WO2020108075A1 - 结合人脸与外观的两阶段行人搜索方法 - Google Patents

结合人脸与外观的两阶段行人搜索方法 Download PDF

Info

Publication number
WO2020108075A1
WO2020108075A1 PCT/CN2019/108502 CN2019108502W WO2020108075A1 WO 2020108075 A1 WO2020108075 A1 WO 2020108075A1 CN 2019108502 W CN2019108502 W CN 2019108502W WO 2020108075 A1 WO2020108075 A1 WO 2020108075A1
Authority
WO
WIPO (PCT)
Prior art keywords
pedestrian
image
target
face
candidate
Prior art date
Application number
PCT/CN2019/108502
Other languages
English (en)
French (fr)
Inventor
杨华
李亮奇
Original Assignee
上海交通大学
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 上海交通大学 filed Critical 上海交通大学
Publication of WO2020108075A1 publication Critical patent/WO2020108075A1/zh
Priority to US17/030,358 priority Critical patent/US11017215B2/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/51—Indexing; Data structures therefor; Storage structures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/55—Clustering; Classification
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • 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
    • 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
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/161—Detection; Localisation; Normalisation
    • 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
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation
    • 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
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/172—Classification, e.g. identification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization

Definitions

  • the present invention relates to the field of computer vision technology, and in particular, to a two-stage pedestrian search method combining face and appearance.
  • pedestrian re-identification Patent Re-identification
  • security monitoring crowd flow monitoring
  • pedestrian behavior analysis etc.
  • computer vision researchers focus on pedestrian re-recognition research on pedestrian matching between multiple cameras in similar scenes.
  • the appearance of the target pedestrian image (such as a portrait photo) and the pedestrian image captured by the camera to be matched are often quite different, and the traditional pedestrian re-identification method cannot be effective Match the target pedestrian.
  • the reliable pedestrian feature should be the face feature, but if only the face is used for matching, the target pedestrian sample without face cannot be matched, and the tracking of its trajectory may be lost.
  • the purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art, and proposes a two-stage pedestrian search method combining face and appearance based on deep learning, which combines the problems of face recognition and pedestrian re-recognition.
  • the present invention extracts more discriminative features based on the convolutional neural network, and uses the DenseNet network structure (see: Gao Huang, Zhuang Liu, Laurens Van der Maaten, and Kilian Q Weinberger, "Densely connected" convolutional networks, "in Proceedings of the IEEE Conference Computer Vision and Pattern Recognition, 2017.);
  • the present invention proposes a two-stage pedestrian search method that combines face and appearance, using face detection and recognition to proceed from the target pedestrian image and from the candidate pedestrian Multi-query is generated in the picture; finally, the present invention designs a compression method for pedestrian re-identification matching sample space to improve the accuracy of pedestrian re-identification.
  • the present invention is achieved by the following technical solutions.
  • a two-stage pedestrian search method combining face and appearance including:
  • Sort according to the similarity score of the face if the similarity score is greater than or equal to the preset threshold, then the corresponding image containing the candidate pedestrian is used as the target image for pedestrian re-recognition IQ (x, y) (Multi-query) ; If the similarity score is less than the preset threshold, the corresponding image containing the candidate pedestrian is used as a panoramic image for pedestrian re-recognition (Gallery);
  • the panoramic image re-recognized by the pedestrian corresponding to each target pedestrian is filtered to obtain the processed candidate pedestrian image I G '(x ,y);
  • the average Euclidean distance between the candidate pedestrian in the processed candidate pedestrian image I G '(x,y) and the candidate pedestrian in the target image IQ (x,y) is taken as the final European distance of the candidate pedestrian, and The final Euclidean distance sorts the candidate pedestrians in the processed candidate pedestrian image I G '(x,y) to obtain the corresponding sorting result.
  • calculating the face similarity scores of the target pedestrian and the candidate pedestrian including:
  • the face image of the candidate pedestrian is included, the face image of the candidate pedestrian is input into the Insight Face network to obtain the face feature vector x G of the candidate pedestrian;
  • the cosine similarity calculation formula calculate the face similarity scores of the candidate pedestrian and the target pedestrian, and the cosine similarity calculation formula is as follows:
  • s C, G is the similarity score of the candidate pedestrian and the target pedestrian
  • x C is the feature corresponding to Cast pedestrian
  • x G is the feature corresponding to Gallery pedestrian
  • the face image of the candidate pedestrian is not included, it is determined that the face similarity score of the candidate pedestrian and the target pedestrian is -1.
  • inputting the first zoomed image into the MTCNN network to obtain the face image of the target pedestrian includes:
  • N is a natural number greater than 1;
  • the face image of the target pedestrian with the highest confidence level and the N key points corresponding to the face image are acquired.
  • the preset threshold is 0.35.
  • the method further includes: if the similarity score is greater than or equal to If the number of pedestrian images with a preset threshold is less than M, then the top M candidate pedestrians with similarity scores are selected as the target image IQ (x, y) for pedestrian re-recognition; where M is a natural number greater than 1.
  • I the similarity between the target pedestrian image and the candidate pedestrian image
  • D Q For normalized candidate pedestrian features
  • D Q For normalized candidate pedestrian features, D Q, G are the distance between the target pedestrian image and the candidate pedestrian image;
  • the calculated Euclidean distance is corrected to obtain the candidate pedestrians in the processed candidate pedestrian image I G '(x,y) and the target image IQ (x,y) The initial Euclidean distance between candidate pedestrians.
  • the image I c (x, y) contains two or more target pedestrians, filter the panoramic image re-recognized by the pedestrian corresponding to each target pedestrian to obtain the processed candidate pedestrian image I G '(x,y), including:
  • image I c (x, y) contains target pedestrian A and target pedestrian B. Since the panoramic image sets corresponding to target pedestrian A and target pedestrian B are the same, when an image in the panoramic image set is determined to be target pedestrian A When the target image of is selected, the image is deleted from the panoramic image set of target pedestrian B.
  • the present invention has the following beneficial effects:
  • FIG. 1 is a schematic diagram of the principle of a two-stage pedestrian search method combining face and appearance provided by the present invention
  • FIG. 2 is a schematic diagram of the effect of a two-stage pedestrian search method combining face and appearance provided by the present invention.
  • FIG. 1 is a schematic diagram of the principle of a two-stage pedestrian search method combining face and appearance provided by the present invention.
  • face features are extracted from the original target set and the original matching set to perform face detection and face recognition To obtain the similarity score of the face between the target pedestrian and the candidate pedestrian. Then, sort according to the face similarity score, and perform threshold separation on the original set to be matched according to a preset threshold to obtain a new multi-target set and a new set to be matched. If the original target set contains two or more target pedestrians, the panoramic image re-recognized by the pedestrians corresponding to each target pedestrian is filtered to obtain the processed candidate pedestrian images.
