CN103970771A - Search method and system for human body - Google Patents
Search method and system for human body Download PDFInfo
- Publication number
- CN103970771A CN103970771A CN201310035131.9A CN201310035131A CN103970771A CN 103970771 A CN103970771 A CN 103970771A CN 201310035131 A CN201310035131 A CN 201310035131A CN 103970771 A CN103970771 A CN 103970771A
- Authority
- CN
- China
- Prior art keywords
- human body
- retrieval
- result
- rise
- checked
- 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
- 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
- G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
Landscapes
- Engineering & Computer Science (AREA)
- Library & Information Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a search method and system for a human body. The method comprises the following steps: a picture existence judging step, a first search step and a second search step; during the picture existence judging step, judging the fact whether a picture of a human body to be inquired exists or not, if yes, entering the first search step, otherwise, entering the second search step; during the first search step, using the apparent characteristics of the human body underlayer of the human body to be inquired for screening the human body in a human body database, and generating a first search result; according to the first search result, using the set of top attribute characteristics of the human body to be inquired for screening the human body in the first search result, generating the final search result, and completing the search; during the second search step, confirming a search mode, using the top attribute characteristics of the human body to be inquired for screening the human body in the human body database according to the search mode, generating the final search result of the human body database, and completing the search. Through the adoption of the search method, the human body is searched more accurately and more quickly.
Description
Technical field
The present invention relates to computer vision field, particularly relate to a kind of human body retrieval technique based on human body bottom appearance features and the high-rise attributive character of human body.
Background technology
In recent years, along with for example appearance of social phobia's event such as " 7/7 London case of explosion " and " airport, Moscow case of explosion ", the security protection problem of public place is thus lifted to a new height.And popularizing of monitoring camera equipment provides a kind of new approach for preventing and processing security protection problem, by detecting the frequency of occurrences and the state of same person in zones of different.On the one hand,, in prevention security protection problem, change by analyzing the frequency of occurrences and the state of appearance of pedestrian on diverse location (such as: from having knapsack to packet loss), object suspicious, that there is criminal tendency can be found; On the other hand, by the place that crime occurs monitoring camera is around investigated, the pedestrian's that search once occurred track route and time of occurrence, can assist search suspect.Because number of videos is very huge, process these problems by the method for manually searching and become very difficult; So, design automatic machine algorithm and under different time, different video camera and imaging environment, retrieve same person and just become an Important Problems.Except safety precaution, the human body search method of this robotization can also be applied in other problem, such as: search for the missing or personnel that wander away.
Human body retrieval under the monitor video of multiple-camera network faces the difficult point of following several respects: first, different cameras environment difference of living in, different time light conditions difference, in the time that pedestrian occurs in different time different location, the apparent illumination effect that is subject to of human body has larger difference; Secondly, under different cameras, because the difference at visual angle causes profile difference larger, in the situation that body part cannot being alignd, the traditional coupling of the human body apparent model based on human body bottom appearance features accuracy is poor.In research aspect human body retrieval under existing monitoring environment, existing work great majority are started with from stronger feature or the learning distance metric of design separating capacity.But these a Feature Extraction Methods often pin are aimed at positive human design, and be not suitable for carrying out the human body coupling across visual angle; Meanwhile, utilize the result for retrieval of human body bottom appearance features often to obtain similar to human body to be checked but visibly different human body, for example: human body to be checked is the male sex, the women in Search Results is certainly undesirable; Finally, the common computational complexity of these methods is higher, is not suitable for processing mass data, online quick obtaining result for retrieval.
Some experiences that use according to Web Search Tools, on extensive search problem, can provide the content similar, relevant with the content that is retrieved fully with result for retrieval as much as possible, but in result for retrieval sequence, to investigate, allow as far as possible result for retrieval that user pays close attention on earlier position.
Summary of the invention
Fundamental purpose of the present invention is mate for human body bottom appearance features in the human body retrieval technique of the existing human body (human body to be checked) based on example inaccurate, the problems such as real-time is poor, a kind of search method and system of the human body based on human body bottom appearance features and the high-rise attributive character of human body have been proposed, in the result of retrieving by traditional human body bottom appearance features, use the high-rise attributive character of human body of visual angle robust to filter out some result for retrieval different from the high-rise attributive character of human body of human body to be checked, and/or according to the matching degree of the high-rise attributive character of human body, result for retrieval is resequenced, obtain result for retrieval more accurately.
The search method that the invention provides a kind of human body, comprising:
There is determining step in picture, judges whether to exist the picture of human body to be checked, enters the first searching step, otherwise enter the second searching step if existed;
The first searching step, uses the human body bottom appearance features of described human body to be checked to screen the human body in somatic data storehouse, generates the first result for retrieval; In described the first result for retrieval, use the set of the high-rise attributive character of human body of described human body to be checked, the human body in described the first result for retrieval is screened, generate final result for retrieval, retrieval finishes;
The second searching step, deterministic retrieval mode, is used the high-rise attributive character of described human body of described human body to be checked to screen the human body in somatic data storehouse according to described retrieval mode, generates the final result for retrieval in described somatic data storehouse, and retrieval finishes.
Described the first searching step is further:
The first result for retrieval generates sub-step, calculate the first similarity of the described human body bottom appearance features of human body in described somatic data storehouse and described human body to be checked, according to described the first similarity, the human body in described somatic data storehouse is screened, generate described the first result for retrieval;
The second result for retrieval generates sub-step, calculate the second similarity of the high-rise attributive character of every kind of human body of described human body in described the first result for retrieval and described human body to be checked, according to the second similarity every kind described, human body in described the first result for retrieval is screened, generate described the second result for retrieval;
Final result for retrieval generates sub-step, calculate first total attributive distance of the set of the high-rise attributive character of described human body of described human body in described the second result for retrieval and described human body to be checked, according to described total attributive distance, the human body in described the second result for retrieval is screened, generate the 3rd result for retrieval; Use described first similarity of the described human body in described the 3rd result for retrieval and the second similarity of the high-rise attributive character of whole described human bodies, calculate total similarity of the described human body in described the 3rd result for retrieval, according to described total similarity, the human body in described the 3rd result for retrieval is screened, generate described final result for retrieval; Wherein, described first total attributive distance is: the high-rise attributive character of whole described human body of the described human body in described the second result for retrieval respectively with the absolute value of the difference of the high-rise attributive character of described human body of described human body to be checked and mean value.
