CN105182350A - Multi-beam sonar target detection method by applying feature tracking - Google Patents

Abstract

The invention provides a multi-beam sonar target detection method by applying feature tracking. The method comprises steps: (1) data are acquired through a sonar system, sonar images are formed through imaging, and the continuous sonar images form a sonar image sequence; (2) pretreatment is carried out; (3) scale invariant feature transform features of each frame of sonar image after pretreatment are extracted; (4) feature matching is carried out on former two frames in the sonar image sequence, features which are successfully matched are marked as potential targets, and features which are not matched successfully are discarded; and (5) features presenting the potential targets are tracked in subsequent frames in the sonar image sequence, real targets are screened in the potential targets after the image sequence is traversed, and a feature locus for the real targets can be obtained. The method is used in a single frame of image to judge whether the target exists or not, multiple target features are tracked at the same in the sonar image sequence, the real targets are screened in the potential targets, and the method is wide in applicability.

Description

The multi-beam sonar object detection method that a kind of application characteristic is followed the tracks of

Technical field

What the present invention relates to is a kind of Underwater Target Detection method, the multi-beam sonar object detection method of specifically a kind of application characteristic tracking.

Background technology

Along with the traction of national marine strategy, Underwater Target Detection technology is developed rapidly, and its application demand comprises keeps away barrier under water and search for navigation, hypostasis and the field such as salvage, pipeline maintenance and bridge pier detection.Researchist, generally by Imaging sonar, forms the acoustic image of scene under water, realizes Underwater Target Detection by Iamge Segmentation, and problems faced comprises the selection of segmentation threshold and the limitation of single image information.(TrackBeforeDetect is followed the tracks of before detecting, TBD) be that the detection of a kind of emerging weak target is theoretical, first do not judge there being driftlessness in single-frame images, in image sequence, multiple target is followed the tracks of simultaneously, then from potential target, filter out real goal according to target trajectory.Carrying out signature tracking based in TBD under water acoustic image sequence, is the method for a kind of novelty realizing Underwater Target Detection.

Feature selecting is the key realizing the tracking of underwateracoustic image sequence, and common feature comprises statistical nature, textural characteristics, shape facility, mathematic(al) manipulation feature etc.Under water in target detection, still there is the problems such as separating capacity is weak, computation complexity large, poor robustness in these feature application.The Scale invariant features transform (ScaleInvariantFeatureTransform, SIFT) that Lowe proposes, based on Scale-space theory, has had in optical imagery coupling and identification field and has successfully applied precedent.The feature that Scale invariant features transform generates maintains the invariance to rotation, scaling, brightness change, to the stability that visual angle change, affined transformation, noise also keep to a certain degree, the potential target that image quality is relatively low and be subject in the sonar image of environmental interference better can be characterized.

Summary of the invention

The object of the present invention is to provide and a kind ofly not be used in single-frame images the presence or absence judging target, by following the tracks of multiple target signature in sonar image sequence simultaneously, the multi-beam sonar object detection method that the application characteristic filtering out real goal from potential target is followed the tracks of.

The object of the present invention is achieved like this:

(1) by Sonar system image data, sonar image is formed to gathered data imaging, by continuous print sonar image composition sonar image sequence;

(2) pre-service is carried out to the every two field picture in described sonar image sequence;

(3) the Scale invariant features transform feature of every frame sonar image after extracting pre-service;

(4) launch characteristic matching at the front cross frame of sonar image sequence, the features localization that the match is successful is potential target, unsuccessful for coupling feature is abandoned;

(5) follow the tracks of at the subsequent frame of sonar image sequence the feature representing potential target, from potential target, filter out real goal after traversal image sequence, and obtain the characteristic locus of real goal.

The present invention can also comprise:

1, described Sonar system is shallow water high-resolution multibeam sounding system, and acoustics basic matrix is T-shape, and launching basic matrix is polynary arc battle array, receives the even linear array that basic matrix is 80 array elements; Frequency of operation is 300kHz, and sample frequency is 40kHz, and the pulse width selected is 1ms.

