CN112163499A - Small target pedestrian detection method based on fusion features - Google Patents

Small target pedestrian detection method based on fusion features Download PDF

Info

Publication number
CN112163499A
CN112163499A CN202011007245.9A CN202011007245A CN112163499A CN 112163499 A CN112163499 A CN 112163499A CN 202011007245 A CN202011007245 A CN 202011007245A CN 112163499 A CN112163499 A CN 112163499A
Authority
CN
China
Prior art keywords
pedestrian detection
small
target
pedestrian
detection frame
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
CN202011007245.9A
Other languages
Chinese (zh)
Inventor
邹腾涛
杨尚明
朱俊林
邓翔文
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.)
University of Electronic Science and Technology of China
Original Assignee
University of Electronic Science and Technology of China
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 University of Electronic Science and Technology of China filed Critical University of Electronic Science and Technology of China
Priority to CN202011007245.9A priority Critical patent/CN112163499A/en
Publication of CN112163499A publication Critical patent/CN112163499A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

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)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Human Computer Interaction (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)
  • Traffic Control Systems (AREA)

Abstract

The invention discloses a small-target pedestrian detection method based on fusion characteristics, which is characterized in that fusion multi-scale convolution characteristics are adopted to detect that pedestrians share characteristic information among convolution layers when analyzing that an SSD algorithm is applied to a place with deficiency in pedestrian detection, and the input of each layer comprises the output of the previous layer and partial convolution information of the previous convolution layer, so that the central point position of the pedestrian is better regressed; the defect that a small-resolution feature map is not easy to detect a small-target pedestrian is considered, the size of the finally generated mask map is increased, and the problem that the resolution of the small-area feature map is insufficient is solved. The method expands the size of the final feature map by fusing the convolution features of each stage, enriches the feature information of the small target, well realizes the identification of the small target, improves the identification accuracy of pedestrian detection of the small target, and has higher recall rate and identification rate compared with the SDD party.