  • the new multi-target set contains target images for pedestrian re-recognition; the new to-be-matched set contains processed candidate pedestrian images for re-recognition of pedestrians. Then, re-identify the new multi-target set and the new set to be matched, and correct the calculated Euclidean distance according to the rearrangement algorithm based on K complementary neighbors, and finally output the candidate pedestrian ranking results.
  • FIG. 2 is a schematic diagram of the effect of a two-stage pedestrian search method combining face and appearance provided by the present invention. As shown in FIG. 2, only the front photo indicated by the solid arrow on the left can be recognized only through face recognition, and the dotted line on the right Arrows indicate that faceless pedestrian images cannot be recognized.
  • the method provided by the present invention combines facial features and appearance features, which can effectively recognize images such as the back view and improve the performance of image recognition.
  • the two-stage pedestrian search method combining face and appearance provided by the present invention includes specific steps including:
  • the original target set may contain images of multiple target pedestrians, or a single image may contain multiple target pedestrians.
  • S102 Acquire a panoramic image (Gallery) to be recognized and pedestrian coordinate information in the panoramic image, and determine an image I G (x, y) containing a candidate pedestrian.
  • the given input target pedestrian image I c (x,y), namely Cast is fed into the neural network, scaled to a fixed scale (112 ⁇ 112), and the current target pedestrian image is detected using the MTCNN network Face, determine the position of the face and send the face part to the Insight Face network to extract features to obtain a 512-dimensional vector
  • the candidate pedestrian image similarity score for which no face is detected is set to -1.
  • the MTCNN network outputs the face area and the corresponding 10 face key points; if MTCNN detects more than one face area in a pedestrian image, only the face area with the highest confidence level and the corresponding face are selected Key points for output.
  • the characteristics of the network output are first normalized, and then the vector dot product is sufficient; in specific operations, matrix multiplication can be used to accelerate the operation.
  • S104 Sort according to the face similarity score. If the similarity score is greater than or equal to a preset threshold, use the corresponding image containing the candidate pedestrian as the target image for pedestrian re-identification IQ (x, y) (Multi-query) ; If the similarity score is less than the preset threshold, the corresponding image containing the candidate pedestrian is used as a panoramic image (Pallery) for pedestrian re-recognition.
  • all candidate pedestrian images are sorted according to the similarity with the target pedestrian, Cast, from large to small, and the candidate pedestrian image similarity scores of which no face is detected are set to -1, and are randomly sorted at the end, thus Get a sorted list of face similarities.
  • the threshold of similarity score is set according to experience. Candidate pedestrian images with similarity scores greater than the threshold are used as Multi-query for subsequent pedestrian re-identification, and candidate almond images with remaining similarity scores less than the threshold are used as Gallery for subsequent pedestrian re-identification.
  • the candidate pedestrian images for which the similarity score is obtained are only a part of all the candidate pedestrian images, and the subsequent multi-query recognition is generated from here.
  • the similarity threshold is generally selected to be 0.35, that is, candidate pedestrians with a face similarity of more than 0.35 are selected as subsequent multi-query for re-recognition; if the number of candidate pedestrians above some target pedestrian threshold is too small, such as less than Five, then select the five candidate pedestrians with the highest similarity score as Multi-query for subsequent recognition.
  • the Multi-query images corresponding to the remaining target pedestrians in the data set are removed from the Gallery of the current target pedestrian, and the matching sample space for pedestrian re-identification is reduced.
  • the candidate pedestrian sets of different target pedestrians are the same, then the Multi-query corresponding to pedestrian A has a high degree of similarity to the face of pedestrian A, which can be considered as pedestrian A with a high probability, so that Remove this part of Multi-query from the Gallery of Pedestrian B, compress the pedestrians of Pedestrian B and then identify the matching space, and improve the matching quasi-group rate.
  • the pedestrian re-identification network must be trained first.
  • the process of pedestrian re-identification network training specifically:
  • Pedestrian re-recognition is regarded as a classification task, all the ID numbers in the training set are counted as the number of classification categories, the input picture is scaled to a size of 288 ⁇ 288, and randomly cropped to a size of 256 ⁇ 256, and data is augmented using random horizontal flips, and then Input the DenseNet network to extract features, and output a feature map with a size of 8 ⁇ 8 ⁇ 1024; perform Average Pooling to obtain a 1024-dimensional vector, and use a fully connected layer to output a vector corresponding to the number of categories. After activation by the Softmax function, the input image corresponds to each Probability of each ID category.
  • p i represents the probability that the current pedestrian belongs to an i-th category
  • p t represents the probability corresponding to the real category
  • the average distance relative to all Multi-query pedestrians is calculated as the final distance; and all Gallery pedestrians are sorted according to the distance from small to large, and the final matching result is output.
  • the existing face detector is used to detect the face of a given pedestrian; the face comparison model is trained based on a large public data set and the face representation vector is output; the matched pedestrian set is sorted and used according to the Euclidean distance
  • the rearrangement algorithm obtains a more robust face ranking result; selects several samples from the ranking result as multiple matching targets in the next stage according to the distance from the original matching target; the next selected by different pedestrians in the same data set Multi-matching targets at the stage are used as negative samples for each other, thereby compressing the sample space for the next stage of matching; finally, multi-target pedestrian recognition is performed, and the image set to be matched is sorted according to the average distance or similarity to multiple targets to output the final result.
  • the invention combines face and pedestrian re-recognition through the DenseNet-based convolutional neural network to search for target pedestrians, and improves the robustness and reduces the amount of calculation through carefully designed constraints, and further improves the performance of pedestrian search.
  • the training data used in this specific example comes from 115 Chinese and Western movie sample frames. Each movie can be regarded as an independent data set, with an average of 6.7 pedestrians to be searched for each movie.
  • the verification data and test data contain 19 and 58 movies, respectively.
  • Table 1 and Table 2 are the results of pedestrian re-recognition on the verification data set and the test data set of this embodiment, respectively, and the evaluation standard adopts mAP (meanAveragePrecision).
  • this embodiment greatly improve the performance of pedestrian re-recognition in the pedestrian search task.
  • this embodiment adopts an integrated processing method, which fully utilizes the result of face detection, and greatly reduces the research complexity in practical applications.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Databases & Information Systems (AREA)
  • Human Computer Interaction (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Library & Information Science (AREA)
  • Medical Informatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