2, described pre-service adopts dynamic brightness to distribute enhancing contrast, realize noise filtering by medium filtering, original image intensity is mapped to linear gradation [0 when distributing by dynamic brightness, 255], be L with minimum of intensity in original image, to 1% be averaged as maximal value H before original image intensity, the mapped function relation of input intensity and output gray level is as follows

$\left\{\begin{array}{cc}{z}_{out}=0& z_{in}=L\\ z_{out}=255{\left(\frac{z_{in}-L}{H-L}\right)}^{\mathrm{\&gamma;}}& L<z_{in}<H\\ z_{out}=255& z_{in}\&GreaterEqual;H\end{array}\right.$

Wherein z _inimage input intensity value, z _outbe image output gray level value, γ is mapping parameters, span is [0.5-2.0].

3, described Scale invariant features transform feature comprises coordinate, yardstick and directional information, and the feature interpretation vector of 128 dimensions.

4, characteristic matching is using the feature interpretation of Scale invariant features transform vector Euclidean distance as similarity measurement, adopt arest neighbors ratioing technigue, find arest neighbors and time neighbour of unique point to be matched, if nearest neighbor distance is less than certain threshold value with the ratio of time nearest neighbor distance, then think nearest neighbor point and this be characterized as optimum matching pair, otherwise be considered as that it fails to match.

The invention provides a kind of Underwater Target Detection method, not be used in single-frame images the presence or absence judging target, and by following the tracks of multiple target signature in sonar image sequence simultaneously, from potential target, filter out real goal.

The present invention is owing to taking above technical scheme, and it has the following advantages

(1) the present invention utilizes medium filtering and dynamic brightness to distribute, and significantly improves underwateracoustic picture quality, for feature extraction and tracking create advantage.

(2) the Scale invariant features transform feature that adopts of the present invention, comprise more frame internal information, frame matching robustness stronger, be applicable to characterize the potential target in underwateracoustic image sequence.

(3) the present invention uses characteristic tracking method, not be used in single-frame images the presence or absence judging target, and by following the tracks of multiple clarification of objective in image sequence simultaneously, from potential target, filter out real goal, for Underwater Target Detection provides a new approach.

Accompanying drawing explanation

The process flow diagram of Fig. 1 Underwater Target Detection method of signature tracking designed by the present invention.

Fig. 2 is that the present invention gathers image sequence the 1st frame original image.

Fig. 3 a to Fig. 3 d is image sequence of the present invention 1st frame dynamic brightness distribution effects, wherein Fig. 3 a when different mappings parameter, γ=0.5; Fig. 3 b, γ=1; Fig. 3 c, γ=1.5; Fig. 3 d, γ=2.

Fig. 4 is image sequence of the present invention 1st frame pretreatment image.

Fig. 5 is image sequence the 1st frame Scale invariant features transform feature that the present invention extracts.

Fig. 6 is image sequence of the present invention 1st frame and the 2nd frame Scale invariant features transform characteristic matching situation.

The algorithm flow of the signature tracking that Fig. 7 designs for the present invention.

Fig. 8 is the target detection experimental result of signature tracking of the present invention.

Embodiment

The overview flow chart of the multi-beam sonar Underwater Target Detection method that application characteristic is followed the tracks of as shown in Figure 1.Below in conjunction with case study on implementation, method of the present invention is described in further detail.

1, by Sonar system image data, sonar image is formed to data imaging, by continuous print sonar image composition sonar image sequence.

Testing equipment is shallow water high-resolution multibeam sounding system.The acoustics basic matrix of system is T-shape, and launching basic matrix is polynary arc battle array, receives the even linear array that basic matrix is 80 array elements.System operating frequency is 300kHz, and sample frequency is 40kHz, and the pulse width selected is 1ms.The depth of water is 20 ~ 30m, using oil drum as dynamic Small object, in test, surveying vessel is rested in bank, and Sonar system fixedly mounts, and uses two hawsers to hang oil drum at topside, and is allowed to condition in water and does vertical and tangential movement slowly.By the data imaging gathered, picture size is, chooses continuous print 28 frame composition image sequence.From image sequence, read the 1st frame, imaging effect as shown in Figure 2.

2, pre-service is carried out to the every two field picture in sonar image sequence.Adopt dynamic brightness to distribute and strengthen contrast, realize noise filtering by medium filtering;

When dynamic brightness is distributed, the mapped function relation of input intensity and output gray level is as follows:

$\left\{\begin{array}{cc}{z}_{out}=0& z_{in}=L\\ z_{out}=255{\left(\frac{z_{in}-L}{H-L}\right)}^{\mathrm{\&gamma;}}& L<z_{in}<H\\ z_{out}=255& z_{in}\&GreaterEqual;H\end{array}\right.$

Setting mapping parameters γ value is respectively 0.5,1,1.5,2.0 to contrast, test findings as shown in Figure 3, compare to find when γ=1.5 contrast strengthen and the balance of details reservation best, therefore Choose for user parameter γ=1.5.Medium filtering adopts sliding window size to be 3 × 3, through pretreated image as shown in Figure 4.

3, the Scale invariant features transform characteristic present potential target in underwateracoustic image is extracted.

Be I (x, y) by underwateracoustic image definition, gaussian kernel function is G (x, y, σ), and expression formula is shown below.

$G(x,y,\mathrm{\&sigma;})=\frac{1}{2{\mathrm{\&pi;\&sigma;}}^2}{e}^{-\frac{x^2+y^2}{2{\mathrm{\&sigma;}}^2}}$

In formula, σ is the metric space factor, reflects image by level and smooth degree.I (x, y) and G (x, y, σ) convolution is obtained the metric space of image under different scale be expressed as follows:

R(x,y,σ)＝G(x,y,σ)*I(x,y)

Difference Gauss (DifferenceofGaussian, the DoG) operator definitions of image is:

D(x,y,σ)＝R(x,y,kσ)-R(x,y,σ)

In formula, k is the scale factor between adjacent yardstick.By the DoG pyramid of DoG operator composing images, each point in DoG pyramid needs to compare with surrounding neighbors 8 points of same yardstick and surrounding neighbors 18 points of adjacent yardstick correspondence position.When this point is local extremum, be judged to be candidate feature point.The candidate feature point set U of definition image ₀, therefrom reject contrast according to following formula low and be positioned at Edge Feature Points, filter out key point S (x, y).

$\left\{\begin{array}{c}u\&Element;S,\left|D\right(\hat{u}\left)\right|\&GreaterEqual;T_c\\ u\&NotElement;S,\left|D\left(\hat{u}\right)\right|<T_c\end{array}\right.$

In formula, u ∈ U ₀, T _cfor contrast threshold, for contrast absolute value.

$\left\{\begin{array}{c}u\&Element;S,\frac{Tr{\left(u\right)}^2}{Det\left(He\right)}\&le;\frac{{(T_{\mathrm{\&gamma;}}+1)}^2}{T_{\mathrm{\&gamma;}}}\\ u\&NotElement;S,\frac{Tr{\left(u\right)}^2}{Det\left(He\right)}>\frac{{(T_{\mathrm{\&gamma;}}+1)}^2}{T_{\mathrm{\&gamma;}}}\end{array}\right.$

In formula, T _γfor principal curvatures fractional threshold, the mark that Tr (u) is u, Det (He) is Hessian determinant of a matrix.

The mould m (x, y) of its gradient and direction θ (x, y) is calculated by following formula.With S (x, y) histogrammic mode is utilized to add up the gradient distribution of neighborhood territory pixel in the neighborhood window centered by, the principal direction of the residing neighborhood gradient of histogrammic peak value reflection S (x, y), the direction of S (x, y) is accurately obtained by parabola interpolation.

$\left\{\begin{array}{c}m(x,y)=\frac{\sqrt{{\&lsqb;R(x+1,y)-R(x-1,y)\&rsqb;}^2}}{+{\&lsqb;R(x,y+1)-R(x,y-1)\&rsqb;}^2}\\ \mathrm{\&theta;}(x,y)={\tan }^{-1}\frac{R(x,y+1)-R(x,y-1)}{R(x+1,y)-R(x-1,y)}\end{array}\right.$

Centered by key point, get 16 × 16 windows, in each 4 × 4 fritters, calculate the gradient orientation histogram in 8 directions, draw the accumulated value of each gradient direction, can form a Seed Points, each key point uses 16 Seed Points to describe.Key point and Scale invariant features transform feature, comprise coordinate, yardstick and directional information, and the feature interpretation vector of 128 dimensions.With image sequence the 1st frame feature extraction effect as shown in Figure 5, feature 79 is extracted.In figure, horizontal ordinate represents the horizontal level of relative basic matrix, and ordinate represents the degree of depth of the distance water surface, arrow points and length distribution representation feature direction and yardstick.Can find out that feature concentrates on the water-bed profile that dynamic Small object that the degree of depth is 17m and the degree of depth are 22 ~ 26m from image.

4, launch characteristic matching at the front cross frame of sonar image sequence, the features localization that the match is successful is potential target, unsuccessful for coupling feature is abandoned.

Characteristic matching, using the feature interpretation of Scale invariant features transform vector Euclidean distance as similarity measurement, adopts arest neighbors ratioing technigue.Find arest neighbors and time neighbour of unique point to be matched, if nearest neighbor distance is less than certain threshold value with the ratio of time nearest neighbor distance, then think nearest neighbor point and this be characterized as optimum matching pair, otherwise be considered as that it fails to match.Image sequence the 1st frame and the 2nd two field picture have 42 matching characteristics, and matching state as shown in Figure 6.As can be seen from the figure the feature characterizing dynamic Small object and water-bed profile all achieves coupling, but also has the characteristic matching of some non-targeted, needs to be followed the tracks of by complete image sequence to determine whether further to represent real goal.

5, follow the tracks of at the subsequent frame of sonar image sequence the feature representing potential target, from potential target, filter out real goal after traversal image sequence, and obtain the characteristic locus of real goal.

The algorithm flow of signature tracking as shown in Figure 7.

1) read a two field picture, obtain Scale invariant features transform feature set and save as a template;

2) read next frame image, each Scale invariant features transform feature extracted mated with template, the match is successful by this features localization is that potential target is kept in

3) continuous 3 frame couplings are unsuccessful, and this feature has greatly may characterize false-alarm, can reject this feature from temporarily providing room;

4) after completing whole tracking, the feature still retained is judged to represent real goal, obtains this clarification of objective track simultaneously.

Target detection result as shown in Figure 8.6 groups of Scale invariant features transform features are successfully traced into, the sign Static Water return pulley exterior feature that wherein 5 groups of side-play amounts are less, and the dynamic Small object of sign that 1 group of side-play amount is larger from image sequence.Water-bed contour feature position distribution is (-13.0,24.4), (-7.0,24.1), (5.8,23.8), (16.2,23.2) and (20.8,23.0), dynamic Small object moves upward from 17m, starts tangential movement to 11m.

Claims (9)

1. a multi-beam sonar object detection method for application characteristic tracking, is characterized in that:

(1) by Sonar system image data, sonar image is formed to gathered data imaging, by continuous print sonar image composition sonar image sequence;

(2) pre-service is carried out to the every two field picture in described sonar image sequence;

(3) the Scale invariant features transform feature of every frame sonar image after extracting pre-service;

(4) launch characteristic matching at the front cross frame of sonar image sequence, the features localization that the match is successful is potential target, unsuccessful for coupling feature is abandoned;

(5) follow the tracks of at the subsequent frame of sonar image sequence the feature representing potential target, from potential target, filter out real goal after traversal image sequence, and obtain the characteristic locus of real goal.

2. the multi-beam sonar object detection method of application characteristic tracking according to claim 1, it is characterized in that: described Sonar system is shallow water high-resolution multibeam sounding system, acoustics basic matrix is T-shape, and launching basic matrix is polynary arc battle array, receives the even linear array that basic matrix is 80 array elements; Frequency of operation is 300kHz, and sample frequency is 40kHz, and the pulse width selected is 1ms.

3. the multi-beam sonar object detection method of application characteristic tracking according to claim 1 and 2, it is characterized in that: described pre-service adopts dynamic brightness to distribute enhancing contrast, realize noise filtering by medium filtering, original image intensity is mapped to linear gradation [0 when distributing by dynamic brightness, 255], be L with minimum of intensity in original image, to 1% be averaged as maximal value H before original image intensity, the mapped function relation of input intensity and output gray level be as follows

$\left\{\begin{array}{cc}{z}_{out}=0& z_{in}=L\\ z_{out}=255{\left(\frac{z_{in}-L}{H-L}\right)}^{\mathrm{\&gamma;}}& L<z_{in}<H\\ z_{out}=255& z_{in}\&GreaterEqual;H\end{array}\right.$

Wherein z _inimage input intensity value, z _outbe image output gray level value, γ is mapping parameters, span is [0.5-2.0].

4. the multi-beam sonar object detection method of application characteristic tracking according to claim 1 and 2, is characterized in that: described Scale invariant features transform feature comprises coordinate, yardstick and directional information, and the feature interpretation vector of 128 dimensions.

5. the multi-beam sonar object detection method of application characteristic tracking according to claim 3, is characterized in that: described Scale invariant features transform feature comprises coordinate, yardstick and directional information, and the feature interpretation vector of 128 dimensions.

6. the multi-beam sonar object detection method of application characteristic tracking according to claim 1 and 2, it is characterized in that: characteristic matching is using the feature interpretation of Scale invariant features transform vector Euclidean distance as similarity measurement, adopt arest neighbors ratioing technigue, find arest neighbors and time neighbour of unique point to be matched, if nearest neighbor distance is less than certain threshold value with the ratio of time nearest neighbor distance, then think nearest neighbor point and this be characterized as optimum matching pair, otherwise be considered as that it fails to match.

7. the multi-beam sonar object detection method of application characteristic tracking according to claim 3, it is characterized in that: characteristic matching is using the feature interpretation of Scale invariant features transform vector Euclidean distance as similarity measurement, adopt arest neighbors ratioing technigue, find arest neighbors and time neighbour of unique point to be matched, if nearest neighbor distance is less than certain threshold value with the ratio of time nearest neighbor distance, then think nearest neighbor point and this be characterized as optimum matching pair, otherwise be considered as that it fails to match.

8. the multi-beam sonar object detection method of application characteristic tracking according to claim 4, it is characterized in that: characteristic matching is using the feature interpretation of Scale invariant features transform vector Euclidean distance as similarity measurement, adopt arest neighbors ratioing technigue, find arest neighbors and time neighbour of unique point to be matched, if nearest neighbor distance is less than certain threshold value with the ratio of time nearest neighbor distance, then think nearest neighbor point and this be characterized as optimum matching pair, otherwise be considered as that it fails to match.

9. the multi-beam sonar object detection method of application characteristic tracking according to claim 5, it is characterized in that: characteristic matching is using the feature interpretation of Scale invariant features transform vector Euclidean distance as similarity measurement, adopt arest neighbors ratioing technigue, find arest neighbors and time neighbour of unique point to be matched, if nearest neighbor distance is less than certain threshold value with the ratio of time nearest neighbor distance, then think nearest neighbor point and this be characterized as optimum matching pair, otherwise be considered as that it fails to match.