Described the second searching step is further:
According to the value type deterministic retrieval mode of the high-rise attributive character of described human body of described human body to be checked, use the high-rise attributive character of described human body of described human body to be checked to screen the human body in somatic data storehouse according to described retrieval mode, generate the final result for retrieval in described somatic data storehouse.
The described value type deterministic retrieval mode according to the high-rise attributive character of described human body of human body to be checked is further: in the time that described value type is two-value output, for the high-rise attributive character of human body every kind described, described human body in described somatic data storehouse is set up to inverted index, adopt the result of one or more described inverted indexs as described final result for retrieval; In the time that described value type is probability output, calculate second total attributive distance of the set of the high-rise attributive character of described human body of described human body in described somatic data storehouse and described human body to be checked, according to described second total attributive distance, the described human body in described somatic data storehouse is screened, generate described final result for retrieval; Wherein, described second total attributive distance is: the high-rise attributive character of whole described human body of the described human body in described somatic data storehouse respectively with the absolute value of the difference of the high-rise attributive character of described human body of described human body to be checked and mean value.
Use the human body attributive classification device of off-line training to extract the high-rise attributive character of described human body.
The present invention also provides a kind of searching system of human body, comprising:
There is judge module in picture, for judging whether to exist the picture of human body to be checked, enters the first retrieval module, otherwise enter the second retrieval module if existed;
The first retrieval module, screens the human body in somatic data storehouse for the human body bottom appearance features that uses described human body to be checked, generates the first result for retrieval; In described the first result for retrieval, use the set of the high-rise attributive character of human body of described human body to be checked, the human body in described the first result for retrieval is screened, generate final result for retrieval, retrieval finishes;
The second retrieval module, for deterministic retrieval mode, uses the high-rise attributive character of described human body of described human body to be checked to screen the human body in somatic data storehouse according to described retrieval mode, generates the final result for retrieval in described somatic data storehouse, and retrieval finishes.
Described the first retrieval module is further:
The first result for retrieval generates submodule, for calculating the first similarity of the human body in described somatic data storehouse and the described human body bottom appearance features of described human body to be checked, according to described the first similarity, the human body in described somatic data storehouse is screened, generate described the first result for retrieval;
The second result for retrieval generates submodule, for calculating the second similarity of the described human body of described the first result for retrieval and the high-rise attributive character of every kind of human body of described human body to be checked, according to the second similarity every kind described, human body in described the first result for retrieval is screened, generate described the second result for retrieval;
Final result for retrieval generates submodule, be used for first total attributive distance of the set of calculating the described human body of described the second result for retrieval and the high-rise attributive character of described human body of described human body to be checked, according to described first total attributive distance, the human body in described the second result for retrieval is screened, generate the 3rd result for retrieval; Use described first similarity of the described human body in described the 3rd result for retrieval and the second similarity of the high-rise attributive character of whole described human bodies, calculate total similarity of the described human body in described the 3rd result for retrieval, according to described total similarity, the human body in described the 3rd result for retrieval is screened, generate described final result for retrieval; Wherein, described first total attributive distance is: the high-rise attributive character of whole described human body of the described human body in described the second result for retrieval respectively with the absolute value of the difference of the high-rise attributive character of described human body of described human body to be checked and mean value.
Described the second retrieval module is further:
According to the value type deterministic retrieval mode of the high-rise attributive character of described human body of described human body to be checked, use the high-rise attributive character of described human body of described human body to be checked to screen the human body in somatic data storehouse according to described retrieval mode, generate the final result for retrieval in described somatic data storehouse.
The described value type deterministic retrieval mode according to the high-rise attributive character of described human body of human body to be checked is further: in the time that described value type is two-value output, for the high-rise attributive character of human body every kind described, described human body in described somatic data storehouse is set up to inverted index, adopt the result of one or more described inverted indexs as described final result for retrieval; In the time that described value type is probability output, calculate second total attributive distance of the set of the high-rise attributive character of described human body of described human body in described somatic data storehouse and described human body to be checked, according to described second total attributive distance, the described human body in described somatic data storehouse is screened, generate described final result for retrieval; Wherein, described second total attributive distance is: the high-rise attributive character of whole described human body of the described human body in described somatic data storehouse respectively with the absolute value of the difference of the high-rise attributive character of described human body of described human body to be checked and mean value.
Use the human body attributive classification device of off-line training to extract the high-rise attributive character of described human body.
Beneficial effect of the present invention is:
(1) utilize the human body image index based on the high-rise attributive character of human body can complete in the short period of time retrieval, make online human body retrieval and extensive human body retrieval become possibility.
(2) end user's height layer attributive character, to being filtered and/or reordered by the result for retrieval of human body bottom appearance features, reduced the quantity of irrelevant Search Results, improved the result rank of relevant human body, reduced user and inquired about burden.
(3) represent human body in conjunction with human body bottom appearance features and the high-rise attributive character of human body, obtained a kind of more complete human body and represented, result for retrieval and human body to be checked have sufficient consistance, have promoted the accuracy of searching system.
(4) in the situation that there is no human body image to be checked, can utilize the description of the high-rise attributive character of human body to carry out human body retrieval, remove the dependence for human body image to be checked.