一种结合人脸与外观的两阶段行人搜索方法，包括：利用人脸检测器检测行人的人脸，基于人脸比对模型输出人脸表征向量；根据欧氏距离对待匹配行人集进行排序获取人脸排序结果；根据排序结果选取若干个样本作为下一阶段的多匹配目标；将同一数据集中不同行人的所挑选出的下一阶段的多匹配目标作为彼此的负样本，从而压缩下一阶段匹配的样本空间；最后进行多目标行人再识别，将待匹配图片集根据与多个目标的平均距离或相似度进行排序输出最终结果。通过基于DenseNet的卷积神经网络将人脸与行人再识别联合起来搜索目标行人，并通过精心设计的限制条件提高了鲁棒性减少了计算量，进一步提升了行人搜索的性能。

Description

结合人脸与外观的两阶段行人搜索方法 技术领域
本发明涉及计算机视觉技术领域，具体地，涉及一种结合人脸与外观的两阶段行人搜索方法。
背景技术
目前，行人再识别(Person Re-identification)技术在实际应用中发挥着越来越重要的作用，比如安全监控，人群流量监测，行人行为分析等。现如今大部分计算机视觉研究者将行人再识别研究聚焦于相似场景下多摄像头间的行人匹配问题。
但是，在实际应用中，例如在对犯罪嫌疑人进行布控时，目标行人图像(如肖像照)与待匹配的摄像机拍摄行人图像之间往往外观差异较大，采用传统的行人再识别方法无法有效地匹配目标行人。该场景下较可靠的行人特征应为人脸特征，但如果只用人脸进行匹配，则无人脸的目标行人样本无法匹配，可能会丢失对其轨迹的追踪。
发明内容
本发明的目的在于克服上述现有技术的不足之处，提出了一种基于深度学习的结合人脸与外观的两阶段行人搜索方法，联合人脸识别与行人再识别问题。首先，本发明基于卷积神经网络提取更具有分辨力的特征，采用了DenseNet网络结构(参见：Gao Huang,Zhuang Liu,Laurens van der Maaten,and Kilian Q Weinberger,“Densely connected convolutional networks,”in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,2017.)；其次，本发明提出了一种结合人脸 与外观的两阶段行人搜索方法，利用人脸检测和识别从目标行人图像出发，从候选行人图片中产生Multi-query；最后，本发明设计了一种行人再识别匹配样本空间压缩方法，提升行人再识别的准确率。
本发明是通过以下技术方案实现的。
一种结合人脸与外观的两阶段行人搜索方法，包括：
获取包含目标行人的图像I c(x,y)，称为Cast；
获取待识别的全景图像(Gallery)，以及全景图像中的行人坐标信息，确定包含候选行人的图像I G(x,y)；
根据所述图像I c(x,y)、图像I G(x,y)，计算目标行人与候选行人的人脸相似性分数；
根据所述人脸相似性分数进行排序，若相似性分数大于或等于预设阈值，则将对应的包含候选行人的图像作为行人再识别的目标图像I Q(x,y)(Multi-query)；若相似性分数小于预设阈值，则将对应的包含候选行人的图像作为行人再识别的全景图像(Gallery)；
若图像I c(x,y)中包含2个及以上目标行人时，对每个所述目标行人对应的行人再识别的全景图像进行过滤处理，得到处理后的候选行人图像I G‘(x,y)；
根据所述目标图像I Q(x,y)、处理后的候选行人图像I G‘(x,y)，计算目标图像I Q(x,y)中的候选行人与处理后的候选行人图像I G‘(x,y)中的候选行人之间的初始欧式距离；
将处理后的候选行人图像I G‘(x,y)中的候选行人与目标图像I Q(x,y)中的候选行人之间的平均欧式距离作为候选行人的最终欧式距离，并按照所述最终欧式距离对处理后的候选行人图像I G‘(x,y)中的候选行人进行排序，得到对应的排序结果。
可选地，根据所述图像I c(x,y)、图像I G(x,y)，计算目标行人与候选行人的人脸相似性分数，包括：
按照第一预设尺度对所述图像I c(x,y)进行缩放处理，得到第一缩放图像；
将所述第一缩放图像输入MTCNN网络中，得到目标行人的人脸图像；
将所述目标行人的人脸图像输入到Insight Face网络中，得到所述目标行人的人脸特征向量x c；
按照第一预设尺度对所述图像I G(x,y)进行缩放处理，得到第二缩放图像；将所述第二缩放图像输入MTCNN网络中，判断是否包含候选行人的人脸图像；
若包含有候选行人的人脸图像，则将所述候选行人的人脸图像输入到Insight Face网络中，得到所述候选行人的人脸特征向量x G；
根据余弦相似度计算公式，计算所述候选行人与所述目标行人的人脸相似性分数，所述余弦相似度计算公式如下：
Figure PCTCN2019108502-appb-000001
其中，s C,G为候选行人与目标行人的人脸相似性分数，x C为Cast行人对应的特征，x G为Gallery行人对应的特征；
若不包含候选行人的人脸图像，则确定所述候选行人与所述目标行人的人脸相似性分数为-1。
可选地，将所述第一缩放图像输入MTCNN网络中，得到目标行人的人脸图像，包括：
通过MTCNN网络获取目标行人的人脸图像，以及人脸图像对应的N个关键点；N为大于1的自然数；
若通过MTCNN网络获取的目标人脸图像的数量大于1，则获取置信度最高的目标行人的人脸图像，以及所述人脸图像对应的N个关键点。
可选地，所述预设阈值为0.35。
可选地，在根据所述图像I c(x,y)、图像I G(x,y)，计算目标行人与候选行人的人脸相似性分数之后，还包括：若相似性分数大于或等于预设阈值的行 人图像的数量小于M，则选择相似性分数排在前M个的候选行人作为行人再识别的目标图像I Q(x,y)；其中，M为大于1的自然数。
可选地，根据所述目标图像I Q(x,y)、处理后的候选行人图像I G‘(x,y)，计算目标行人与候选行人的初始欧式距离，包括：
按照第二预设尺度对所述目标图像I Q(x,y)进行缩放处理，得到第三缩放图像；
将所述第三缩放图像输入基于DenseNet的行人再识别网络中，得到目标行人的特征向量
