CN110414558A - Characteristic point matching method based on event camera - Google Patents

Abstract

The invention discloses a kind of methods to the feature point extraction having detected that description, and are matched using description of generation to characteristic point.It is an object of the invention to the feature point description subalgorithms for solving the problems, such as traditional possibly can not stablize suitable for event camera, the present invention, which describes son to feature point extraction using the timestamp information of event camera, can preferably utilize the advantage of event camera, keep description sub-information more abundant, keeps matching result more acurrate.

Description

Characteristic point matching method based on event camera

Technical field

The invention belongs to field of image processings, for generating feature point description based on event camera, and carry out feature Point matching.

Background technique

Machine vision depends on the camera based on frame, these traditional cameras are come with regular time exposure and frame rate Entire frame is obtained, stores and handle image information in the matrix form.This simple image storage format for image procossing and May be unsatisfactory for feature extraction, this is largely because the image based on gray scale and contains many redundancies. Pixel intensity information is particularly useful for the identification and retrieval of human eye, however but increases the difficulty of the image procossing based on machine Degree.Meanwhile sequence image reading may adversely affect image processing hardware, because necessary before the feature needed for obtaining Handle a large amount of unnecessary data.General camera is more sensitive to illumination variation simultaneously, is easy in face of high dynamic illumination scene Dark or overexposure, under high-speed motion state, motion blur or imaging effect caused by ordinary optical image are bad, these are all serious Influence picture quality.Ordinary optical camera imaging is as shown in Figure 1 under high-speed motion state.

Event camera has obtained more and more concerns in field of machine vision, it be it is a kind of imitate human retina it is new Type visual sensor, pixel array face by light intensity change triggers output occur light intensity variation pixel position, the time with Intensity, therefore the sequence of frames of video of its output not instead of standard camera, a series of asynchronous event streams.More specifically, in t_jWhen It carves, u_j=(x_j, y_j) brightness gain on location of pixels reaches threshold value ± c (c > 0), then an event e_j=(x_j, y_j, t_j, p_j) will be triggered, p_j∈ {+1, -1 } indicates that the polarity positive sign of event indicates that brightness increases, and negative sign indicates that brightness reduces, therefore thing The output of part camera is asynchronous event stream, since only recording increment changes event, so the absolute brightness value of scene is no longer Directly visible.This camera is not limited by traditional time for exposure and frame rate, and time coordinate precision can reach microsecond Rank effectively filter background redundancy, absorbed capture camera motion information can save data transfer bandwidth and deposit with data are reduced Store up pressure, with high dynamic, low delay, low-power consumption advantage, this is allowed them to during high-speed motion or with height Reliable visual information is provided in the scene that dynamic range is characterized.Event camera imaging is as shown in Figure 2 under high-speed motion state.

In certain difficult scenes, in the state of high-speed motion or illumination condition variation acutely, ordinary optical image is produced The bad detection for leading to characteristic point of raw motion blur or imaging effect, feature point description and Feature Points Matching exist great It is difficult.However, the real-time status of the available high-speed moving object of event camera, and possess high dynamic range.

What it is due to the output of event camera is point isolated one by one, for same object, different moments different motion state Under be presented on point in event frame may and it is unstable.Therefore, traditional feature point description subalgorithm possibly can not be stablized applicable. Meanwhile compared to ordinary optical image, the case point of event output further comprises respective temporal information.And directly utilize tradition Description subalgorithm does not use this information then well.Therefore, affair character point is extracted from the angle of timestamp information Description can preferably utilize the advantage of event camera.

Summary of the invention

The present invention generates feature point description based on event camera, and carries out Feature Points Matching according to description.In shape Under the inspiration of shape contextual algorithms, a kind of sub- generation method of feature point description for event image imaging characteristics, and root are proposed Characteristic point is matched according to description of generation.Wherein, the method for generating feature point description is slightly different with conventional method, Since time camera can obtain position, time and the polarity of case point, the present invention has used timestamp information, this The advantage of event camera can preferably be used.

Technical solution provided by the invention is a kind of sub- matching process of the feature point description based on event camera, specific steps It is as follows:

Step 1, with characteristic point p_iFor in the center of circle, R1 are outermost radius of circle, R2 is most inner circle radius local by logarithm Distance interval establishes N number of concentric circles, i.e., (log10 (R1), log10 (R2)) is divided into N according to by logarithm (i.e. logspace) A element generates grid then by this region along the circumferential direction M equal part, i.e. bins, as shown in Figure 4.

Step 2, comparative feature point p_iThe timestamp t at place_piWith the timestamp t of case point each in local_qiSize, if t_pi < t_qi, then the point is set to " 1 ", if t_pi> t_qi, then the point is set to " 0 ".

Step 3, statistical nature point p_iIn local in each bins " 1 " number, i.e. the statistical distribution of these points in bins Histogram h_i(k), referred to as characteristic point p_iDescription son, wherein describe sub- size be M*N.

Step 4, all characteristic points are traversed, corresponding description of all characteristic points is obtained.

Step 5, according to h_i(k) similitude, i.e. cost value between every two feature point set are calculated, is then calculated using Hungary Method counts one group of minimum point set corresponding relationship of overall cost value, to obtain Feature Points Matching relationship.Calculate cost value Formula are as follows:

Wherein, k is kth position in description, h_i(k), h_j(k) characteristic point p is respectively indicated_iAnd q_jDescription son.

Step 6, error hiding is removed using vector coherence method, obtains best match.

Compared with prior art, the advantages of the present invention: the case point of event camera output further comprises respectively From timestamp information, and in existing matching algorithm describe son extract do not use this information.Utilize event camera Timestamp information, which describes son to feature point extraction, can preferably utilize the advantage of event camera, keep description sub-information more abundant, Keep matching result more acurrate.

Detailed description of the invention

Fig. 1 is ordinary optical camera imaging.

Fig. 2 is event camera imaging.

Fig. 3 is the sub- generation method of description and Feature Points Matching flow chart.

Fig. 4 is that feature point description sub-grid divides schematic diagram.

Fig. 5 is Feature Points Matching result.

Specific embodiment

Present invention is primarily based on event cameras, handle eventstream data.In view of the characteristic of event camera, propose Sub- extracting method is described in a kind of pair of characteristic point, and carries out Feature Points Matching according to description.In order to make mesh of the invention , technical solution and advantage be more clearly understood, below in conjunction with example, the present invention will be described in further detail.It should manage Solution, specific example described herein are used only for explaining the present invention, be not intended to limit the present invention.

As shown in figure 3, the sub- matching process of a kind of feature point description based on event camera provided in an embodiment of the present invention, tool Body realizes that steps are as follows:

Step 1, N=8, M=12, R1=12, R2=1 are set.With characteristic point p_iIt is outermost radius of circle, R2 for the center of circle, R1 To establish N number of concentric circles by logarithm distance interval in the local of most inner circle radius, this region along the circumferential direction M equal part generates Grid.

Step 2, comparative feature point p_iThe timestamp t at place_piWith the timestamp t of case point each in local_qiSize, if t_pi < t_qi, then the point is set to " 1 ", if t_pi> t_qi, then the point is set to " 0 ", i.e. grid interior coding.

Step 3, statistical nature point p_iIn local in each bins " 1 " number, i.e. the statistical distribution of these points in bins Histogram h_i(k), referred to as characteristic point p_iDescription son.

Step 4, all characteristic points are traversed, corresponding description of all characteristic points is obtained.

Step 5, calculate every two feature point set between cost value, using be X2 test statistics (Chi-square Test count Amount, the departure degree between the actual observed value and theoretical implications value of statistical sample).Totality is counted using Hungary Algorithm One group of minimum point set corresponding relationship of cost value, to obtain Feature Points Matching relationship.

Step 6, error hiding is removed using vector coherence method, obtains best match.

As shown in figure 5, initial matching is 1642 pairs, removal misses result after removing error hiding using vector coherence method Number of matches is 1101 pairs after matching, ratio 0.6705, as shown in Figure 5, the feature after being matched by the method for the invention Point is more, and line is matched between two width figures substantially in same direction, and matching result is accurate.

Specific embodiment described herein is only an example for the spirit of the invention.The neck of technology belonging to the present invention The technical staff in domain can make various modifications or additions to the described embodiments or replace by a similar method In generation, however, it does not deviate from the spirit of the invention or beyond the scope of the appended claims.

Claims (2)

1. the characteristic point matching method based on event camera, which comprises the steps of:

Step 1, with characteristic point p_iFor in the center of circle, R1 are outermost radius of circle, R2 is most inner circle radius local by between logarithm distance Every establishing N number of concentric circles, then by this region along the circumferential direction M equal part, grid is generated；

Step 2, comparative feature point p_iThe timestamp t at place_piWith the timestamp t of case point each in local_qiSize, if t_pi< t_qi, Then the point is set to " 1 ", if t_pi> t_qi, then the point is set to " 0 "；

Step 3, statistical nature point p_iIn local in each bins " 1 " number, i.e. the statistical distribution histogram of these points in bins h_i(k), referred to as characteristic point p_iDescription son；

Step 4, all characteristic points are traversed, corresponding description of all characteristic points is obtained；

Step 5, according to h_i(k) similitude, i.e. cost value between every two feature point set are calculated, is then united using Hungary Algorithm One group of minimum point set corresponding relationship of totality cost value is counted out, to obtain Feature Points Matching relationship；

Step 6, error hiding is removed using vector coherence method, obtains best match.

2. the characteristic point matching method as described in claim 1 based on event camera, it is characterised in that: calculate cost in step 5 The formula of value is,

Wherein, k is kth position in description, h_i(k), h_j(k) characteristic point p is respectively indicated_iAnd q_jDescription son.