Brief description of the drawings
Fig. 1 is the schematic diagram of the search method of a kind of human body of the present invention.
Fig. 2 is the schematic diagram of the searching system of a kind of human body of the present invention.
Fig. 3 utilizes human body bottom appearance features to retrieve the result obtaining.
Fig. 4, Fig. 5, Fig. 6 and Fig. 7 progressively increase the result that the high-rise attributive character of human body is filtered.
Fig. 8 and Fig. 9 are the result of filtering after the value of the high-rise attributive character of the human body of given human body to be checked.
Embodiment
For making object of the present invention, technical scheme and advantage clearer, the embodiment that develops simultaneously with reference to the accompanying drawings, is described in further details the present invention.
The present invention is taking intelligent video monitoring as application background, given input monitoring video data, use an online human body detector to detect the pedestrian that in video, each frame occurs, and record pedestrian occur picture position, then, calculate off-line these pedestrians' that detect human body bottom appearance features and the high-rise attributive character of human body, store high-rise to pedestrian's picture position, human body bottom appearance features and human body attributive character into somatic data storehouse, this pedestrian is a human body detecting in somatic data storehouse.
Fig. 1 is the schematic diagram of the search method of a kind of human body of the present invention.Comprise: picture exists determining step (S10), the first searching step (S20) and the second searching step (S30).Wherein,
There is determining step (S10) in picture, judges whether to exist the picture of human body to be checked, enters the first searching step, otherwise enter the second searching step if existed.
The first searching step (S20), uses the human body bottom appearance features of human body to be checked to screen the human body in somatic data storehouse, generates the first result for retrieval; In the first result for retrieval, use the set of the high-rise attributive character of human body of human body to be checked, the human body in the first result for retrieval is screened, generate final result for retrieval, retrieval finishes.Comprise:
The first result for retrieval generates sub-step (S201), and the first similarity of the human body in calculating somatic data storehouse and the human body bottom appearance features of human body to be checked, screens the human body in somatic data storehouse according to the first similarity, generates the first result for retrieval;
The second result for retrieval generates sub-step (S202), calculate the second similarity of the high-rise attributive character of every kind of human body of human body in the first result for retrieval and human body to be checked, according to every kind of second similarity, the human body in the first result for retrieval is screened, generate the second result for retrieval;
Final result for retrieval generates sub-step (S203), calculate first total attributive distance of the set of the high-rise attributive character of human body of human body in the second result for retrieval and human body to be checked, according to first total attributive distance, the human body in the second result for retrieval is screened, generate the 3rd result for retrieval; Use the first similarity of the human body in the 3rd result for retrieval and the second similarity of the high-rise attributive character of whole human bodies, calculate total similarity of the human body in the 3rd result for retrieval, according to total similarity, the human body in the 3rd result for retrieval is screened, generate final result for retrieval; Wherein, first total attributive distance is: the high-rise attributive character of whole human bodies of the human body in the second result for retrieval respectively with the absolute value of the difference of the high-rise attributive character of human body of human body to be checked and mean value.
The second searching step (S30), deterministic retrieval mode, is used the high-rise attributive character of human body of human body to be checked to screen the human body in somatic data storehouse according to retrieval mode, generates the final result for retrieval in somatic data storehouse, and retrieval finishes.
According to the value type deterministic retrieval mode of the high-rise attributive character of the human body of human body to be checked, use the high-rise attributive character of human body of human body to be checked to screen the human body in somatic data storehouse according to retrieval mode, generate the final result for retrieval in somatic data storehouse.
According to the value type deterministic retrieval mode of the high-rise attributive character of the human body of human body to be checked be further: in the time that value type is two-value output, for the high-rise attributive character of every kind of human body, human body in somatic data storehouse is set up to inverted index, adopt the result of one or more inverted indexs as final result for retrieval; In the time that value type is probability output, second total attributive distance of the set of the high-rise attributive character of human body of the human body in calculating somatic data storehouse and human body to be checked, according to second total attributive distance, the human body in somatic data storehouse is screened, generate final result for retrieval; Wherein, second total attributive distance is: the high-rise attributive character of whole human bodies of the human body in somatic data storehouse respectively with the absolute value of the difference of the high-rise attributive character of human body of human body to be checked and mean value.Wherein, use the human body attributive classification device of off-line training to extract the high-rise attributive character of human body.
Below in conjunction with Fig. 3, Fig. 4, Fig. 5, Fig. 6, Fig. 7, Fig. 8 and Fig. 9, the present invention is described in detail.
There is determining step (S10) in picture, judges whether to exist the picture of human body to be checked, enters the first searching step, otherwise enter the second searching step if existed.
Generate in sub-step (S201) at the first result for retrieval of the first searching step (S20), human body bottom appearance features represents with color histogram, specifically, for the human body in human body to be checked and somatic data storehouse, represent with a human detection frame respectively, the color value of the each pixel in its corresponding human detection frame is transformed into HSV space; Again human detection frame is carried out to piecemeal processing.Piecemeal processing, namely human detection frame is on average divided into the gridiron pattern piece of m*n, be preferably the gridiron pattern piece of 6*6, calculate again each corresponding hsv color histogram, wherein, each Color Channel (H Color Channel, S Color Channel or V Color Channel) is quantified as p discrete magnitude, is preferably 8 discrete magnitudes.Color histogram has recorded in lattice piece each Color Channel in the frequency of occurrences (being the quantity that is distributed in the pixel of each discrete magnitude of each Color Channel in lattice piece) of its each discrete magnitude.Wherein, m, n and p are natural number.
