CN1825959A - Image tracking algorithm based on adaptive multi-step searching - Google Patents

Abstract

This invention relates to an image tracking method based on an adaptive multi-step search algorithm, which forecasts the possible deviation volume of a target in the current frame to dynamically adjust the search sphere of the algorithm, in which, the tracking algorithm can make the search sphere to cover the regions the target may appear possibly when the target speed is high and the searching sphere is reduced to reduce the computing volume and reduce the possibility of getting in the local extremum point when the target speed is small.

Description

A kind of image tracking algorithm based on adaptive multi-step searching

Technical field

The invention belongs to computer vision and pattern analysis technical field, be specifically related to a kind of image tracking algorithm based on adaptive multi-step searching.

Background technology

Image tracking system can be discerned in a series of video frame images and locate specific target, is widely used in a large amount of fields such as Robotics, guidance technology, automatic supervision.

The algorithm that image tracking system is the most commonly used is based on searching algorithm (Block Matching Algorithm, BMA) [1] [2] of piece coupling.This algorithm is set up template at tracked target, and seeks the zone that has optimum Match with template in current image frame, as the estimation of target location in the present frame, promptly

$x_{\mathrm{tgt}}=\underset{x\∈\mathrm{\Ω}}{\arg \max }\mathrm{sim}\left(x\right)---\left(1\right)$

Wherein x is the centre coordinate of candidate region; Ω is the coordinate set of picture frame; Sim () is at the zone at a certain coordinate place and the matching degree function between the template, and this zone and template be coupling more, and functional value is just big more; x _TgtEstimation for target location in the present frame.

Because the frame per second all higher (20～30fps) of video generally speaking, the position of target in adjacent two frames can be not a good distance away, therefore in order to reduce amount of calculation, to the search of target in each frame all be one with previous frame in the position of target be to carry out on certain neighborhood at center.Even like this, if search for each point on this neighborhood, amount of calculation is still very huge.For example search for one 15 * 15 field, need carry out 255 matching operations, this gives and realizes that real-time tracking has caused very big difficulty.

In the estimation of field of video compression, also run into similar problem.In order to reduce amount of calculation significantly under the prerequisite that does not influence search quality substantially, many fast search algorithms have been carried [3].One of algorithm that wherein is most widely used is three step searching algorithm (Three Step Search, TSS) [3].The concrete seek mode of TSS algorithm as shown in Figure 1.The first step of TSS algorithm is the center with the position of target in former frame, and computer center's point and spaced around thereof are the object matching degree of eight points of 4 pixel distances; Second step of TSS algorithm is the center with that the highest point of matching degree in nine points of the first step, and calculating its spaced around is the object matching degree of eight points of 2 pixel distances; The 3rd step of TSS algorithm is the center with that the highest point of matching degree in related being had a few of second step, calculate the object matching degree of eight points of next-door neighbour around it, related institute have a few the highest that of middle matching degree and put and just think the target loca the 3rd step.For one 15 * 15 field, the TSS algorithm only need carry out 25 matching operations, has saved 88.9% amount of calculation, and search quality decline is very little.

Yet the TSS algorithm is to propose at the estimation of video compression.In the estimation of video compression, do not need accurately track and localization moving target, therefore fixing hunting zone (the TSS algorithm is 15 * 15) is an acceptable.Yet in image was followed the tracks of, the variation of the movement velocity of target can be very big.When target velocity was big, the location interval that target is in two frames may surpass 15 pixels, surpassed the hunting zone of TSS algorithm; When target velocity was very little, if background is comparatively complicated, the hunting zone of TSS algorithm was excessive again, has carried out many unnecessary matching operations, and might make search be absorbed in Local Extremum.More than two kinds of situations all can cause losing of target.

(Adaptive Multi-step Search AMS), dynamically adjusts the hunting zone according to the movement velocity of target to the present invention proposes a kind of adaptive multi-step searching algorithm.No matter experiment showed, that AMS compares with TSS, be under target high speed or low-speed motion situation, all has tracking effect and lower amount of calculation preferably.

List of references

[1]C.Smith，C.Richards，S.Brandt，and N.Papanikolopoulos，“Visual tracking forintelligent vehicle-highway systems”，in IEEE Trans.on Vehicle Technology，1996.

[2]T.Kaneko，and O.Hori，“Template update criterion for template matching of imagesequences”，in Proceedings of 16th International Conference on Pattern Recognition，vol.2，pages 1-5，2002.

