CN109583355A - A kind of device and method of stream of people's statistics based on boundary selection - Google Patents
A kind of device and method of stream of people's statistics based on boundary selection Download PDFInfo
- Publication number
- CN109583355A CN109583355A CN201811406445.4A CN201811406445A CN109583355A CN 109583355 A CN109583355 A CN 109583355A CN 201811406445 A CN201811406445 A CN 201811406445A CN 109583355 A CN109583355 A CN 109583355A
- Authority
- CN
- China
- Prior art keywords
- people
- yolo
- stream
- boundary
- neural 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
- G06V20/53—Recognition of crowd images, e.g. recognition of crowd congestion
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of device and method of stream of people's statistics based on boundary selection, belong to deep learning and field of image processing.The present invention is by improving YOLO neural network, the division of YOLO unit is increased to 9 × 9 from 7 × 7, the amount detection of each unit increases to 3, substitute the 16th in YOLO-PC neural network respectively with the Fire module in SqueezeNet again, 18th and the 24th 3 × 3 convolutional layer, and the convolution nuclear volume of the compression section in Fire module is reduced to 96 by 128, re -training network, new S-YOLO-PC neural network can be obtained, S-YOLO-PC neural network is recycled to carry out stream of people's statistics of boundary selection, utilize new neural network, so that it improves its accuracy in the case where model is substantially reduced, it can be used in stream of people's detection of a variety of occasions.
Description
Technical field
The present invention relates to a kind of device and method of stream of people's statistics based on boundary selection, belong at deep learning and image
Reason field.
Background technique
In video monitoring picture, pedestrian is very important detection target, pedestrian detection, crowd density estimation and the stream of people
The work such as statistics are the key components of wisdom security protection, intelligent building.How the background of changeable complexity is being supervised in video
Control the problem of differentiation detection pedestrian and other classification objects and effectively difference pedestrian and the proposition of its background in picture.This work
Often occur detection inaccuracy in specific scene, counts inaccuracy, detection and count results delay are longer, and depth model accounts for
It is larger with memory space, it is unfavorable for the problems such as transmitting and downloading.
Depth learning technology has started the tide of artificial intelligence in recent years, in image classification and object detection field, depth
Convolutional neural networks are even more to achieve a series of great breakthroughs, and accuracy rate has very big promotion.But robot controls, is automatic
This kind of relevant work of driving technology will not only pursue accuracy rate, it is often more important that speed, it is necessary to the system for relying on low delay.Closely
Several years R-CNN, Fast R-CNN and Faster R-CNN these types method represent the state-of-the-art level of target detection, and YOLO
(You Only Look Once) is a kind of method that wherein real-time performance is best.
YOLO is a kind of advanced real-time target detection algorithm, and accuracy rate is relatively high, but in actual application environment
Still many problems are encountered, such as occupy larger memory space, are unfavorable for being transmitted and being downloaded, detected in practical applications,
Serious to limit its practical application, the YOLOv2 that update after 2017 and YOLOv3 version later have used for reference Faster
The thought of R-CNN introduces the testing mechanism of anchor, cause new YOLO neural network be not available boundary selection method into
Row pedestrian counting.
It is, thus, sought for a kind of memory space is small and the method for the higher stream of people's statistics of real-time.
SqueezeNet is to be based on convolutional neural networks (Convolutional Neural using existing
Networks, CNN) model and the network structure of a kind of bench model compressed in a manner of damaging.It is instructed using a small amount of parameter
Practice network model, the compression of implementation model.It uses Fire Modle model structure, is divided into compression section and expansion, benefit
It is connected to be formed in a kind of Fire module with compression section with expansion and organizes Convolution Filter.Common SqueezeNet is opened
An independent convolutional layer (conv1), followed by 8 Fire modules are started from, are finally a final conversion layers (conv10).
Summary of the invention
To solve the above-mentioned problems, the present invention provides a kind of method of stream of people's statistics based on boundary selection, the present invention
It is improved on the basis of YOLOv1 neural network, so that its memory space is obviously reduced, in the feelings for not influencing its accuracy
Realize that the stream of people compared with high real-time counts under condition.