Illustrate the histogrammic method of lattice piece hsv color that generates.The image of one frame 720*576 is divided into the gridiron pattern piece of 6*6, the size of so each lattice piece is 120*96, and each lattice piece has 11520 pixels.(i, j) lattice piece of image is handled as follows to calculate to the color histogram of every kind of Color Channel, wherein, i, j are the natural number in 1 to 6:
Step 1, H, the S of each pixel and the pixel value of V Color Channel of acquisition lattice piece;
Step 2, is converted into the pixel value of every kind of Color Channel of pixel the discrete magnitude of corresponding Color Channel;
Step 3: for every kind of Color Channel, add up respectively the quantity of pixel corresponding to each discrete magnitude in 8 discrete magnitudes, the quantity of pixel corresponding to 8 discrete magnitudes that every kind of Color Channel is corresponding forms the color histogram of this kind of Color Channel, is 8 tuples.For example: the color histogram of the H Color Channel of (i, j) lattice piece of the human detection frame of human body to be checked is (X
h1, X
h2, X
h3, X
h4, X
h5, X
h6, X
h7, X
h8), the color histogram of the H Color Channel of (i, j) lattice piece of the human detection frame in somatic data storehouse is (Y
h1, Y
h2, Y
h3, Y
h4, Y
h5, Y
h6, Y
h7, Y
h8), the rest may be inferred for the color histogram of the color histogram of S Color Channel and V Color Channel.
The specific implementation process that generates sub-step (S201) at the first result for retrieval of the first searching step (S20) is as follows: according to the human body bottom appearance features of the human body in human body to be checked and somatic data storehouse, service range metric function (for example: Euclidean distance or complementary chord angle distance) is calculated the first similarity between the human body bottom appearance features (color histogram) of the human body in human body to be checked and somatic data storehouse, and from big to small the human body in somatic data storehouse is sorted according to the first similarity, generate first testing result in somatic data storehouse.
Taking the Euclidean distance of distance metric function as example, illustrate that service range metric function calculates the process of the first similarity of the human body bottom appearance features (color histogram) of the human body in human body to be checked and somatic data storehouse.Wherein, the color histogram of the H Color Channel of (i, j) lattice piece of the human detection frame of human body to be checked is (X
h1, X
h2, X
h3, X
h4, X
h5, X
h6, X
h7, X
h8), the color histogram of the S Color Channel of (i, j) lattice piece of human detection frame is (X
s1, X
s2, X
s3, X
s4, X
s5, X
s6, X
s7, X
s8), the color histogram of the V Color Channel of (i, j) lattice piece of human detection frame is (X
v1, X
v2, X
v3, X
v4, X
v5, X
v6, X
v7, X
v8); The color histogram of the H Color Channel of (i, j) lattice piece of the human detection frame of the human body in somatic data storehouse is (Y
h1, Y
h2, Y
h3, Y
h4, Y
h5, Y
h6, Y
h7, Y
h8), the color histogram of the S Color Channel of (i, j) lattice piece of human detection frame is (Y
s1, Y
s2, Y
s3, Y
s4, Y
s5, Y
s6, Y
s7, Y
s8), the color histogram of the V Color Channel of (i, j) lattice piece of human detection frame is (Y
v1, Y
v2, Y
v3, Y
v4, Y
v5, Y
v6, Y
v7, Y
v8).
So, the Euclidean distance of 3 kinds of Color Channels of (i, j) lattice piece of the human detection frame of the human body in human body to be checked and somatic data storehouse is:
The distance d of H Color Channel
lh(formula 1):
The distance d of S Color Channel
ls(formula 2):
The distance d of V Color Channel
lv(formula 3):
Total distance B of the human body bottom appearance features of the human detection frame of the human body in human body to be checked and somatic data storehouse
l(formula 4):
Wherein, w
lh, w
lsand w
lvbe respectively d
lh, d
lsand d
lvweight, is not less than 0 real number, and w
lh+ w
ls+ w
lv=1; I is the natural number between 1 to m*n.
The first similarity S between the human body bottom appearance features of the human body in human body to be checked and somatic data storehouse
1for (formula 5):
S
1=1/D
L。
According to S
1human body in somatic data storehouse is sorted, generate first result for retrieval in somatic data storehouse.Wherein, sortord is for from big to small.
Fig. 3 utilizes human body bottom appearance features to retrieve the result obtaining, i.e. the first result for retrieval.
Generate in sub-step (S202) at the second result for retrieval of the first searching step (S20), for the extraction of the high-rise attributive character of human body, the conduct oneself classification of height layer attributive character of the human body attributive classification device of multiple off-line trainings for the present invention.Wherein, for all probable values of the high-rise attributive character of every kind of human body, train a sorter, the positive example of training data comes from the high-rise attributive character of human body and is labeled as positive above-mentioned human detection frame, negative example comes from the high-rise attributive character of human body and is labeled as the above-mentioned human detection frame of other values, and the input feature vector of human body attributive classification device is the gradient orientation histogram (HOG) of human detection frame.In one embodiment of the invention, using sex, age, upper lower part of the body garment, knapsack type as the high-rise attributive character of human body, adopt SVM(Support Vector Machine) sorter is as human body attributive classification device.(the i of image, j) computing method of the gradient orientation histogram of lattice piece (HOG) and the computing method of color histogram are similar, difference is: in the step 1 of the computing method of above-mentioned color histogram, obtain the gradient direction value of each pixel of lattice piece, then gradient direction value is carried out to follow-up step 2 and 3 operation.Gradient orientation histogram (HOG) as the human detection frame of the input feature vector of human body attributive classification device is to generate according to the gradient orientation histogram of m*n lattice piece (HOG).