Figure PCTCN2019108502-appb-000002
按照第二预设尺度对所述处理后的候选行人图像I G‘(x,y)进行缩放处理，得到第四缩放图像；
将所述第四缩放图像输入基于DenseNet的行人再识别网络中，得到候选行人的特征向量
Figure PCTCN2019108502-appb-000003
计算处理后的候选行人图像I G‘(x,y)中的候选行人与目标图像I Q(x,y)中的候选行人之间的欧式距离，计算公式如下：
Figure PCTCN2019108502-appb-000004
Figure PCTCN2019108502-appb-000005
其中：
Figure PCTCN2019108502-appb-000006
为目标行人图像与候选行人图像间的相似度，
Figure PCTCN2019108502-appb-000007
为归一化的目标行人特征，
Figure PCTCN2019108502-appb-000008
为归一化的候选行人特征，D Q,G为目标行人图像与候选行人图像之间的距离；
根据基于K互补近邻的重排算法，对计算到的欧式距离进行修正，得到处理后的候选行人图像I G‘(x,y)中的候选行人与目标图像I Q(x,y)中的候选行人之间的初始欧式距离。
可选地，若图像I c(x,y)中包含2个及以上目标行人时，对每个所述目标行人对应的行人再识别的全景图像进行过滤处理，得到处理后的候选行人图像I G‘(x,y)，包括：
假设图像I c(x,y)中包含目标行人A和目标行人B，由于目标行人A和目标行人B所对应的全景图像集相同，因此，当确定全景图像集中的某一图像为目标行人A的目标图像时，则将该图像从目标行人B的全景图像集中删除。
与现有技术相比，本发明具有如下的有益效果：
1)利用基于DenseNet的卷积神经网络提取具有更高鲁棒性的特征；
2)有效利用人脸信息，首先获取高置信度的含人脸的目标图像，再根据外观特征行人再识别，获得最终的候选行人排序结果，识别效果更好；
3)设计多个目标行人之间的行人再识别的全景图像进行过滤处理，压缩行人再识别匹配样本空间，减少计算量，提高准确率。
附图说明
通过阅读参照以下附图对非限制性实施例所作的详细描述，本发明的其它特征、目的和优点将会变得更明显：
图1为本发明提供的结合人脸与外观的两阶段行人搜索方法的原理示意图；
图2为本发明提供的结合人脸与外观的两阶段行人搜索方法的效果示意图。
具体实施方式
下面结合具体实施例对本发明进行详细说明。以下实施例将有助于本领域的技术人员进一步理解本发明，但不以任何形式限制本发明。应当指出的是，对本领域的普通技术人员来说，在不脱离本发明构思的前提下，还可以做出若干变化和改进。这些都属于本发明的保护范围。
图1为本发明提供的结合人脸与外观的两阶段行人搜索方法的原理示意图，如图1所示，对原目标集、原待匹配集提取人脸特征，进行人脸检测及人脸识别，获得目标行人与候选行人的人脸相似性分数。然后，根据人脸相似性分数进行排序，根据预设阈值对原待匹配集进行阈值分离，得到新多目标集、新待匹配集。若原目标集中包含2个及以上目标行人时，对每个目标行人对应的行人再识别的全景图像进行过滤处理，得到处理后的候选行人图像。其中，新多目标集包含行人再识别的目标图像；新待匹配集包含行人再识别的处理后的候选行人图像。然后，对新多目标集、新待匹配集进行行人再识别，并根据基于K互补近邻的重排算法，对计算到的欧式距离进行修正，最终输出候选行人的排序结果。图2为本发明提供的结合人脸与外观的两阶段行人搜索方法的效果示意图，如图2所示，仅通过人脸识别只能识别左侧实线箭头表示的正面照片，对右侧虚线箭头表示不露脸的行人图像不能识别。而本发明提供的方法将人脸特征和外观特征相结合，可以有效地识别背影等图像，提高了图像识别的性能。
本发明提供的结合人脸与外观的两阶段行人搜索方法，具体步骤包括：
S101、获取包含目标行人的图像I c(x,y)。
本实施例中，原目标集中可以包含多个目标行人的图像，也可以一个图像中包含多个目标行人。
S102、获取待识别的全景图像(Gallery)，以及全景图像中的行人坐标信息，确定包含候选行人的图像I G(x,y)。
S103、根据图像I c(x,y)、图像I G(x,y)，计算目标行人与候选行人的人脸相似性分数。
本实施例中，将给定的输入目标行人图片I c(x,y)，即Cast，送入神经网络，按比例缩放到固定尺度(112×112)，利用MTCNN网络检测当前目标行人 图像中的人脸，确定人脸位置并将人脸部分送入Insight Face网络提取特征得到512维向量
Figure PCTCN2019108502-appb-000009
根据给定的输入全景图片(Gallery)和行人坐标信息获取行人图像I G(x,y)，然后将其送入神经网络，按比例缩放到固定尺度(112×112)，利用MTCNN网络检测当前行人图像中是否含有人脸。若含有人脸，确定人脸位置并将人脸部分送入Insight Face网络提取特征得到512维向量
Figure PCTCN2019108502-appb-000010
计算目标行人与候选行人的人脸相似度，利用如下的余弦相似度计算公式得到相似性分数：
Figure PCTCN2019108502-appb-000011
没有检测到人脸的候选行人图片相似性分数设为-1。
优选地，MTCNN网络输出人脸所在区域以及对应的10个人脸关键点；若MTCNN在一张行人图像中检测到超过一个人脸区域，则只选择最高置信度的人脸区域及对应的人脸关键点进行输出。
优选地，计算相似性分数时首先对网络输出的特征进行归一化，而后进行向量点乘即可；具体操作时可利用矩阵乘法加速运算。
S104、根据人脸相似性分数进行排序，若相似性分数大于或等于预设阈值，则将对应的包含候选行人的图像作为行人再识别的目标图像I Q(x,y)(Multi-query)；若相似性分数小于预设阈值，则将对应的包含候选行人的图像作为行人再识别的全景图像(Gallery)。