The technical solution of the present invention is as follows: a kind of method of stream of people's statistics based on boundary selection, the method includes following
Step:
Step 1: the height and angle for setting camera allow the picture of acquisition to cover the region of the stream of people to be determined, lead to
Cross camera acquisition stream of people's picture；
Step 2: the detection confidence level of S-YOLO-PC neural network is set by the equipment containing GPU；
Step 3:S-YOLO-PC neural network reads camera acquired image；
Step 4: the boundary of S-YOLO-PC neural network being set by the equipment containing GPU and detects pedestrian behavior；
Step 5:S-YOLO-PC neural network counts people information；
Step 6: the result information that the stream of people counts is output to by computer screen or camera by the equipment containing GPU in real time
Included screen；
The S-YOLO-PC neural network is improved YOLO neural network: the division of YOLO unit is increased from 7 × 7
To 9 × 9, the amount detection of each unit increases to 3, obtains YOLO-PC neural network, then with the Fire mould in SqueezeNet
Block substitutes the 16th, the 18th and the 24th 3 × 3 convolutional layer in YOLO-PC neural network respectively, and will be in Fire module
The convolution nuclear volume of compression section be reduced to 96 by 128, S-YOLO-PC neural network can be obtained in re -training network.
In one embodiment of the invention, detection confidence level described in step 2 is 0.2-0.4.
In one embodiment of the invention, the trained network only trains " people " this kind of targets.
In one embodiment of the invention, the boundary is that one or more nets are selected from 243 YOLO units
Lattice unit selects different boundaries according to the actual conditions of people as zone boundary, when people are from somewhere turning left,
The boundary of video left area is selected, the value on boundary is the numerical value in 81-89,108-116 or 135-143；When people from some
Place is turned right, and the value on boundary is the numerical value in 99-107,126-134 or 153-161；When people's straight trip, the value on boundary is
Numerical value in 90-98,117-125 or 144-152.
In one embodiment of the invention, the statistics people information specifically: stream of people's statistical value is S/n, wherein
The S indicates the number in the moment t detection block detected, and n is the pedestrian in the confidence level and selected borderline region of setting
The number being detected repeatedly.
In one embodiment of the invention, when detecting confidence level is 0.2, the number that pedestrian is detected repeatedly is
18, stream of people's statistical value is S/18；When detecting confidence level is 0.3, the number that pedestrian is detected repeatedly is 16, stream of people's statistical value
For S/16；When detecting confidence level is 0.4, the number that pedestrian is detected repeatedly is 13, and stream of people's statistical value is S/13.
In one embodiment of the invention, the signal output of the step 6 is directly to show real-time stream of people's quantity
On the camera of acquisition image or the equipment containing GPU, or real time information is saved and is constantly accumulated and is updated, then is passed through
Other interfaces are shown.
The present invention also provides it is a kind of based on boundary selection the stream of people statistics device,
Described device includes image collecting device, computing module and output module, wherein computing module includes calculating network
And hardware device, for acquiring data, calculating network is run on hardware device to be acquired described image acquisition module with reading
To image and pedestrian behavior is identified and is counted, later by stream of people's statistical information pass through hardware device or image collector
Set output；
Wherein, the hardware device is the equipment containing GPU, and the calculating network is S-YOLO-PC neural network, described
S-YOLO-PC neural network is improved YOLO neural network: the division of YOLO unit is increased to 9 × 9, Mei Gedan from 7 × 7
The amount detection of member increases to 3, obtains YOLO-PC neural network, then substituted respectively with the Fire module in SqueezeNet
The 16th, the 18th and the 24th 3 × 3 convolutional layer in YOLO-PC neural network, and by the compression section in Fire module
Convolution nuclear volume be reduced to 96 by 128, S-YOLO-PC neural network can be obtained in re -training network.
Advantageous effects of the invention:
(1) division of former YOLO unit is increased to 9 × 9 from 7 × 7 by S-YOLO-PC neural network of the invention, and will be every
The amount detection of a unit increases to 3, can significantly improve the mean accuracy of pedestrian detection, the more former YOLO nerve of mean accuracy
Network improves 14.5%；
(2) present invention is substituted by the Fire module being introduced into SqueezeNet specific 3 in former YOLO neural network
× 3 convolutional layers, and the convolution nuclear volume of the compression section in Fire module is reduced to 96 by 128, it can guarantee accuracy rate
In the case where, it will be apparent that the storage size for compressing neural network, the size than former YOLO neural network reduce 36.5%, significantly
Requirement of the network mechanism to GPU performance is reduced, so that it is with broader applicable scene.
(3) method that the present invention uses limited boundary, so that S-YOLO-PC can ignore that some meaningless interference, such as
People and the incoherent back side in billboard, therefore S-YOLO-PC can more accurately detect the flowing of pedestrian.