The specific implementation process that generates sub-step (S202) at the second result for retrieval of the first searching step (S20) is as follows: according to the high-rise attributive character of the human body of the human body in human body to be checked and the first result for retrieval (the high-rise attributive character of human body that uses human body attributive classification device to calculate), service range metric function is calculated the second similarity of the high-rise attributive character of every kind of human body of the human body in human body to be checked and the first result for retrieval; The second similarity that filters out the high-rise attributive character of human body is less than the human body in first result for retrieval of first threshold of the high-rise attributive character of human body, generate second result for retrieval in somatic data storehouse (in experiment of the present invention, in the first testing result, have in the human body of different human body bottom appearance features with human body to be checked, more than 50% human body can be used as flase drop and filter out).
Taking the Euclidean distance of distance metric function as example, the process of calculating the high-rise attributive character of human body of the human body in the first result for retrieval and the second similarity of human body to be checked is described.Calculate the high-rise attributive character of human body apart from d
h, d
hthe absolute value (formula 6) of=(Output rusults of the human body attributive classification device of the human body in Output rusults-the first result for retrieval of the human body attributive classification device of human body to be checked).According to apart from d
hcalculate the second similarity s of the high-rise attributive character of human body of the human body in the first result for retrieval and the high-rise attributive character of human body of human body to be checked
h, s
h=1/d
h(formula 7).If the second similarity of the high-rise attributive character of human body is less than the first threshold of the high-rise attributive character of human body, it is filtered out from the first result for retrieval, generate second result for retrieval in somatic data storehouse.
The specific implementation process that generates sub-step (S203) at the final result for retrieval of the first searching step (S20) is as follows: the human body in calculating the second result for retrieval and first total attributive distance between human body to be checked; Filter out the human body in the second result for retrieval that first total attributive distance is greater than Second Threshold, generate the 3rd result for retrieval in somatic data storehouse; Use the first similarity of the human body in each the 3rd result for retrieval and the second similarity of the high-rise attributive character of whole described human bodies to calculate total similarity, according to total similarity, the human body in the 3rd result for retrieval is sorted and obtains final result for retrieval.
First, the human body in calculating the second result for retrieval and first total attributive distance between human body to be checked.First total attributive distance of the human body in the second result for retrieval and human body to be checked is: all the high-rise attributive character of human body is apart from d
hand mean value.
Then, filter out the human body in the second result for retrieval that first total attributive distance is greater than Second Threshold, generate the 3rd result for retrieval.
Then, the human body in calculating the 3rd result for retrieval and total similarity S of human body to be checked
2.Computing method are: the second similarity of the first similarity of human body bottom appearance features and whole high-rise attributive character of human bodies and linear weighted function and,
(formula 8), wherein q is the species number of people's height layer attributive character, w
1and w
2be respectively the first similarity and the second similarity and weight, w
1+ w
2=1, w
1and w
2for being not less than 0 real number, q is natural number.In one embodiment of this invention, w
1and w
2equate to be 0.5.
Finally, according to total similarity S
2human body in the 3rd result for retrieval is sorted and obtains final result for retrieval.Wherein, sortord is for from big to small.
Fig. 4, Fig. 5, Fig. 6 and Fig. 7 progressively increase the result that the high-rise attributive character of human body is filtered.Wherein, the high-rise attributive character of human body of the human body to be checked of Fig. 4 is " male sex ", the high-rise attributive character of human body of the human body to be checked of Fig. 5 is " male sex " and " cotta ", the high-rise attributive character of human body of the human body to be checked of Fig. 6 is " male sex ", " cotta " and " trousers ", and the high-rise attributive character of human body of the human body to be checked of Fig. 7 is " male sex ", " cotta ", " trousers " and " knapsack ".
The specific implementation process of the second searching step (S30) is as follows: for the situation that there is no human body picture to be checked, need the high-rise attributive character of human body of given human body to be checked, adopt different retrieval modes according to the value type of the high-rise attributive character of human body, if value type is two-value output, according to the high-rise attributive character of every kind of human body, for the human body in somatic data storehouse is set up inverted index, and adopt the result of inverted index of the high-rise attributive character of one or more human bodies as final result for retrieval; If value type is probability output, the human body image in somatic data storehouse is carried out to second total attributive distance and calculate, and from small to large the human body in somatic data storehouse is sorted according to second total attributive distance.
Wherein, the extracting method of the high-rise attributive character of the human body of the human body in somatic data storehouse generates the identical of sub-step (S202) with the second result for retrieval at the first searching step (S20).
Second total attributive distance of the human body in somatic data storehouse and human body to be checked is: the high-rise attributive character of whole human bodies of the human body in somatic data storehouse respectively with the absolute value of the difference of the high-rise attributive character of corresponding human body of human body to be checked and mean value.
Fig. 8 and Fig. 9 are the result of after the high-rise attributive character of the human body of given human body to be checked, human body in somatic data storehouse being filtered.Wherein, the high-rise attributive character of the human body of the human body to be checked of Fig. 8 is " red upper garment " and " black trousers ", and the high-rise attributive character of human body of the human body to be checked of Fig. 9 is " white upper garment ", " black trousers " and " women ".
Fig. 2 is the schematic diagram of the searching system of a kind of human body of the present invention.Comprise: picture exists judge module (M1), the first retrieval module (M2) and the second retrieval module (M3).Wherein,
There is judge module (M1) in picture, for judging whether to exist the picture of human body to be checked, enters the first retrieval module (M2), otherwise enter the second retrieval module (M3) if existed.