[3]Y.Wang，J.Ostermann，Y.Q.Zhang，“Video Processing and Communications”，pages159-161，Prentice Hall，2002.

Summary of the invention

The objective of the invention is to propose a kind ofly can when target velocity is big, can not catch up with target, hour save amount of calculation greatly and increase the image tracking algorithm based on adaptive multi-step searching algorithm (AMS) of tracking stability in target velocity.

The image tracking algorithm that the present invention proposes based on the adaptive multi-step searching algorithm, its step is as follows: when the speed of tracked target was big, corresponding increase hunting zone was with the zone that guarantees that its coverage goal may occur; When the speed of tracked target hour, then should reduce the hunting zone, to reduce amount of calculation, reduce the situation that search is absorbed in local extremum simultaneously.Specifically, determine based on three step searching algorithm (Three Step Search, TSS) hunting zones of the adaptive multi-step searching algorithm of [1] according to the predicted value of present frame target possibility side-play amount.

The present invention can be by changing first step search dot spacing and searching for the hunting zone that step number is regulated the adaptive multi-step searching algorithm.

1, the estimation of target velocity

In image was followed the tracks of, the speed of target can change arbitrarily, but because the frame per second of general video is higher, and the time interval of consecutive frame is very short (only to be 1/30～1/20s), therefore can be similar to the uniform motion model and to come the speed of target of prediction at present frame.The position of note target in the n frame is x (n)=[x _h(n), x _v(v)] ^T, x wherein _h(n) and x _v(n) be respectively the abscissa and the ordinate of target position in the n frame.Then target is approximately in the speed of n frame

$v\left(n\right)\≈\frac{x\left(n\right)-x(n-1)}{\mathrm{\Δt}}---\left(2\right)$

Wherein Δ t is the time interval of consecutive frame, v (n)=[v _h(n), v _v(n)] ^TThe velocity vector of forming by the speed of the speed of horizontal direction and vertical direction.

If present frame is the c frame, then the speed of present frame and former frame is approximately respectively

$v\left(c\right)\≈\frac{x\left(c\right)-x(c-1)}{\mathrm{\Δt}}---\left(3\right)$

$v(c-1)\≈\frac{x(c-1)-x(c-2)}{\mathrm{\Δt}}---\left(4\right)$

By the even hypothesis of quickening, the speed of present frame equates with the speed of former frame, promptly

v(c)＝v(c-1) (5)

With (3), (4) formula substitution (5) Shi Kede

x(c)-x(c-1)＝Δx(c)≈x(c-1)-x(c-2) (6)

Wherein Δ x (c) for target in the present frame possible depart from the motion vector of searching for starting point.The amplitude of Δ x (c)

$d\left(c\right)=\sqrt{{(x_h(c-1)-x_h(c-2))}^2+{(x_v(c-1)-x_v(c-2))}^2}---\left(7\right)$

Promptly can be used as the foundation of determining the hunting zone.

2, the adaptive multi-step searching algorithm (Adaptive Multi-step Search, AMS)

After having obtained the side-play amount that target is possible in the present frame, just can be set in the size of hunting zone in the present frame.Be advisable with the possible bias size that suitably surpasses target in the hunting zone, the position that like this can either coverage goal may arrive can not cause the hunting zone excessive again.

The search pattern of AMS algorithm and the TSS class of algorithms seemingly still can dynamically be adjusted the hunting zone.Adjust the hunting zone of AMS algorithm, can change the spacing of first step search point, the also corresponding change of Sou Suo step number simultaneously.Because the search point spacing distance in each step all reduces by half, and the search of final step point is spaced apart 1 pixel distance, so the search point spacing distance s of the first step ₁And the pass between the search step number N is

s ₁＝2 ^N-1 (8)

The AMS hunting zone is defined as the maximum distance that point that AMS can search leaves search center, promptly

$R\left(N\right)=\underset{p\∈B_N}{\max }\left|\right|p-p_c\left|\right|---\left(9\right)$

Wherein p is search point, p _cBe search center, B _NFor the search step number is the set of N all search points during the step, R (N) is the hunting zone corresponding to N step search.