Detailed description of the invention
Fig. 1: the flow diagram of the method for stream of people's statistics based on boundary selection.
Fig. 2: experimental result picture when measurement left margin.
Specific embodiment
AP: i.e. Average Precision, mean accuracy, calculation formula are as follows:
Wherein, P (i) refer to given threshold value i when precision, Δ r (i) refers to the recall changing value between k and k-1.
The data set of training and test derives from PASCAL VOC (The pattern analysis, statistical
Modelling and computational learning Visual Object Classes Project), it is divided into VOC
2007 and VOC2012 two parts, this classification of the present invention training of human, using 2007 data set of VOC 5011 pictures and
11540 pictures of 2012 data set of VOC do training data, and training data totally 16551, test set is the test of VOC 2007
Collect, totally 4952 test pictures.
Embodiment 1
The technical solution of the present invention is as follows: a kind of method of stream of people's statistics based on boundary selection, the method includes following
Step:
Step 1: the height and angle for setting camera allow the picture of acquisition to cover the region of the stream of people to be determined, lead to
Cross camera acquisition stream of people's picture；
Step 2: the detection confidence level of S-YOLO-PC neural network is set by computer；
Step 3: the division of YOLO unit being increased to 9 × 9 from 7 × 7, the amount detection of each unit increases to 3, obtains
YOLO-PC neural network, then substituted respectively with the Fire module in SqueezeNet the 16th in YOLO-PC neural network,
18th and the 24th 3 × 3 convolutional layer, and the convolution nuclear volume of the compression section in Fire module is reduced to 96 by 128,
Trained network is re-started for 2007+2012 data set, only this kind of targets of training " people ", obtain S-YOLO-PC nerve net
Network reads camera acquired image；
Step 4: passing through the boundary that computer sets S-YOLO-PC neural network: when people are from somewhere turning left, selecting
The boundary of video left area, the value 113 on boundary；When people are from somewhere turning right, the value on boundary is 129；When people are straight
When row, the value on boundary is 121, and detects pedestrian behavior；
Step 5:S-YOLO-PC neural network counts people information, and when confidence level is 0.2, pedestrian is detected repeatedly
Number be 18, corresponding stream of people's statistical value is S/18, and wherein S value indicates the number of detection block detected in moment t；
Step 6: including current persons count, confidence value by the result information that the stream of people counts, output display in real time is on computers.
Same stream of people's detection and statistics are carried out using YOLO-PC and original YOLO neural network after the same method, and
Three kinds of different neural networks and the data for detecting and counting for the stream of people are compared.
(1) storage size of neural network
The memory of three kinds of neural network models is compared, the results are shown in Table 1, it is seen that S-YOLO-PC neural network
The more former YOLO of storage size reduce 36.5%, compared with YOLO-PC reduce 9.1%.
The storage size of the different neural network models of table 1
Fire module is replaced the convolutional layer in YOLO-PC by S-YOLO-PC neural network, is 512 for input channel number
Or when 1024, the results are shown in Table 2, and after replacement, obtained parameter greatly reduces compared with reel lamination, such as when input is logical
When road number is 512, the number of parameters compared to reel lamination reduces about 84.7%, and then by the convolution kernel number in Fire module
96 are reduced to by 128, parameter will continue to reduction 25%, and parameter further decreases, it is seen that using in Fire module replacement YOLO
Convolutional layer, corresponding parameter can greatly reduce.But parameter largely reduces and will lead to its accuracy rate and substantially reduce, therefore needs
It finds specific convolutional layer and carries out the replacement of Fire module, be possible to keep accuracy rate constant.
2 difference s of table1Different ciBring Parameters variation
Wherein s1For the convolution kernel number of Fire module, ciFor input channel number.
(2) mean accuracy AP
Stream of people's inspection is carried out after the same method for three kinds of different neural network YOLO, YOLO-PC and S-YOLO-PC
It surveying, mean accuracy is shown in Table 3, it is seen that the more former YOLO of the precision of S-YOLO-PC is significantly improved, and improves about 14.5%,
The precision of YOLO-PC and S-YOLO-PC is almost unchanged, it is seen that the present invention is by rolling up the 16th, 18,24 in YOLO-PC
Lamination carries out the replacement of Fire model, and adjusts its convolution kernel number and be reduced to 96 from 128, so that its model storage size obviously subtracts
Small (reducing about 9%), but can guarantee that precision is almost consistent simultaneously.
The mean accuracy of 3 distinct methods of table progress pedestrian detection
(3) the value of the confidence
It is provided with the experiment of three groups of 4 minutes real-time videos using different confidence levels (i.e. 0.2,0.3 and 0.4), is surveyed
Examination, the results are shown in Table 4, and S-YOLO-PC compares YOLO and detects to obtain more detection blocks, and average the value of the confidence be 50% with
On, hence it is evident that it is higher than original YOLO model, but relative to YOLO-PC, average the value of the confidence decreases, this is because Fire module
Caused by being compressed to data, but it is little to the influence of its accuracy rate on the whole.
4 distinct methods of table detection block quantity and average confidence comparison under different threshold values
Comparative example 1
With the Fire module in SqueezeNet substitute respectively the input in YOLO neural network be 1024 the 15th, 17,
24 layers, the storage size of neural network model becomes larger, and is 738.1M, and its accuracy rate becomes smaller, and is 71%.
The present invention replaces the 16th layer, the 18th layer with Fire module it can be seen from comparative example, and replacement is input convolution kernel
The layer that quantity is 512, compared to layer of whole inputs for 1024 is replaced with, model is smaller, and accuracy rate is higher, it is seen then that Fire mould
The replacement position of block must specifically can be only achieved effect of the invention.
Comparative example 2
Substitute the 16th, 18,19 layer in YOLO neural network respectively with the Fire module in SqueezeNet, the 19th layer same
Sample is to input the convolutional layer that convolution nuclear volume is 1024, but the storage size of the neural network model obtained after replacing is constant,
But its accuracy rate decreases.
Comparative example 3
When the convolution nuclear volume holding 128 of compression section in Fire module is constant, You Shangbiao 2 is it is found that work as Fire module
In the convolution nuclear volume of compression section when being 128, the number of parameter is obvious when compared with the convolution nuclear volume of compression section being 96
Increase, can significantly increase the storage size of neural network model, and its accuracy rate is without significantly improving.
Although the present invention has been described by way of example and in terms of the preferred embodiments, it is not intended to limit the invention, any to be familiar with this skill
The people of art can do various change and modification, therefore protection model of the invention without departing from the spirit and scope of the present invention
Enclosing subject to the definition of the claims.
Claims (10)
1. a kind of method of stream of people's statistics based on boundary selection, which is characterized in that the described method comprises the following steps:
Step 1: the height and angle for setting camera allow the picture of acquisition to cover the region of the stream of people to be determined, by taking the photograph
As head acquires stream of people's picture；
Step 2: the detection confidence level of S-YOLO-PC neural network is set by the equipment containing GPU；
Step 3:S-YOLO-PC neural network reads camera acquired image；
Step 4: the boundary of S-YOLO-PC neural network being set by the equipment containing GPU and detects pedestrian behavior；
Step 5:S-YOLO-PC neural network counts people information；
Step 6: the result information that the stream of people counts being output to by computer screen by the equipment containing GPU in real time or camera carries
Screen；
The S-YOLO-PC neural network is improved YOLO neural network: the division of YOLO unit is increased to 9 from 7 × 7 ×
9, the amount detection of each unit increases to 3, obtains YOLO-PC neural network, then divided with the Fire module in SqueezeNet
Not Ti Dai the 16th, the 18th and the 24th 3 × 3 convolutional layer in YOLO-PC neural network, and by the pressure in Fire module
The convolution nuclear volume of contracting part is reduced to 96 by 128, and S-YOLO-PC neural network can be obtained in re -training network.
2. a kind of method of stream of people's statistics based on boundary selection according to claim 1, which is characterized in that in step 2
The detection confidence level is 0.2-0.4.
3. a kind of method of stream of people's statistics based on boundary selection according to claim 1 or 2, which is characterized in that described
Training network only trains " people " this kind of targets.
4. a kind of method of stream of people's statistics based on boundary selection according to claim 1 to 3, which is characterized in that institute
Boundary is stated to select one or more grid cells as zone boundary from 243 YOLO units, and according to the reality of people
Situation selects different boundaries, when people are from somewhere turning left, selects the boundary of video left area, the value on boundary is
Numerical value in 81-89,108-116 or 135-143；When people are from somewhere turning right, the value on boundary is 99-107,126-
Numerical value in 134 or 153-161；When people's straight trip, the value on boundary is the numerical value in 90-98,117-125 or 144-152.