The first retrieval module (M2), screens the human body in somatic data storehouse for the human body bottom appearance features that uses human body to be checked, generates the first result for retrieval; In the first result for retrieval, use the set of the high-rise attributive character of human body of human body to be checked, the human body in the first result for retrieval is screened, generate final result for retrieval, retrieval finishes.Comprise:
The first result for retrieval generates submodule (M21), for calculating the first similarity of the human body in somatic data storehouse and the human body bottom appearance features of human body to be checked, according to the first similarity, the human body in somatic data storehouse is screened, generate the first result for retrieval;
The second result for retrieval generates submodule (M22), for calculating the second similarity of the human body of the first result for retrieval and the high-rise attributive character of every kind of human body of human body to be checked, according to every kind of second similarity, the human body in the first result for retrieval is screened, generate the second result for retrieval;
Final result for retrieval generates submodule (M23), be used for first total attributive distance of the set of calculating the human body of the second result for retrieval and the high-rise attributive character of the human body of human body to be checked, according to first total attributive distance, the human body in the second result for retrieval is screened, generate the 3rd result for retrieval; Use the first similarity of the human body in the 3rd result for retrieval and the second similarity of the high-rise attributive character of whole human bodies, calculate total similarity of the human body in the 3rd result for retrieval, according to total similarity, the human body in the 3rd result for retrieval is screened, generate final result for retrieval; Wherein, first total attributive distance is: the high-rise attributive character of whole human bodies of the human body in the second result for retrieval respectively with the absolute value of the difference of the high-rise attributive character of human body of human body to be checked and mean value.
The second retrieval module (M3), for deterministic retrieval mode, uses the high-rise attributive character of human body of human body to be checked to screen the human body in somatic data storehouse according to retrieval mode, generates the final result for retrieval in somatic data storehouse, and retrieval finishes.
According to the value type deterministic retrieval mode of the high-rise attributive character of the human body of human body to be checked, use the high-rise attributive character of human body of human body to be checked to screen the human body in somatic data storehouse according to retrieval mode, generate the final result for retrieval in somatic data storehouse.
According to the value type deterministic retrieval mode of the high-rise attributive character of the human body of human body to be checked be further: in the time that value type is two-value output, for the high-rise attributive character of every kind of human body, human body in somatic data storehouse is set up to inverted index, adopt the result of one or more inverted indexs as final result for retrieval; In the time that value type is probability output, second total attributive distance of the set of the high-rise attributive character of human body of the human body in calculating somatic data storehouse and human body to be checked, according to second total attributive distance, the human body in somatic data storehouse is screened, generate final result for retrieval; Wherein, second total attributive distance is: the high-rise attributive character of whole human bodies of the human body in somatic data storehouse respectively with the absolute value of the difference of the high-rise attributive character of human body of human body to be checked and mean value.
Wherein, use the human body attributive classification device of off-line training to extract the high-rise attributive character of human body.
The above, it is only preferred embodiment of the present invention, not the present invention is done to any pro forma restriction, any person of ordinary skill in the field, if not departing from the scope of technical characterictic proposed by the invention, utilize technology contents disclosed in this invention to do the local equivalent embodiment that changes or revise, and do not depart from technical characterictic content of the present invention, all still belong in the scope of the technology of the present invention feature.
Claims (10)
1. a search method for human body, is characterized in that, comprising:
There is determining step in picture, judges whether to exist the picture of human body to be checked, enters the first searching step, otherwise enter the second searching step if existed;
The first searching step, uses the human body bottom appearance features of described human body to be checked to screen the human body in somatic data storehouse, generates the first result for retrieval; In described the first result for retrieval, use the set of the high-rise attributive character of human body of described human body to be checked, the human body in described the first result for retrieval is screened, generate final result for retrieval, retrieval finishes;
The second searching step, deterministic retrieval mode, is used the high-rise attributive character of described human body of described human body to be checked to screen the human body in somatic data storehouse according to described retrieval mode, generates the final result for retrieval in described somatic data storehouse, and retrieval finishes.
2. the search method of human body as claimed in claim 1, is characterized in that, described the first searching step is further:
The first result for retrieval generates sub-step, calculate the first similarity of the described human body bottom appearance features of human body in described somatic data storehouse and described human body to be checked, according to described the first similarity, the human body in described somatic data storehouse is screened, generate described the first result for retrieval;
The second result for retrieval generates sub-step, calculate the second similarity of the high-rise attributive character of every kind of human body of described human body in described the first result for retrieval and described human body to be checked, according to the second similarity every kind described, human body in described the first result for retrieval is screened, generate described the second result for retrieval;
Final result for retrieval generates sub-step, calculate first total attributive distance of the set of the high-rise attributive character of described human body of described human body in described the second result for retrieval and described human body to be checked, according to described first total attributive distance, the human body in described the second result for retrieval is screened, generate the 3rd result for retrieval; Use described first similarity of the described human body in described the 3rd result for retrieval and the second similarity of the high-rise attributive character of whole described human bodies, calculate total similarity of the described human body in described the 3rd result for retrieval, according to described total similarity, the human body in described the 3rd result for retrieval is screened, generate described final result for retrieval; Wherein, described first total attributive distance is: the high-rise attributive character of whole described human body of the described human body in described the second result for retrieval respectively with the absolute value of the difference of the high-rise attributive character of described human body of described human body to be checked and mean value.
3. the search method of human body as claimed in claim 1, is characterized in that, described the second searching step is further:
According to the value type deterministic retrieval mode of the high-rise attributive character of described human body of described human body to be checked, use the high-rise attributive character of described human body of described human body to be checked to screen the human body in somatic data storehouse according to described retrieval mode, generate the final result for retrieval in described somatic data storehouse.
4. the search method of human body as claimed in claim 3, it is characterized in that, the described value type deterministic retrieval mode according to the high-rise attributive character of described human body of human body to be checked is further: in the time that described value type is two-value output, for the high-rise attributive character of human body every kind described, described human body in described somatic data storehouse is set up to inverted index, adopt the result of one or more described inverted indexs as described final result for retrieval; In the time that described value type is probability output, calculate second total attributive distance of the set of the high-rise attributive character of described human body of described human body in described somatic data storehouse and described human body to be checked, according to described second total attributive distance, the described human body in described somatic data storehouse is screened, generate described final result for retrieval; Wherein, described second total attributive distance is: the high-rise attributive character of whole described human body of the described human body in described somatic data storehouse respectively with the absolute value of the difference of the high-rise attributive character of described human body of described human body to be checked and mean value.