本实施例中，按照与目标行人即Cast的相似度从大到小对全部候选行人图片进行排序，没有检测到人脸的候选行人图片相似性分数设为-1，并随机排序在最后，从而得到人脸相似度排序列表。
根据经验设置相似性分数阈值，取相似性分数大于阈值的候选行人图片作为后续行人再识别的Multi-query，剩余的相似性分数小于阈值的候选杏仁图片作为后续行人再识别的Gallery。
优选地，得到相似性分数的候选行人图像只是全部候选行人图像的一部分，后续再识别的Multi-query从这里产生。
优选地，一般选择相似性阈值为0.35，即具有人脸相似度0.35以上的候选行人被选作后续再识别的Multi-query；若某些目标行人阈值以上的候选行人数量过少，如少于5个，则选择相似性分数最高的5个候选行人作为后续再识别的Multi-query。
S105、若图像I c(x,y)中包含2个及以上目标行人时，对每个目标行人对应的行人再识别的全景图像进行过滤处理，得到处理后的候选行人图像I G‘(x,y)。
本实施例中，从当前目标行人的Gallery中移除该数据集中其余目标行人对应的Multi-query图像，缩减行人再识别的匹配样本空间。
优选地，在同一个数据集里，不同的目标行人的候选行人集合相同，则行人A对应的Multi-query因为与行人A的人脸相似度较高，可以大概率认为就是行人A，从而可以将这部分Multi-query从行人B的Gallery中移除，压缩行人B的行人再识别匹配空间，提高匹配准群率。
S106、根据目标图像I Q(x,y)、处理后的候选行人图像I G‘(x,y)，计算目标图像I Q(x,y)中的候选行人与处理后的候选行人图像I G‘(x,y)中的候选行人之间的初始欧式距离。
本实施例中，将当前行人的所有Multi-query图像I Q(x,y)送入神经网络，按比例缩放到固定尺度(256×256)，利用基于DenseNet的行人再识别网络提取Multi-query图像的特征得到向量
Figure PCTCN2019108502-appb-000012
将当前行人的所有Gallery图像I G(x,y)送入神经网络，按比例缩放到固定尺度(256×256)，利用基于DenseNet的行人再识别网络提取Gallery图像的特征得到向量
Figure PCTCN2019108502-appb-000013
计算Multi-query行人与Gallery行人的相似度，利用如下的余弦相似度计 算公式得到相似性分数：
Figure PCTCN2019108502-appb-000014
将
Figure PCTCN2019108502-appb-000015
作为Multi-query行人与Gallery行人的距离，并根据基于K互补近邻的重排算法重新计算距离Multi-query行人与Gallery行人的距离；
优选地：首先要对行人再识别网络进行训练。
优选地：行人再识别网络进行训练的过程，具体为：
将行人再识别视为分类任务，统计训练集中所有的ID数作为分类类别数，将输入图片缩放到288×288大小，并随机裁剪到256×256大小，利用随机水平翻转进行数据增广，随后输入DenseNet网络提取特征，输出8×8×1024大小的特征图；对其进行Average Pooling得到1024维向量，并用全连接层输出对应于类别数维度的向量，经过Softmax函数激活得到输入图像对应于每个ID类别的概率。
行人再识别结果由如下的Loss函数监督：
L(x,y)＝-logp t,
Figure PCTCN2019108502-appb-000016
其中，p i表示当前行人属于某第i个类别的概率，p t表示对应于真实类别的概率。
S107、将处理后的候选行人图像I G‘(x,y)中的候选行人与目标图像I Q(x,y)中的候选行人之间的平均欧式距离作为候选行人的最终欧式距离，并按照最终欧式距离对处理后的候选行人图像I G‘(x,y)中的候选行人进行排序，得到对应的排序结果。
本实施例中，对某一Gallery行人，计算其相对于所有Multi-query行人的平均距离作为最终距离；并按照此距离从小到大对所有Gallery行人进行排序， 输出最终匹配结果。
本实施例，利用现有的人脸检测器检测所给行人的人脸；基于公开大型数据集训练人脸比对模型并输出人脸表征向量；根据欧氏距离对待匹配行人集进行排序并利用重排算法获取更鲁棒的人脸排序结果；根据与原匹配目标的距离从排序结果中选取若干个样本作为下一阶段的多匹配目标；将同一数据集中不同行人的所挑选出的下一阶段的多匹配目标作为彼此的负样本，从而压缩下一阶段匹配的样本空间；最后进行多目标行人再识别，将待匹配图片集根据与多个目标的平均距离或相似度进行排序输出最终结果。本发明通过基于DenseNet的卷积神经网络将人脸与行人再识别联合起来搜索目标行人，并通过精心设计的限制条件提高了鲁棒性减少了计算量，进一步提升了行人搜索的性能。
下面结合具体实例对本实施例进一步描述。
本具体实例采用的训练数据来自115部中西方电影采样帧，每部电影可以视作一个独立的数据集，平均每部电影待搜索行人为6.7个。验证数据和测试数据分别包含19部和58部电影。
通过实验证明，本实施例方法能很好联合人脸识别和行人再识别完成行人搜索任务。表1和表2为本实施例分别在验证数据集和测试数据集上的行人再识别结果，评价标准采用mAP(mean Average Precision)。
表1
Figure PCTCN2019108502-appb-000017
Figure PCTCN2019108502-appb-000018
表2
Figure PCTCN2019108502-appb-000019
可以看出由本实施例得到的结果较大程度了提升了行人搜索任务中行人再识别的性能。此外，本实施例采用一体化的处理方式，充分地利用了人脸检测的结果，极大地降低了实际应用中的研究复杂度。
以上对本发明的具体实施例进行了描述。需要理解的是，本发明并不局限于上述特定实施方式，本领域技术人员可以在权利要求的范围内做出各种变形或修改，这并不影响本发明的实质内容。