5. a kind of method of stream of people's statistics based on boundary selection according to claim 4, which is characterized in that the boundary
To select one or more grid cells as zone boundary from 243 units, and not according to the selection of the actual conditions of people
Same boundary, when people are from somewhere turning left, the boundary of selection video left area, the value on boundary is 113；When people exist
It somewhere turns right, the value on boundary is 129；When people directly walk somewhither, the value on boundary is 121.
6. the method for -5 any a kind of stream of people's statistics based on boundary selection according to claim 1, which is characterized in that institute
State statistics people information specifically: stream of people's statistical value is S/n, wherein the S indicates the number in the moment t detection block detected
Mesh, n are the number that pedestrian is detected repeatedly in the confidence level and selected borderline region of setting.
7. a kind of method of stream of people's statistics based on boundary selection according to claim 6, which is characterized in that when detection is set
When reliability is 0.2, the number that pedestrian is detected repeatedly is 18, and stream of people's statistical value is S/18；When detecting confidence level is 0.3,
The number that pedestrian is detected repeatedly is 16, and stream of people's statistical value is S/16；When detecting confidence level is 0.4, pedestrian is repeated inspection
The number measured is 13, and stream of people's statistical value is S/13.
8. a kind of device of stream of people's statistics based on boundary selection, which is characterized in that described device includes image collecting device, meter
Calculate module and output module, wherein computing module includes calculating network and hardware device, and described image acquisition module is for acquiring
Data calculate network and are run on hardware device to read the image collected and pedestrian behavior is identified and counted,
Stream of people's statistical information is exported by hardware device or image collecting device later；Wherein,
The hardware device is the equipment containing GPU, and the calculating network is S-YOLO-PC neural network, the S-YOLO-PC
Neural network is improved YOLO neural network: the division of YOLO unit being increased to 9 × 9 from 7 × 7, the detection of each unit
Quantity increases to 3, obtains YOLO-PC neural network, then substitute YOLO-PC nerve respectively with the Fire module in SqueezeNet
The 16th, the 18th and the 24th 3 × 3 convolutional layer in network, and by the convolution nucleus number of the compression section in Fire module
Amount is reduced to 96 by 128, and S-YOLO-PC neural network can be obtained in re -training network.
9. a kind of device of stream of people's statistics based on boundary selection according to claim 8 is detected or security protection, is built in the stream of people
Build the application in field.
10. a kind of method of -7 any stream of people's statistics based on boundary selection is in safety-security area according to claim 1
Using.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811406445.4A CN109583355B (en) | 2018-11-23 | 2018-11-23 | People flow counting device and method based on boundary selection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811406445.4A CN109583355B (en) | 2018-11-23 | 2018-11-23 | People flow counting device and method based on boundary selection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109583355A true CN109583355A (en) | 2019-04-05 |
| CN109583355B CN109583355B (en) | 2020-10-09 |
Family
ID=65924204
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811406445.4A Active CN109583355B (en) | 2018-11-23 | 2018-11-23 | People flow counting device and method based on boundary selection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109583355B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110110627A (en) * | 2019-04-24 | 2019-08-09 | 江南大学 | A method of the real-time target detection towards computing resource limitation Platform deployment |
| CN110222673A (en) * | 2019-06-21 | 2019-09-10 | 杭州宇泛智能科技有限公司 | A kind of passenger flow statistical method based on head detection |
| CN110974179A (en) * | 2019-12-20 | 2020-04-10 | 山东大学齐鲁医院 | Auxiliary diagnosis system for stomach precancer under electronic staining endoscope based on deep learning |
| WO2020215241A1 (en) * | 2019-04-24 | 2020-10-29 | 江南大学 | Real-time object detection method deployed in platform having limited computing resource |
| CN112068792A (en) * | 2020-09-09 | 2020-12-11 | 安徽宇烁光电科技有限公司 | External detection system of display screen |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8897492B2 (en) * | 2011-06-23 | 2014-11-25 | UDP Technology Ltd. | People counter including setting interface and method for setting the same |