5. the search method of human body as claimed in claim 1, is characterized in that, uses the human body attributive classification device of off-line training to extract the high-rise attributive character of described human body.
6. a searching system for human body, is characterized in that, comprising:
There is judge module in picture, for judging whether to exist the picture of human body to be checked, enters the first retrieval module, otherwise enter the second retrieval module if existed;
The first retrieval module, screens the human body in somatic data storehouse for the human body bottom appearance features that uses described human body to be checked, generates the first result for retrieval; In described the first result for retrieval, use the set of the high-rise attributive character of human body of described human body to be checked, the human body in described the first result for retrieval is screened, generate final result for retrieval, retrieval finishes;
The second retrieval module, for deterministic retrieval mode, uses the high-rise attributive character of described human body of described human body to be checked to screen the human body in somatic data storehouse according to described retrieval mode, generates the final result for retrieval in described somatic data storehouse, and retrieval finishes.
7. the searching system of human body as claimed in claim 6, is characterized in that, described the first retrieval module is further:
The first result for retrieval generates submodule, for calculating the first similarity of the human body in described somatic data storehouse and the described human body bottom appearance features of described human body to be checked, according to described the first similarity, the human body in described somatic data storehouse is screened, generate described the first result for retrieval;
The second result for retrieval generates submodule, for calculating the second similarity of the described human body of described the first result for retrieval and the high-rise attributive character of every kind of human body of described human body to be checked, according to the second similarity every kind described, human body in described the first result for retrieval is screened, generate described the second result for retrieval;
Final result for retrieval generates submodule, be used for first total attributive distance of the set of calculating the described human body of described the second result for retrieval and the high-rise attributive character of described human body of described human body to be checked, according to described total attributive distance, the human body in described the second result for retrieval is screened, generate the 3rd result for retrieval; Use described first similarity of the described human body in described the 3rd result for retrieval and the second similarity of the high-rise attributive character of whole described human bodies, calculate total similarity of the described human body in described the 3rd result for retrieval, according to described total similarity, the human body in described the 3rd result for retrieval is screened, generate described final result for retrieval; Wherein, described first total attributive distance is: the high-rise attributive character of whole described human body of the described human body in described the second result for retrieval respectively with the absolute value of the difference of the high-rise attributive character of described human body of described human body to be checked and mean value.
8. the searching system of human body as claimed in claim 6, is characterized in that, described the second retrieval module is further:
According to the value type deterministic retrieval mode of the high-rise attributive character of described human body of described human body to be checked, use the high-rise attributive character of described human body of described human body to be checked to screen the human body in somatic data storehouse according to described retrieval mode, generate the final result for retrieval in described somatic data storehouse.
9. the searching system of human body as claimed in claim 8, it is characterized in that, the described value type deterministic retrieval mode according to the high-rise attributive character of described human body of human body to be checked is further: in the time that described value type is two-value output, for the high-rise attributive character of human body every kind described, described human body in described somatic data storehouse is set up to inverted index, adopt the result of one or more described inverted indexs as described final result for retrieval; In the time that described value type is probability output, calculate second total attributive distance of the set of the high-rise attributive character of described human body of described human body in described somatic data storehouse and described human body to be checked, according to described second total attributive distance, the described human body in described somatic data storehouse is screened, generate described final result for retrieval; Wherein, described second total attributive distance is: the high-rise attributive character of whole described human body of the described human body in described somatic data storehouse respectively with the absolute value of the difference of the high-rise attributive character of described human body of described human body to be checked and mean value.
10. the search method of human body as claimed in claim 6, is characterized in that, uses the human body attributive classification device of off-line training to extract the high-rise attributive character of described human body.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201310035131.9A CN103970771B (en) | 2013-01-29 | 2013-01-29 | The search method of a kind of human body and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201310035131.9A CN103970771B (en) | 2013-01-29 | 2013-01-29 | The search method of a kind of human body and system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN103970771A true CN103970771A (en) | 2014-08-06 |
CN103970771B CN103970771B (en) | 2016-12-28 |
Family
ID=51240283
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201310035131.9A Active CN103970771B (en) | 2013-01-29 | 2013-01-29 | The search method of a kind of human body and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN103970771B (en) |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2016134666A1 (en) * | 2015-02-28 | 2016-09-01 | 华为技术有限公司 | Method and device for discovering friend through image matching |
CN106033549A (en) * | 2015-03-16 | 2016-10-19 | 北京大学 | Reordering method in vehicle retrieval and apparatus thereof |
CN106033443A (en) * | 2015-03-16 | 2016-10-19 | 北京大学 | Method and device for expansion query in vehicle retrieval |
CN106528568A (en) * | 2015-09-11 | 2017-03-22 | 中兴通讯股份有限公司 | Method and system for information supply |
WO2018170695A1 (en) * | 2017-03-20 | 2018-09-27 | 华为技术有限公司 | Method and device for recognizing descriptive attributes of appearance feature |
CN108960412A (en) * | 2018-06-29 | 2018-12-07 | 北京京东尚科信息技术有限公司 | Image-recognizing method, device and computer readable storage medium |
CN109101922A (en) * | 2018-08-10 | 2018-12-28 | 广东电网有限责任公司 | Operating personnel device, assay, device and electronic equipment |
CN110163106A (en) * | 2019-04-19 | 2019-08-23 | 中国科学院计算技术研究所 | Integral type is tatooed detection and recognition methods and system |
CN111078993A (en) * | 2019-09-24 | 2020-04-28 | 上海依图网络科技有限公司 | Method and system for improving retrieval recall rate through extended query |
CN111126102A (en) * | 2018-10-30 | 2020-05-08 | 富士通株式会社 | Personnel searching method and device and image processing equipment |
CN112148924A (en) * | 2019-06-28 | 2020-12-29 | 杭州海康威视数字技术股份有限公司 | Luggage case retrieval method and device and electronic equipment |
CN114202710A (en) * | 2020-08-26 | 2022-03-18 | 中国电信股份有限公司 | Target identification method, device and system |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20090316983A1 (en) * | 2008-06-20 | 2009-12-24 | Feng Han | Real-Time Action Detection and Classification |
CN101650728A (en) * | 2009-08-26 | 2010-02-17 | 北京邮电大学 | Video high-level characteristic retrieval system and realization thereof |
-
2013
- 2013-01-29 CN CN201310035131.9A patent/CN103970771B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20090316983A1 (en) * | 2008-06-20 | 2009-12-24 | Feng Han | Real-Time Action Detection and Classification |
CN101650728A (en) * | 2009-08-26 | 2010-02-17 | 北京邮电大学 | Video high-level characteristic retrieval system and realization thereof |
Non-Patent Citations (2)
Title |
---|
***: "基于时空分析和多粒度特征表示的人体检测方", 《中国博士学位论文全文数据库 信息科技辑》 * |
李浩: "跨摄像机的人体重现检测", 《NSTL国家科技图书文献中心》 * |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2016134666A1 (en) * | 2015-02-28 | 2016-09-01 | 华为技术有限公司 | Method and device for discovering friend through image matching |
CN106033549A (en) * | 2015-03-16 | 2016-10-19 | 北京大学 | Reordering method in vehicle retrieval and apparatus thereof |
CN106033443A (en) * | 2015-03-16 | 2016-10-19 | 北京大学 | Method and device for expansion query in vehicle retrieval |
CN106033443B (en) * | 2015-03-16 | 2019-05-07 | 北京大学 | A kind of expanding query method and device in vehicle retrieval |
CN106033549B (en) * | 2015-03-16 | 2019-05-07 | 北京大学 | Method for reordering and device in a kind of vehicle retrieval |
CN106528568A (en) * | 2015-09-11 | 2017-03-22 | 中兴通讯股份有限公司 | Method and system for information supply |
US11410411B2 (en) | 2017-03-20 | 2022-08-09 | Huawei Technologies Co., Ltd. | Method and apparatus for recognizing descriptive attribute of appearance feature |
WO2018170695A1 (en) * | 2017-03-20 | 2018-09-27 | 华为技术有限公司 | Method and device for recognizing descriptive attributes of appearance feature |
CN108960412A (en) * | 2018-06-29 | 2018-12-07 | 北京京东尚科信息技术有限公司 | Image-recognizing method, device and computer readable storage medium |
CN109101922A (en) * | 2018-08-10 | 2018-12-28 | 广东电网有限责任公司 | Operating personnel device, assay, device and electronic equipment |
CN111126102A (en) * | 2018-10-30 | 2020-05-08 | 富士通株式会社 | Personnel searching method and device and image processing equipment |
CN110163106A (en) * | 2019-04-19 | 2019-08-23 | 中国科学院计算技术研究所 | Integral type is tatooed detection and recognition methods and system |
CN112148924A (en) * | 2019-06-28 | 2020-12-29 | 杭州海康威视数字技术股份有限公司 | Luggage case retrieval method and device and electronic equipment |
CN111078993A (en) * | 2019-09-24 | 2020-04-28 | 上海依图网络科技有限公司 | Method and system for improving retrieval recall rate through extended query |
CN114202710A (en) * | 2020-08-26 | 2022-03-18 | 中国电信股份有限公司 | Target identification method, device and system |
Also Published As
Publication number | Publication date |
---|---|
CN103970771B (en) | 2016-12-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN103970771A (en) | Search method and system for human body | |
Toulouse et al. | Computer vision for wildfire research: An evolving image dataset for processing and analysis | |
Nam et al. | Pest detection on traps using deep convolutional neural networks | |
CN108520226B (en) | Pedestrian re-identification method based on body decomposition and significance detection | |
CN102982313B (en) | The method of Smoke Detection | |
CN107301380A (en) | One kind is used for pedestrian in video monitoring scene and knows method for distinguishing again | |
CN107315795B (en) | The instance of video search method and system of joint particular persons and scene | |
CN109359625A (en) | The method and system of customer identification is judged based on head and shoulder detection and face recognition technology | |
CN104392232B (en) | The bearing calibration of video metadata under a kind of traffic scene | |
CN110852222A (en) | Campus corridor scene intelligent monitoring method based on target detection | |
CN109190475A (en) | A kind of recognition of face network and pedestrian identify network cooperating training method again | |
CN108960124A (en) | The image processing method and device identified again for pedestrian | |
TW201140502A (en) | System and method for monitoring objects | |
CN109460787A (en) | IDS Framework method for building up, device and data processing equipment | |
CN106295585A (en) | A kind of filtration system of selection taking into account real-time and face quality and system | |
TW201140470A (en) | System and method for monitoring objects and key persons of the objects | |
CN111539257B (en) | Person re-identification method, device and storage medium | |
KR101092472B1 (en) | Video indexing system using surveillance camera and the method thereof | |
CN110245592A (en) | A method of for promoting pedestrian's weight discrimination of monitoring scene | |
CN102156881A (en) | Method for detecting salvage target based on multi-scale image phase information | |
KR101113515B1 (en) | Video index system using surveillance camera and the method thereof | |
CN111462310A (en) | Bolt defect space positioning method based on multi-view geometry | |
Jangade et al. | Study on deep learning models for human pose estimation and its real time application | |
CN107123126A (en) | A kind of stream of people's moving scene temperature method of estimation | |
CN102244769A (en) | Object and key person monitoring system and method thereof |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C14 | Grant of patent or utility model | ||
GR01 | Patent grant |