Description

Small target pedestrian detection method based on fusion features
Technical Field
The invention belongs to the technical field of target detection, and particularly relates to a small target pedestrian detection method based on fusion characteristics.
Background
Pedestrian detection has wide application in the fields of intelligent monitoring, image retrieval, environmental perception and the like, and thus becomes an important research topic in the field of computer vision. Accurate pedestrian detection provides better preconditions for subsequent analysis, such as pedestrian tracking, people counting, re-recognition of people, attitude estimation, face recognition and the like.
Pedestrian detection has been extensively studied for many years. Most of the traditional detection methods achieve the aim by manually designing a characteristic and constructing a detection model of the load. The research of this type of target detection method focuses on the skillful design of suitable features and powerful classifiers, such as HoG + SVM, HoG + DPM, DOT + RF, etc. With the rapid development of deep learning, many studies combine feature extraction and feature classification using deep neural networks, resulting in a stage of target detection algorithms. The algorithm directly inputs the identification picture into a neural network, and performs target identification by predicting the position coordinates of the target. Representative algorithms for a stage are YOLO, YOLO v2, SSD, etc. The SSD, which is the most commonly used one-stage detection algorithm in the field of object detection, has a high detection rate and detection speed, and is also widely used in the field of object detection in the industry.
In a target detection method called SSD, a Single Shot MultiBox Detector, the method consists of two parts in total, a feature extraction network and an identification network. The feature extraction network is composed of a full convolution network, two common structures are respectively designed based on VGG and VGG improvements, a feature map of extracted features is input into a recognition network to predict the specific position and the category of a target, the recognition network is also composed of the full convolution network and is not used for YOLO to predict the target on the last feature map, the SSD respectively extracts feature maps of Conv Conv4_3, Conv7, Conv8_2, Conv9_2, Conv10_2 and Conv11_2, and then 6 default boxes with different dimensions are respectively constructed at each point on the feature maps. And then, respectively carrying out fine adjustment and prediction on the detection frames generated on the different feature maps, and deleting the detection frame with the identification result as the background so as to generate a plurality of detection frames which preliminarily accord with the conditions. Finally, the detection frames generated on different feature maps are combined, a part of overlapped or incorrect detection frames are suppressed by adopting an NMS (non-maximum suppression algorithm), and a final detection frame set (namely a detection result) is generated, although the SSD is a general target detection algorithm, in the industry, the SSD is often used in the field of pedestrian detection and monitoring safety, and a good detection effect can also be obtained. SSD-based pedestrian detection algorithms do not share features between different scale feature maps, although the detection rate for small target pedestrians is improved by predicting pedestrians on different feature maps. The convolution characteristics of the small target pedestrian are rich in the shallow characteristic diagram, but redundant information is more and difficult to identify; fewer deep profiles exist and are difficult to identify. Target frames with multiple sizes are predicted on multiple convolution feature maps, position fine adjustment and category prediction need to be carried out on each target frame, the calculated amount is large, and the number of detection frames meeting requirements is large, so that the detection time of an NMS algorithm is increased, and the step of obtaining the final recognition result is complicated.
Disclosure of Invention
Aiming at the defects in the prior art, the method for detecting the small target pedestrian based on the fusion features solves the problems that the small target pedestrian cannot be identified or the identification rate is low in the background technology.
In order to achieve the purpose of the invention, the invention adopts the technical scheme that: a small target pedestrian detection method based on fusion features comprises the following steps:
s1, extracting the features of the image to be detected through a Unet full convolution neural network to generate a corresponding feature map;
s2, inputting the feature maps into a center prediction network and a position prediction network respectively to obtain corresponding prediction results, and constructing a pedestrian detection frame set according to the prediction results;
and S3, deleting the detection frames with the overlapping rate larger than the set value in the pedestrian detection frame set by adopting a non-maximum inhibition algorithm to obtain a final pedestrian detection result, and realizing small-target pedestrian detection.
Further, the sizes of the images to be detected input into the Unet full convolution neural network in the step S1 are the same;
the feature map generated by the Unet full convolution neural network in the step S1
Figure BDA0002696353490000031
Comprises the following steps:
Figure BDA0002696353490000032
where f (-) is the RELU activation function;
Figure BDA0002696353490000033
is the ith characteristic diagram of the l-1 layer;
Figure BDA0002696353490000034
a convolution kernel of the first convolution layer in the Unet full convolution neural network;
Figure BDA0002696353490000035
the translation parameter of the first layer convolution layer in the Unet full convolution neural network.
Further, in step S2, the central prediction network is an independent convolution layer with a convolution kernel of 1 × 1 and a computation function of Sigmod function, and is used to fuse the extracted feature maps and generate a pedestrian position mask map M.
