CN105389567A - Group anomaly detection method based on a dense optical flow histogram - Google Patents
Group anomaly detection method based on a dense optical flow histogram Download PDFInfo
- Publication number
- CN105389567A CN105389567A CN201510786736.0A CN201510786736A CN105389567A CN 105389567 A CN105389567 A CN 105389567A CN 201510786736 A CN201510786736 A CN 201510786736A CN 105389567 A CN105389567 A CN 105389567A
- Authority
- CN
- China
- Prior art keywords
- sample
- pixel
- light stream
- optical flow
- dense optical
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
- G06V20/53—Recognition of crowd images, e.g. recognition of crowd congestion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a group anomaly detection method based on a dense optical flow histogram, comprising the following steps: step 1, calculating the dense optical flow field of a video frame image; step 2, carrying out blocked processing according to the size of the dense optical flow field and acquiring the optical flow vectors of all the pixels in each image block; step 3, carrying out calculation to obtain the features of an optical flow direction histogram corresponding to the frame image; step 4, selecting a training sample of an SVM classifier to train the SVM classifier; step 5, using the SVM classifier after training to classify a sample to be detected, and judging whether the crowd in the frame image is abnormal; and step 6, optimizing and correcting the classification result of the SVM classifier by taking the crowd behavior state duration as a basis for judging whether the crowd in the frame image is abnormal. According to the invention, an optimal flow vector field can be well analyzed, crowd motion can be reflected more objectively and accurately, the classification result of the SVM classifier is optimized, and the error rate of judgment of the classifier is reduced.
Description
Technical field
The present invention relates to group behavior method for detecting abnormality, particularly, relate to utilize dense optical flow and histogram feature to divide based on the histogrammic group abnormality detection method of dense optical flow.
Background technology
Along with growing stronger day by day of urban population, the flow of the people of public domain increases severely, and the monitor and managment of crowd is faced with challenge and pressure greatly, and the attack of terrorism, the crowd security incident such as to have a fist fight is of common occurrence.If Video Supervision Technique can be utilized to monitor in time the group behavior abnormal occurrence in important area and report to the police, relevant departments just can take corresponding behave for early warning or warning phenomenon within the shortest time, and to make, security incident possibility occurrence minimizes, accident causes damage minimumization.Therefore, increasing video monitoring system is applied to public place to maintain public order, to improve public domain safety, and group movement analysis and abnormality detection also receive the concern of more and more people.
The key of unusual checking is to utilize all kinds of behavior pattern to go to analyze to identify in scene and have visibly different or that probability of happening is lower in this scenario behavior with other behaviors.In public safety region, crowd density is excessive, vehicle flowrate is excessive, pedestrian's unusual aggregation or dispersion and pedestrian all belong to the category of abnormal behaviour in the suspicious conduct of special scenes.
Crowd behaviour angle, detect abnormal typical method and exactly target detection and tracking are carried out to crowd, analyze current behavior by the movement locus acquired, differentiate and judge normal movement locus or belong to abnormal, namely traditional method for tracking target; Whether another kind of mode is by whole image scene as a whole, by extracting the feature that can be used for event classification in picture, as utilized optical flow method to extract Global movement feature etc., differentiating and having anomalous event to occur.Light stream is as a kind of expression way of simple and practical image motion, effectively can avoid due to crowd density and block the impact brought to trajectory track, colony is carried out to analysis and the expression of motion conditions from whole and part angle, be widely used in method for detecting abnormality.
Light stream is the form of expression of moving object movement velocity of corresponding pixel in moving image, it utilizes every piece image be adjacent the corresponding relation between frame and contact the movable information calculating object, obtain the situation of change of each pixel in time domain in image sequence, reflect the instantaneous velocity of moving object pixel corresponding in two-dimensional image space in this moment, embody velocity magnitude and the direction of its motion.
Light stream direction histogram, as the feature of two field picture, can reflect the movable information of crowd in image to a certain extent.Light stream direction histogram the earliest carries out cumulative statistics to the pixel number belonging to same angular interval in region, have ignored the change of crowd movement's velocity magnitude.And amplitude size and the severity of this motion can be reflected to a certain extent due to the size of movement velocity, the amplitude adding pixel light stream can show motor behavior better, the amplitude size different according to pixel gives the sum weight of its different size, and the amplitude of crowd movement can be made also can to embody to some extent in histogram.
The people such as IEEE member TianWang proposed the computing method of image block and light stream direction histogram feature in 2014, utilized SVM classifier to carry out tagsort and judged abnormal, reach higher Detection accuracy.In contrast, maximum difference is in the present invention, and the present invention is after SVM tagsort, utilize anomalous identification to optimize thought to adjust classification results, behavior minimum length in time is set in 0.5s, has both decreased the error rate of sorter classification, in turn ensure that real-time to a certain extent.
Patent documentation: based on the local anomaly behavioral value method (patent publication No.: CN104820824A of light stream and spatio-temporal gradient; Patent publication date: 2015.08.05) in utilize light stream and semi-parameter model statistical method to detect to find the block that abnormal behaviour may occur in entire image, and utilize light stream energy to carry out further having detected final abnormality detection to this region.This invention judges that abnormal foundation utilizes likelihood ratio measurement target area whether similar between other regions, need artificially to set judgment threshold, the present invention then utilizes sorter automatic learning exception and non-abnormal cut-off rule, decrease human intervention, and optimize by abnormal the error rate reducing sorter classification, improve the accuracy of detection.On the other hand, split the block obtained in this invention and there is not overlapping situation, do not consider the correlativity between adjacent block, and the present invention is while concern colony local anomaly, taken into account the correlativity and continuity of move between local and local, make the analysis of motion and expression more accurate.
Summary of the invention
For defect of the prior art, the object of this invention is to provide a kind of based on the histogrammic group abnormality detection method of dense optical flow.
According to provided by the invention based on the histogrammic group abnormality detection method of dense optical flow, it is characterized in that, comprise the steps:
Step 1: the dense optical flow field obtaining at least one width two field picture in video image;
Step 2: obtain multiple image block after carrying out piecemeal process according to the size of described dense optical flow field, obtains the light stream vector of each pixel in each image block;
Step 3: by the direction of motion discretize of pixel, by the light stream discrete--direction of described pixel, number in all angles interval of the light stream vector adding up in each image block the pixel comprised in circumference range, and the amplitude of pixel light stream vector is carried out as this pixel the weights that add up on light stream direction, embody the pixel number of motion pixel and size of instantaneous velocity in each angular interval of current time in image block in the mode of accumulative amplitude; For each width two field picture, all pixels in all images block of combined frame image, in the motion vector statistics in this moment, obtain the light stream direction histogram feature that described two field picture is corresponding;
Step 4: the training sample selecting support vector machines sorter, and described training sample is carried out to the mark of class label, by performing the operation of step I 1, step I 2, step I 3 to the described training sample with class label, obtain the light stream direction histogram feature of the described training sample corresponding with class label.The parameter of adjustment SVM classifier and kernel function, utilize the class label of described training sample and the light stream direction histogram feature of correspondence, SVM classifier is trained, in the feature space that kernel function maps, obtain optimal separating hyper plane, the training sample of inhomogeneity label makes a distinction by this lineoid;
Wherein:
Step I 1: the dense optical flow field obtaining at least one width two field picture in the video image in sample;
Step I 2: obtain multiple image block after carrying out piecemeal process according to the size of described dense optical flow field, obtains the light stream vector of each pixel in each image block;
Step I 3: by the direction of motion discretize of pixel, by the light stream discrete--direction of described pixel, number in all angles interval of the light stream vector adding up in each image block the pixel comprised in circumference range, and the amplitude of pixel light stream vector is carried out as this pixel the weights that add up on light stream direction, embody the pixel number of motion pixel and size of instantaneous velocity in each angular interval of current time in image block in the mode of accumulative amplitude; For each width two field picture, all pixels in all images block of combined frame image, in the motion vector statistics in this moment, obtain the light stream direction histogram feature that described two field picture is corresponding;
Step 5: to sample to be tested carry out step I 1, step I 2, step I 3 operation obtain the light stream direction histogram feature of sample to be tested afterwards, utilize kernel function that histogram feature is mapped to feature space, in this feature space, judge this sample to be tested feature is positioned at which side of the SVM classifier optimal separating hyper plane of having trained, determine sample to be tested generic, whether occur exception with the crowd in this judgment frame image;
Step 6: whether crowd behaviour state duration is occurred an abnormal basis for estimation as the crowd in two field picture, and SVM classifier is corrected the sample to be tested classification results obtained of classifying.
Preferably, described step 1 comprises:
Step 1.1: the video image of input is processed frame by frame, and by every color image frame converting gradation image, and represent by the Mat data structure storage that OpenCV increases income in storehouse;
Step 1.2: call OpenCV and to increase income built-in function calcOpticalFlowFarneback, namely the optical flow algorithm of GunnarFarneback calculates the dense optical flow field of every two field picture; The size of described dense optical flow field is identical with the resolution of video image, and the bivector of each pixel is an all corresponding mark movable information, namely comprises movement velocity size and the direction of motion of pixel.
Preferably, obtain after carrying out piecemeal process according to the size of described dense optical flow field in described step 2 often organizing in multiple image block to include between two adjacent blocks 50% area overlapping.
Preferably, described step 3 comprises:
Step 3.1: by the direction of motion discretize of pixel, by the light stream discrete--direction of described pixel, with the horizontal left direction of two field picture for angle of circumference 8 decile being obtained 8 angular interval within the scope of initial direction counterclockwise-180 ° ~ 180 °, each angular interval represents a direction.
Wherein ,-180 ° all represent the horizontal left direction of two field picture with 180 °, all represent with-180 °.A kth angular interval is designated as b
k, 1≤k≤8, the angular range comprised is 45 ° of (k-1)-180 °≤θ <45 ° of k-180 °.
Step 3.2: dividing in each image block obtained, add up the number of all pixel light stream vectors in all angles interval comprised in described image block, in statistic processes, the amplitude of each pixel light stream vector is carried out as this pixel the weights that add up on the light stream direction of correspondence, embody the pixel number of motion pixel and size of instantaneous velocity in each angular interval of current time in image block in the mode of accumulative amplitude;
Suppose the light stream amplitude of pixel (x, y), direction is designated as R (x, y) respectively, α (x, y), then the sum weight G of this pixel (x, y) on various discrete direction
k(x, y) is:
Wherein b
krepresent the kth angular interval after dividing, 1≤k≤8, x represents the X-axis coordinate figure of pixel under rectangular coordinate system, and y represents the Y-axis coordinate figure of pixel under rectangular coordinate system;
Step 3.3: the accumulated result of each image block is linked in sequence together by block number, all pixels motion vector statistics at a time in all images block of i.e. combined frame image, obtain the light stream direction histogram feature that described two field picture is corresponding, namely obtain the higher-dimension histogram feature of view picture two field picture;
Preferably, described step 4 comprises:
Step 4.1: the training sample selecting SVM classifier, and selected training sample is carried out to the mark of class label, by the training sample containing crowd's abnormal behaviour as negative sample, marking its class label is-1; All the other training samples are then positive sample, and marking its class label is+1;
Step 4.2: by performing step 1 to the operation of step 3 to the described training sample with class label, obtain the light stream direction histogram feature of the described training sample corresponding with class label, for training SVM classifier;
Step 4.3: select gaussian kernel function to carry out the map classification of Nonlinear separability feature, the adjustment exceptional value penalty factor of SVM classifier and the width parameter δ of gaussian kernel function.
Wherein, the process adjusting parameter comprises:
The exceptional value penalty factor of SVM classifier is initialized as 1, the width parameter δ of gaussian kernel function is initialized as 0.01, utilize the class label of described sample and the light stream direction histogram feature of correspondence, call the OpenCV storehouse class function CvSVM::train_auto that increases income and Automatic Optimal is carried out to parameter C and δ, C and δ after being optimized;
To parameter C and δ with 0.1 interval carry out upper and lower value, observe the probability that sample under the different value condition of parameter is correctly classified, the parameter value of selection sort best results carries out final assignment respectively to parameter C and δ.
Step 4.4: utilize the class label of described training sample and the light stream direction histogram feature of correspondence, the parameter value selected after Use Adjustment, call the OpenCV storehouse class function CvSVM::train that increases income to train SVM classifier, optimal separating hyper plane (w, b) is obtained in the feature space that kernel function maps.The training sample of inhomogeneity label makes a distinction by this lineoid, and makes in feature space, and the distance between all samples and this lineoid reaches maximum, uses function w
tx+b=0 represents.Wherein w, b are vector, w
tfor the transposition of vector.
Preferably, described step 5 comprises: to sample to be tested carry out step I 1, step I 2, step I 3 operation obtain the light stream direction histogram feature of sample to be tested afterwards, it can be used as the input of the SVM classifier of having trained, this sample to be tested classified;
Wherein, described classification refers to: by sample to be tested through performing step I 1, step I 2, sample to be tested characteristic use gaussian kernel function that the operation of step I 3 obtains be mapped to high-dimensional feature space, in this feature space, judge this sample to be tested feature is positioned at which side of the SVM classifier optimal separating hyper plane of having trained, determine sample to be tested generic, carry out the judgement of class label;
The judgement formula of described sample characteristics class label is as follows:
In formula: α
irepresent Lagrange multiplier, b is optimal separating hyper plane vector parameter, K (F
i, F
j) represent gaussian kernel function, F
irepresent the histogram feature of i-th training sample in sample space, F
jrepresent the histogram feature of a jth sample to be tested, N is the training sample sum in sample space; S (F
j) represent the decision content of a jth sample to be tested class label, sgn () representative function sign of operation, S (F
j) when being+1, represent that sample to be tested belongs to positive sample class, crowd does not exist abnormal behaviour, S (F
j) when being-1, then representing and belong to negative sample class, in crowd, there is abnormal behaviour.
Preferably, described step 6 comprises: when the duration in video is no more than 0.5s, then think and the behavior that this action does not belong to complete cannot judge that whether this action is normal.
When SVM classifier to the class label decision content of successive frame sequence classification results from+1 become-1 or become+1 from-1 time, the crowd operating state of being designated as changes, think that the crowd behaviour in video just can be confirmed as belonging to current state only when described operating state changes lasting multiple consecutive frame.
Preferably, the alternative condition of described video image is as follows:
Condition A: there is the sport people that individual amount is more than or equal to 2 in the picture of two field picture;
Condition B: comprise crowd's normal behaviour situation and abnormal behaviour situation in the sequence of video images that two field picture is formed;
Feature C: the video camera gathering video image is stationary state.
Preferably, namely dense optical flow field, is divided into the junior unit of multiple equivalent size by described piecemeal process according to width pixel count, each block comprises adjacent four junior units up and down, and there are two junior units to be shared between adjacent block, namely ensure that the area of 50% is overlapping.
Compared with prior art, the present invention has following beneficial effect:
1, the present invention utilizes SVM classifier automatic learning exception and non-abnormal cut-off rule, this cut-off rule is positive optimal separating hyper plane between sample and negative sample in feature space, accurately two class sample area can be separated, decrease human intervention, and optimize by abnormal the error rate reducing sorter classification, improve the accuracy of detection.
2, the present invention is after SVM tagsort, utilizes anomalous identification to optimize thought and adjusts classification results, behavior minimum length in time is set in 0.5s, has both decreased the error rate of sorter classification, in turn ensure that real-time to a certain extent.
3, the present invention is while concern colony local anomaly, has taken into account the correlativity and continuity of move between local and local, make the analysis of motion and expression more accurate.
Accompanying drawing explanation
By reading the detailed description done non-limiting example with reference to the following drawings, other features, objects and advantages of the present invention will become more obvious:
Fig. 1 is the process flow diagram based on the histogrammic group abnormality detection method of dense optical flow provided by the invention;
Fig. 2 is provided by the invention based on two field picture block segmentation schematic diagram in the histogrammic group abnormality detection method of dense optical flow;
Fig. 3 is provided by the invention based on abnormal conditions identification optimization process figure in the histogrammic group abnormality detection method of dense optical flow.
Embodiment
Below in conjunction with specific embodiment, the present invention is described in detail.Following examples will contribute to those skilled in the art and understand the present invention further, but not limit the present invention in any form.It should be pointed out that to those skilled in the art, without departing from the inventive concept of the premise, some distortion and improvement can also be made.These all belong to protection scope of the present invention.
Provided by the invention based on the histogrammic group abnormality detection method of dense optical flow, when the behavior pattern of characterizing motility colony, use based on dense optical flow and calculate motion feature, whether differentiate dyskinesia with this.Method of the present invention is mainly divided into training stage in advance and real-time test phase.For each two field picture in video file, utilize its prior image frame information, adopt GunnarFarneback light stream detection method to calculate the dense optical flow field of image, after piecemeal process is carried out to image, calculate its light stream direction histogram feature, the group behavior be used in description and reflection image.In the training stage, calculate the positive and negative sample characteristics for training and mark class label corresponding to each sample, selected kernel function, trains SVM classifier; At test phase, the calculating that feature carried out to sample to be detected with extract after, this sample belongs to positive class still negative class to utilize the svm classifier model of having trained to judge, then by anomalous identification optimization, classification results is adjusted, obtain final abnormality detection result, when testing result is abnormal, carry out alarm.
Particularly, the video file of input is processed frame by frame, first by every color image frame converting gradation image, and represent by the Mat data structure storage that OpendCV increases income in storehouse, call OpenCV to increase income built-in function calcOpticalFlowFarneback (), i.e. the optical flow algorithm calcOpticalFlowFarneback of GunnarFarneback.Calculate the optical flow field of every two field picture dense optical flow.Analytic target is afterwards the dense optical flow field of acquisition, and described dense optical flow field size is identical with video file resolution, and each element is the bivector of mark movable information, comprises movement velocity size and the direction of motion of pixel.Gained direction of motion is that it, by direction of motion discretize, is on average divided into eight angular interval by the present invention, and each interval represents a direction with the horizontal left direction of two field picture for the arbitrary value within the scope of initial direction counterclockwise-180 ° ~ 180 °.Angle of circumference 8 decile is obtained 8 angular interval, and each angular interval represents a direction.Wherein ,-180 ° all represent the horizontal left direction of two field picture with 180 °, therefore all represent with-180 °.A kth angular interval is designated as b
k, 1≤k≤8, the angular range comprised is 45 ° of (k-1)-180 °≤θ <45 ° of k-180 °.
Particularly, as shown in Figure 2, overlapped some spaces block is obtained to Image Segmentation Using.If image size is w × h, after piecemeal is carried out to it, obtain { B
1, B
2..., B
nbe total to n=n
bw× n
bhblock region, and the overlapping area having 50% between all adjacent block.Note entire image can be divided into n
cw× n
chindividual junior unit, in each junior unit, transverse and longitudinal comprises w respectively
cand h
cindividual pixel, every block region comprises four adjacent junior units and amounts to 2w
c× 2h
cindividual pixel, and share two junior units between adjacent block, then have:
w＝w
c×n
cw,h＝h
c×n
ch；
n
bw＝n
cw-1,n
bh＝n
ch-1；
Wherein, w represents image contained number of pixels in the horizontal, and h represents that image vertically goes up contained number of pixels, B
1, B
2, B
nrespectively represent the 1st, the 2nd, the n-th image block, n
bwrepresent number of blocks transversely after image block, n
bhrepresent the number of blocks that image is vertical, n
cwrepresent the junior unit number that entire image can be got in the horizontal, n
chrepresent the junior unit number that image is vertical, w
crepresent that each junior unit is at the pixel number laterally comprised, h
crepresent the vertical contained pixel number of junior unit.
Then the light stream direction histogram feature of the dense optical flow field computed image calculated is utilized.The present invention adds up based on direction in the statistic processes of light stream, and with the size of pixel light stream amplitude for sum weight is to embody the amplitude of crowd movement.If the light stream amplitude of pixel (x, y) and direction are designated as R (x, y) and α (x, y) respectively, then the sum weight of each pixel on various discrete direction is:
Wherein b
krepresent kth (1≤k≤8) the individual angular interval after dividing.Therefore the light stream direction in each block region
Histogram is all the feature of one 8 dimension.By the integrate features in this n block region together, 8 × n dimensional feature vector F of kth two field picture is combined to form
k.
In the training stage, extract the light stream direction histogram proper vector of the 8 × n dimension calculating its correspondence after obtaining the dense optical flow of each pixel of training video two field picture, note training set proper vector is { F
1, F
2..., F
t, wherein F
1, F
2, F
trepresent the 1st respectively, the 2nd, the proper vector of a t training sample.Each training video frame is carried out to the mark of class label, by the training sample containing crowd's abnormal behaviour as negative sample, marking its class label is-1; All the other training samples are then positive sample, and marking its class label is+1.When utilizing training sample to train SVM classifier, because this is characterized as multi-C vector, be also Nonlinear separability simultaneously, therefore use gaussian kernel function to carry out the map classification of feature, maps feature vectors in the approximately linear feature space that can divide, then can classify to more higher-dimension.Wherein, gaussian kernel function form is:
In formula: gaussian kernel function is expressed as K (x, y), the exponent arithmetic that it is the truth of a matter with constant e that exp () represents, δ represents the width parameter of Gaussian function, any point in x representation feature space, y is the center of Gaussian function, || the norm computing that x-y|| represents (x-y), the i.e. tolerance of amount of orientation.
Gaussian kernel function is selected to carry out the map classification of Nonlinear separability feature, the adjustment exceptional value penalty factor of SVM classifier and the width parameter δ of gaussian kernel function.
Wherein, the process adjusting parameter comprises:
The exceptional value penalty factor of SVM classifier is initialized as 1, the width parameter δ of gaussian kernel function is initialized as 0.01, utilize the class label of described sample and the light stream direction histogram feature of correspondence, call the OpenCV storehouse class function CvSVM::train_auto that increases income and Automatic Optimal is carried out to parameter C and δ, C and δ after being optimized; To parameter C and δ with 0.1 interval carry out upper and lower value, observe the probability that sample under the different value condition of parameter is correctly classified, the parameter value of selection sort best results carries out final assignment respectively to parameter C and δ.
Utilize the class label of described training sample and the light stream direction histogram feature of correspondence, the parameter value selected after Use Adjustment, call the OpenCV storehouse class function CvSVM::train that increases income to train SVM classifier, in the feature space that kernel function maps, obtain one can compared with the division of the optimal separating hyper plane (w, b) of Accurate classification for positive negative sample.The training sample of inhomogeneity label makes a distinction by this lineoid, and makes in feature space, and the distance between all samples and this lineoid reaches maximum, uses function w
tx+b=0 represents.Wherein w, b are vector, w
tfor the transposition of vector.
At test phase, after the calculating of light stream extraction and proper vector is carried out to each two field picture to be detected, it can be used as the input of the SVM classifier of having trained, this sample to be tested is classified;
Wherein, described classification refers to: by sample to be tested through performing step 1, step 2, sample to be tested characteristic use gaussian kernel function that the operation of step 3 obtains be mapped to high-dimensional feature space, in this feature space, judge this sample to be tested feature is positioned at which side of the SVM classifier optimal separating hyper plane of having trained, determine sample to be tested generic, carry out the judgement of class label;
The judgement formula of described sample characteristics class label is as follows:
In formula: α
irepresent Lagrange multiplier, b is optimal separating hyper plane vector parameter, K (F
i, F
j) represent gaussian kernel function, F
irepresent the histogram feature of i-th training sample in sample space, F
jrepresent the histogram feature of a jth sample to be tested, N is the training sample sum in sample space, S (F
j) represent the decision content of a jth sample to be tested class label, the functional operation of sgn () expression symbol, S (F
j) when being+1, represent that sample to be tested belongs to positive sample class, crowd do not exist run away on a large scale, panic four abnormal behaviours such as loose, S (F
j) when being-1, then representing and belong to negative sample class, in crowd, there is abnormal behaviour.Wherein the distinguishing rule of abnormal conditions is as follows: when crowd occur suddenly fairly large acceleration run away or alarmed four fall apart, think that present image comprises abnormal behaviour situation; Otherwise then think and belong to normal behaviour situation.
In order to solve the inevitable misclassification problem of svm classifier, the present invention proposes anomalous identification optimization and carries out adjusting and optimizing to svm classifier result.Because the present invention is applicable to continuous print sequence of frames of video; abnormal or non-abnormal behavior all will continue a bit of time; and the situation of misclassification may cause the class label of this segment successive video frames to occur jumping phenomena back and forth; therefore the present invention is in whether abnormal decision process; the behavior state of colony is defined as "abnormal" and " non-exception " two classes; class label value corresponding to "abnormal" class is+1, and class label value corresponding to " non-exception " class is-1.When SVM classifier to the class label decision content of successive frame sequence classification results from+1 become-1 or become+1 from-1 time, the crowd operating state of being designated as changes, think and only have when this this change continues some consecutive frames (N frame), the group behavior in video formally just can be confirmed as belonging to current state.Especially, behavior state is continued to the determination of frame number, it is considered herein that the action that any duration is no more than 0.5s does not belong to complete behavior, more cannot judge that it is normal behaviour or abnormal behaviour, therefore when final abnormality detection, first the present invention determines the frame per second of video to be measured inputted, according to frame per second to calculate in 0.5s time range comprise the number of frame, namely this frame number is set as the size of N value.
So, for in video sequence, " non-exception " that due to erroneous judgement, intermittence appears in the middle of a succession of abnormal frame can be adjusted, and in non-abnormal frame, emergent several interruption "abnormal" also can be left in the basket, and substantially increases the accuracy that group abnormality behavioral value judges.In this case, when abnormal behaviour occurs in guarded region, only have as its lasting more than 0.5s, just can carry out alarm to it.
Above specific embodiments of the invention are described.It is to be appreciated that the present invention is not limited to above-mentioned particular implementation, those skilled in the art can make various distortion or amendment within the scope of the claims, and this does not affect flesh and blood of the present invention.
Claims (9)
1., based on the histogrammic group abnormality detection method of dense optical flow, it is characterized in that, comprise the steps:
Step 1: the dense optical flow field obtaining at least one width two field picture in video image;
Step 2: obtain multiple image block after carrying out piecemeal process according to the size of described dense optical flow field, obtains the light stream vector of each pixel in each image block;
Step 3: by the direction of motion discretize of pixel, by the light stream discrete--direction of described pixel, number in all angles interval of the light stream vector adding up in each image block the pixel comprised in circumference range, and the amplitude of pixel light stream vector is carried out as this pixel the weights that add up on light stream direction, embody the pixel number of motion pixel and size of instantaneous velocity in each angular interval of current time in image block in the mode of accumulative amplitude; For each width two field picture, all pixels in all images block of combined frame image, in the motion vector statistics in this moment, obtain the light stream direction histogram feature that described two field picture is corresponding;
Step 4: the training sample selecting support vector machines sorter, and described training sample is carried out to the mark of class label, by performing the operation of step I 1, step I 2, step I 3 to the described training sample with class label, obtain the light stream direction histogram feature of the described training sample corresponding with class label.The parameter of adjustment SVM classifier and kernel function, utilize the class label of described training sample and the light stream direction histogram feature of correspondence, SVM classifier is trained, in the feature space that kernel function maps, obtain optimal separating hyper plane, the training sample of inhomogeneity label makes a distinction by this lineoid;
Wherein:
Step I 1: the dense optical flow field obtaining at least one width two field picture in the video image in sample;
Step I 2: obtain multiple image block after carrying out piecemeal process according to the size of described dense optical flow field, obtains the light stream vector of each pixel in each image block;
Step I 3: by the direction of motion discretize of pixel, by the light stream discrete--direction of described pixel, number in all angles interval of the light stream vector adding up in each image block the pixel comprised in circumference range, and the amplitude of pixel light stream vector is carried out as this pixel the weights that add up on light stream direction, embody the pixel number of motion pixel and size of instantaneous velocity in each angular interval of current time in image block in the mode of accumulative amplitude; For each width two field picture, all pixels in all images block of combined frame image, in the motion vector statistics in this moment, obtain the light stream direction histogram feature that described two field picture is corresponding;
Step 5: to sample to be tested carry out step I 1, step I 2, step I 3 operation obtain the light stream direction histogram feature of sample to be tested afterwards, utilize kernel function that histogram feature is mapped to feature space, in this feature space, judge this sample to be tested feature is positioned at which side of the SVM classifier optimal separating hyper plane of having trained, determine sample to be tested generic, whether occur exception with the crowd in this judgment frame image;
Step 6: whether crowd behaviour state duration is occurred an abnormal basis for estimation as the crowd in two field picture, and SVM classifier is corrected the sample to be tested classification results obtained of classifying.
2. according to claim 1 based on the histogrammic group abnormality detection method of dense optical flow, it is characterized in that, described step 1 comprises:
Step 1.1: the video image of input is processed frame by frame, and by every color image frame converting gradation image, and represent by the Mat data structure storage that OpenCV increases income in storehouse;
Step 1.2: call OpenCV and to increase income built-in function calcOpticalFlowFarneback, namely the optical flow algorithm of GunnarFarneback calculates the dense optical flow field of every two field picture; The size of described dense optical flow field is identical with the resolution of video image, and the bivector of each pixel is an all corresponding mark movable information, namely comprises movement velocity size and the direction of motion of pixel.
3. according to claim 1 based on the histogrammic group abnormality detection method of dense optical flow, it is characterized in that, obtain after carrying out piecemeal process according to the size of described dense optical flow field in described step 2 often organizing in multiple image block to include between two adjacent blocks 50% area overlapping.
4. according to claim 1 based on the histogrammic group abnormality detection method of dense optical flow, it is characterized in that, described step 3 comprises:
Step 3.1: by the direction of motion discretize of pixel, by the light stream discrete--direction of described pixel, with the horizontal left direction of two field picture for angle of circumference 8 decile being obtained 8 angular interval within the scope of initial direction counterclockwise-180 ° ~ 180 °, each angular interval represents a direction.
Wherein ,-180 ° all represent the horizontal left direction of two field picture with 180 °, all represent with-180 °.A kth angular interval is designated as b
k, 1≤k≤8, the angular range comprised is 45 ° of (k-1)-180 °≤θ <45 ° of k-180 °.
Step 3.2: dividing in each image block obtained, add up the number of all pixel light stream vectors in all angles interval comprised in described image block, in statistic processes, the amplitude of each pixel light stream vector is carried out as this pixel the weights that add up on the light stream direction of correspondence, embody the pixel number of motion pixel and size of instantaneous velocity in each angular interval of current time in image block in the mode of accumulative amplitude;
Suppose the light stream amplitude of pixel (x, y), direction is designated as R (x, y) respectively, α (x, y), then the sum weight G of this pixel (x, y) on various discrete direction
k(x, y) is:
Wherein b
krepresent the kth angular interval after dividing, 1≤k≤8, x represents the X-axis coordinate figure of pixel under rectangular coordinate system, and y represents the Y-axis coordinate figure of pixel under rectangular coordinate system;
Step 3.3: the accumulated result of each image block is linked in sequence together by block number, all pixels motion vector statistics at a time in all images block of i.e. combined frame image, obtain the light stream direction histogram feature that described two field picture is corresponding, namely obtain the higher-dimension histogram feature of view picture two field picture.
5. according to claim 1 based on the histogrammic group abnormality detection method of dense optical flow, it is characterized in that, described step 4 comprises:
Step 4.1: the training sample selecting SVM classifier, and selected training sample is carried out to the mark of class label, by the training sample containing crowd's abnormal behaviour as negative sample, marking its class label is-1; All the other training samples are then positive sample, and marking its class label is+1;
Step 4.2: by performing step 1 to the operation of step 3 to the described training sample with class label, obtain the light stream direction histogram feature of the described training sample corresponding with class label, for training SVM classifier;
Step 4.3: select gaussian kernel function to carry out the map classification of Nonlinear separability feature, the adjustment exceptional value penalty factor of SVM classifier and the width parameter δ of gaussian kernel function.
Wherein, the process adjusting parameter comprises:
The exceptional value penalty factor of SVM classifier is initialized as 1, the width parameter δ of gaussian kernel function is initialized as 0.01, utilize the class label of described sample and the light stream direction histogram feature of correspondence, call the OpenCV storehouse class function CvSVM::train_auto that increases income and Automatic Optimal is carried out to parameter C and δ, C and δ after being optimized;
To parameter C and δ with 0.1 interval carry out upper and lower value, observe the probability that sample under the different value condition of parameter is correctly classified, the parameter value of selection sort best results carries out final assignment respectively to parameter C and δ.
Step 4.4: utilize the class label of described training sample and the light stream direction histogram feature of correspondence, the parameter value selected after Use Adjustment, call the OpenCV storehouse class function CvSVM::train that increases income to train SVM classifier, optimal separating hyper plane (w, b) is obtained in the feature space that kernel function maps.The training sample of inhomogeneity label makes a distinction by this lineoid, and makes in feature space, and the distance between all samples and this lineoid reaches maximum, uses function w
tx+b=0 represents.Wherein w, b are vector, w
tfor the transposition of vector.
6. according to claim 5 based on the histogrammic group abnormality detection method of dense optical flow, it is characterized in that, described step 5 comprises: to sample to be tested carry out step I 1, step I 2, step I 3 operation obtain the light stream direction histogram feature of sample to be tested afterwards, it can be used as the input of the SVM classifier of having trained, this sample to be tested is classified;
Wherein, described classification refers to: by sample to be tested through performing step I 1, step I 2, sample to be tested characteristic use gaussian kernel function that the operation of step I 3 obtains be mapped to high-dimensional feature space, in this feature space, judge this sample to be tested feature is positioned at which side of the SVM classifier optimal separating hyper plane of having trained, determine sample to be tested generic, carry out the judgement of class label;
The judgement formula of described sample characteristics class label is as follows:
In formula: α
irepresent Lagrange multiplier, b is optimal separating hyper plane vector parameter, K (F
i, F
j) represent gaussian kernel function, F
irepresent the histogram feature of i-th training sample in sample space, F
jrepresent the histogram feature of a jth sample to be tested, N is the training sample sum in sample space; S (F
j) represent the decision content of a jth sample to be tested class label, sgn () representative function sign of operation, S (F
j) when being+1, represent that sample to be tested belongs to positive sample class, crowd does not exist abnormal behaviour, S (F
j) when being-1, then representing and belong to negative sample class, in crowd, there is abnormal behaviour.
7. according to claim 1 based on the histogrammic group abnormality detection method of dense optical flow, it is characterized in that, described step 6 comprises: when the duration in video is no more than 0.5s, then think and the behavior that this action does not belong to complete cannot judge that whether this action is normal.
When SVM classifier to the class label decision content of successive frame sequence classification results from+1 become-1 or become+1 from-1 time, the crowd operating state of being designated as changes, think that the crowd behaviour in video just can be confirmed as belonging to current state only when described operating state changes lasting multiple consecutive frame.
8. according to claim 1 based on the histogrammic group abnormality detection method of dense optical flow, it is characterized in that, the alternative condition of described video image is as follows:
Condition A: there is the sport people that individual amount is more than or equal to 2 in the picture of two field picture;
Condition B: comprise crowd's normal behaviour situation and abnormal behaviour situation in the sequence of video images that two field picture is formed;
Feature C: the video camera gathering video image is stationary state.
9. according to claim 3 based on the histogrammic group abnormality detection method of dense optical flow, it is characterized in that, described piecemeal process namely, dense optical flow field is divided into the junior unit of multiple equivalent size according to width pixel count, each block comprises adjacent four junior units up and down, and there are two junior units to be shared between adjacent block, namely ensure that the area of 50% is overlapping.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510786736.0A CN105389567B (en) | 2015-11-16 | 2015-11-16 | Group abnormality detection method based on dense optical flow histogram |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510786736.0A CN105389567B (en) | 2015-11-16 | 2015-11-16 | Group abnormality detection method based on dense optical flow histogram |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105389567A true CN105389567A (en) | 2016-03-09 |
| CN105389567B CN105389567B (en) | 2019-01-25 |
Family
ID=55421839
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510786736.0A Active CN105389567B (en) | 2015-11-16 | 2015-11-16 | Group abnormality detection method based on dense optical flow histogram |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105389567B (en) |
Cited By (28)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105913008A (en) * | 2016-04-07 | 2016-08-31 | 杭州电子科技大学 | Crowd exceptional event detection method based on hypothesis examination |
| CN106022229A (en) * | 2016-05-11 | 2016-10-12 | 北京航空航天大学 | Abnormal behavior identification method in error BP Adaboost network based on video motion information feature extraction and adaptive boost algorithm |
| CN106022234A (en) * | 2016-05-13 | 2016-10-12 | 中国人民解放军国防科学技术大学 | Abnormal crowd behavior detection algorithm based on optical flow computation |
| CN107194360A (en) * | 2017-05-25 | 2017-09-22 | 智慧航安（北京）科技有限公司 | Inversely pass through object identifying method, apparatus and system |
| CN107230218A (en) * | 2016-03-24 | 2017-10-03 | 德尔菲技术公司 | Method and apparatus for generating the confidence level measurement to estimating derived from the image of the cameras capture on delivery vehicle |
| CN107230226A (en) * | 2017-05-15 | 2017-10-03 | 深圳奥比中光科技有限公司 | Determination methods, device and the storage device of human body incidence relation |
| CN107424170A (en) * | 2017-06-17 | 2017-12-01 | 复旦大学 | Motion feature for detecting local anomaly behavior in monitor video automatically describes method |
| CN107506734A (en) * | 2017-08-28 | 2017-12-22 | 中国民航大学 | One kind of groups unexpected abnormality event detection and localization method |
| CN107886067A (en) * | 2017-11-07 | 2018-04-06 | 中国矿业大学 | A kind of pedestrian detection method of the multiple features fusion based on HIKSVM graders |
| CN108229422A (en) * | 2018-01-25 | 2018-06-29 | 浙江中控技术股份有限公司 | A kind of method and apparatus of image identification |
| CN108509889A (en) * | 2018-03-27 | 2018-09-07 | 哈尔滨工业大学深圳研究生院 | A kind of close shot anomaly detection method and device based on skin color segmentation |
| CN108596045A (en) * | 2018-04-02 | 2018-09-28 | 四川大学 | A kind of group abnormality behavioral value method based on aerial monitor supervision platform |
| CN108596157A (en) * | 2018-05-14 | 2018-09-28 | 三峡大学 | A kind of crowd's agitation scene detection method and system based on motion detection |
| CN109376696A (en) * | 2018-11-28 | 2019-02-22 | 北京达佳互联信息技术有限公司 | Method, apparatus, computer equipment and the storage medium of video actions classification |
