Personal identification method and system based on uniform characteristic
Technical field
The present invention relates to a kind of personal identification method and system.
Existing identification system generally all is applicable to more in-plant situation, for example recognition of face, fingerprint recognition, iris recognition etc. all only at object to be identified nearer or contact just work when obtaining equipment, and all be the identification of carrying out the individual.Under remote and non-contacting condition, identity recognizing technology commonly used all can't play a role for object to be identified.In the application as public security analysis at the prison, battlefield enemy and we's situation differentiation etc., need discern the identity of the personnel in the target area under the condition of remote and large scene, this moment, recognition methods commonly used can't be used.
The technical problem to be solved in the present invention provides a kind of personal identification method and system based on uniform characteristic, can enlarge based on the scope of application and the identification of the identification identity of uniform characteristic more accurate.
In order to address the above problem, the invention provides a kind of personal identification method of multidimensional uniform characteristic, be used for the occasion that the multiclass recognition objective is worn different uniforms, described uniform characteristic comprises color, pattern and the pattern of uniform; This method may further comprise the steps:
(a) gather video image, detect recognition target image wherein and it is split;
(b) extract the uniform characteristic of described recognition target image; Wherein, the described uniform characteristic of extraction comprises the main feature of uniform, perhaps comprises the main feature and time feature of uniform;
(c) described uniform characteristic is classified, determine the identity of recognition objective according to classification results.
Further, above-mentioned personal identification method also can have following characteristics:
Step (a) is to detect moving target and it is split from video image, as described recognition target image.
Further, above-mentioned personal identification method also can have following characteristics:
The uniform characteristic that extracts in the step (b) comprises the main feature and time feature of uniform, in the step (c) described main feature and time feature is classified respectively, and two classification results to output are weighted combination then, obtain final classification results.
Further, above-mentioned personal identification method also can have following characteristics:
The master of described uniform is characterized as color, the inferior pattern that is characterized as of described uniform.
Further, above-mentioned personal identification method also can have following characteristics:
Step (c) is divided the assembled classifier of time-like employing based on neural network, earlier this assembled classifier is trained before in step (a), and this training process may further comprise the steps:
(I) input comprises the sample image of the training usefulness of different uniforms, therefrom extracts main feature of uniform and uniform time feature;
(II) the primary and secondary feature is imported main feature neural network,, carried out classification based training respectively inferior feature input time feature neural network;
(III) discrimination as obtaining expecting under given sample, then the combination neural net that this is trained is as the assembled classifier of reality use.
Further, above-mentioned personal identification method also can have following characteristics:
Step (b) also extracts the positional information of recognition objective, and the identity information that step (c) obtains is together as the foundation of making a strategic decision.
The identification system of multidimensional uniform characteristic provided by the invention comprises video image acquisition module, recognition objective detection and cuts apart module, characteristic extracting module and identification module that described uniform characteristic comprises color, pattern and the pattern of uniform; Wherein:
Described video image acquisition module is used for video image is gathered and outputed to described recognition objective detection and cuts apart module;
Described recognition objective detects and cuts apart module, is used for detecting the recognition target image that comprises uniform and it being split from video image, outputs to described characteristic extracting module; Wherein, the described uniform characteristic of extraction comprises the main feature of uniform, perhaps comprises the main feature and time feature of uniform;
Described characteristic extracting module is used to extract the uniform characteristic of described recognition target image, outputs to described sorter;
Described identification module is used for described uniform characteristic is classified, and determines the identity of recognition objective again according to classification results.
Further, above-mentioned identification system also can have following characteristics:
It is to detect moving target and it is split from video image that described recognition objective detects with cutting apart module, as described recognition target image.
Further, above-mentioned identification system also can have following characteristics:
The uniform characteristic that described characteristic extracting module is extracted comprises the main feature and time feature of uniform; Described identification module is an assembled classifier based on neural network, be used for described main feature and time feature extracted are classified respectively, two classification results to output are weighted combination then, obtain final classification results, determine the identity of recognition objective again according to this classification results.
Further, above-mentioned identification system also can have following characteristics:
The uniform master that described characteristic extracting module is extracted is characterized as color, and uniform is characterized as pattern.
The present invention is by in monitor video, the video of gathering as the monitor video in the prison, scounting aeroplane on the battlefield etc., specific uniform crowd and position thereof are worn in identification, can detect whether there is specific behavior, in time provide advisory opinion accurately for the decision maker.
Description of drawings
Fig. 1 is the synoptic diagram of the embodiment of the invention based on the assembled classifier of a plurality of neural networks.
Fig. 2 is the synoptic diagram of embodiment of the invention training process.
Fig. 3 is the synoptic diagram of embodiment of the invention identifying.
Fig. 4 is the module map of embodiment of the invention identification system.
Embodiment
In remote large scene, there are the lower characteristics of many people and resolution, we only need carry out roughly Classification and Identification to their identity, as police or prisoner etc.Under the lower situation of this resolution, features such as the color of clothes, pattern with respect to people's face, iris etc. still clearly; And in this class application scenarios, personnel wear unified uniform, and we just can utilize the feature such as color, pattern of uniform to discern, distinguish personnel's identity like this.The identity recognizing technology that the present invention that Here it is proposes based on uniform characteristic.
Be example with prison supervision automated system below, the present invention is described in detail in conjunction with the accompanying drawings.
Present embodiment has at first defined a kind of combination uniform characteristic.Uniform, for example the clothes of supervisor's clothes and person under surveillance such as convict exist very big difference on design and color.Therefore under the situation of remote large scene, color characteristic is topmost feature, therefore extracts color average in the region unit of target or main color etc. as main feature (Main Feature, F
M).The profile of some patterns also can be distinguished different objects under remote situation, number on grid, striped, the football shirt etc. for example, but because the projection distortion of pattern on camera plane tends to bring mistake to know too greatly, therefore pattern characteristics is subdued inferior feature (Sub Feature, the F of identification as us
S).Then, two kinds of features are made up obtain assemblage characteristic, be defined as " combination multidimensional uniform characteristic ":
F={F
M,F
S}.
The module that is used for identification can adopt multiple sorter, as linear range sorter, neural network or the like, can design proper classifier as the case may be.A kind of assembled classifier based on a plurality of neural networks is proposed in the present embodiment, the classification that realizes making up the multidimensional uniform characteristic, the sorter structure is seen Fig. 1.Wherein, main feature and time feature are imported main feature neural network classifier, inferior feature neural network classifier respectively, then the output of two sorters are weighted combination, have just obtained last classification output.Here can adopt BP network etc.μ among the figure
M, μ
SBe weight factor, can set according to application scenarios.
Before discerning, need be that assembled classifier is trained to identity recognition device, please refer to Fig. 2, this training process specifically may further comprise the steps:
Step 210, input comprises the sample image of the training usefulness of different uniforms;
Can be the sample image of the clothes of supervisor's clothes and convict in the present embodiment.
Step 220, extracting color characteristic from the sample image of input uniform is main feature, extracts pattern characteristics and is time feature;
Step 230 is carried out classification based training to main feature neural network and time feature neural network respectively;
Carrying out classification based training is in order to obtain optimum network parameter by adjustment.
Step 240, the discrimination as obtaining expecting under given sample, then the combination neural net that this is trained is as the sorter of reality use.
In identifying, because liveried number is generally many in the large scene, so goes out the target of motion according to the feature detection of video in the present embodiment, from moving target, leach human body then, extract the uniform characteristic of human body, the input assembled classifier is discerned.
With reference to Fig. 3, this identifying may further comprise the steps:
Step 310 is gathered video image;
Step 320 detects moving target and it is split from video image;
Under the present embodiment environment, moving target is generally the personnel that will discern its identity.Therefore need it is detected, splits and carry out follow-up processing.In other embodiments, also can detect moving target by alternate manner.
Step 330 is carried out human body target and is detected, and extracting the main feature of uniform from the image of each moving target respectively is that color and uniform time feature are pattern;
Step 340 is input to assembled classifier based on nerve net respectively with the main feature of uniform and time feature, obtains classification results, can determine the identity of moving target according to classification results.
Wherein assembled classifier is that the main feature and time feature of input are classified respectively, again two classification results is weighted combination, gets classification results to the end.
In addition, can also extract the positional information of moving target and its identity information together as the foundation of decision-making in step 330.
Accordingly, present embodiment comprises based on the identification system that makes up the multidimensional uniform characteristic as shown in Figure 4 with lower module:
The video image acquisition module is used for video image is gathered and outputed to moving object detection and cuts apart module;
Moving object detection with cut apart module, be used for detecting moving target and it being split from video image, output to characteristic extracting module;
Characteristic extracting module, be used to extract the main feature of uniform and the uniform time feature of each movement destination image, be input to assembled classifier then, wherein main feature is input to main feature neural network classifier wherein, and inferior feature is input to inferior feature neural network classifier wherein;
Assembled classifier, comprise main feature neural network classifier, inferior feature neural network classifier and assembled unit, be respectively applied for the main feature and time feature of input are classified, the classification results of two sub-classifier outputs is weighted combination at assembled unit again, obtain final classification results, can determine the identity of moving target according to classification results.
Utilize above-mentioned recognition methods, can judge any the class people that has who occurs in the target area, also can learn the residing position of these people, thereby can judge whether to have taken place abnormal conditions.
What need describe is in another embodiment, can for example be that color master feature is extracted, trained and discerns only to a main feature also.In addition, under different application scenarios, subdue main feature and also be not limited to color, inferior feature also is not limited to pattern, and both can exchange or adopt the further feature of uniform, as shape facility.