Further, in step S2, the position prediction network is an independent convolution layer with a convolution kernel of 1 × 1, and is configured to reconstruct the extracted feature map and obtain a four-channel position feature map N.
Further, the position mask map M is:
outcen＝f(o)
in the formula, outcenThe expression of the position mask map M, f (o) is the function value of 0 in the Sigmod function, and the expression of the Sigmod function is
Figure BDA0002696353490000036
Further, a plurality of pedestrian detection boxes are included in the position feature map N, and each pedestrian detection box is marked with an upper left corner coordinate (x1, y1) and a lower right corner coordinate (x2, y 2).
Further, the step S2 is specifically:
s21, when a position with a pixel value larger than 0.5 exists in the arbitrary position mask image M, determining that the pedestrian detection frame where the position is located in the position feature image N corresponding to the position mask image M is valid;
and S22, constructing a pedestrian detection frame set by taking all the confirmed effective pedestrian detection frames as the constituent elements in the pedestrian detection frame set.
Further, the step S3 is specifically:
s31, taking the pixel value of each pedestrian detection frame in the pedestrian detection frame set B in the corresponding position mask image M as the confidence score of the pedestrian detection frame, thereby constructing a score set S of all pedestrian detection frames;
s32, deleting the pedestrian detection frame corresponding to the highest confidence score S in the score set S from the pedestrian detection frame set B;
s33, deleting other pedestrian detection frames with the pedestrian detection frame overlapping area larger than 0.5 corresponding to the highest confidence score S from the pedestrian detection frame set B;
and S34, repeating the steps S31-S33 until the pedestrian detection frame set B is empty, and finally taking the current pedestrian detection frame set as a pedestrian detection result to realize small-target pedestrian detection.
The invention has the beneficial effects that:
the invention provides a small-target pedestrian detection method based on fusion features, wherein fusion multi-scale convolution features are adopted to detect that pedestrians share feature information among convolution layers when analyzing that an SSD algorithm is applied to a place with defects in pedestrian detection, and the input of each layer comprises the output of the previous layer and partial convolution information of the previous convolution layer, so that the central point position of the pedestrian is better regressed; the defect that a small-resolution feature map is not easy to detect a small-target pedestrian is considered, the size of the finally generated mask map is increased, and the problem that the resolution of the small-area feature map is insufficient is solved. The method expands the size of the final feature map by fusing the convolution features of each stage, enriches the feature information of the small target, well realizes the identification of the small target, improves the identification accuracy of pedestrian detection of the small target, and has higher recall rate and identification rate compared with the SDD party.
Drawings
Fig. 1 is a flow chart of a small-target pedestrian detection method based on fusion features provided by the invention.
Fig. 2 is a schematic diagram of the calculation composition of the Unet full convolution neural network provided by the present invention.
Fig. 3 is a structural diagram of a pnet full convolution neural network provided by the present invention.
Fig. 4 is a schematic diagram of the time convolution operation in the Unet full convolution neural network provided by the present invention.
Fig. 5 is a schematic diagram of Max pool operation process in the Unet full convolution neural network provided by the present invention.
Detailed Description
The following description of the embodiments of the present invention is provided to facilitate the understanding of the present invention by those skilled in the art, but it should be understood that the present invention is not limited to the scope of the embodiments, and it will be apparent to those skilled in the art that various changes may be made without departing from the spirit and scope of the invention as defined and defined in the appended claims, and all matters produced by the invention using the inventive concept are protected.
As shown in fig. 1, a method for detecting a small target pedestrian based on fusion features includes the following steps:
s1, extracting the features of the image to be detected through a Unet full convolution neural network to generate a corresponding feature map;
s2, inputting the feature maps into a center prediction network and a position prediction network respectively to obtain corresponding prediction results, and constructing a pedestrian detection frame set according to the prediction results;
and S3, deleting the detection frames with the overlapping rate larger than the set value in the pedestrian detection frame set by adopting a non-maximum inhibition algorithm to obtain a final pedestrian detection result, and realizing small-target pedestrian detection.
The size of the images to be detected input into the Unet full convolution neural network in the step S1 is the same;
when the feature extraction is performed in the Unet full convolution neural network in step S1, the feature map is generated by repeating the convolution calculation, the down-sampling calculation and the up-sampling calculation, and the specific composition is as shown in fig. 2
Figure BDA0002696353490000051
Comprises the following steps:
Figure BDA0002696353490000061
where f (-) is the RELU activation function;
Figure BDA0002696353490000062
is the ith characteristic diagram of the l-1 layer;
Figure BDA0002696353490000063
a convolution kernel of the first convolution layer in the Unet full convolution neural network;
Figure BDA0002696353490000064
the translation parameter of the first layer convolution layer in the Unet full convolution neural network is obtained;
in the above process, the expression of the downsampling calculation is:
Figure BDA0002696353490000065
the expression for the upsampling calculation is:
Figure BDA0002696353490000066