Interval by each step search point all reduces by half as can be known, and the hunting zone of AMS can be expressed as

$R\left(N\right)=s_1+\frac{1}{2}{s}_1+\frac{1}{2^2}{s}_1+\·\·\·+\frac{1}{2^{N-1}}{s}_1=(2-\frac{1}{2^{N-1}})s_1---\left(10\right)$

(8) formula substitution following formula can be got hunting zone and the relation of searching for step number:

R(N)＝2 ^N-1 (11)

Suitably above the principle of target possibility side-play amount, can define hunting zone R, first step search dot spacing s according to the hunting zone ₁, the relation between the possible side-play amount d of search step number N and target is as follows:

R=2 ^N-1, s ₁=2 ^N-1When 2 ^N-1-1≤d≤2 ^N-2, N=2 wherein, 3 ... (12)

When d=0, to be on the safe side, the search step number still is decided to be 2, and first step search dot spacing is 2, and the hunting zone is 3.

For common target possibility side-play amount, its corresponding AMS algorithm first step search dot spacing, hunting zone and search step number are listed as follows:

Target may side-play amount	AMS first step search dot spacing	The AMS hunting zone	AMS searches for step number
				0～2	2	3	2
3～6	4	7	3
				7～14	8	15	4
15～30	16	31	5

Table 1 target possibility side-play amount, the search of the AMS first step dot spacing, hunting zone and search step number relation table

When practical application, can after having obtained target possibility side-play amount, utilize lookup table mode (LUT) to obtain first step search point interval, hunting zone and the search step number of AMS algorithm, increase any amount of calculation hardly.

Also dynamically adjust the hunting zone in view of the above by target of prediction in the velocity magnitude of present frame, can significantly improve the effect of following the tracks of high-speed target, and reduce the amount of calculation of following the tracks of slower-velocity target.

Description of drawings

Fig. 1: three step searching algorithm (TSS) schematic diagrames.

Fig. 2 a: the 17th frame sectional drawing of using the video flowing 1 of AMS algorithm.

Fig. 2 b: the 25th frame sectional drawing of using the video flowing 1 of AMS algorithm.

Fig. 2 c: the 33rd frame sectional drawing of using the video flowing 1 of AMS algorithm.

Fig. 2 d: the 36th frame sectional drawing of using the video flowing 1 of AMS algorithm.

Fig. 3: the target prediction speed of video flowing 1 and actual speed curve.

Fig. 4: the AMS algorithm search scope curve of video flowing 1.

Fig. 5 a: the 17th frame sectional drawing of using the video flowing 1 of TSS algorithm.

Fig. 5 b: the 25th frame sectional drawing of using the video flowing 1 of TSS algorithm.

Fig. 5 c: the 33rd frame sectional drawing of using the video flowing 1 of TSS algorithm.

Fig. 5 d: the 36th frame sectional drawing of using the video flowing 1 of TSS algorithm.

Fig. 6 a: the 1st frame sectional drawing of using the video flowing 2 of AMS algorithm.

Fig. 6 b: the 40th frame sectional drawing of using the video flowing 2 of AMS algorithm.

Fig. 6 c: the 90th frame sectional drawing of using the video flowing 2 of AMS algorithm.

Fig. 6 d: the 130th frame sectional drawing of using the video flowing 2 of AMS algorithm.

Fig. 7: the target prediction speed of video flowing 1 and actual speed curve.

Fig. 8: the AMS algorithm search scope curve of video flowing 1.

Fig. 9 a: the 1st frame sectional drawing of using the video flowing 2 of TSS algorithm.

Fig. 9 b: the 40th frame sectional drawing of using the video flowing 2 of TSS algorithm.

Fig. 9 c: the 90th frame sectional drawing of using the video flowing 2 of TSS algorithm.

Fig. 9 d: the 130th frame sectional drawing of using the video flowing 2 of TSS algorithm.

Embodiment

Further introduce embodiments of the present invention below by the emulation of some being followed the tracks of example.The platform of emulation is Matlab6.5.

1, the target (video flowing 1) that applies the present invention to follow the tracks of motion at a relatively high speed and constantly quicken.

Tracking target in the video flowing 1 is the flying saucer that aloft flies.At first selected target in first frame.In second frame, owing to also do not have and therefore to suppose that the hunting zone is 15, to guarantee coverage goal for the target speed of prediction.In ensuing each frame,, and determine the hunting zone of AMS algorithm according to (12) formula (or by question blank 1) according to the side-play amount of (7) formula target of prediction.

