CN105678730A - Camera movement self-detecting method on the basis of image identification - Google Patents

Abstract

The present invention discloses a camera movement self-detecting method on the basis of image identification. The reliability and the accuracy of camera self detection may be effectively improved through an image matching identification technology. Through combination of partial characteristics and a global model, the camera movement self-detecting method on the basis of image identification may accurately estimate the state of the camera so as to determine whether the camera is moved or not, and the camera movement self-detecting method on the basis of image identification comprises an image obtaining module, a feature extraction module, an image retrieval module, an accurate matching module, a filtering module and a movement determination module. The method comprises the steps: (1) collecting static scenes at different times and different illuminations, and taking the collected static scenes as a set of reference images; (2) performing feature extraction of the set of reference images and real-time images; (3) retrieving images being the most similar to the real-time images in the set of the reference images; (4) performing accurate matching of the real-time images and the retrieved images; (5) effectively identifying the matching points, removing noise interference and choosing appropriate feature points; and (6) calculating corresponding transformation matrixes through adoption matching key points, and determining whether the camera is moved or not.

Description

A kind of camera based on image recognition moves self-sensing method

Technical field

The invention belongs to technical field of video monitoring, relate to a kind of monitoring system, more particularly, relate to a kind of camera movement detection system in monitoring system; Meanwhile, the invention still further relates to camera movement detection method in a kind of monitoring system. By the collection of real time imaging, the slow of camera in the method Inspection and monitoring system of image recognition is adopted to move.

Background technology

Monitoring camera is widely used in public safety, the effect escort is played for social security, continuous increase along with monitoring grid, the camera disposed in network is on the increase, camera in monitoring system gets more and more, it is impossible to arrange enough security protection personnel to carry out the duty of each camera is checked. A lot of cameras due to long-time unmanned maintenance cannot normal operation, as interference, empty burnt, block, thus causing that monitoring cannot be carried out. Check the duty of camera in real time, need the intellectuality of camera itself, the content of this Intelligent Measurement includes: camera lens lose Jiao, the slowly mobile of camera lens causes scene fade, camera lens have accumulated substantial amounts of dust, shakiness, the flating that strong wind or vibrations cause, swinging of signal that loose contact causes or signal disturbing etc. are installed.

Camera slowly movement is the most important also modal detection of one of which. In the intelligent monitor system of smart city, owing to wind, rain, vehicle are through factors such as vibrations, installations, CCTV camera is after using after a while, and what its position and attitude can occur slowly drifts about so that camera monitoring region changes. Especially for the collaborative supervision system of multiple-camera, camera moves and can cause whole monitoring thrashing. In this case it is necessary to camera moves detection.

In the practical application of monitoring camera, some extraneous factors can affect the performance of automatic system. These factors include: weather (wind, snow and rain), camera are arranged and other environmental conditions.Generally, typical background subtracting method (backgroundsubtractionmethod) is (such as, " AdaptiveBackgroundMixtureModelsforReal-timeTracking.CVPR; 1999; 2:246-252; StaufferC, GrimsonWEL., disclosed in background subtracting method) need camera absolute stability. But, due to the reason of wind, storm, truck process or even manual operation, it is impossible to ensure that camera is static. In this case, it is also desirable to camera moves detection.

Owing to outdoor camera is fixed on The Cloud Terrace, long-term wind and weather Exposure to Sunlight, screw corrosion causes the fixing shakiness of camera, and slightly wind or earth shock cause picture moving, have impact on the effect of video record, more affect the intelligent video analysis in later stage. Some cameras are fixed on the shelf that some are higher, are subject to support attachment vibration influence, are such as arranged on bridge or camera overhead, railway limit, it is easy to be subject to through automobile, the earth shock that brings of train and shake occurs. It addition, some criminal is in order to be able to avoid camera video recording, have a mind to slowly move to outside monitored picture camera so that monitored picture cannot monitor crime fact. Present invention is generally directed to the slow mobile of camera, adopt image processing techniques, be susceptible to the interference of light change, the movement of camera can be used for quickly detecting, and make warning in time.