fig. 3 is a structural diagram of a net full convolution neural network according to the present invention, conv represents a convolution operation, and copy and crop represents a convolution operation in which a convolution feature obtained earlier and a convolution feature obtained currently are connected and output to a next neural layer. The activation function expressed by ReLU is expressed as:
Figure BDA0002696353490000067
in fig. 3, Up indicates the feature size expansion, a linear interpolation algorithm is used to expand the feature size, Max pool indicates the down-sampling operation, fig. 4 is a schematic diagram of the time convolution operation in the above process, and fig. 5 is a schematic diagram of the Max pool operation process in the above process.
In step S2, the central prediction network is an independent convolution layer with a convolution kernel of 1 × 1 and a computation function of Sigmod function, and is used to fuse the extracted feature maps and generate a pedestrian position mask map M:
outcen＝f(o)
in the formula, outcenThe expression of the position mask map M, f (o) is the function value of 0 in the Sigmod function, and the expression of the Sigmod function is
Figure BDA0002696353490000068
In the step S2, the position prediction network is an independent convolution layer with a convolution kernel of 1 × 1, and is configured to reconstruct the extracted feature map and obtain a four-channel position feature map N, where the position feature map N includes a plurality of pedestrian detection frames, and each pedestrian detection frame is labeled with an upper-left coordinate (x1, y1) and a lower-right coordinate (x2, y 2).
The step S2 is specifically:
s21, when a position with a pixel value larger than 0.5 exists in the arbitrary position mask image M, determining that the pedestrian detection frame where the position is located in the position feature image N corresponding to the position mask image M is valid;
and S22, constructing a pedestrian detection frame set by taking all the confirmed effective pedestrian detection frames as the constituent elements in the pedestrian detection frame set.
The step S3 is specifically:
s31, taking the pixel value of each pedestrian detection frame in the pedestrian detection frame set B in the corresponding position mask image M as the confidence score of the pedestrian detection frame, thereby constructing a score set S of all pedestrian detection frames;
s32, deleting the pedestrian detection frame corresponding to the highest confidence score S in the score set S from the pedestrian detection frame set B;
s33, deleting other pedestrian detection frames with the pedestrian detection frame overlapping area larger than 0.5 corresponding to the highest confidence score S from the pedestrian detection frame set B;
and S34, repeating the steps S31-S33 until the pedestrian detection frame set B is empty, and finally taking the current pedestrian detection frame set as a pedestrian detection result to realize small-target pedestrian detection.
In one embodiment of the present invention, the experimental comparison between the method of the present invention and the SDD method in the background art for small target pedestrian identification is provided:
compared with the background technology, the pedestrian detection method has the advantages that the multi-scale convolution characteristics are fused to identify the pedestrian, the pedestrian is detected by generating the mask image with the same size as the detection image, and the detection rate of the small target pedestrian is improved. Experiments show that compared with the original SDD, the method provided by the invention has higher recall rate and recognition rate. We chose two data sets, INRIA and Caltech, to implement the improved scheme and compare experiments with the original SSD. As can be seen from tables 1 and 2, the method has better embodiment in recall rate and accuracy rate. This also demonstrates the effectiveness of our approach.
Table 1 comparative experimental results of SSD algorithm and our proposed algorithm on INRIA row human data set
Method of producing a composite material Recall rate Rate of accuracy
SSD 13.3％ 88.6％
Ours 9.8％ 92.5％
Table 2 SSD algorithm and comparative experimental results of our proposed algorithm on Caltech pedestrian dataset
Method of producing a composite material Recall rate Rate of accuracy
SSD 21.1％ 85.3％
Ours 13.7％ 90.2％
In addition, in order to verify that the invention can effectively improve the pedestrian detection accuracy of small targets in the data set, we performed improvement scheme experiments on the small target data set of Caltech, which divides the pedestrian mark frames into 3 different subsets according to their pixel heights, wherein the pedestrian pixels smaller than 30 pixels are regarded as small targets. By adopting the scheme and the SSD algorithm, comparison is carried out, the omission factor is used as a judgment index, and table 3 shows that the omission factor of the Caltech small target data set is obviously improved by the method.
Table 3 SSD algorithm and comparative experimental results of our proposed algorithm on Caltech pedestrian dataset
Method of producing a composite material Rate of missed examination
SSD 48.5％
Ours 35.2％
The invention has the beneficial effects that:
the invention provides a small-target pedestrian detection method based on fusion features, wherein fusion multi-scale convolution features are adopted to detect that pedestrians share feature information among convolution layers when analyzing that an SSD algorithm is applied to a place with defects in pedestrian detection, and the input of each layer comprises the output of the previous layer and partial convolution information of the previous convolution layer, so that the central point position of the pedestrian is better regressed; the defect that a small-resolution feature map is not easy to detect a small-target pedestrian is considered, the size of the finally generated mask map is increased, and the problem that the resolution of the small-area feature map is insufficient is solved. The method expands the size of the final feature map by fusing the convolution features of each stage, enriches the feature information of the small target, well realizes the identification of the small target, improves the identification accuracy of pedestrian detection of the small target, and has higher recall rate and identification rate compared with the SDD party.