| CN106886795A (en) * | 2017-02-17 | 2017-06-23 | 北京维弦科技有限责任公司 | Object identification method based on the obvious object in image |
| CN108805070A (en) * | 2018-06-05 | 2018-11-13 | 合肥湛达智能科技有限公司 | A kind of deep learning pedestrian detection method based on built-in terminal |
-
2018
- 2018-11-23 CN CN201811406445.4A patent/CN109583355B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8897492B2 (en) * | 2011-06-23 | 2014-11-25 | UDP Technology Ltd. | People counter including setting interface and method for setting the same |
| CN106886795A (en) * | 2017-02-17 | 2017-06-23 | 北京维弦科技有限责任公司 | Object identification method based on the obvious object in image |
| CN108805070A (en) * | 2018-06-05 | 2018-11-13 | 合肥湛达智能科技有限公司 | A kind of deep learning pedestrian detection method based on built-in terminal |
Non-Patent Citations (4)
| Title |
|---|
| PEIMING REN 等: "A Novel Squeeze YOLO-based Real-time People Counting Approach", 《HTTPS://WWW.INDERSCIENCE.COM/INFO/INGENERAL/FORTHCOMING.PHP?JCODE=IJBIC》 * |
| PEIMING REN 等: "A Novel YOLO-based Real-time People Counting Approach", 《2017 INTERNATIONAL SMART CITIES CONFERENCE (ISC2)》 * |
| WONG, A. 等: "Tiny SSD: a tiny single-shot detection deep convolutional neural network for real-time embedded object detection", 《ARXIV》 * |
| 廖辉: "基于轻量级卷积神经网络的人脸检测算法", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110110627A (en) * | 2019-04-24 | 2019-08-09 | 江南大学 | A method of the real-time target detection towards computing resource limitation Platform deployment |
| WO2020215241A1 (en) * | 2019-04-24 | 2020-10-29 | 江南大学 | Real-time object detection method deployed in platform having limited computing resource |
| US11410035B2 (en) | 2019-04-24 | 2022-08-09 | Jiangnan University | Real-time target detection method deployed on platform with limited computing resources |
| CN110222673A (en) * | 2019-06-21 | 2019-09-10 | 杭州宇泛智能科技有限公司 | A kind of passenger flow statistical method based on head detection |
| CN110222673B (en) * | 2019-06-21 | 2021-04-06 | 杭州宇泛智能科技有限公司 | Passenger flow statistical method based on head detection |
| CN110974179A (en) * | 2019-12-20 | 2020-04-10 | 山东大学齐鲁医院 | Auxiliary diagnosis system for stomach precancer under electronic staining endoscope based on deep learning |
| CN112068792A (en) * | 2020-09-09 | 2020-12-11 | 安徽宇烁光电科技有限公司 | External detection system of display screen |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109583355B (en) | 2020-10-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109583355A (en) | A kind of device and method of stream of people's statistics based on boundary selection | |
| CN106096577B (en) | A kind of target tracking method in camera distribution map | |
| CN109697726B (en) | Event camera-based end-to-end target motion estimation method | |
| CN104106260B (en) | Control based on geographical map | |
| CN106022244B (en) | Unsupervised crowd's exception monitoring and localization method based on recurrent neural network modeling | |
| CN111626128A (en) | Improved YOLOv 3-based pedestrian detection method in orchard environment | |
| CN105678811A (en) | Motion-detection-based human body abnormal behavior detection method | |
| CN106204638A (en) | A kind of based on dimension self-adaption with the method for tracking target of taking photo by plane blocking process | |
| CN107408303A (en) | System and method for Object tracking | |
| CN108133172A (en) | Method, the analysis method of vehicle flowrate and the device that Moving Objects are classified in video | |
| CN111260738A (en) | Multi-scale target tracking method based on relevant filtering and self-adaptive feature fusion | |
| CN114332749B (en) | Sand production monitoring method of sand production ship based on deep learning | |
| CN112836640A (en) | Single-camera multi-target pedestrian tracking method | |
| CN112036381B (en) | Visual tracking method, video monitoring method and terminal equipment | |
| CN110114801B (en) | Image foreground detection device and method and electronic equipment | |
| CN114639075A (en) | Method and system for identifying falling object of high altitude parabola and computer readable medium | |
| CN113763424A (en) | Real-time intelligent target detection method and system based on embedded platform | |
| CN106384359A (en) | Moving target tracking method and television set | |
| CN102314591A (en) | Method and equipment for detecting static foreground object | |
| CN111260687A (en) | Aerial video target tracking method based on semantic perception network and related filtering | |
| CN111027440A (en) | Crowd abnormal behavior detection device and method based on neural network | |
| CN101877135B (en) | Moving target detecting method based on background reconstruction | |
| CN111339950B (en) | Remote sensing image target detection method | |
| CN113076808A (en) | Method for accurately acquiring bidirectional pedestrian flow through image algorithm | |
| CN102054278B (en) | Object tracking method based on grid contraction |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |