Detection method based on the illegal mining of video river courseTechnical field
The invention belongs to the technical field of video monitoring is and in particular to detection method based on the illegal mining of video river course.
Background technology
In safety monitoring, the application of video monitoring is quite varied, leads to multipair video and is analyzed it can be determined that going out motionThe way of act of target, and then warning reminding is carried out to the Deviant Behavior of target, to prevent accident.
It is used widely based on the warning function of video, but due to by complicated outdoor environment, light, day and night changingInterference, the false alarm that simple function produces and to fail to report alert problem ratio more serious, which prevent the illegal mining of video river course and steal and unload inspectionThe large-scale use surveyed.
Content of the invention
The present invention is to solve the problems, such as that prior art proposes, and its objective is to provide one kind to be based on the illegal mining of video river courseDetection method.
The technical scheme is that a kind of detection method based on the illegal mining of video river course, comprise the following steps:
() carries out the time domain modeling of surrounding neighbors to each pixel in image, judges foreground image；
() utilizes the foreground image in step (), carries out mass detection, adjacent foreground point is carried out subject fusion, obtainsCandidate target region；
() detects target number and type in step () target area with model method；
According to default rule, movement locus and the time of staying, () judges whether each target is alarm target.
Described step () when time domain modeling is carried out to each pixel, by this pixel value and its time domain Gauss model pointCloth compares, and when this pixel value changes exceedes the standard deviation of 3 times of Gauss model, judges this pixel for foreground pixel point, otherwise forBackground pixel；Update the Model in Time Domain of each pixel by fixing frame per second simultaneously.
According to the spatial relationship between foreground pixel point, each foreground pixel point is clustered nearby；The result of clusterThen as the result of agglomerate, i.e. candidate target.
Using polymorphic type grader, candidate target is carried out with traversal detection in described step () model method and obtain eachThe target number of candidate target and target type.
The result that described polymorphic type grader produces is: sand dredger type, floating thing type, other types；Traversal detectionMethod as follows:
A trains: carries out characteristic vector pickup to sand dredger sample, floating thing sample and negative sample, and is input to polymorphic type classificationCarry out machine learning in device, preserve sample pattern；
B identifies: input picture, and extracts the characteristic vector of candidate target, and grader loads sample pattern the feature to inputVector is identified, to judge the type of target.
Judge whether the target type being drawn by step () suits the requirements the target type of warning, target type meetsWhen, continue to judge this target trajectory and default rule relation, when the target stay time meeting alert if, produce and report to the police.
Described rule, with the polygonal profile of Mulit-point Connection one-tenth as basis for estimation, judges including target type, target is movedTrack and the judgement of target stay time.
The river course illegal mining based on video for the present invention is stolen and is unloaded in detection method, each pixel in image is carried out time domain modeling withJudge foreground point, neighbouring foreground point is carried out subject fusion and obtains candidate target region, by polymorphic type grader to targetCarry out target classification, then sorted target and its movement locus and the time of staying are judged, thus improve targetInvade the accuracy judging, it is to avoid in alarm detection, report and fail to report the generation of situation by mistake.
Brief description
Fig. 1 is method of the present invention flow chart；
Fig. 2 is the pixel value schematic diagram in the present invention, a certain pixel being carried out during time domain modeling；
Fig. 3 is the traversal detection process schematic diagram of polymorphic type grader in the present invention；
Fig. 4 is the judgement schematic diagram that the present invention implements to report to the police.
Specific embodiment
Hereinafter, referring to the drawings and embodiment the present invention is described in detail:
As shown in figure 1, a kind of detection method based on the illegal mining of video river course, comprise the following steps:
() carries out the time domain modeling of surrounding neighbors to each pixel in image, judges foreground image；
() utilizes the foreground image in step (), carries out mass detection, adjacent foreground point is carried out subject fusion, obtainsCandidate target region；
() detects target number and type in step () target area with model method；
According to default rule, movement locus and the time of staying, () judges whether each target is alarm target.
Described step () when time domain modeling is carried out to each pixel, by this pixel value and its time domain Gauss model pointCloth compares, and when this pixel value changes exceedes the standard deviation of 3 times of Gauss model, judges this pixel for foreground pixel point, otherwise forBackground pixel；Update the Model in Time Domain of each pixel by fixing frame per second simultaneously.
As shown in Fig. 2 this pixel value changes is 80 in this example, and the standard deviation of its Gauss modeling is 20, visually this picturePlain value is undergone mutation, and regards as foreground pixel point；Update the Model in Time Domain of each pixel by fixing frame per second simultaneously, up-to-date to obtainPixel value.
According to the spatial relationship between foreground pixel point, each foreground pixel point is clustered nearby；The result of clusterThen as the result of agglomerate, i.e. candidate target.
Using polymorphic type grader, candidate target is carried out with traversal detection in described step () model method and obtain eachThe target number of candidate target and target type.
As shown in figure 3, the result that described polymorphic type grader produces is: sand dredger type, floating thing type, other classesType；The method of traversal detection is as follows:
A trains: carries out characteristic vector pickup to sand dredger sample, floating thing sample and negative sample, and is input to polymorphic type classificationCarry out machine learning in device, preserve sample pattern；
B identifies: input picture, and extracts the characteristic vector of candidate target, and grader loads sample pattern the feature to inputVector is identified, to judge the type of target.
Judge whether the target type being drawn by step () suits the requirements the target type of warning, target type meetsWhen, continue to judge this target trajectory and default rule relation, when the target stay time meeting alert if, produce and report to the police.
The polygonal profile that default rule is become with Mulit-point Connection, as basis for estimation, judges including target type and target fortuneFlowing mode judges, target motion mode judges to include judging whether the movement locus of target enter above-mentioned polygonal profile, sentenceWhether the movement locus of disconnected target leave above-mentioned polygonal profile or movement locus the stopping in polygonal profile judging targetStay the time.
As shown in figure 4, judging target trajectory in the range of rule settings, during the stop of the type of target and targetBetween all reach alert if, target can produce illegal mining report to the police.
The river course illegal mining based on video for the present invention is stolen and is unloaded in detection method, each pixel in image is carried out time domain modeling withJudge foreground point, neighbouring foreground point is carried out subject fusion and obtains candidate target region, by polymorphic type grader to targetCarry out target classification, then sorted target and its movement locus and the time of staying are judged, thus improve targetInvade the accuracy judging, it is to avoid in alarm detection, report and fail to report the generation of situation by mistake.