Claims (8)

1. A small target pedestrian detection method based on fusion features is characterized by comprising the following steps:
s1, extracting the features of the image to be detected through a Unet full convolution neural network to generate a corresponding feature map;
s2, inputting the feature maps into a center prediction network and a position prediction network respectively to obtain corresponding prediction results, and constructing a pedestrian detection frame set according to the prediction results;
and S3, deleting the detection frames with the overlapping rate larger than the set value in the pedestrian detection frame set by adopting a non-maximum inhibition algorithm to obtain a final pedestrian detection result, and realizing small-target pedestrian detection.
2. The fusion feature-based small-target pedestrian detection method according to claim 1, wherein the images to be detected input into the Unet full convolution neural network in the step S1 are the same in size;
the feature map generated by the Unet full convolution neural network in the step S1
Figure FDA0002696353480000011
Comprises the following steps:
Figure FDA0002696353480000012
where f (-) is the RELU activation function;
Figure FDA0002696353480000013
is the ith characteristic diagram of the l-1 layer;
Figure FDA0002696353480000014
a convolution kernel of the first convolution layer in the Unet full convolution neural network;
Figure FDA0002696353480000015
for layer I volume in Unet full convolution neural networkTranslation parameters of the stack.
3. The method for detecting small-target pedestrians according to the fused feature of claim 1, wherein in the step S2, the central prediction network is an independent convolution layer with convolution kernel of 1 × 1 and calculation function of Sigmod function, and is used for fusing the extracted feature maps and generating the position mask map M of the pedestrian.
4. The method according to claim 3, wherein in step S2, the position prediction network is an independent convolution layer with convolution kernel of 1 × 1, and is used to reconstruct the extracted feature map and obtain a four-channel position feature map N.
5. The fusion-feature-based small-target pedestrian detection method according to claim 4, wherein the position mask map M is:
outcen＝f(o)
in the formula, outcenThe expression of the position mask map M, f (o) is the function value of 0 in the Sigmod function, and the expression of the Sigmod function is
Figure FDA0002696353480000021
6. The fusion-feature-based small-target pedestrian detection method according to claim 5, wherein a plurality of pedestrian detection boxes are included in the position feature map N, and each pedestrian detection box is labeled with an upper-left corner coordinate (x1, y1) and a lower-right corner coordinate (x2, y 2).
7. The method for detecting small-target pedestrians based on fusion characteristics according to claim 6, wherein the step S2 is specifically that:
s21, when a position with a pixel value larger than 0.5 exists in the arbitrary position mask image M, determining that the pedestrian detection frame where the position is located in the position feature image N corresponding to the position mask image M is valid;
and S22, constructing a pedestrian detection frame set by taking all the confirmed effective pedestrian detection frames as the constituent elements in the pedestrian detection frame set.
8. The method for detecting small-target pedestrians based on fusion characteristics according to claim 7, wherein the step S3 is specifically that:
s31, taking the pixel value of each pedestrian detection frame in the pedestrian detection frame set B in the corresponding position mask image M as the confidence score of the pedestrian detection frame, thereby constructing a score set S of all pedestrian detection frames;
s32, deleting the pedestrian detection frame corresponding to the highest confidence score S in the score set S from the pedestrian detection frame set B;
s33, deleting other pedestrian detection frames with the pedestrian detection frame overlapping area larger than 0.5 corresponding to the highest confidence score S from the pedestrian detection frame set B;
and S34, repeating the steps S31-S33 until the pedestrian detection frame set B is empty, and finally taking the current pedestrian detection frame set as a pedestrian detection result to realize small-target pedestrian detection.
CN202011007245.9A 2020-09-23 2020-09-23 Small target pedestrian detection method based on fusion features Pending CN112163499A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011007245.9A CN112163499A (en) 2020-09-23 2020-09-23 Small target pedestrian detection method based on fusion features

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011007245.9A CN112163499A (en) 2020-09-23 2020-09-23 Small target pedestrian detection method based on fusion features

Publications (1)

Publication Number Publication Date
CN112163499A true CN112163499A (en) 2021-01-01

Family

ID=73863354

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011007245.9A Pending CN112163499A (en) 2020-09-23 2020-09-23 Small target pedestrian detection method based on fusion features

Country Status (1)

Country Link
CN (1) CN112163499A (en)

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107273870A (en) * 2017-07-07 2017-10-20 郑州航空工业管理学院 The pedestrian position detection method of integrating context information under a kind of monitoring scene
CN107563299A (en) * 2017-08-07 2018-01-09 盐城禅图智能科技有限公司 A kind of pedestrian detection method using ReCNN integrating context informations
CN108830205A (en) * 2018-06-04 2018-11-16 江南大学 Based on the multiple dimensioned perception pedestrian detection method for improving full convolutional network
CN109214241A (en) * 2017-07-03 2019-01-15 中国科学院文献情报中心 Pedestrian detection method based on deep learning algorithm
CN109271888A (en) * 2018-08-29 2019-01-25 汉王科技股份有限公司 Personal identification method, device, electronic equipment based on gait
US20190057507A1 (en) * 2017-08-18 2019-02-21 Samsung Electronics Co., Ltd. System and method for semantic segmentation of images
CN109784291A (en) * 2019-01-23 2019-05-21 电子科技大学 Pedestrian detection method based on multiple dimensioned convolution feature
CN109886100A (en) * 2019-01-14 2019-06-14 苏州工业园区职业技术学院 A kind of pedestrian detecting system based on Area generation network
CN110008853A (en) * 2019-03-15 2019-07-12 华南理工大学 Pedestrian detection network and model training method, detection method, medium, equipment
CN111539402A (en) * 2020-07-13 2020-08-14 平安国际智慧城市科技股份有限公司 Deep learning-based lane line detection method, device, terminal and storage medium

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109214241A (en) * 2017-07-03 2019-01-15 中国科学院文献情报中心 Pedestrian detection method based on deep learning algorithm
CN107273870A (en) * 2017-07-07 2017-10-20 郑州航空工业管理学院 The pedestrian position detection method of integrating context information under a kind of monitoring scene
CN107563299A (en) * 2017-08-07 2018-01-09 盐城禅图智能科技有限公司 A kind of pedestrian detection method using ReCNN integrating context informations
US20190057507A1 (en) * 2017-08-18 2019-02-21 Samsung Electronics Co., Ltd. System and method for semantic segmentation of images
CN108830205A (en) * 2018-06-04 2018-11-16 江南大学 Based on the multiple dimensioned perception pedestrian detection method for improving full convolutional network
CN109271888A (en) * 2018-08-29 2019-01-25 汉王科技股份有限公司 Personal identification method, device, electronic equipment based on gait
CN109886100A (en) * 2019-01-14 2019-06-14 苏州工业园区职业技术学院 A kind of pedestrian detecting system based on Area generation network
CN109784291A (en) * 2019-01-23 2019-05-21 电子科技大学 Pedestrian detection method based on multiple dimensioned convolution feature
CN110008853A (en) * 2019-03-15 2019-07-12 华南理工大学 Pedestrian detection network and model training method, detection method, medium, equipment
CN111539402A (en) * 2020-07-13 2020-08-14 平安国际智慧城市科技股份有限公司 Deep learning-based lane line detection method, device, terminal and storage medium

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
A. NEUBECK; L. VAN GOOL: "Efficient non-maximum suppression", 《18TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION (ICPR"06)》 *
SONGTAO LIU, DI HUANG, YUNHONG WANG: "Adaptive NMS: Refining Pedestrian Detection in a Crowd", 《ARXIV.ORG 》 *
付新川: "图像中的行人检测关键技术研究", 《中国优秀博硕士学位论文全文数据库(博士)信息科技辑》 *
邹腾涛: "道路交通的行人检测若干问题研究与实现", 《中国优秀博硕士学位论文全文数据库(硕士)工程科技Ⅱ辑》 *

Similar Documents

Publication Publication Date Title
CN109840556B (en) Image classification and identification method based on twin network
CN109583340B (en) Video target detection method based on deep learning
CN110334762B (en) Feature matching method based on quad tree combined with ORB and SIFT
CN112288008B (en) Mosaic multispectral image disguised target detection method based on deep learning
CN110782420A (en) Small target feature representation enhancement method based on deep learning
CN105809651B (en) Image significance detection method based on the comparison of edge non-similarity
CN110490913B (en) Image matching method based on feature description operator of corner and single line segment grouping
CN111898432B (en) Pedestrian detection system and method based on improved YOLOv3 algorithm
WO2023082784A1 (en) Person re-identification method and apparatus based on local feature attention
CN110517270B (en) Indoor scene semantic segmentation method based on super-pixel depth network
CN109886159B (en) Face detection method under non-limited condition
CN109299303B (en) Hand-drawn sketch retrieval method based on deformable convolution and depth network
CN111768415A (en) Image instance segmentation method without quantization pooling
CN105069447A (en) Facial expression identification method
CN111353544A (en) Improved Mixed Pooling-Yolov 3-based target detection method
CN110263731B (en) Single step human face detection system
Li et al. A complex junction recognition method based on GoogLeNet model
Zang et al. Traffic lane detection using fully convolutional neural network
CN109325407B (en) Optical remote sensing video target detection method based on F-SSD network filtering
CN111126275A (en) Pedestrian re-identification method and device based on multi-granularity feature fusion
CN110807463B (en) Image segmentation method and device, computer equipment and storage medium
Mo et al. PVDet: Towards pedestrian and vehicle detection on gigapixel-level images
CN111881803B (en) Face recognition method based on improved YOLOv3
CN113627481A (en) Multi-model combined unmanned aerial vehicle garbage classification method for smart gardens
CN111797704B (en) Action recognition method based on related object perception

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: 20210101