Use simulation result such as Fig. 2 a to Fig. 2 d and Fig. 3, shown in Figure 4 of AMS algorithm.Wherein Fig. 2 a to Fig. 2 d is respectively a sectional drawing (being respectively the 17th, 25,33,36 frames) of following the tracks of video, and Fig. 3 is target prediction speed and actual speed curve, and Fig. 4 is an AMS algorithm search scope curve.

By Fig. 2 a to Fig. 2 d as seen, no matter how the speed of target changes, and has all obtained desirable tracking effect based on the image tracking algorithm of AMS.As seen from Figure 3, the predicated error of target velocity is very little, and the consensus forecast error is 0.87 pixel distance, so the target that the present invention did is rational in the approximate hypothesis of making uniform motion of interframe.As seen from Figure 4, along with the increase of target velocity, the hunting zone of AMS has guaranteed that also along with increase the covering of position may appear in the hunting zone to target.

As a comparison, the AMS algorithm is replaced with the TSS algorithm, once more identical target is followed the tracks of.Simulation result is shown in Fig. 5 a to Fig. 5 d.By these figure as seen, when target velocity increases when acquiring a certain degree, surpassed the hunting zone of TSS algorithm, thereby finally lost target.

2, apply the present invention to follow the tracks of the target (video flowing 2) of low-speed motion.

Tracking target in the video flowing 2 is the people who slowly walks on the ground.It is described with precedent that target is chosen the principle of selecting with the AMS hunting zone in the tracing process.

Use simulation result such as Fig. 6 a to Fig. 6 d and Fig. 7, shown in Figure 8 of AMS algorithm.Wherein Fig. 6 a to Fig. 6 d is respectively a sectional drawing (being respectively the 1st, 40,90,130 frames) of following the tracks of video, and Fig. 7 is target prediction speed and actual speed curve, and Fig. 8 is an AMS algorithm search scope curve.

As seen from Figure 7, the translational speed of target is very low always, and the prediction of speed error is very little, only is 0.77 pixel distance.As seen from Figure 8, Jue Daduoshuo AMS hunting zone all is 3.So little hunting zone does not a bit influence tracking effect, and this point can be found out by Fig. 6 a to Fig. 6 d.Moreover, because the hunting zone diminishes, saved amount of calculation greatly.

As a comparison, the simulation result of TSS algorithm is shown in Fig. 9 a to Fig. 9 d.Although its tracking effect is also good, amount of calculation rises greatly.The matching operation of AMS algorithm adds up to 2335 times, and the matching operation of TSS algorithm adds up to 3375 times, has increased by 44.5%.

To sum up, determine the size of hunting zone according to the target offset amount of prediction based on the image tracking algorithm of AMS algorithm, with compare based on the image tracking algorithm of TSS algorithm, the effect of following the tracks of high-speed target has tangible raising, and can farthest reduce amount of calculation.

Claims (4)

1, a kind of image tracking algorithm based on the adaptive multi-step searching algorithm is characterized in that corresponding increase hunting zone is with the zone that guarantees that its coverage goal may occur when the speed of tracked target is big; When the speed of tracked target hour, then should reduce the hunting zone, to reduce amount of calculation, reduce the situation that search is absorbed in local extremum simultaneously; And determine hunting zone based on the adaptive multi-step searching algorithm of three step searching algorithms according to the predicted value that the present frame target may side-play amount.

2, the image tracking algorithm based on the adaptive multi-step searching algorithm according to claim 1 is characterized in that by changing first step search dot spacing and searching for the hunting zone that step number is regulated the adaptive multi-step searching algorithm.

3, the image tracking algorithm based on the adaptive multi-step searching algorithm according to claim 2 is characterized in that hunting zone R, first step search dot spacing s _l, the relation between the possible side-play amount d of search step number N and target is as follows:

R=2 ^N-1, s _l=2 ^N-1When 2 ^N-1-1≤d≤2 ^N-2, N=2 wherein, 3 ... (4)

When d=0, the search step number is 2, and first step search dot spacing is 2, and the hunting zone is 3.

4, the image tracking algorithm based on the adaptive multi-step searching algorithm according to claim 3 is characterized in that target possibility side-play amount, first step search dot spacing, hunting zone and search step number are listed as follows: Target may side-play amount AMS first step search dot spacing The AMS hunting zone AMS searches for step number 0～2 2 3 2 3～6 4 7 3 7～14 8 15 4 15～30 16 31 5

Images