Claims (7)

  1. 一种结合人脸与外观的两阶段行人搜索方法，其特征在于，包括:
    获取包含目标行人的图像I c(x,y)；
    获取待识别的全景图像，以及全景图像中的行人坐标信息，确定包含候选行人的图像I G(x,y)；
    根据所述图像I c(x,y)、图像I G(x,y)，计算目标行人与候选行人的人脸相似性分数；
    根据所述人脸相似性分数进行排序，若相似性分数大于或等于预设阈值，则将对应的包含候选行人的图像作为行人再识别的目标图像I Q(x,y)；若相似性分数小于预设阈值，则将对应的包含候选行人的图像作为行人再识别的全景图像；
    若图像I c(x,y)中包含2个及以上目标行人时，对每个所述目标行人对应的行人再识别的全景图像进行过滤处理，得到处理后的候选行人图像I G‘(x,y)；
    根据所述目标图像I Q(x,y)、处理后的候选行人图像I G‘(x,y)，计算目标图像I Q(x,y)中的候选行人与处理后的候选行人图像I G‘(x,y)中的候选行人之间的初始欧式距离；
    将处理后的候选行人图像I G‘(x,y)中的候选行人与目标图像I Q(x,y)中的候选行人之间的平均欧式距离作为候选行人的最终欧式距离，并按照所述最终欧式距离对处理后的候选行人图像I G‘(x,y)中的候选行人进行排序，得到对应的排序结果。
  2. 根据权利要求1所述的结合人脸与外观的两阶段行人搜索方法，其特征在于，根据所述图像I c(x,y)、图像I G(x,y)，计算目标行人与候选行人的人脸相似性分数，包括：
    按照第一预设尺度对所述图像I c(x,y)进行缩放处理，得到第一缩放图像；
    将所述第一缩放图像输入MTCNN网络中，得到目标行人的人脸图像；
    将所述目标行人的人脸图像输入到Insight Face网络中，得到所述目标行人的人脸特征向量x c；
    按照第一预设尺度对所述图像I G(x,y)进行缩放处理，得到第二缩放图像；将所述第二缩放图像输入MTCNN网络中，判断是否包含候选行人的人脸图像；
    若包含有候选行人的人脸图像，则将所述候选行人的人脸图像输入到Insight Face网络中，得到所述候选行人的人脸特征向量x G；
    根据余弦相似度计算公式，计算所述候选行人与所述目标行人的人脸相似性分数，所述余弦相似度计算公式如下：
    Figure PCTCN2019108502-appb-100001
    其中，s C,G为候选行人与目标行人的人脸相似性分数，x C为Cast行人对应的特征，x G为Gallery行人对应的特征；
    若不包含候选行人的人脸图像，则确定所述候选行人与所述目标行人的人脸相似性分数为-1。
  3. 根据权利要求2所述的结合人脸与外观的两阶段行人搜索方法，其特征在于，将所述第一缩放图像输入MTCNN网络中，得到目标行人的人脸图像，包括：
    通过MTCNN网络获取目标行人的人脸图像，以及人脸图像对应的N个关键点；N为大于1的自然数；
    若通过MTCNN网络获取的目标人脸图像的数量大于1，则获取置信度最高的目标行人的人脸图像，以及所述人脸图像对应的N个关键点。
  4. 根据权利要求1所述的结合人脸与外观的两阶段行人搜索方法，其特征在于，所述预设阈值为0.35。
  5. 根据权利要求1所述的结合人脸与外观的两阶段行人搜索方法，其特征在于，在根据所述图像I c(x,y)、图像I G(x,y)，计算目标行人与候选行人的人脸相似性分数之后，还包括：若相似性分数大于或等于预设阈值的行人图像的数量小于M，则选择相似性分数排在前M个的候选行人作为行人再识别的目标图像I Q(x,y)；其中，M为大于1的自然数。
  6. 根据权利要求1所述的结合人脸与外观的两阶段行人搜索方法，其特征在于，根据所述目标图像I Q(x,y)、处理后的候选行人图像I G‘(x,y)，计算目标行人与候选行人的初始欧式距离，包括：
    按照第二预设尺度对所述目标图像I Q(x,y)进行缩放处理，得到第三缩放图像；
    将所述第三缩放图像输入基于DenseNet的行人再识别网络中，得到目标行人的特征向量
    Figure PCTCN2019108502-appb-100002
    按照第二预设尺度对所述处理后的候选行人图像I G‘(x,y)进行缩放处理，得到第四缩放图像；
    将所述第四缩放图像输入基于DenseNet的行人再识别网络中，得到候选行人的特征向量
    Figure PCTCN2019108502-appb-100003
    计算处理后的候选行人图像I G‘(x,y)中的候选行人与目标图像I Q(x,y)中的候选行人之间的欧式距离，计算公式如下：
    Figure PCTCN2019108502-appb-100004
    Figure PCTCN2019108502-appb-100005
    其中：
    Figure PCTCN2019108502-appb-100006
    为目标行人图像与候选行人图像间的相似度，
    Figure PCTCN2019108502-appb-100007
    为归一化的目标行人特征，
    Figure PCTCN2019108502-appb-100008
    为归一化的候选行人特征，D Q,G为目标行人图像与候选行人图像之间的距离；
    根据基于K互补近邻的重排算法，对计算到的欧式距离进行修正，得到处理后的候选行人图像I G‘(x,y)中的候选行人与目标图像I Q(x,y)中的候选行人之间的初始欧式距离。
  7. 根据权利要求1所述的结合人脸与外观的两阶段行人搜索方法，其特征在于，若图像I c(x,y)中包含2个及以上目标行人时，对每个所述目标行人对应的行人再识别的全景图像进行过滤处理，得到处理后的候选行人图像I G‘(x,y)，包括：
    假设图像I c(x,y)中包含目标行人A和目标行人B，由于目标行人A和目标行人B所对应的全景图像集相同，因此，当确定全景图像集中的某一图像为目标行人A的目标图像时，则将该图像从目标行人B的全景图像集中删除。
PCT/CN2019/108502 2018-11-29 2019-09-27 结合人脸与外观的两阶段行人搜索方法 WO2020108075A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/030,358 US11017215B2 (en) 2018-11-29 2020-09-23 Two-stage person searching method combining face and appearance features

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811447035.4 2018-11-29
CN201811447035.4A CN109635686B (zh) 2018-11-29 2018-11-29 结合人脸与外观的两阶段行人搜索方法

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/030,358 Continuation US11017215B2 (en) 2018-11-29 2020-09-23 Two-stage person searching method combining face and appearance features

Publications (1)

Publication Number Publication Date
WO2020108075A1 true WO2020108075A1 (zh) 2020-06-04

Family

ID=66070310

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/108502 WO2020108075A1 (zh) 2018-11-29 2019-09-27 结合人脸与外观的两阶段行人搜索方法

Country Status (3)

Country Link
US (1) US11017215B2 (zh)
CN (1) CN109635686B (zh)
WO (1) WO2020108075A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111782856A (zh) * 2020-07-15 2020-10-16 上海依图网络科技有限公司 一种人体图像处理方法、装置、设备及介质
CN112699810A (zh) * 2020-12-31 2021-04-23 中国电子科技集团公司信息科学研究院 一种提升室内监控系统人物识别精度的方法及装置

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109635686B (zh) 2018-11-29 2021-04-23 上海交通大学 结合人脸与外观的两阶段行人搜索方法
CN110084157B (zh) * 2019-04-12 2021-08-24 北京旷视科技有限公司 图像重识别的数据处理方法及装置
AU2020329148A1 (en) * 2019-08-09 2022-03-17 Clearview Ai, Inc. Methods for providing information about a person based on facial recognition
CN110796072B (zh) * 2019-10-28 2023-04-07 桂林电子科技大学 一种基于双任务学习的目标跟踪及身份识别方法
CN111368772B (zh) * 2020-03-11 2023-08-22 杭州海康威视系统技术有限公司 身份识别方法、装置、设备及存储介质
CN111738062B (zh) * 2020-05-09 2024-05-17 广智微芯(扬州)有限公司 一种基于嵌入式平台的自动再识别系统
CN111951229A (zh) * 2020-07-22 2020-11-17 国网安徽省电力有限公司电力科学研究院 一种基于全连接层增广的小金具图像数据集训练方法
CN112270269B (zh) * 2020-10-30 2023-04-18 湖南快乐阳光互动娱乐传媒有限公司 一种人脸图像质量的评估方法及装置
CN112612956B (zh) * 2020-12-23 2022-11-25 陇东学院 一种基于全景视图的艺术设计方法和平台

