CN109948455B - Detection method and device for left-behind object - Google Patents
Detection method and device for left-behind object Download PDFInfo
- Publication number
- CN109948455B CN109948455B CN201910140397.7A CN201910140397A CN109948455B CN 109948455 B CN109948455 B CN 109948455B CN 201910140397 A CN201910140397 A CN 201910140397A CN 109948455 B CN109948455 B CN 109948455B
- Authority
- CN
- China
- Prior art keywords
- target object
- pixel
- target
- value
- contour
- 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.)
- Active
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a method and a device for detecting a left object, which are characterized in that firstly, according to the real-time rapidity requirement of the detection of the left object, the motion characteristic of a pixel is judged by adopting a frame difference method, and a target object which changes from dynamic to static, namely a suspected left object, is found; then, HOG features are combined with SVM for classification, people and articles are distinguished, and people are filtered; judging whether the stay time of the target object exceeds a preset time by adopting an evidence accumulation mode; and finally, performing feature matching on the people near the target object and the owner of the target object, and if the people are not matched, determining that the target object is a left-behind object. On one hand, the method has the characteristics of low calculation complexity and real-time and high efficiency, and on the other hand, the method solves the problems of false alarm and shielding in the detection process of the left-over object, also solves the problem of distinguishing whether the left-over object is a person or an object, and also solves the problem of distinguishing whether the left-over object is really without a master, so that the detection accuracy is also high, and the detection problem of the left-over object is effectively solved.
Description
Technical Field
The embodiment of the invention relates to a video image processing technology, in particular to a method and a device for detecting a left-behind object.
Background
In a place with dense people flow, such as a transportation hub like a subway, objects are sometimes left behind or objects are called to stay abnormally, such as packages without owners. The left-over articles may be ordinary articles left by the owner or dangerous articles such as bombs, so that the real-time monitoring of occasions such as transportation hubs becomes very important for timely finding the left-over articles. In addition, detection of carry-over objects has other widespread applications, such as dangerous obstacles on railways, dropped rocks, litter, pieces of material on the main/highway/runway, slippery liquid left on the ground (which may lead to accidents such as pedestrian falls, car accidents, etc.), and the like.
In the prior art, the traditional video monitoring only collects and stores videos, information in the videos cannot be intelligently and automatically judged, and the discovery of the left articles is realized by naked eyes of video monitors, so that the efficiency is low. In this case, intelligent monitoring arises. The intelligent monitoring means that the image and video are processed by means of a computer technology to replace manual work to analyze and understand real-time video information, so that monitoring personnel are assisted to further respond and process the video information on the basis of the image and video, and the purpose of reducing or even replacing the work of the monitoring personnel is achieved.
However, the detection of the left-behind object is a particularly complex problem because the environment in which the detection is performed is usually complex, and there are many shadows in public places such as traffic junctions. There are currently some schemes for legacy object detection, such as: the problem of occlusion easily existing in an intelligent video monitoring scene can be solved by fusing multi-camera information, but an accurate camera model is often difficult to obtain, so that the effect of the algorithm in practical application is not ideal; the mixed gaussians can be used for background modeling, and on the basis, the state migration of each gaussians distribution is observed to detect the target which becomes static after moving into a scene, however, the mixed gaussians background modeling needs large training amount and calculation amount, and the real-time requirement of the monitoring system is difficult to meet.
The inventor finds that the solutions in the prior art are not accurate enough or large in calculation amount, so that the problems of shielding and false alarm cannot be solved, the problems of abnormal staying of people or objects cannot be judged whether people or the owners of people nearby the left-over objects are not judged, and the problems are not complete, and therefore the problem of detecting the left-over objects cannot be effectively solved in the prior art.
Disclosure of Invention
The embodiment of the invention provides a method and a device for detecting a left article, which are used for effectively solving the problem of detecting the left article.
According to a first aspect of embodiments of the present invention, there is provided an legacy object detection method, the method comprising:
for the collected video, determining whether a target object appears according to whether the value of a pixel at the same position of an adjacent frame changes, wherein the target object is an object which moves firstly and then is static;
classifying the found target object by combining HOG (histogram of oriented gradient) features with a Support Vector Machine (SVM) so as to eliminate the condition that the target object is a person;
for each found target object, judging whether the stay time of the target object after the movement and the stillness exceeds a time threshold value or not by accumulating the times of the target object appearing in each frame;
and if the stay time length exceeds a time length threshold value, matching the people near the target object with the owner of the target object, and if the stay time length does not exceed the time length threshold value, determining that the target object is a left-behind object, wherein the owner of the target object is the people closest to the target object and detected when the target object is calmed.
Optionally, determining whether the target object appears according to whether the value of the pixel at the same position in the adjacent frames changes includes:
when the variation amplitude of the value of the pixel exceeds a pixel threshold value, setting the attribute value corresponding to the pixel as a preset attribute value, wherein the attribute value represents the longest time for allowing the pixel to be continuously unchanged without being regarded as a background;
according to the change of the pixel value and the attribute value, executing the following strategies:
when the variation amplitude of the value of the pixel is smaller than the pixel threshold value and the attribute value is 0, the pixel is indicated as the current background and no operation is performed; when the variation amplitude of the pixel value is smaller than the pixel threshold value but the attribute value is larger than 0, the pixel position is indicated to have a motion foreground, but the motion is in a static state or a slow motion state at the moment, and the attribute value is subjected to self-subtraction operation every time; when the attribute value is reduced from 0, indicating that the pixel position has remained still from moving to static, it is determined that the target object is present.
Optionally, the method further includes an original context, and the policy further includes:
if It(x,y)-BCt(x,y)>T&&It(x,y)-BOt(x,y)>T&&Dt(x, y) is 0, indicating that a new stationary object is present, then order
BOt(x,y)＝BCt(x,y)
BCt(x,y)＝It(x,y)
Updating the current background and storing the original current background into the original background;
if It(x,y)-BCt(x,y)>T&&It(x,y)-BOt(x,y)<T, indicating that the old target leaves, then order
BCt(x,y)＝BOt(x,y)
Restoring the information stored in the original background to the current background;
if It(x,y)-BCt(x,y)<T&&It(x,y)-BOt(x,y)<T&&Dt(x, y) ═ 0, which indicates a clear background, let
BCt(x,y)＝a×It(x,y)+(1-a)×BCt(x,y)
BOt(x,y)＝a×It(x,y)+(1-a)×BOt(x,y)
Wherein the content of the first and second substances,&&is represented by the general formula It(x, y) represents the value of the pixel in the current frame, BCt(x, y) stands for Current background, BOt(x, y) represents the original background, Dt(x, y) represents the attribute value corresponding to the pixel, T is the pixel threshold value, and a is the preset learning rate.
Optionally, the method for classifying the found target object by using the histogram of oriented gradients HOG feature in combination with a support vector machine SVM includes:
extracting HOG characteristics of a positive sample and a negative sample in advance, wherein the positive sample is an image sample of a person, and the negative sample is an image sample of an object;
putting the extracted HOG features into an SVM classifier for training to obtain a model;
generating a filter using the model;
detecting the negative sample with the screener to obtain a difficult sample;
extracting HOG characteristics of the difficult samples and putting the HOG characteristics and the HOG characteristics of the positive and negative samples into training to obtain a final screener;
extracting HOG characteristics of the target object;
and classifying the human and object of the target object by using the pre-obtained final filter.
Optionally, for each found target object, determining whether the staying time of the target object after moving and standing exceeds a time threshold by accumulating the number of times that the target object appears in each frame includes:
matching the contour of a first target object with the contour of a second target object, wherein the first target object is a target object selected from a current target object set, and the second target object is a target object selected from an existing target object set;
if the contour of the first target object is matched with the contour of the second target object, adding 1 to the found position and found count value of the second target object; when the found position of the second target object is true, clearing a lost count value of the second target object by 0, and if the found count value reaches a time threshold, determining that the stay time of the second target object after the second target object is changed from dynamic to static exceeds the time threshold;
if the contour of the first target object is not matched with the contour of each second target object in the existing target object set, adding the first target object into the existing target object set as a new second target object;
if the contour of a second target object does not match the contour of each first target object in the current frame target object set, the found position of the second target object is false; and when the found position of the second target object is false, adding 1 to the loss count value, and if the loss count value meets a preset condition, removing the second target object from the existing target object set.
Optionally, before matching the contour of the first target object with the contour of the second target object, the method further includes:
performing morphological processing on the differential image to obtain the outline of each target object in the current frame target object set;
judging whether the number of pixels of the outline of each target object exceeds a pixel number threshold value or not;
and filtering out the target object with the number of the pixels of the contour not exceeding the threshold value of the number of the pixels.
Optionally, if the staying time length exceeds a time length threshold, matching the person near the target object with the owner of the target object, including:
when the target object changes from dynamic to static, all people existing in an interested area with the center of mass of the target object as the center are detected through the SVM;
calculating the distance between the mass center of each person and the mass center of the target object, and taking the closest person as a target person, wherein the target person is the host of the target object;
learning and storing characteristics of the target person;
and when the staying time of the target object from moving to static exceeds the time threshold, detecting all the people in the same region of interest by using the SVM again, and matching the detected people with the originally stored characteristics of the target people.
According to a second aspect of embodiments of the present invention, there is provided an legacy object detection apparatus, the apparatus including:
the target finding module is used for determining whether a target object appears according to whether the value of a pixel at the same position of an adjacent frame changes or not for the acquired video, wherein the target object is an object which moves firstly and then is static;
the target elimination module is used for classifying the found target object by adopting the HOG (histogram of oriented gradients) characteristic in combination with a Support Vector Machine (SVM) so as to eliminate the condition that the target object is a person;
the target stay time length judging module is used for judging whether the stay time length of the target object after the target object is changed from moving to static exceeds a time length threshold value or not by accumulating the times of the target object appearing in each frame for each found target object;
and the target subordinate judgment module is used for matching the people near the target object with the owner of the target object when the stay time exceeds a time threshold, and determining that the target object is a left-behind object if the stay time does not match the owner of the target object, wherein the owner of the target object is the person closest to the target object detected when the target object is changed from moving to static.
Optionally, when the target finding module determines whether the target object appears according to whether the value of the pixel at the same position of the adjacent frame changes, the target finding module is specifically configured to:
when the variation amplitude of the value of the pixel exceeds a pixel threshold value, setting the attribute value corresponding to the pixel as a preset attribute value, wherein the attribute value represents the longest time for allowing the pixel to be continuously unchanged without being regarded as a background; according to the change of the pixel value and the attribute value, executing the following strategies:
when the variation amplitude of the value of the pixel is smaller than the pixel threshold value and the attribute value is 0, the pixel is indicated as the current background and no operation is performed; when the variation amplitude of the pixel value is smaller than the pixel threshold value but the attribute value is larger than 0, the pixel position is indicated to have a motion foreground, but the motion is in a static state or a slow motion state at the moment, and the attribute value is subjected to self-subtraction operation every time; when the attribute value is reduced from 0, indicating that the pixel position has remained still from moving to static, it is determined that the target object is present.
Optionally, the method further includes an original context, and the policy further includes:
if It(x,y)-BCt(x,y)>T&&It(x,y)-BOt(x,y)>T&&Dt(x, y) is 0, indicating that a new stationary object is present, then order
BOt(x,y)＝BCt(x,y)
BCt(x,y)＝It(x,y)
Updating the current background and storing the original current background into the original background;
if It(x,y)-BCt(x,y)>T&&It(x,y)-BOt(x,y)<T, indicating that the old target leaves, then order
BCt(x,y)＝BOt(x,y)
Restoring the information stored in the original background to the current background;
if It(x,y)-BCt(x,y)<T&&It(x,y)-BOt(x,y)<T&&Dt(x,y)If the background is clear, the order is 0
BCt(x,y)＝a×It(x,y)+(1-a)×BCt(x,y)
BOt(x,y)＝a×It(x,y)+(1-a)×BOt(x,y)
Wherein the content of the first and second substances,&&is represented by the general formula It(x, y) represents the value of the pixel in the current frame, BCt(x, y) stands for Current background, BOt(x, y) represents the original background, Dt(x, y) represents the attribute value corresponding to the pixel, T is the pixel threshold value, and a is the preset learning rate.
Optionally, the target elimination module is specifically configured to:
extracting HOG characteristics of a positive sample and a negative sample in advance, wherein the positive sample is an image sample of a person, and the negative sample is an image sample of an object; putting the extracted HOG features into an SVM classifier for training to obtain a model; generating a filter using the model; detecting the negative sample with the screener to obtain a difficult sample; extracting HOG characteristics of the difficult samples and putting the HOG characteristics and the HOG characteristics of the positive and negative samples into training to obtain a final screener; extracting HOG characteristics of the target object; and classifying the human and object of the target object by using the pre-obtained final filter.
Optionally, the target staying time length determining module is specifically configured to:
matching the contour of a first target object with the contour of a second target object, wherein the first target object is a target object selected from a current target object set, and the second target object is a target object selected from an existing target object set;
if the contour of the first target object is matched with the contour of the second target object, adding 1 to the found position and found count value of the second target object; when the found position of the second target object is true, clearing a lost count value of the second target object by 0, and if the found count value reaches a time threshold, determining that the stay time of the second target object after the second target object is changed from dynamic to static exceeds the time threshold; if the contour of the first target object is not matched with the contour of each second target object in the existing target object set, adding the first target object into the existing target object set as a new second target object; if the contour of a second target object does not match the contour of each first target object in the current frame target object set, the found position of the second target object is false; and when the found position of the second target object is false, adding 1 to the loss count value, and if the loss count value meets a preset condition, removing the second target object from the existing target object set.
Optionally, the target staying time length determining module is further configured to:
before the contour of the first target object is matched with the contour of the second target object, performing morphological processing on the difference image to obtain the contour of each target object in the current frame target object set; judging whether the number of pixels of the outline of each target object exceeds a pixel number threshold value or not; and filtering out the target object with the number of the pixels of the contour not exceeding the threshold value of the number of the pixels.
Optionally, the target slave determining module is specifically configured to:
when the target object changes from dynamic to static, all people existing in an interested area with the center of mass of the target object as the center are detected through the SVM; calculating the distance between the mass center of each person and the mass center of the target object, and taking the closest person as a target person, wherein the target person is the host of the target object; learning and storing characteristics of the target person; and when the stay time of the target object from the dynamic state to the static state exceeds the time threshold, detecting all the people in the same region of interest by using the SVM again, matching the detected people with the originally stored characteristics of the target person, and if the detected people are not matched with the characteristics of the target person, determining that the target object is a left-behind object.
The technical scheme provided by the embodiment of the invention has the following beneficial effects:
the method comprises the steps of judging and analyzing various problems possibly occurring in the detection process of the left-behind object to form a complete solution, firstly determining the motion characteristic of pixels judged by adopting a frame difference method according to the real-time rapidity requirement of the detection of the left-behind object, and finding out a target object which changes from dynamic to static, namely a suspected left-behind object; then, classifying the target object by using HOG characteristics and SVM to distinguish people and articles, and filtering the people; judging whether the stay time of the target object exceeds a preset time by adopting an evidence accumulation mode; and finally, carrying out feature matching on the people near the target object and the owner of the target object, and if the people are not matched, determining that the target object is a left-behind object. On one hand, the scheme of the invention has the characteristics of low calculation complexity and real-time high efficiency, and on the other hand, the invention solves the problems of false alarm and shielding in the detection process of the left-over object, also solves the problem of distinguishing whether the left-over object is a person or an object, and also solves the problem of distinguishing whether the left-over object is really without a master, so the detection accuracy is also very high, and the detection problem of the left-over object is finally effectively solved.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings without inventive exercise. Furthermore, these descriptions should not be construed as limiting the embodiments, wherein elements having the same reference number designation are identified as similar elements throughout the figures, and the drawings are not to scale unless otherwise specified.
FIG. 1 is a flow chart illustrating a method of legacy object detection according to an exemplary embodiment of the present invention;
FIG. 2 is a system framework diagram shown in accordance with an exemplary embodiment of the present invention; (ii) a
FIG. 3 is a flow chart illustrating a method of legacy object detection according to an exemplary embodiment of the present invention;
FIG. 4 is a flow chart illustrating a method of legacy object detection according to an exemplary embodiment of the present invention;
FIG. 5 is a flow chart illustrating a method of legacy object detection according to an exemplary embodiment of the present invention;
FIG. 6 is a flow chart illustrating a method of legacy object detection according to an exemplary embodiment of the present invention;
FIG. 7 is a flow chart illustrating a method of legacy object detection according to an exemplary embodiment of the present invention;
FIG. 8 is a flow chart illustrating a method of legacy object detection in accordance with an exemplary embodiment of the present invention;
fig. 9 is a schematic diagram illustrating an legacy object detection apparatus according to an exemplary embodiment of the present invention.
Detailed Description
Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the invention, as detailed in the appended claims.
Fig. 1 is a flowchart illustrating a legacy object detection method according to an exemplary embodiment of the present invention. As an example, the method can be used in hardware devices such as DSPs (digital signal processing).
The system in which the method of the invention is located will be illustrated below. FIG. 2 is a system framework diagram shown in accordance with an exemplary embodiment of the present invention. The system mainly comprises a high-definition camera, a remote control platform, embedded equipment, a switch, external storage equipment and the like. The embedded device of the system can select a multi-core system integrating an ARM, a DSP and a video coprocessor, such as DM8168 of TI, and the like, so as to improve the real-time computing capability of the system while keeping low power consumption.
The main process of the system in use is as follows:
1) sending video data collected by camera to embedded equipment
2) In the embedded device, ARM compresses received video data, and divides the video data into two paths, one path is used for algorithm processing of the DSP, and the other path is stored after being zoomed.
3) After the DSP is processed, the data is sent to the ARM, and the ARM sends the data containing the processing result to the PC end through the TCP to be displayed.
Referring to fig. 1, the method may include the steps of:
step S101, for the collected video, determining whether a target object appears according to whether the value of the pixel at the same position of the adjacent frame changes, wherein the target object is an object which moves firstly and then is static.
The method mainly solves the problem of how to preliminarily judge the object as the left object. In this step, according to the real-time rapidity requirement of the detection of the left-behind object, the motion characteristics of the pixels are determined by adopting the frame difference method (namely, the interframe difference method), and the target object changed from dynamic to static is found.
Step S102, the HOG features of the histogram of oriented gradients are combined with a Support Vector Machine (SVM) to classify the found target object by people and objects so as to eliminate the situation that the target object is a person.
The method mainly solves the problem of how to judge whether the left object is an object rather than a person, and the specific classification method adopts a mode of combining HOG (histogram of oriented gradients) features with SVM (support vector machine).
Step S103, for each found target object, judging whether the stay time of the target object after the movement and the stillness exceeds a time threshold value by accumulating the times of the target object appearing in each frame.
The step further judges whether the article is possible to be left by accumulating the stay time of the article.
And step S104, if the stay time length exceeds a time length threshold value, matching the people near the target object with the owner of the target object, and if the stay time length does not exceed the time length threshold value, determining that the target object is a left-behind object, wherein the owner of the target object is the people closest to the target object and detected when the target object is changed from moving to static.
This step finally determines whether the object is a left-behind object by judging whether the target object has a subordinate relationship, i.e., whether the object belongs to a person.
The method comprises the steps of firstly, determining the motion characteristic of pixels judged by adopting a frame difference method according to the real-time rapidity requirement of the detection of the left-behind object, and finding out a target object which changes from dynamic to static, namely a suspected left-behind object; then, classifying the target object by using HOG characteristics and SVM to distinguish people and articles, and filtering the people; determining whether the stay time of the target object exceeds a preset time by adopting an evidence accumulation mode; and finally, carrying out feature matching on the people near the target object and the owner of the target object, and if the people are not matched, determining that the target object is a left-behind object. On one hand, the scheme of the invention has the characteristics of low calculation complexity and real-time high efficiency, and on the other hand, the invention solves the problems of false alarm and shielding in the detection process of the left-over object, also solves the problem of distinguishing whether the left-over object is a person or an object, and also solves the problem of distinguishing whether the left-over object is really without a master, so the detection accuracy is also very high, and the detection problem of the left-over object is finally effectively solved.
The invention is further described below. Of course, the following application scenarios are only exemplary, and in practical applications, other application scenarios may also be applicable.
As an example, in this embodiment or some other embodiments of the present invention, the step of determining whether the target object appears according to whether the value of the pixel at the same position in the adjacent frame changes may specifically include:
when the variation amplitude of the value of the pixel exceeds a pixel threshold value T, setting the attribute value corresponding to the pixel as a preset attribute value lambda, wherein the attribute value represents the longest time for allowing the pixel to be continuously unchanged without being regarded as a background;
according to the change of the pixel value and the attribute value, executing the following strategies:
when the variation amplitude of the value of the pixel is smaller than the pixel threshold value and the attribute value is 0, the pixel is indicated as the current background and no operation is performed; when the variation amplitude of the pixel value is smaller than the pixel threshold value but the attribute value is larger than 0, the pixel position is indicated to have a motion foreground, but the motion is in a static state or a slow motion state at the moment, and the attribute value is subjected to self-subtraction operation every time; when the attribute value is reduced from 0, indicating that the pixel position has remained still from moving to static, it is determined that the target object is present.
Dynamic matrix: each pixel corresponds to an attribute value, and the attribute values of all pixels can form a matrix, which is called a dynamic matrix. In the embodiment of the invention, a frame difference method is adopted to determine the motion characteristic of the pixel, and a dynamic matrix is introduced on the basis of the motion characteristic to determine when a certain pixel should be updated into the background. And judging whether the value of a certain pixel is changed or not by adopting an interframe difference method, namely whether a moving target appears or not. The value at a pixel in the dynamic matrix is set to λ if the pixel change value is larger than a threshold value, which means the maximum time a pixel is allowed to continue unchanged without being updated into the background. And λ is reduced to 0, it can indicate that a motion foreground appears at the pixel, but the motion foreground is still, i.e. the motion foreground is currently in a state from motion to still, and preliminarily conforms to the characteristics of the left-behind object.
Further, when there is a change in the image, it is difficult to distinguish whether there is an object entering or an object leaving, and if the judgment is wrong, a false alarm may occur. Therefore, aiming at the problems of false alarm and the like, in order to improve the judgment accuracy, the embodiment of the invention can further adopt a double-background detection strategy on the basis of detecting the foreground by a simple background difference method, and simultaneously maintain the double-background BC on the basis of the background differencet(x, y) and BOt(x, y) to determine if a new stationary object is present.
Thus, by way of example, in this embodiment or in certain other embodiments of the invention, the invention is capable of operation in addition to or in lieu of the present contextThe method may also include an original background, and use It(x,y)-BCt(x,y)>T denotes the presence of an object, It(x,y)-BCt(x,y)<T denotes no object present, It(x,y)-BOt(x,y)>T denotes the presence of an object, It(x,y)-BOt(x,y)<T denotes no object present, DtIf (x, y) ═ 0 indicates that the object is stable, the strategy may further include:
if It(x,y)-BCt(x,y)>T&&It(x,y)-BOt(x,y)>T&&Dt(x, y) is 0, indicating that a new stationary object is present, then order
BOt(x,y)＝BCt(x,y)
BCt(x,y)＝It(x,y)
Updating the current background and storing the original current background into the original background;
if It(x,y)-BCt(x,y)>T&&It(x,y)-BOt(x,y)<T, indicating that there is an old object left (e.g., there is a shelter before it disappears), then let
BCt(x,y)＝BOt(x,y)
Restoring the information stored in the original background to the current background;
if It(x,y)-BCt(x,y)<T&&It(x,y)-BOt(x,y)<T&&DtIf (x, y) ═ 0, indicating a clean background, then the dual background is updated directly, i.e. let
BCt(x,y)＝a×It(x,y)+(1-a)×BCt(x,y)
BOt(x,y)＝a×It(x,y)+(1-a)×BOt(x,y)
Wherein the content of the first and second substances,&&is represented by the general formula It(x, y) represents the value of the pixel in the current frame, BCt(x, y) stands for Current background, BOt(x, y) represents the original background, Dt(x, y) represents the attribute value corresponding to the pixel, T is the pixel threshold value, and a is the preset learning rate.
Regarding the learning rate: even if there is no object movement, the image of the scene may change gradually, for example, in an outdoor day, the light is changing, so that the background needs to be adjusted adaptively through a learning rate a, where the learning rate a may be obtained according to experience/experiment and is not described again.
According to the above two-background updating strategy, when a new static object is updated to the current background BC, the previous current background BC information at the same pixel is stored in the original background BO, and the current frame I pixel information is stored in the current background BC at the same time, so that the new static object can be ensured to exist only in the current background; when an old still object leaves, the information originally stored in the original background BO needs to be restored to the current background BC, so that the problem of false alarm caused by the object being wrongly updated into the background can be solved, and the original background BO stores the background information before the new still object is still, so that the detection of the left-behind object is facilitated. When judging that no new or old target appears or leaves (namely, the new or old target is a pure background), the double backgrounds are directly updated at a certain rate a.
Referring to fig. 3, in this embodiment or some other embodiments of the present invention, the step of classifying the found target object by using the histogram of oriented gradients HOG feature in combination with the support vector machine SVM may specifically include:
step S301, HOG characteristics of a positive sample and a negative sample are extracted in advance, wherein the positive sample is an image sample of a person, and the negative sample is an image sample of an object.
Step S302, putting the extracted HOG features into an SVM classifier for training so as to obtain a model (model).
In step S303, a filter (i.e., a detector) is generated using the model.
And step S304, detecting the negative sample by using the screener to obtain a difficult sample.
In step S305, the HOG features of the difficult samples are extracted and trained together with the HOG features of the positive and negative samples to obtain the final filter (i.e., the final detector).
Step S306, extracting HOG characteristics of the target object;
and step S307, classifying the people and objects of the target object by using the pre-obtained final filter. The case where the target object is a person is excluded, and finally, only the case where the target object is an article is retained.
In order to determine whether the target object in the current frame is the same object as the previous target object, an evidence accumulation method may be used. Referring to fig. 4, in this embodiment or some other embodiments of the present invention, for each found target object, the step of determining whether the staying time length of the target object after the movement and the stillness exceeds the time length threshold by accumulating the number of times that the target object appears in each frame may include:
step S401, match the contour of a first target object with the contour of a second target object, where the first target object is a target object selected from the current target object set, and the second target object is a target object selected from the existing target object set.
As an example, the matching method may include comparing the area (S), perimeter (P), and dispersion (P) of the two profiles2/s), etc., and if so, both can be considered to be the same profile.
Step S402, if the contour of the first target object is matched with the contour of the second target object, adding 1 to the found position of the second target object and the found count value; and when the found position of the second target object is true, clearing 0 the lost count value of the second target object, and if the found count value reaches a time threshold value, confirming that the stay time of the second target object after the second target object is changed from moving to static exceeds the time threshold value. That is, the second target object is a suspected remaining object.
Step S403, if the contour of the first target object does not match the contour of each second target object in the existing target object set, adding the first target object as a new second target object to the existing target object set.
Step S404, if the contour of the second target object is not matched with the contour of each first target object in the current frame target object set, the found position of the second target object is false; and when the found position of the second target object is false, adding 1 to the loss count value, and if the loss count value meets a preset condition, removing the second target object from the existing target object set.
As an example, the preset condition may be whether the loss count value is greater than a prescribed value, or whether the loss rate is greater than a fixed value, or the like. That is, if one target object is lost several times in each frame and is difficult to track or even completely unable to track, processing such as discarding may be adopted.
In a specific implementation of the program, as shown in fig. 5, one object may correspond to one object structure. Traversing each object structure in the existing target object set, during traversal, if the found bit of the current object structure is true, clearing 0 the lost count value of the object structure, and then judging whether the found count value is greater than a time threshold (namely a warning value), if so, considering that the stay time of the target object represented by the object structure after the object structure is changed from static to dynamic exceeds the time threshold, namely, judging that the target object is a suspected left-over object; if the found bit of the current object structure is not true, the loss count value is added by 1, and whether the loss count value is greater than a specified value or the loss rate is greater than a fixed value is judged, if so, the object structure is emptied, which is equivalent to discarding the object structure.
Furthermore, in order to improve the detection efficiency and accuracy, particularly small objects can be filtered. Therefore, referring to fig. 6, in this embodiment or some other embodiments of the present invention, before matching the target contour with the contours in the contour set one by one, the method may further include:
step S601, performing morphological processing on the difference image to obtain a contour of each target object in the current frame target object set.
The difference image is an image obtained by subtracting the current background from the current frame. The morphological treatment may include, for example, swelling, erosion, and the like, and will not be described in detail.
In step S602, it is determined whether the number of pixels of the contour of each target object exceeds a threshold value of the number of pixels.
Step S603, filtering out the target object whose contour pixel number does not exceed the pixel number threshold.
The left-over object is detected immediately after moving into the scene and changing from motion to static, and the owner of the left-over object does not leave, and usually, people are nearest to the wrapped pedestrians, so that the people nearest to the wrapped people can be considered as the owner of the left-over object. Referring to fig. 7, in this embodiment or some other embodiments of the present invention, if the staying time exceeds the time threshold, the step of matching the person near the target object with the owner of the target object may specifically include:
and step S701, when the target object is changed from static to dynamic, detecting all persons in the region of interest with the center of mass of the target object as the center through an SVM.
Wherein the region of interest may be a pre-set size range.
Step S702, calculating the distance between the mass center of each person and the mass center of the target object, and taking the closest person as a target person, wherein the target person is the host of the target object.
Step S703, learning and storing the characteristics of the target person.
Wherein the features may be color, texture, shape, etc.
Step S704, after the staying time of the target object from moving to static exceeds the time threshold, detecting all people in the same region of interest again by using the SVM, and matching the detected people with the originally stored characteristics of the target person.
If the match is found, the owner of the object is not away, and the object is not a left-behind object; if not, this indicates that the owner of the object has left the scene or has been further away from the object, and eventually may be deemed to be a left-behind object.
Referring to fig. 8, for an object, on one hand, after the object stays, an SVM may be used to detect people near the object, and then learn features of colors, textures, and the like of people nearest to the object, on the other hand, it may be determined through evidence accumulation whether the features satisfy suspected abnormal object features, if the object is suspected abnormal, the learned features of the owner may be used to match with people near the current time, and if the object is not matched, it is determined that the object is abnormal, that is, the object is left behind.
The method comprises the steps of firstly, determining the motion characteristic of pixels judged by adopting a frame difference method according to the real-time rapidity requirement of the detection of the left-behind object, and finding out a target object which changes from dynamic to static, namely a suspected left-behind object; then, classifying the target object by using HOG characteristics and SVM to distinguish people and articles, and filtering the people; determining whether the stay time of the target object exceeds a preset time by adopting an evidence accumulation mode; and finally, carrying out feature matching on the people near the target object and the owner of the target object, and if the people are not matched, determining that the target object is a left-behind object. On one hand, the scheme of the invention has the characteristics of low calculation complexity and real-time high efficiency, and on the other hand, the invention solves the problems of false alarm and shielding in the detection process of the left-over object, also solves the problem of distinguishing whether the left-over object is a person or an object, and also solves the problem of distinguishing whether the left-over object is really without a master, so the detection accuracy is also very high, and the detection problem of the left-over object is finally effectively solved.
The following are embodiments of the apparatus of the present invention that may be used to perform embodiments of the method of the present invention. For details which are not disclosed in the embodiments of the apparatus of the present invention, reference is made to the embodiments of the method of the present invention.
Fig. 9 is a schematic diagram illustrating an legacy object detection apparatus according to an exemplary embodiment of the present invention. As an example, the apparatus may be used in hardware devices such as a DSP (digital signal processing).
Referring to fig. 9, the apparatus may include:
a target finding module 901, configured to determine whether a target object appears according to whether a value of a pixel at the same position of an adjacent frame changes for the acquired video, where the target object is an object that moves first and then is stationary.
The target elimination module 902 classifies the found target object by using histogram of oriented gradients HOG features in combination with support vector machine SVM to eliminate the case that the target object is a person.
And a target staying time length judging module 903, configured to judge, for each found target object, whether a staying time length of the target object after the target object changes from static to dynamic exceeds a time length threshold by accumulating the number of times that the target object appears in each frame.
And a target subordinate judgment module 904, configured to match, when the staying time length exceeds a time length threshold, a person near the target object with an owner of the target object, and if the staying time length does not exceed the time length threshold, determine that the target object is a left-behind object, where the owner of the target object is a person closest to the target object detected when the target object is still.
For example, in this embodiment or some other embodiments of the present invention, when the target finding module determines whether the target object appears according to whether the value of the pixel at the same position in the adjacent frame changes, the target finding module is specifically configured to:
when the variation amplitude of the value of the pixel exceeds a pixel threshold value, setting the attribute value corresponding to the pixel as a preset attribute value, wherein the attribute value represents the longest time for allowing the pixel to be continuously unchanged without being regarded as a background; according to the change of the pixel value and the attribute value, executing the following strategies:
when the variation amplitude of the value of the pixel is smaller than the pixel threshold value and the attribute value is 0, the pixel is indicated as the current background and no operation is performed; when the variation amplitude of the pixel value is smaller than the pixel threshold value but the attribute value is larger than 0, the pixel position is indicated to have a motion foreground, but the motion is in a static state or a slow motion state at the moment, and the attribute value is subjected to self-subtraction operation every time; when the attribute value is reduced from 0, indicating that the pixel position has remained still from moving to static, it is determined that the target object is present.
As an example, in this embodiment or some other embodiments of the present invention, the original context is further included, and the policy further includes:
if It(x,y)-BCt(x,y)>T&&It(x,y)-BOt(x,y)>T&&Dt(x, y) is 0, indicating that a new stationary object is present, then order
BOt(x,y)＝BCt(x,y)
BCt(x,y)＝It(x,y)
Updating the current background and storing the original current background into the original background;
if It(x,y)-BCt(x,y)>T&&It(x,y)-BOt(x,y)<T, indicating that the old target leaves, then order
BCt(x,y)＝BOt(x,y)
Restoring the information stored in the original background to the current background;
if It(x,y)-BCt(x,y)<T&&It(x,y)-BOt(x,y)<T&&Dt(x, y) ═ 0, which indicates a clear background, let
BCt(x,y)＝a×It(x,y)+(1-a)×BCt(x,y)
BOt(x,y)＝a×It(x,y)+(1-a)×BOt(x,y)
Wherein the content of the first and second substances,&&is represented by the general formula It(x, y) represents the value of the pixel in the current frame, BCt(x, y) stands for Current background, BOt(x, y) represents the original background, Dt(x, y) represents the attribute value corresponding to the pixel, T is the pixel threshold value, and a is the preset learning rate.
As an example, in this embodiment or some other embodiments of the present invention, the target excluding module is specifically configured to:
extracting HOG characteristics of a positive sample and a negative sample in advance, wherein the positive sample is an image sample of a person, and the negative sample is an image sample of an object; putting the extracted HOG features into an SVM classifier for training to obtain a model; generating a filter using the model; detecting the negative sample with the screener to obtain a difficult sample; extracting HOG characteristics of the difficult samples and putting the HOG characteristics and the HOG characteristics of the positive and negative samples into training to obtain a final screener; extracting HOG characteristics of the target object; and classifying the human and object of the target object by using the pre-obtained final filter.
As an example, in this embodiment or some other embodiments of the present invention, the target staying time period determination module is specifically configured to:
matching the contour of a first target object with the contour of a second target object, wherein the first target object is a target object selected from a current target object set, and the second target object is a target object selected from an existing target object set;
if the contour of the first target object is matched with the contour of the second target object, adding 1 to the found position and found count value of the second target object; when the found position of the second target object is true, clearing a lost count value of the second target object by 0, and if the found count value reaches a time threshold, determining that the stay time of the second target object after the second target object is changed from dynamic to static exceeds the time threshold; if the contour of the first target object is not matched with the contour of each second target object in the existing target object set, adding the first target object into the existing target object set as a new second target object; if the contour of a second target object does not match the contour of each first target object in the current frame target object set, the found position of the second target object is false; and when the found position of the second target object is false, adding 1 to the loss count value, and if the loss count value meets a preset condition, removing the second target object from the existing target object set.
As an example, in this embodiment or some other embodiments of the present invention, the target staying time period determination module is further configured to:
before the contour of the first target object is matched with the contour of the second target object, performing morphological processing on the difference image to obtain the contour of each target object in the current frame target object set; judging whether the number of pixels of the outline of each target object exceeds a pixel number threshold value or not; and filtering out the target object with the number of the pixels of the contour not exceeding the threshold value of the number of the pixels.
As an example, in this embodiment or some other embodiments of the present invention, the target slave determining module is specifically configured to:
when the target object changes from dynamic to static, all people existing in an interested area with the center of mass of the target object as the center are detected through the SVM; calculating the distance between the mass center of each person and the mass center of the target object, and taking the closest person as a target person, wherein the target person is the host of the target object; learning and storing characteristics of the target person; and when the stay time of the target object from the dynamic state to the static state exceeds the time threshold, detecting all the people in the same region of interest by using the SVM again, matching the detected people with the originally stored characteristics of the target person, and if the detected people are not matched with the characteristics of the target person, determining that the target object is a left-behind object.
The method comprises the steps of firstly, determining the motion characteristic of pixels judged by adopting a frame difference method according to the real-time rapidity requirement of the detection of the left-behind object, and finding out a target object which changes from dynamic to static, namely a suspected left-behind object; then, classifying the target object by using HOG characteristics and SVM to distinguish people and articles, and filtering the people; determining whether the stay time of the target object exceeds a preset time by adopting an evidence accumulation mode; and finally, carrying out feature matching on the people near the target object and the owner of the target object, and if the people are not matched, determining that the target object is a left-behind object. On one hand, the scheme of the invention has the characteristics of low calculation complexity and real-time high efficiency, and on the other hand, the invention solves the problems of false alarm and shielding in the detection process of the left-over object, also solves the problem of distinguishing whether the left-over object is a person or an object, and also solves the problem of distinguishing whether the left-over object is really without a master, so the detection accuracy is also very high, and the detection problem of the left-over object is finally effectively solved.
Regarding the apparatus in the foregoing embodiments, the specific manner in which each unit \ module executes operations has been described in detail in the embodiments of the related method, and is not described herein again.
Other embodiments of the invention will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the invention and including such departures from the present disclosure as come within known or customary practice within the art to which the invention pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the invention being indicated by the following claims.
It will be understood that the invention is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the invention is limited only by the appended claims.
Claims (12)
1. A method of legacy object detection, the method comprising:
for the collected video, determining whether a target object appears according to whether the value of a pixel at the same position of an adjacent frame changes, wherein the target object is an object which moves firstly and then is static;
classifying the found target object by combining HOG (histogram of oriented gradient) features with a Support Vector Machine (SVM) so as to eliminate the condition that the target object is a person;
for each found target object, judging whether the stay time of the target object after the movement and the stillness exceeds a time threshold value or not by accumulating the times of the target object appearing in each frame;
if the stay time length exceeds a time length threshold value, matching people near the target object with the owner of the target object, and if the stay time length does not exceed the time length threshold value, determining that the target object is a left-behind object, wherein the owner of the target object is a person which is detected to be closest to the target object when the target object is calmed;
the method for determining whether the target object appears according to whether the value of the pixel at the same position of the adjacent frames changes comprises the following steps:
when the variation amplitude of the value of the pixel exceeds a pixel threshold value, setting the attribute value corresponding to the pixel as a preset attribute value, wherein the attribute value represents the longest time for allowing the pixel to be continuously unchanged without being regarded as a background;
according to the change of the pixel value and the attribute value, executing the following strategies:
when the variation amplitude of the value of the pixel is smaller than the pixel threshold value and the attribute value is 0, the pixel is indicated as the current background and no operation is performed; when the variation amplitude of the pixel value is smaller than the pixel threshold value but the attribute value is larger than 0, the pixel position is indicated to have a motion foreground, but the motion is in a static state or a slow motion state at the moment, and the attribute value is subjected to self-subtraction operation every time; when the attribute value is reduced from 0, indicating that the pixel position has remained still from moving to static, it is determined that the target object is present.
2. The method of claim 1, further comprising an original context, the policy further comprising:
if It(x,y)-BCt(x,y)>T&&It(x,y)-BOt(x,y)>T&&Dt(x, y) is 0, indicating that a new stationary object is present, then order
BOt(x,y)＝BCt(x,y)
BCt(x,y)＝It(x,y)
Updating the current background and storing the original current background into the original background;
if It(x,y)-BCt(x,y)>T&&It(x,y)-BOt(x,y)<T, indicating that the old target leaves, then order
BCt(x,y)＝BOt(x,y)
Restoring the information stored in the original background to the current background;
if It(x,y)-BCt(x,y)<T&&It(x,y)-BOt(x,y)<T&&Dt(x, y) ═ 0, which indicates a clear background, let
BCt(x,y)＝a×It(x,y)+(1-a)×BCt(x,y)
BOt(x,y)＝a×It(x,y)+(1-a)×BOt(x,y)
Wherein the content of the first and second substances,&&is represented by the general formula It(x, y) represents the value of the pixel in the current frame, BCt(x, y) stands for Current background, BOt(x, y) represents the original background, Dt(x, y) represents the attribute value corresponding to the pixel, T is the pixel threshold value, and a is the preset learning rate.
3. The method of claim 1, wherein the step of classifying the found target object by using Histogram of Oriented Gradients (HOG) feature in combination with Support Vector Machine (SVM) comprises:
extracting HOG characteristics of a positive sample and a negative sample in advance, wherein the positive sample is an image sample of a person, and the negative sample is an image sample of an object;
putting the extracted HOG features into an SVM classifier for training to obtain a model;
generating a filter using the model;
detecting the negative sample with the screener to obtain a difficult sample;
extracting HOG characteristics of the difficult samples and putting the HOG characteristics and the HOG characteristics of the positive and negative samples into training to obtain a final screener;
extracting HOG characteristics of the target object;
and classifying the human and object of the target object by using the pre-obtained final filter.
4. The method of claim 1, wherein for each target object found, determining whether the stay time of the target object after the target object is made static exceeds a time threshold by accumulating the number of times the target object appears in each frame comprises:
matching the contour of a first target object with the contour of a second target object, wherein the first target object is a target object selected from a current target object set, and the second target object is a target object selected from an existing target object set;
if the contour of the first target object is matched with the contour of the second target object, adding 1 to the found position and found count value of the second target object; when the found position of the second target object is true, clearing a lost count value of the second target object by 0, and if the found count value reaches a time threshold, determining that the stay time of the second target object after the second target object is changed from dynamic to static exceeds the time threshold;
if the contour of the first target object is not matched with the contour of each second target object in the existing target object set, adding the first target object into the existing target object set as a new second target object;
if the contour of a second target object does not match the contour of each first target object in the current frame target object set, the found position of the second target object is false; and when the found position of the second target object is false, adding 1 to the loss count value, and if the loss count value meets a preset condition, removing the second target object from the existing target object set.
5. The method of claim 4, wherein prior to matching the contour of the first target object with the contour of the second target object, the method further comprises:
performing morphological processing on the differential image to obtain the outline of each target object in the current frame target object set;
judging whether the number of pixels of the outline of each target object exceeds a pixel number threshold value or not;
and filtering out the target object with the number of the pixels of the contour not exceeding the threshold value of the number of the pixels.
6. The method of claim 1, wherein matching the person near the target object with the owner of the target object if the length of stay exceeds a length of time threshold comprises:
when the target object changes from dynamic to static, all people existing in an interested area with the center of mass of the target object as the center are detected through the SVM;
calculating the distance between the mass center of each person and the mass center of the target object, and taking the closest person as a target person, wherein the target person is the host of the target object;
learning and storing characteristics of the target person;
and when the staying time of the target object from moving to static exceeds the time threshold, detecting all the people in the same region of interest by using the SVM again, and matching the detected people with the originally stored characteristics of the target people.
7. An carryover object detection apparatus, comprising:
the target finding module is used for determining whether a target object appears according to whether the value of a pixel at the same position of an adjacent frame changes or not for the acquired video, wherein the target object is an object which moves firstly and then is static;
the target elimination module is used for classifying the found target object by adopting the HOG (histogram of oriented gradients) characteristic in combination with a Support Vector Machine (SVM) so as to eliminate the condition that the target object is a person;
the target stay time length judging module is used for judging whether the stay time length of the target object after the target object is changed from moving to static exceeds a time length threshold value or not by accumulating the times of the target object appearing in each frame for each found target object;
the target subordinate judgment module is used for matching the people near the target object with the owner of the target object when the stay time exceeds a time threshold, and determining that the target object is a left-behind object if the stay time does not match the owner of the target object, wherein the owner of the target object is the person closest to the target object detected when the target object changes from moving to static;
when the target finding module determines whether a target object appears according to whether the value of the pixel at the same position of the adjacent frames changes, the target finding module is specifically configured to:
when the variation amplitude of the value of the pixel exceeds a pixel threshold value, setting the attribute value corresponding to the pixel as a preset attribute value, wherein the attribute value represents the longest time for allowing the pixel to be continuously unchanged without being regarded as a background; according to the change of the pixel value and the attribute value, executing the following strategies:
when the variation amplitude of the value of the pixel is smaller than the pixel threshold value and the attribute value is 0, the pixel is indicated as the current background and no operation is performed; when the variation amplitude of the pixel value is smaller than the pixel threshold value but the attribute value is larger than 0, the pixel position is indicated to have a motion foreground, but the motion is in a static state or a slow motion state at the moment, and the attribute value is subjected to self-subtraction operation every time; when the attribute value is reduced from 0, indicating that the pixel position has remained still from moving to static, it is determined that the target object is present.
8. The apparatus of claim 7, further comprising an original context, the policy further comprising:
if It(x,y)-BCt(x,y)>T&&It(x,y)-BOt(x,y)>T&&Dt(x, y) is 0, indicating that a new stationary object is present, then order
BOt(x,y)＝BCt(x,y)
BCt(x,y)＝It(x,y)
Updating the current background and storing the original current background into the original background;
if It(x,y)-BCt(x,y)>T&&It(x,y)-BOt(x,y)<T, indicating that the old target leaves, then order
BCt(x,y)＝BOt(x,y)
Restoring the information stored in the original background to the current background;
if It (x, y) -BCt (x, y) < T & & It (x, y) -BOt (x, y) < T & & Dt (x, y) & 0 indicates a clear background, let us order
BCt(x,y)＝a×It(x,y)+(1-a)×BCt(x,y)
BOt(x,y)＝a×It(x,y)+(1-a)×BOt(x,y)
Where & & represents and, It (x, y) represents the value of the pixel in the current frame, BCt (x, y) represents the current background, BOt (x, y) represents the original background, Dt (x, y) represents the attribute value corresponding to the pixel, T is the pixel threshold, and a is the preset learning rate.
9. The apparatus of claim 7, wherein the target exclusion module is specifically configured to:
extracting HOG characteristics of a positive sample and a negative sample in advance, wherein the positive sample is an image sample of a person, and the negative sample is an image sample of an object; putting the extracted HOG features into an SVM classifier for training to obtain a model; generating a filter using the model; detecting the negative sample with the screener to obtain a difficult sample; extracting HOG characteristics of the difficult samples and putting the HOG characteristics and the HOG characteristics of the positive and negative samples into training to obtain a final screener; extracting HOG characteristics of the target object; and classifying the human and object of the target object by using the pre-obtained final filter.
10. The apparatus according to claim 7, wherein the target staying time period determining module is specifically configured to:
matching the contour of a first target object with the contour of a second target object, wherein the first target object is a target object selected from a current target object set, and the second target object is a target object selected from an existing target object set;
if the contour of the first target object is matched with the contour of the second target object, adding 1 to the found position and found count value of the second target object; when the found position of the second target object is true, clearing a lost count value of the second target object by 0, and if the found count value reaches a time threshold, determining that the stay time of the second target object after the second target object is changed from dynamic to static exceeds the time threshold; if the contour of the first target object is not matched with the contour of each second target object in the existing target object set, adding the first target object into the existing target object set as a new second target object; if the contour of a second target object does not match the contour of each first target object in the current frame target object set, the found position of the second target object is false; and when the found position of the second target object is false, adding 1 to the loss count value, and if the loss count value meets a preset condition, removing the second target object from the existing target object set.
11. The apparatus of claim 10, wherein the target dwell time determination module is further configured to:
before the contour of the first target object is matched with the contour of the second target object, performing morphological processing on the difference image to obtain the contour of each target object in the current frame target object set; judging whether the number of pixels of the outline of each target object exceeds a pixel number threshold value or not; and filtering out the target object with the number of the pixels of the contour not exceeding the threshold value of the number of the pixels.
12. The apparatus of claim 7, wherein the target slave determination module is specifically configured to:
when the target object changes from dynamic to static, all people existing in an interested area with the center of mass of the target object as the center are detected through the SVM; calculating the distance between the mass center of each person and the mass center of the target object, and taking the closest person as a target person, wherein the target person is the host of the target object; learning and storing characteristics of the target person; and when the stay time of the target object from the dynamic state to the static state exceeds the time threshold, detecting all the people in the same region of interest by using the SVM again, matching the detected people with the originally stored characteristics of the target person, and if the detected people are not matched with the characteristics of the target person, determining that the target object is a left-behind object.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910140397.7A CN109948455B (en) | 2019-02-22 | 2019-02-22 | Detection method and device for left-behind object |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910140397.7A CN109948455B (en) | 2019-02-22 | 2019-02-22 | Detection method and device for left-behind object |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109948455A CN109948455A (en) | 2019-06-28 |
| CN109948455B true CN109948455B (en) | 2020-12-18 |
Family
ID=67006796
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910140397.7A Active CN109948455B (en) | 2019-02-22 | 2019-02-22 | Detection method and device for left-behind object |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109948455B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110705461B (en) * | 2019-09-29 | 2022-11-01 | 北京百度网讯科技有限公司 | Image processing method and device |
| CN110796017A (en) * | 2019-09-30 | 2020-02-14 | 北京三快在线科技有限公司 | Method and device for determining article drop-out and method and device for training model |
| CN110674787A (en) * | 2019-10-09 | 2020-01-10 | 山东浪潮人工智能研究院有限公司 | Video decompression method and system based on Hog feature and lgb classifier |
| CN111079621B (en) * | 2019-12-10 | 2023-10-03 | 北京百度网讯科技有限公司 | Method, device, electronic equipment and storage medium for detecting object |
| CN111160187B (en) * | 2019-12-20 | 2023-05-02 | 浙江大华技术股份有限公司 | Method, device and system for detecting left-behind object |
| CN111797727B (en) * | 2020-06-18 | 2023-04-07 | 浙江大华技术股份有限公司 | Method and device for detecting road surface sprinkled object and storage medium |
| WO2023039781A1 (en) * | 2021-09-16 | 2023-03-23 | 华北电力大学扬中智能电气研究中心 | Method for detecting abandoned object, apparatus, electronic device, and storage medium |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101727672A (en) * | 2008-10-24 | 2010-06-09 | 云南正卓信息技术有限公司 | Method for detecting, tracking and identifying object abandoning/stealing event |
| CN101715111B (en) * | 2009-11-16 | 2011-12-14 | 南京邮电大学 | Method for automatically searching abandoned object in video monitoring |
| JP6295122B2 (en) * | 2014-03-27 | 2018-03-14 | 株式会社メガチップス | State estimation device, program, and integrated circuit |
| CN104156942B (en) * | 2014-07-02 | 2017-02-15 | 华南理工大学 | Detection method for remnants in complex environment |
| CN105404847B (en) * | 2014-09-16 | 2019-01-29 | 北京计算机技术及应用研究所 | A kind of residue real-time detection method |
| CN104933710B (en) * | 2015-06-10 | 2018-06-19 | 华南理工大学 | Based on the shop stream of people track intelligent analysis method under monitor video |
-
2019
- 2019-02-22 CN CN201910140397.7A patent/CN109948455B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN109948455A (en) | 2019-06-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109948455B (en) | Detection method and device for left-behind object | |
| EP1974326B1 (en) | Video signal analysis | |
| CN110188807A (en) | Tunnel pedestrian target detection method based on cascade super-resolution network and improvement Faster R-CNN | |
| CN102982313B (en) | The method of Smoke Detection | |
| CN101258512A (en) | Method and image evaluation unit for scene analysis | |
| CN104881643B (en) | A kind of quick remnant object detection method and system | |
| CN110852179B (en) | Suspicious personnel invasion detection method based on video monitoring platform | |
| CN103729858A (en) | Method for detecting article left over in video monitoring system | |
| CN112071084A (en) | Method and system for judging illegal parking by utilizing deep learning | |
| Malhi et al. | Vision based intelligent traffic management system | |
| CN111079621A (en) | Method and device for detecting object, electronic equipment and storage medium | |
| Su et al. | A new local-main-gradient-orientation HOG and contour differences based algorithm for object classification | |
| Subaweh et al. | Implementation of pixel based adaptive segmenter method for tracking and counting vehicles in visual surveillance | |
| JP7125843B2 (en) | Fault detection system | |
| Filonenko et al. | Illegally parked vehicle detection using adaptive dual background model | |
| CN110677619A (en) | Intelligent monitoring video processing method | |
| CN104268548A (en) | Accumulated snow detection method based on road images | |
| Wang et al. | Vision-based highway traffic accident detection | |
| CN103177248A (en) | Rapid pedestrian detection method based on vision | |
| Li et al. | Robust real-time detection of abandoned and removed objects | |
| Liu et al. | Analysis and design of public places crowd stampede early-warning simulating system | |
| CN116311166A (en) | Traffic obstacle recognition method and device and electronic equipment | |
| CN110647831A (en) | Court trial patrol method and system | |
| CN115810161A (en) | Transformer substation fire identification method and system | |
| KR101407394B1 (en) | System for abandoned and stolen object detection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |