insulator defect detection neural network | |
| CN113989487A (en) | Fault defect detection method and system for live-action scheduling | |
| Li et al. | Method research on ship detection in remote sensing image based on Yolo algorithm | |
| CN111563525A (en) | Moving target detection method based on YOLOv3-Tiny | |
| CN112084897A (en) | Rapid traffic large-scene vehicle target detection method of GS-SSD | |
| CN116820131A (en) | Unmanned aerial vehicle tracking method based on target perception ViT | |
| CN110674845B (en) | Dish identification method combining multi-receptive-field attention and characteristic recalibration | |
| CN117593623A (en) | Lightweight vehicle detection method based on improved YOLOv8n model | |
| CN112132207A (en) | Target detection neural network construction method based on multi-branch feature mapping | |
| CN109815999B (en) | Simple, convenient and efficient automatic coal gangue identification method | |
| CN116580324A (en) | Yolov 5-based unmanned aerial vehicle ground target detection method | |
| CN113780462B (en) | Vehicle detection network establishment method based on unmanned aerial vehicle aerial image and application thereof | |
| CN115861595A (en) | Multi-scale domain self-adaptive heterogeneous image matching method based on deep learning | |
| Pillai et al. | Fine-Tuned EfficientNetB4 Transfer Learning Model for Weather Classification |
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 |