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160307561A1 (en) * 2015-04-17 2016-10-20 Lakdas Nanayakkara System for Providing Assistance to the Visually Impaired
CN107145826A (zh) * 2017-04-01 2017-09-08 浙江大学 基于双约束度量学习和样本重排序的行人再识别方法
CN107704811A (zh) * 2017-09-14 2018-02-16 云南大学 一种基于模糊鲁棒特征的行人再识别方法及模块装置
CN109635686A (zh) * 2018-11-29 2019-04-16 上海交通大学 结合人脸与外观的两阶段行人搜索方法

Family Cites Families (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8503800B2 (en) * 2007-03-05 2013-08-06 DigitalOptics Corporation Europe Limited Illumination detection using classifier chains
US8379940B2 (en) 2009-06-02 2013-02-19 George Mason Intellectual Properties, Inc. Robust human authentication using holistic anthropometric and appearance-based features and boosting
US8648959B2 (en) * 2010-11-11 2014-02-11 DigitalOptics Corporation Europe Limited Rapid auto-focus using classifier chains, MEMS and/or multiple object focusing
US9396412B2 (en) * 2012-06-21 2016-07-19 Siemens Aktiengesellschaft Machine-learnt person re-identification
US10120879B2 (en) * 2013-11-29 2018-11-06 Canon Kabushiki Kaisha Scalable attribute-driven image retrieval and re-ranking
US10318576B2 (en) * 2013-12-12 2019-06-11 Nant Holdings Ip, Llc Image recognition verification
CN105718882B (zh) * 2016-01-19 2018-12-18 上海交通大学 一种分辨率自适应特征提取与融合的行人重识别方法
US10007707B1 (en) 2016-01-23 2018-06-26 Digerati, Inc. System and method for modification, personalization and customizable filtering of search results and search result ranking in an internet-based search engine
CN106874898B (zh) * 2017-04-08 2021-03-30 复旦大学 基于深度卷积神经网络模型的大规模人脸识别方法
CN108875507B (zh) * 2017-11-22 2021-07-23 北京旷视科技有限公司 行人跟踪方法、设备、系统和计算机可读存储介质
CN108133230A (zh) * 2017-12-14 2018-06-08 西北工业大学 一种面向目标人物距离量度学习的人物再识别方法
CN108280411A (zh) * 2018-01-10 2018-07-13 上海交通大学 一种具有空间变换能力的行人搜索方法
CN108875534B (zh) * 2018-02-05 2023-02-28 北京旷视科技有限公司 人脸识别的方法、装置、系统及计算机存储介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160307561A1 (en) * 2015-04-17 2016-10-20 Lakdas Nanayakkara System for Providing Assistance to the Visually Impaired
CN107145826A (zh) * 2017-04-01 2017-09-08 浙江大学 基于双约束度量学习和样本重排序的行人再识别方法
CN107704811A (zh) * 2017-09-14 2018-02-16 云南大学 一种基于模糊鲁棒特征的行人再识别方法及模块装置
CN109635686A (zh) * 2018-11-29 2019-04-16 上海交通大学 结合人脸与外观的两阶段行人搜索方法

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111782856A (zh) * 2020-07-15 2020-10-16 上海依图网络科技有限公司 一种人体图像处理方法、装置、设备及介质
CN112699810A (zh) * 2020-12-31 2021-04-23 中国电子科技集团公司信息科学研究院 一种提升室内监控系统人物识别精度的方法及装置
CN112699810B (zh) * 2020-12-31 2024-04-09 中国电子科技集团公司信息科学研究院 一种提升室内监控系统人物识别精度的方法及装置

Also Published As

Publication number Publication date
CN109635686A (zh) 2019-04-16
US11017215B2 (en) 2021-05-25
US20210012094A1 (en) 2021-01-14
CN109635686B (zh) 2021-04-23

Similar Documents

Publication Publication Date Title
WO2020108075A1 (zh) 结合人脸与外观的两阶段行人搜索方法
CN111126360B (zh) 基于无监督联合多损失模型的跨域行人重识别方法
Kawewong et al. Online and incremental appearance-based SLAM in highly dynamic environments
WO2019134327A1 (zh) 一种基于边缘检测与sift的人脸表情识别特征提取方法
CN110866466B (zh) 一种人脸识别方法、装置、存储介质和服务器
Kang et al. Pairwise relational networks for face recognition
CN105184238A (zh) 一种人脸识别方法及系统
CN111126122B (zh) 人脸识别算法评估方法及装置
WO2019042195A1 (zh) 一种人体目标身份识别方法及装置
TWI704505B (zh) 人臉辨識系統、建立人臉辨識之資料之方法及其人臉辨識之方法
Du et al. Face association across unconstrained video frames using conditional random fields
WO2023123923A1 (zh) 人体重识别方法、人体重识别装置、计算机设备及介质
CN109359530B (zh) 一种智能视频监控方法及装置
WO2024074075A1 (zh) 重识别方法、存储介质、数据库编辑方法、存储介质
JP2018194956A (ja) 画像認識装置、方法、及びプログラム
CN110490027B (zh) 人脸特征提取训练方法及系统
Cui et al. A face alignment method based on SURF features
Shivakumara et al. New texture-spatial features for keyword spotting in video images
Cheng et al. Automatic Data Cleaning System for Large-Scale Location Image Databases Using a Multilevel Extractor and Multiresolution Dissimilarity Calculation
Annadani et al. Selfie detection by synergy-constraint based convolutional neural network
Alfattama et al. Face Recognition from Partial Face Data
CN111353353A (zh) 跨姿态的人脸识别方法及装置
Yang et al. Misleading Video Detection Using Deep Image Retrieval and Dual-stage Confidence Filtering.
Ma et al. Effective camera motion analysis approach
Meng et al. Object instance search in videos

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19891551

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19891551

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS ONOTING OF LOSS OF RIGHTS (EPO FORM 1205A DATED 27.09.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 19891551

Country of ref document: EP

Kind code of ref document: A1