Existing being used for detects whether picture occurs that the method for motion includes: Chinese patent CN200580032022(detects the system of camera movement and includes detecting the photographing unit of the system of camera movement), adopt optical spectroscopic principle detection civil camera rocking in shooting; The method of Chinese patent CN01124045(digital camera detection moving) it is used for detecting in scene whether have moving object, thus triggering some events; Chinese patent CN200410101568(device for detecting motion vector and moving picture camera) it is carry out moving region detection to reduce reference frame image transmission.

But, whether detection picture occurs that the classic algorithm of motion is to adopt adjacent image calculus of finite differences, and this algorithm speed is fast, but is susceptible to the impact of light change, therefore there has been proposed a lot of innovatory algorithm. wherein, the patent No. 200710165357, name is called: detection suppress camera to move the equipment of impact and the method for generation in monitoring system, belongs to Beijing Samsung communication Technology Research Co., Ltd of Samsung Electronics Co., Ltd. this patent of invention is for the false alarm problem in motion detection, when being detected as the object of which movement in scene due to the movement of camera and produce false alarm, it is judged that whether the type of this warning is cause owing to camera moves, thus reaching to suppress this false alarm. although the method that this invention uses is that in tracing figure picture, the excursion of pixel value judges whether to be moved, but whether the same position pixel intensity value that the basis of its detection method is still based between two width images changes to judge, simply this invention make use of Gaussian statistics model, a statistics extent of competence is set up for each pixel, what change within the specific limits is still used as background, although therefore this invention can partly suppress noise, illumination etc. affect, but light change is more sensitive, especially when the change of environment light is too fast, such as thunder and lightning, sunlight changes, car light irradiation, street lamps etc. all can affect its algorithm performance, cause erroneous judgement.

It is that the change according to pixel value judges whether camera moves that traditional camera moves self-sensing method major part, this method to illumination, noise, moving target move in and out all very sensitive, it is easy to occur erroneous judgement. In order to solve this problem, the mode that the present invention proposes to adopt local feature to combine with world model, to estimate camera status, thus judging whether mobile, has superiority compared to traditional approach in robustness and adaptability.

Summary of the invention

The present invention is directed to the deficiency of prior art, the present invention provides a kind of camera movement detection system in monitoring system based on image recognition, the mode combined by local feature and world model estimates camera status accurately, detect that in monitoring system, the slow of camera moves, environmental light intensity change is insensitive.Compared with traditional method, the present invention has robustness and adaptive feature, it is not necessary to judges from the change of pixel scale whether camera occurs mobile, also avoids the problem that context update arrives in Pixel-level change-detection.

For solving above-mentioned technical problem, the present invention adopts the following technical scheme that, described system includes:

(1) image capture module, in order to catch current real time imaging;

(2) characteristic extracting module, in order to carry out feature extraction to the image caught and reference picture;

(3) image retrieval module, in order to retrieve the image the highest with real time imaging similarity in reference set image;

(4) accurate matching module, in order to real time imaging accurately to be mated with the image that is retrieved, provides corresponding matching double points;

(5) filtration module, in order to filter out the characteristic point of erroneous matching, removes noise jamming.

(6) move judge module, utilize the key point of coupling to find out corresponding transformation matrix, judge whether phase machine moves by world model.

This method has considered long illumination variation, sets up the Multi reference images of this scene under different time points different illumination conditions, then passes through retrieval, it is thus achieved that the global state of camera. This method is that the mode combined with world model by local feature estimates camera status accurately, and without judging whether camera slow movement occurs from the change of pixel scale, its advantage shows the following aspects:

(1) illumination variation, effect of noise are not easy erroneous judgement occur by this method, it is achieved robustly camera moves detection, and adaptability is fine;

(2) compared to traditional method, it is to avoid context update brings in Pixel-level change-detection problem;

(3) feature-based matching algorithm is affected little by geometry deformation and grey scale change, and the characteristic type that can mate is many and matching accuracy rate is high.

Accompanying drawing explanation

Fig. 1 is the whole invention flow chart that the present invention detects the slow moving method of camera.

The words tree that Fig. 2 is the present invention sets up block diagram.

Fig. 3 is the real time imaging retrieval block diagram of the present invention.

Specific implementation method

Consulting Fig. 1, present invention is disclosed a kind of camera movement detection system in monitoring system, described system includes the image capture module being sequentially connected with, characteristic extracting module, image retrieval module, accurate matching module, filtration module, mobile judge module. Below the technology of the present invention method is explained in further detail, comprises the steps:

Step one: image capture module

(1.1) the static scene figure under different illumination is gathered every given interval, it is thus achieved that the Multi reference images of this scene is built as reference picture set, the data base under vertical static scene monitoring system;

(1.2) real time imaging of this collected by camera is gathered every set time section (such as a hour);

Step 2: characteristic extracting module.

(2.1) real time imaging and reference set image are extracted characteristic vector respectively, as SIFT feature vector, SURF characteristic vector etc. can be adopted, and save as tag file.

Step 3: image retrieval module

(3.1) consulting Fig. 2, read in the tag file of all images in reference set, characteristic vector carries out layering k-means cluster, arranges the size of the height L of branching factor K and tree, each cluster centre is a visual vocabulary, generates the words tree of not weighting.

(3.2) utilize term frequency-inverse document frequency (TF-IDF) to add inverted index to each visual vocabulary, obtain the inverted index file of words tree visual vocabulary, including vocabulary and Inverted List, generate the words tree of weighting.

(3.3) consult Fig. 3, each characteristic vector of real time imaging is quantified the visual vocabulary to words tree, the Inverted List according to visual vocabulary, calculate the similarity of respective image in reference set.

(3.4) reference picture of highest scoring, namely the highest with real time imaging similarity image are selected.

Step 4: accurately matching module.

(4.1) real time imaging and the reference picture that is retrieved accurately are mated.

Step 5: filtration module.

(5.1) randomly select 4 matching double points, calculate transformation matrix M according to these 4 matching double points.

(5.2) according to transformation matrix M, calculate the consistent collection meeting current transform matrix with error metrics function, and element number is unanimously concentrated in return.

(5.3) judge whether that optimum (maximum) unanimously collects according to current consistent concentration element number, if then updating currently most consistent collection.

(5.4) update current erroneous Probability p, if p is more than the minimum error probability allowed, repeats (5.1) to (5.4) and continue iteration, until current erroneous Probability p is less than minimum error probability.

(5.5) deletion error match point, exports correct match point set.

Step 6: mobile judge module.

(6.1) key point of above coupling is utilized to find out corresponding transformation matrix.

(6.2) input picture is projected to reference picture place plane according to transformation matrix, calculate the overlapping area of projection image and reference picture, if overlapping area is less than giving less than setting threshold value, then judge that camera does not move, otherwise judge that camera moves, the direction of output mobile and angle.

Claims (3)

1. one kind is moved self-check system based on the camera of image recognition, it is characterised in that: the mode combined by local feature and world model estimates camera status accurately, detects that in monitoring system, the slow of camera moves, and environmental light intensity change is insensitive.

2. need not judge from the change of pixel scale whether camera occurs mobile, also avoid the problem that context update arrives in Pixel-level change-detection.

3. comprise the steps:

Image capture module, in order to catch current real time imaging;

Characteristic extracting module, in order to carry out feature extraction to the image caught and reference picture;

Image retrieval module, in order to retrieve the image the highest with real time imaging similarity in reference set image;

Accurate matching module, in order to real time imaging accurately to be mated with the image that is retrieved, provides corresponding match point;

Filtration module, in order to filter out the characteristic point of erroneous matching, removes noise jamming;

Mobile judge module, utilizes the key point of coupling to find out corresponding transformation matrix, judges whether phase machine moves by world model.