| CN109492640A (en) * | 2017-09-12 | 2019-03-19 | 中国移动通信有限公司研究院 | Licence plate recognition method, device and computer readable storage medium |
| CN109697387A (en) * | 2017-10-23 | 2019-04-30 | 北京京东尚科信息技术有限公司 | Direction of motion prediction technique, device, electronic equipment and storage medium |
| CN110490148A (en) * | 2019-08-22 | 2019-11-22 | 四川自由健信息科技有限公司 | A kind of recognition methods for behavior of fighting |
| CN110781853A (en) * | 2019-11-01 | 2020-02-11 | 浙江大华技术股份有限公司 | Crowd abnormality detection method and related device |
| CN110991307A (en) * | 2019-11-27 | 2020-04-10 | 北京锐安科技有限公司 | Face recognition method, device, equipment and storage medium |
| TWI691940B (en) * | 2019-04-02 | 2020-04-21 | 奇美車電股份有限公司 | Vehicle photography system and object detection method |
| CN111401225A (en) * | 2020-03-13 | 2020-07-10 | 河海大学常州校区 | Crowd abnormal behavior detection method based on improved logistic regression classification |
| CN111539301A (en) * | 2020-04-20 | 2020-08-14 | 贵州安防工程技术研究中心有限公司 | Scene chaos degree judging method based on video analysis technology |
| CN112560039A (en) * | 2020-12-25 | 2021-03-26 | 河南交通职业技术学院 | Computer safety protection method |
| CN112800975A (en) * | 2021-01-29 | 2021-05-14 | 中国船舶重工集团公司第七一三研究所 | Behavior identification method in security check channel based on image processing |
| CN112883861A (en) * | 2021-02-07 | 2021-06-01 | 同济大学 | Feedback type bait casting control method based on fine-grained classification of fish school feeding state |
| CN113051971A (en) * | 2019-12-27 | 2021-06-29 | 沈阳新松机器人自动化股份有限公司 | Method and system for estimating number of people in abnormal emergency of group |
| CN114332762A (en) * | 2021-12-24 | 2022-04-12 | 杭州电子科技大学 | Crowd abnormity detection method based on pedestrian group ring block feature extraction |
| CN115880784A (en) * | 2023-02-22 | 2023-03-31 | 武汉商学院 | Scenic spot multi-person action behavior monitoring method based on artificial intelligence |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102682303A (en) * | 2012-03-13 | 2012-09-19 | 上海交通大学 | Crowd exceptional event detection method based on LBP (Local Binary Pattern) weighted social force model |
| CN103745230A (en) * | 2014-01-14 | 2014-04-23 | 四川大学 | Adaptive abnormal crowd behavior analysis method |
| US20140270363A1 (en) * | 2013-03-15 | 2014-09-18 | Sri International | 3D Visual Proxemics: Recognizing Human Interactions in 3D From a Single Image |
| US20150193649A1 (en) * | 2012-02-05 | 2015-07-09 | Apple Inc. | Enhanced face detection using depth information |
| CN104820824A (en) * | 2015-04-23 | 2015-08-05 | 南京邮电大学 | Local abnormal behavior detection method based on optical flow and space-time gradient |
-
2015
- 2015-11-16 CN CN201510786736.0A patent/CN105389567B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20150193649A1 (en) * | 2012-02-05 | 2015-07-09 | Apple Inc. | Enhanced face detection using depth information |
| CN102682303A (en) * | 2012-03-13 | 2012-09-19 | 上海交通大学 | Crowd exceptional event detection method based on LBP (Local Binary Pattern) weighted social force model |
| US20140270363A1 (en) * | 2013-03-15 | 2014-09-18 | Sri International | 3D Visual Proxemics: Recognizing Human Interactions in 3D From a Single Image |
| CN103745230A (en) * | 2014-01-14 | 2014-04-23 | 四川大学 | Adaptive abnormal crowd behavior analysis method |
| CN104820824A (en) * | 2015-04-23 | 2015-08-05 | 南京邮电大学 | Local abnormal behavior detection method based on optical flow and space-time gradient |
Non-Patent Citations (1)
| Title |
|---|
| 余昊等: "基于光流块统计特征的视频异常行为检测算法", 《上海交通大学学报》 * |
Cited By (45)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107230218A (en) * | 2016-03-24 | 2017-10-03 | 德尔菲技术公司 | Method and apparatus for generating the confidence level measurement to estimating derived from the image of the cameras capture on delivery vehicle |
| CN107230218B (en) * | 2016-03-24 | 2020-12-01 | 德尔菲技术公司 | Method and apparatus for generating confidence measures for estimates derived from images captured by vehicle-mounted cameras |
| CN105913008A (en) * | 2016-04-07 | 2016-08-31 | 杭州电子科技大学 | Crowd exceptional event detection method based on hypothesis examination |
| CN105913008B (en) * | 2016-04-07 | 2019-03-01 | 杭州电子科技大学 | Based on the assumption that the crowd's accident detection method examined |
| CN106022229A (en) * | 2016-05-11 | 2016-10-12 | 北京航空航天大学 | Abnormal behavior identification method in error BP Adaboost network based on video motion information feature extraction and adaptive boost algorithm |
| CN106022229B (en) * | 2016-05-11 | 2019-08-27 | 北京航空航天大学 | The abnormal behaviour recognition methods with the Back propagation neural networks of self-adaptive enhancement algorithm is extracted based on video motion information characteristics |
| CN106022234A (en) * | 2016-05-13 | 2016-10-12 | 中国人民解放军国防科学技术大学 | Abnormal crowd behavior detection algorithm based on optical flow computation |
| CN106022234B (en) * | 2016-05-13 | 2021-06-22 | 中国人民解放军国防科学技术大学 | Abnormal congestion behavior detection algorithm based on optical flow calculation |
| CN107230226A (en) * | 2017-05-15 | 2017-10-03 | 深圳奥比中光科技有限公司 | Determination methods, device and the storage device of human body incidence relation |
| CN107194360B (en) * | 2017-05-25 | 2018-07-20 | 智慧航安（北京）科技有限公司 | Reverse current object identifying method, apparatus and system |
| CN107194360A (en) * | 2017-05-25 | 2017-09-22 | 智慧航安（北京）科技有限公司 | Inversely pass through object identifying method, apparatus and system |
| CN107424170A (en) * | 2017-06-17 | 2017-12-01 | 复旦大学 | Motion feature for detecting local anomaly behavior in monitor video automatically describes method |
| CN107424170B (en) * | 2017-06-17 | 2020-07-28 | 复旦大学 | Motion characteristic description method for automatically detecting local abnormal behaviors in monitoring video |
| CN107506734A (en) * | 2017-08-28 | 2017-12-22 | 中国民航大学 | One kind of groups unexpected abnormality event detection and localization method |
| CN107506734B (en) * | 2017-08-28 | 2020-07-24 | 中国民航大学 | Group emergency abnormal event detection and positioning method |
| CN109492640A (en) * | 2017-09-12 | 2019-03-19 | 中国移动通信有限公司研究院 | Licence plate recognition method, device and computer readable storage medium |
| CN109697387B (en) * | 2017-10-23 | 2021-07-30 | 北京京东尚科信息技术有限公司 | Motion direction prediction method and device, electronic equipment and storage medium |
| CN109697387A (en) * | 2017-10-23 | 2019-04-30 | 北京京东尚科信息技术有限公司 | Direction of motion prediction technique, device, electronic equipment and storage medium |
| CN107886067A (en) * | 2017-11-07 | 2018-04-06 | 中国矿业大学 | A kind of pedestrian detection method of the multiple features fusion based on HIKSVM graders |
| CN107886067B (en) * | 2017-11-07 | 2020-04-28 | 中国矿业大学 | Pedestrian detection method based on multi-feature fusion of HIKSVM classifier |
| CN108229422A (en) * | 2018-01-25 | 2018-06-29 | 浙江中控技术股份有限公司 | A kind of method and apparatus of image identification |
| CN108509889A (en) * | 2018-03-27 | 2018-09-07 | 哈尔滨工业大学深圳研究生院 | A kind of close shot anomaly detection method and device based on skin color segmentation |
| CN108596045B (en) * | 2018-04-02 | 2021-08-20 | 四川大学 | Group abnormal behavior detection method based on aerial monitoring platform |
| CN108596045A (en) * | 2018-04-02 | 2018-09-28 | 四川大学 | A kind of group abnormality behavioral value method based on aerial monitor supervision platform |
| CN108596157B (en) * | 2018-05-14 | 2020-08-07 | 三峡大学 | Crowd disturbance scene detection method and system based on motion detection |
| CN108596157A (en) * | 2018-05-14 | 2018-09-28 | 三峡大学 | A kind of crowd's agitation scene detection method and system based on motion detection |
| CN109376696A (en) * | 2018-11-28 | 2019-02-22 | 北京达佳互联信息技术有限公司 | Method, apparatus, computer equipment and the storage medium of video actions classification |
| TWI691940B (en) * | 2019-04-02 | 2020-04-21 | 奇美車電股份有限公司 | Vehicle photography system and object detection method |
| CN110490148A (en) * | 2019-08-22 | 2019-11-22 | 四川自由健信息科技有限公司 | A kind of recognition methods for behavior of fighting |
| CN110781853A (en) * | 2019-11-01 | 2020-02-11 | 浙江大华技术股份有限公司 | Crowd abnormality detection method and related device |
| CN110991307A (en) * | 2019-11-27 | 2020-04-10 | 北京锐安科技有限公司 | Face recognition method, device, equipment and storage medium |
| CN110991307B (en) * | 2019-11-27 | 2023-09-26 | 北京锐安科技有限公司 | Face recognition method, device, equipment and storage medium |
| CN113051971A (en) * | 2019-12-27 | 2021-06-29 | 沈阳新松机器人自动化股份有限公司 | Method and system for estimating number of people in abnormal emergency of group |
| CN111401225A (en) * | 2020-03-13 | 2020-07-10 | 河海大学常州校区 | Crowd abnormal behavior detection method based on improved logistic regression classification |
| CN111401225B (en) * | 2020-03-13 | 2022-08-30 | 河海大学常州校区 | Crowd abnormal behavior detection method based on improved logistic regression classification |
| CN111539301A (en) * | 2020-04-20 | 2020-08-14 | 贵州安防工程技术研究中心有限公司 | Scene chaos degree judging method based on video analysis technology |
| CN111539301B (en) * | 2020-04-20 | 2023-04-18 | 贵州安防工程技术研究中心有限公司 | Scene chaos degree discrimination method based on video analysis technology |
| CN112560039B (en) * | 2020-12-25 | 2023-04-18 | 河南交通职业技术学院 | Computer safety protection method |
| CN112560039A (en) * | 2020-12-25 | 2021-03-26 | 河南交通职业技术学院 | Computer safety protection method |
| CN112800975A (en) * | 2021-01-29 | 2021-05-14 | 中国船舶重工集团公司第七一三研究所 | Behavior identification method in security check channel based on image processing |
| CN112883861A (en) * | 2021-02-07 | 2021-06-01 | 同济大学 | Feedback type bait casting control method based on fine-grained classification of fish school feeding state |
| CN112883861B (en) * | 2021-02-07 | 2022-06-14 | 同济大学 | Feedback type bait casting control method based on fine-grained classification of fish school feeding state |
| CN114332762A (en) * | 2021-12-24 | 2022-04-12 | 杭州电子科技大学 | Crowd abnormity detection method based on pedestrian group ring block feature extraction |
| CN115880784A (en) * | 2023-02-22 | 2023-03-31 | 武汉商学院 | Scenic spot multi-person action behavior monitoring method based on artificial intelligence |
| CN115880784B (en) * | 2023-02-22 | 2023-05-09 | 武汉商学院 | Scenic spot multi-person action behavior monitoring method based on artificial intelligence |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105389567B (en) | 2019-01-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105389567A (en) | Group anomaly detection method based on a dense optical flow histogram | |
| US9058744B2 (en) | Image based detecting system and method for traffic parameters and computer program product thereof | |
| Bas et al. | Automatic vehicle counting from video for traffic flow analysis | |
| JP7272533B2 (en) | Systems and methods for evaluating perceptual systems | |
| CN102289686B (en) | Method for identifying classes of moving targets based on transfer learning | |
| CN108230364B (en) | Foreground object motion state analysis method based on neural network | |
| CN103902966B (en) | Video interactive affair analytical method and device based on sequence space-time cube feature | |
| CN105608431A (en) | Vehicle number and traffic flow speed based highway congestion detection method | |
| KR101735365B1 (en) | The robust object tracking method for environment change and detecting an object of interest in images based on learning | |
| CN108009473A (en) | Based on goal behavior attribute video structural processing method, system and storage device | |
| CN104820824A (en) | Local abnormal behavior detection method based on optical flow and space-time gradient | |
| CN104156942A (en) | Detection method for remnants in complex environment | |
| CN102496058B (en) | Passenger flow density detection method | |
| CN105138982A (en) | Crowd abnormity detection and evaluation method based on multi-characteristic cluster and classification | |
| CN103034870B (en) | The boats and ships method for quickly identifying of feature based | |
| CN111626170B (en) | Image recognition method for railway side slope falling stone intrusion detection | |
| CN105788269A (en) | Unmanned aerial vehicle-based abnormal traffic identification method | |
| CN111507235B (en) | Railway perimeter foreign matter intrusion detection method based on video | |
| CN105405150A (en) | Abnormal behavior detection method and abnormal behavior detection device based fused characteristics | |
| Gong et al. | Local distinguishability aggrandizing network for human anomaly detection | |
| CN102982313A (en) | Smog detecting method | |
| KR101472674B1 (en) | Method and apparatus for video surveillance based on detecting abnormal behavior using extraction of trajectories from crowd in images | |
| CN110674887A (en) | End-to-end road congestion detection algorithm based on video classification | |
| CN102087790A (en) | Method and system for low-altitude ground vehicle detection and motion analysis | |
| CN107909009A (en) | A kind of obstacle detection method and device based on road surface study |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |