CN102253391B - Multi-laser-radar-based pedestrian target tracking method - Google Patents

Abstract

The invention discloses a multi-laser-radar-based pedestrian target tracking method. The multi-laser-radar-based pedestrian target tracking method comprises the following steps that: a sensor network consisting of a plurality of single-line laser radars monitors a whole scene; at every sampling moment, the radars output footstep point distance data of pedestrians, and the data among different radars are unified at a coordinate system through system calibration; a footstep movement model of the pedestrians is built and pedestrian movement states are tracked through Kalman filtering so as to finally acquire a stable target trace and relevant statistical information. In the method, the laser radars are used as sensors, so that the defects of influence of illumination variance on an image sensor in the tracking method can be overcome and reliability is enhanced greatly; furthermore, the tracking range is wide, the algorithm speed is high, the number of tracked pedestrians is large, all movement modes of the pedestrians are compatible, such as standing, walking and jogging, and positions of the pedestrians can be detected effectively and movement traces of the pedestrians can be tracked in real time. The multi-laser-radar-based pedestrian target tracking method is applicable to occasions in which main movement targets are pedestrians.

Description

A kind of pedestrian's method for tracking target based on multilasered optical radar

Technical field

The present invention relates to a kind of pedestrian's method for tracking target, especially relate to a kind of pedestrian's method for tracking target based on multilasered optical radar.

The present invention relates to a kind of pedestrian's method for tracking target, be applicable to and in various public arenas, realize a plurality of pedestrians are being stood, walking, the tracking under the motor pattern such as jog based on multilasered optical radar.

Background technology

At subway, the station, in the crowded public arena such as department store, statistical informations such as pedestrian's flow, direction of motion can help engineering design side to carry out more reasonably gateway layout, and safe emergency prediction schemes reasonable in design etc. have huge commercial value.Want to obtain automatically these real-time statistics information, just need follow the tracks of accurately all pedestrian's targets in the scene.Present target following technology is followed the tracks of the moving target in the image through certain image processing algorithm mainly based on vision sensors such as video cameras.Yet the tracking technique based on image processing algorithm is all relatively responsive to the variation of ambient lighting, can't adapt to the requirement of round-the-clock tracking.The visual field of a video camera covering is limited simultaneously, though can lean on a plurality of video cameras to realize covering on a large scale, the target following of striding camera realizes difficulty, and is still at the experimental stage.By contrast, laser radar can directly obtain the range information of target, and does not receive the influence of illumination variation.Extensive gradually along with what use, price also progressively reduces.

Summary of the invention

The object of the present invention is to provide a kind of pedestrian's method for tracking target based on multilasered optical radar; Not only following range is big, and accuracy is high, and algorithm speed is fast; And the compatible pedestrian's of ability various motor patterns; Comprise and standing that walking and jog etc. can very detect pedestrian's position and real-time follow-up pedestrian's movement locus effectively.

The technical scheme that the present invention adopts is may further comprise the steps:

(1) setting of radar site and system calibrating;

(2) data acquisition and fusion;

(3) pairing of the cluster in motor point and step point;

(4) based on the target following of manikin and Kalman filtering.

The setting and the system calibrating of said step (1) radar site: be about to the diverse location that a plurality of radars are arranged on scene, the position choice principle is set is: (a) make any position of target in scene all be not easy to be blocked; (b) evenly arrange laser radar at the scene edge, with the possibility that the raising target is detected, radar scanning plan range floor level 23～25cm, this guarantees that highly radar in most of the cases can both detect and obtain the range data of two pin of pedestrian; Demarcation to a plurality of radars realizes through following method: some bar-shaped objects are set in scene, and the coordinate of establishing i the bar-shaped object that from N radar, obtains is:

$(x_i^1,y_i^1),(x_i^2,y_i^2),...,(x_i^N,y_i^N);$

And suppose that the coordinate with the 1st radar is a reference coordinates, on two dimensional surface, establish the transition matrix R of 1 radar fix of a n radar fix to the _nWith translational movement T _nFor:

$R_n=\left[\begin{array}{cc}{r}_{11}^n& r_{12}^n\\ r_{12}^n& r_{22}^n\end{array}\right]$ $T_n=\left[\begin{array}{cc}{t}_1^n& t_2^n\end{array}\right]$

Then obtain i o'clock transfer equation from 1 radar fix of a n radar fix to the:

$(x_i^1,y_i^1)=R_n(x_i^n,y_i^n)+T_n$

Transition matrix and translational movement are unique when i＞3 confirms, through demarcating data-switching that different radar fixs system is obtained down under same radar fix, in the formula: x representes abscissa value a little, and y representes ordinate value a little.

(2) data acquisition of said step and fusion: promptly through on data transmission to a main frame of hub with a plurality of radars collections, the time of while record acquisition; Data acquisition between a plurality of radars is by unified outer synchronous triggering; Extract earlier the background data of each radar automatically through the data of gathering the above frame number of 50 frames, the method for distilling of background data has adopted histogram method, promptly adds up in institute's acquisition frame the range information of data point on each direction; And do statistics with histogram; The frequency of occurrences is the highest is the background dot data on this direction, preserves this background data, and range data on all directions that then the same time collected and background data are relatively; If then think background dot less than the threshold quantity that sets and remove apart from difference, the foreground point that remaining promptly is in the scene is the motor point; Coordinate conversion is carried out through calibrating parameters in the foreground point that each radar of synchronization obtains, promptly obtain the exercise data point after this fusion that collects constantly under the unified global coordinate system.

The cluster in said step (3) motor point and the pairing of step point: the exercise data point that promptly fusion is obtained; The phase mutual edge distance is carried out cluster less than the point of first distance threshold; And with the position of type center of interior point as candidate's step; Then the step candidate point is carried out the coupling on time domain and the spatial domain, to confirm whether certain two candidate's step point belongs to same pedestrian, and two kinds of situation are arranged: near (a) search matched two step point prediction positions of the existing target that Kalman filtering obtains; If two candidate's step points in this frame respectively with two predicted positions less than the second distance threshold value, judge that then two candidate's step points are the observation station that has target; (b) if current candidate's step point fails in existing target, to find coupling, then belong to fresh target or noise point; The step candidate point that in present frame, does not mate with the search of the 3rd distance threshold then is if find then with match point associating becoming together two the step points that belong to same fresh target that find.

Said step (4) is based on the target following of modelling of human body motion and Kalman filtering: promptly the pedestrian has been set up motion model, and adopted Kalman filtering to carry out target following according to this model:

1) foundation of pedestrian dummy

Research shows that pedestrian's state is divided into: stand, walking, jog, the pedestrian always has a feet in the walking and the motion process of jogging; A swing pin; And both periodically exchange, and feet was always not liftoff when different was walking, and remained static; And feet can be liftoff and move distance can be arranged in the cycle when jogging, based on above mention stand, the modeling of walking, the three kinds of pedestrian's states of jogging respectively as follows:

The pattern of standing: do not have the driving of physiological action power, the speed of left and right sides pin is zero, and acceleration also is zero; Force=0, v _{L, k}=0, v _{R, k}=0, a _{L, k}=0, a _{R, k}=0, (1)

Walking and jogging pattern: feet has an invariable speed; The swing pin has an invariable physiological action power; Produce the constant physiological action power of constant magnitude of a forward and negative sense in preceding semiperiod and later half cycle respectively, thereby the motion pin is formed acceleration;

With T is the time interval between two frames, and right crus of diaphragm is in the preceding semiperiod of feet:

force＝F，v _L，k＝v _L，k-1+A*t，v _R，k＝V，a _L，k＝A，a _R，k＝0，(2)

Right crus of diaphragm is in the later half cycle of feet:

force＝-F，v _L，k＝v _L，k-1+A*t，v _R，k＝V，a _L，k＝-A，a _R，k＝0，(3)

Left foot is in the preceding semiperiod of feet:

force＝F，v _L，k＝V，v _R，k＝v _R，k-1+A*t?a _L，k＝0，a _R，k＝A，(4)

Left foot is in the later half cycle of feet:

force＝-F，v _L，k＝V，v _R，k＝v _R，k-1+A*t?a _L，k＝0，a _R，k＝-A，(5)

Force representes physiological action power in the formula, and F representes the physiological action power of the actual generation of pedestrian, and v representes movement velocity; V representes the speed of generation under the pedestrian F acting force, and a representes acceleration, and A representes the acceleration that the pedestrian produces under the F acting force; L representes left foot, and R representes right crus of diaphragm, and k representes k constantly;

When V=0, the pedestrian promptly is under the walking state, and when V＞0, the pedestrian promptly is under the state of jogging;

2) Kalman filtering is followed the tracks of

Be used as state variable to pedestrian's position P and speed V, acceleration A is used as the input quantity of system, and the state equation that then obtains system is:

$S_k^i=\mathrm{\Φ}{S}_{k-1}^i+\mathrm{\Ψ}{A}_k^i+\mathrm{\Ω}---\left(6\right)$

Wherein S representes state variable, and i representes pedestrian's mark, and Ω representes the state model error;

$S_k^i=\left(\begin{array}{c}{p}_{R\_x,k}^i\\ p_{R\_y,k}^i\\ v_{R\_x,k}^i\\ v_{R\_y,k}^i\\ p_{L\_x,k}^i\\ p_{L\_y,k}^i\\ v_{L\_x,k}^i\\ v_{L\_y,k}^i\end{array}\right)$ $A_k^i=\begin{array}{c}\left(\begin{array}{c}{a}_{R\_x,k}^i\\ a_{R\_y,k}^i\\ a_{L\_x,k}^i\\ a_{L\_y,k}^i\end{array}\right)\end{array}---\left(7\right)$

Represent state-transition matrix with Φ, Ψ representes the transition matrix of acceleration to speed and displacement conversion, and when the radar scanning frequency was 37.5Hz, T was 0.026 second; Then have:

$\mathrm{\Φ}=\left(\begin{array}{cccccccc}1& 0& T& 0& 0& 0& 0& 0\\ 0& 1& 0& T& 0& 0& 0& 0\\ 0& 0& 1& 0& 0& 0& 0& 0\\ 0& 0& 0& 1& 0& 0& 0& 0\\ 0& 0& 0& 0& 1& 0& T& 0\\ 0& 0& 0& 0& 0& 1& 0& T\\ 0& 0& 0& 0& 0& 0& 1& 0\\ 0& 0& 0& 0& 0& 0& 0& 1\end{array}\right)---\left(8\right)$

$\mathrm{\Ψ}=\left(\begin{array}{cccc}0.5T^2& 0& 0& 0\\ 0& 0.5T^2& 0& 0\\ T& 0& 0& 0\\ 0& T& 0& 0\\ 0& 0& 0.5T^2& 0\\ 0& 0& 0& 0.5T^2\\ 0& 0& T& 0\\ 0& 0& 0& T\end{array}\right)---\left(9\right)$

The systematic observation equation is following:

$M_k^i={\mathrm{HS}}_k^i+\mathrm{\ζ}---\left(10\right)$

Wherein M is an observation vector, and H is the transition matrix of S to M, and ζ is an observational error;

$H=\left(\begin{array}{cccccccc}1& 0& 0& 0& 0& 0& 0& 0\\ 0& 1& 0& 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 1& 0& 0& 0\\ 0& 0& 0& 0& 0& 1& 0& 0\end{array}\right)---\left(11\right)$

$M_k^i=\begin{array}{c}\left(\begin{array}{c}{p}_{R\_x,k}^i\\ p_{R\_y,k}^i\\ p_{L\_x,k}^i\\ p_{L\_y,k}^i\end{array}\right)\end{array}---\left(12\right)$

Equation (10) and equation (11) are the state equation and the observation equation of Kalman filtering; Employing standard Kalman Filter Technology comes state vector is carried out the continuous real-time optimal estimation, obtains the estimated position and the variance of each pedestrian's target under each sampling instant;

3) renewal of speed and acceleration

To convert feet into be one-period to definition swing pin from beginning to swing to; In one-period; The speed of swing pin and the speed of feet are upgraded by Kalman filtering automatically; The then every one-period of crossing of the acceleration of swing pin upgrades once; Upgrade with the average acceleration value in the last cycle, as the acceleration input quantity in next cycle, direction is all the time by swinging toe to feet; The distance of swing pin motion is s in the assumption period, and the time is t, and then more new formula is following for acceleration:

$a=\frac{4s-4\mathrm{vt}}{t^2}---\left(13\right)$

The beneficial effect that has of the present invention is:

The present invention is a sensor with the laser radar, has overcome the shortcoming that is subject to the illumination variation influence that exists in the tracking of imageing sensor, and reliability is greatly improved; Not only following range is big, and algorithm speed is fast, and it is many to follow the tracks of quantity, and various motor patterns that can compatible pedestrian, comprises standing, walking, jogging, and can very detect pedestrian's position and real-time follow-up pedestrian's movement locus effectively.The present invention is applicable to that the main moving target in the scene is pedestrian's a occasion.

Description of drawings

Fig. 1 is that radar system is arranged synoptic diagram.

Fig. 2 is pedestrian's motion model synoptic diagram.

Fig. 3 is pedestrian's target tracking algorism process flow diagram.

Embodiment

Below in conjunction with accompanying drawing and embodiment the present invention is further described.

Embodiment of the present invention mainly may further comprise the steps:

(1) setting of radar site and system calibrating;

(2) data acquisition and fusion;

(3) pairing of the cluster in motor point and step point;

(4) based on the target following of manikin and Kalman filtering.

The setting and the system calibrating of said step (1) radar site: be about to the diverse location that a plurality of radars are arranged on scene, as shown in Figure 1, through hub with the data transmission of gathering on the different radars to same computing machine, carry out data processing.The position choice principle is set is: (a) make any position of target in scene all be not easy to be blocked; (b) evenly arrange laser radar at the scene edge; To improve the possibility that target is detected; Radar scanning plan range floor level 23～25cm; This guarantees that highly radar in most of the cases can both detect and obtain the range data of two pin of pedestrian, and this excessive height then laser is blocked by the entrained object of pedestrian easily, crosses the low step range data that obtains under the running state of then can't stablizing.Demarcation to a plurality of radars realizes through following method: some bar-shaped objects are set in scene, and the coordinate of establishing i the bar-shaped object that from N radar, obtains is:

$(x_i^1,y_i^1),(x_i^2,y_i^2),...,(x_i^N,y_i^N);$

And suppose that the coordinate with the 1st radar is a reference coordinates, on two dimensional surface, establish the transition matrix R of 1 radar fix of a n radar fix to the _nWith translational movement T _nFor:

$R_n=\left[\begin{array}{cc}{r}_{11}^n& r_{12}^n\\ r_{12}^n& r_{22}^n\end{array}\right]$ $T_n=\left[\begin{array}{cc}{t}_1^n& t_2^n\end{array}\right]$

Then obtain i o'clock transfer equation from 1 radar fix of a n radar fix to the:

$(x_i^1,y_i^1)=R_n(x_i^n,y_i^n)+T_n$

Transition matrix and translational movement are unique when i＞3 confirms, through demarcating data-switching that different radar fixs system is obtained down under same radar fix, in the formula: x representes abscissa value a little, and y representes ordinate value a little.

(2) data acquisition of said step and fusion: promptly through on data transmission to a main frame of hub with a plurality of radars collections, the time of record acquisition is as shown in Figure 1 simultaneously; Data acquisition between a plurality of radars is by unified outer synchronous triggering.During the experiment beginning, through the first background data that extracts each radar automatically of the data of gathering the above frame number of 50 frames, the method for distilling of background data has adopted histogram method; Promptly add up in institute's acquisition frame the range information of data point on each direction; And do statistics with histogram, the frequency of occurrences is the highest is the background dot data on this direction, preserves this background data; Continue experiment then; Range data on all directions that the same time is collected and background data relatively, if then think background dot less than the threshold quantity that sets and remove apart from difference, the foreground point that remaining promptly is in the scene is the motor point; Coordinate conversion is carried out through calibrating parameters R and T in the foreground point that each radar of synchronization obtains, promptly obtain the exercise data point after this fusion that collects constantly under the unified global coordinate system.

The cluster in said step (3) motor point and the pairing of step point: the exercise data point that promptly fusion is obtained; The phase mutual edge distance is carried out cluster less than the point of first distance threshold; And with the position of type center of interior point as candidate's step; As shown in Figure 1, radar scans often more than one of the point that obtains on a pin, and these methods through cluster are found out the step that central point is scanned with representative.Then the step candidate point is carried out the coupling on time domain and the spatial domain, the time domain coupling is meant that with the candidate point in the present frame be the candidate point in the preceding frame of search center search; Promptly search can the matched candidate point in present frame for spatial domain coupling; To confirm whether certain two candidate's step point belongs to same pedestrian; Two kinds of situation are arranged: near (a) search matched two step point prediction positions of the existing target that Kalman filtering obtains; If two candidate's step points in this frame respectively with two predicted positions distances less than the second distance threshold value, judge that then two candidate's step points are the observation station that has target; (b) if current candidate's step point fails in existing target, to find coupling, then belong to fresh target or noise point; The step candidate point that in present frame, does not mate with the search of the 3rd distance threshold then is if find then with match point associating becoming together two the step points that belong to same fresh target that find.

Said step (4) is based on the target following of modelling of human body motion and Kalman filtering: promptly the pedestrian is set up motion model, and adopt Kalman filtering to carry out target following according to this model:

1) foundation of pedestrian dummy

Research shows that pedestrian's state is divided into: stand, walking, jog, hurry up, jogging is meant velocity less than the 3m/ running state of second, because situation is very complicated when hurrying up (velocity is greater than the 3m/ running of second), more difficult modeling puts aside here; The pedestrian always has a feet in the walking and the motion process of jogging, a swing pin, and both periodically exchange, and feet was always not liftoff when different was walking, and remained static; And feet can be liftoff and move distance can be arranged in the cycle when jogging, based on above mention stand, the modeling of walking, the three kinds of pedestrian's states of jogging respectively as follows:

The pattern of standing: do not have the driving of physiological action power, the speed of left and right sides pin is zero, and acceleration also is zero; Force=0, v _{L, k}=0, v _{R, k}=0, a _{L, k}=0, a _{R, k}=0, (1)

Walking and jogging pattern: feet has an invariable speed; The swing pin has an invariable physiological action power; Produce the constant physiological action power of constant magnitude of a forward and negative sense in preceding semiperiod (the swing pin is near the process of feet) and later half cycle (the swing pin is away from the process of feet) respectively, thereby the motion pin is formed acceleration;

With T is the time interval between two frames, and right crus of diaphragm is in the preceding semiperiod of feet:

force＝F，v _L，k＝v _L，k-1+A*t，v _R，k＝V，a _L，k＝A，a _R，k＝0，(2)

Right crus of diaphragm is in the later half cycle of feet:

force＝-F，v _L，k＝v _L，k-1+A*t，v _R，k＝V，a _L，k＝-A，a _R，k＝0，(3)

Left foot is in the preceding semiperiod of feet:

force＝F，v _L，k＝V，v _R，k＝v _R，k-1+A*t?a _L，k＝0，a _R，k＝A，(4)

Left foot is in the later half cycle of feet:

force＝-F，v _L，k＝V，v _R，k＝v _R，k-1+A*t?a _L，k＝0，a _R，k＝-A，(5)

Force representes physiological action power in the formula, and F representes the physiological action power of the actual generation of pedestrian, and v representes movement velocity; V representes the speed of generation under the pedestrian F acting force, and a representes acceleration, and A representes the acceleration that the pedestrian produces under the F acting force; L representes left foot, and R representes right crus of diaphragm, and k representes k constantly;

When V=0, the pedestrian promptly is under the walking state, and when V＞0, the pedestrian promptly is under the state of jogging; As shown in Figure 2, the acceleration scalar value has obvious periodic to change consistently in the motion process, the direction of acceleration then all the time by the swing toe to feet, the speed of feet then is at the uniform velocity to increase progressively or at the uniform velocity successively decrease.

2) Kalman filtering is followed the tracks of

Be used as state variable to pedestrian's position P and speed V, acceleration A is used as the input quantity of system, and the state equation that then obtains system is:

$S_k^i={\mathrm{\ΦS}}_{k-1}^i+{\mathrm{\ΨA}}_k^i+\mathrm{\Ω}---\left(6\right)$

Wherein S representes state variable, and i representes pedestrian's mark, and Ω representes the state model error;

$S_k^i=\left(\begin{array}{c}{p}_{R\_x,k}^i\\ p_{R\_y,k}^i\\ v_{R\_x,k}^i\\ v_{R\_y,k}^i\\ p_{L\_x,k}^i\\ p_{L\_y,k}^i\\ v_{L\_x,k}^i\\ v_{L\_y,k}^i\end{array}\right)$ $A_k^i=\left(\begin{array}{c}{a}_{R\_x,k}^i\\ a_{R\_y,k}^i\\ a_{L\_x,k}^i\\ a_{L\_y,k}^i\end{array}\right)---\left(7\right)$

Represent state-transition matrix with Φ, Ψ representes the transition matrix of acceleration to speed and displacement conversion, and when the radar scanning frequency was 37.5Hz, T was 0.026 second; Then have:

$\mathrm{\Φ}=\left(\begin{array}{cccccccc}1& 0& T& 0& 0& 0& 0& 0\\ 0& 1& 0& T& 0& 0& 0& 0\\ 0& 0& 1& 0& 0& 0& 0& 0\\ 0& 0& 0& 1& 0& 0& 0& 0\\ 0& 0& 0& 0& 1& 0& T& 0\\ 0& 0& 0& 0& 0& 1& 0& T\\ 0& 0& 0& 0& 0& 0& 1& 0\\ 0& 0& 0& 0& 0& 0& 0& 1\end{array}\right)---\left(8\right)$

$\mathrm{\Ψ}=\left(\begin{array}{cccc}0.5T^2& 0& 0& 0\\ 0& 0.5T^2& 0& 0\\ T& 0& 0& 0\\ 0& T& 0& 0\\ 0& 0& 0.5T^2& 0\\ 0& 0& 0& 0.5T^2\\ 0& 0& T& 0\\ 0& 0& 0& T\end{array}\right)---\left(9\right)$

The systematic observation equation is following:

$M_k^i={\mathrm{HS}}_k^i+\mathrm{\ζ}---\left(10\right)$

Wherein M is an observation vector, and H is the transition matrix of S to M, and ζ is an observational error;

$H=\begin{array}{}\end{array}\left(\begin{array}{cccccccc}1& 0& 0& 0& 0& 0& 0& 0\\ 0& 1& 0& 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 1& 0& 0& 0\\ 0& 0& 0& 0& 0& 1& 0& 0\end{array}\right)---\left(11\right)$

$M_k^i=\left(\begin{array}{c}{p}_{R\_x,k}^i\\ p_{R\_y,k}^i\\ p_{L\_x,k}^i\\ p_{L\_y,k}^i\end{array}\right)---\left(12\right)$

Equation (10) and equation (11) are the state equation and the observation equation of Kalman filtering; Employing standard Kalman Filter Technology comes state vector is carried out the continuous real-time optimal estimation, obtains the estimated position and the variance of each pedestrian's target under each sampling instant;

3) renewal of speed and acceleration

To convert feet into be one-period to definition swing pin from beginning to swing to; In one-period; The speed of swing pin and the speed of feet are upgraded by Kalman filtering automatically; The then every one-period of crossing of the acceleration of swing pin upgrades once; Upgrade with the average acceleration value in the last cycle, as the acceleration input quantity in next cycle, direction is all the time by swinging toe to feet; The distance of swing pin motion is s in the assumption period, and the time is t, and then more new formula is following for acceleration:

$a=\frac{4s-4\mathrm{vt}}{t^2}---\left(13\right)$

As shown in Figure 3; After each frame data is through fusion and cluster; Just begin existing tracking target (being recorded in the variable) is carried out model prediction; At first with the predicted position of existing tracking target be search center present frame with the step candidate point in certain this frame of threshold range search as observation station, successfully promptly found the observation station of tracking target in this frame if search for, then carry out filtering; Obtain the current location of filtered, simultaneously renewal speed, acceleration (if one-period just finishes) as tracking target; Promptly do not find the step candidate point as observation station if follow the tracks of failure; Then with the current location of model prediction result as tracking target; And this type of tracking target is labeled as lose follows the tracks of, count, when losing when following the tracks of frame number and reaching 20 frames; We think that this tracking target follows the tracks of failure or left target place, stop to follow the tracks of.Step candidate point for the observation station that is not identified as existing tracking target; We think that it possibly be new pedestrian or noise; We adopt simple means of tracking under the situation that can not confirm its authenticity at once, and these points are carried out the tracking in certain frame number.Because the radar scanning frequency is high, the move distance between the consecutive point is less, is that setting each similar point constantly of appropriate threshold search in center just can find new candidate point so need only with the candidate point; If do not find, then this candidate point is a noise probably, follows the tracks of if found then continued; Carry out the Kalman filtering tracking after determining that it is real pedestrian; So repeat above step, all leave target place up to all tracking targets, this experiment finishes.

Claims (1)

1. pedestrian's method for tracking target based on multilasered optical radar may further comprise the steps:

(1) setting of radar site and system calibrating;

(2) data acquisition and fusion;

(3) pairing of the cluster in motor point and step point;

(4) based on the target following of manikin and Kalman filtering;

It is characterized in that:

The setting and the system calibrating of said step (1) radar site: be about to the diverse location that a plurality of radars are arranged on scene, the position choice principle is set is: (a) make any position of target in scene all be not easy to be blocked; (b) evenly arrange laser radar at the scene edge, with the possibility that the raising target is detected, radar scanning plan range floor level 23～25cm, this guarantees that highly radar in most of the cases can both detect and obtain the range data of two pin of pedestrian; Demarcation to a plurality of radars realizes through following method: some bar-shaped objects are set in scene, and the coordinate of establishing i the bar-shaped object that from N radar, obtains is:

$(x_i^1,y_i^1),(x_i^2,y_i^2),\·\·\·,(x_i^N,y_i^N);$

And suppose that the coordinate with the 1st radar is a reference coordinates, on two dimensional surface, establish the transition matrix R of 1 radar fix of a n radar fix to the _nWith translational movement T _nFor:

$R_n=\left[\begin{array}{cc}{r}_{11}^n& r_{12}^n\\ r_{12}^n& r_{22}^n\end{array}\right]$ $T_n=\left[\begin{array}{cc}{t}_1^n& t_2^n\end{array}\right]$

Then obtain i o'clock transfer equation from 1 radar fix of a n radar fix to the:

$(x_i^1,y_i^1)=R_n(x_i^n,y_i^n)+T_n$

Transition matrix and translational movement are unique when i＞3 confirms, through demarcating data-switching that different radar fixs system is obtained down under same radar fix, in the formula: x representes abscissa value a little, and y representes ordinate value a little;

(2) data acquisition of said step and fusion: promptly through on data transmission to a main frame of hub with a plurality of radars collections, the time of while record acquisition; Data acquisition between a plurality of radars is by unified outer synchronous triggering; Extract earlier the background data of each radar automatically through the data of gathering the above frame number of 50 frames, the method for distilling of background data has adopted histogram method, promptly adds up in institute's acquisition frame the range information of data point on each direction; And do statistics with histogram; The frequency of occurrences is the highest is the background data on this direction, preserves this background data, and range data on all directions that then the same time collected and background data are relatively; If then think background dot less than the threshold quantity that sets and remove apart from difference, the foreground point that remaining promptly is in the scene is the motor point; Coordinate conversion is carried out through calibrating parameters in the foreground point that each radar of synchronization obtains, promptly obtain the exercise data point after this fusion that collects constantly under the unified global coordinate system;

The cluster in said step (3) motor point and the pairing of step point: the exercise data point that promptly fusion is obtained; The phase mutual edge distance is carried out cluster less than the point of first distance threshold; And with the position of type center of interior point as candidate's step; Then the step candidate point is carried out the coupling on time domain and the spatial domain, to confirm whether certain two candidate's step point belongs to same pedestrian, and two kinds of situation are arranged: near (a) search matched two step point prediction positions of the existing target that Kalman filtering obtains; If two candidate's step points in this frame respectively with two predicted positions less than the second distance threshold value, judge that then two candidate's step points are the observation station that has target; (b) if current candidate's step point fails in existing target, to find coupling, then belong to fresh target or noise point; The step candidate point that in present frame, does not mate with the search of the 3rd distance threshold then is if find then with match point associating becoming together two the step points that belong to same fresh target that find;

Said step (4) is based on the target following of modelling of human body motion and Kalman filtering: promptly the pedestrian has been set up motion model, and adopted Kalman filtering to carry out target following according to this model:

5.1) foundation of pedestrian dummy

Research shows that pedestrian's state is divided into: stand, walking, jog, the pedestrian always has a feet in the walking and the motion process of jogging; A swing pin; And both periodically exchange, and feet was always not liftoff when different was walking, and remained static; And feet can be liftoff and move distance can be arranged in the cycle when jogging, based on above mention stand, the modeling of walking, the three kinds of pedestrian's states of jogging respectively as follows:

The pattern of standing: do not have the driving of physiological action power, the speed of left and right sides pin is zero, and acceleration also is zero;

force＝0，v _L，k＝0，v _R，k＝0，a _L，k＝0，a _R，k＝0，(1)

Walking and jogging pattern: feet has an invariable speed; The swing pin has an invariable physiological action power; Produce the constant physiological action power of constant magnitude of a forward and negative sense in preceding semiperiod and later half cycle respectively, thereby the motion pin is formed acceleration;

With T is the time interval between two frames, and right crus of diaphragm is in the preceding semiperiod of feet:

force＝F，v _L，k＝v _L，k-1+A*t，v _R，k＝V，a _L，k＝A，a _R，k＝0，(2)

Right crus of diaphragm is in the later half cycle of feet:

force＝-F，v _L，k＝v _L，k-1-A*t，v _R，k＝V，a _L，k＝-A，a _R，k＝0，(3)

Left foot is in the preceding semiperiod of feet:

force＝F，v _L，k＝V，v _R，k＝v _R，k-1+A*t a _L，k＝0，a _R，k＝A，(4)

Left foot is in the later half cycle of feet:

force＝-F，v _L，k＝V，v _R，k＝v _R，k-1-A*t，a _L，k＝0，a _R，k＝-A，(5)

Force representes physiological action power in the formula, and F representes the physiological action power of the actual generation of pedestrian, and v representes movement velocity; V representes the speed of generation under the pedestrian F acting force, and a representes acceleration, and A representes the acceleration that the pedestrian produces under the F acting force; L representes left foot, and R representes right crus of diaphragm, and k representes k constantly;

When V=0, the pedestrian promptly is under the walking state, and when V＞0, the pedestrian promptly is under the state of jogging;

5.2) the Kalman filtering tracking

Be used as state variable to pedestrian's position P and speed V, acceleration A is used as the input quantity of system, and the state equation that then obtains system is:

$S_k^i=\mathrm{\Φ}{S}_{k-1}^i+\mathrm{\Ψ}{A}_k^i+\mathrm{\Ω}---\left(6\right)$

Wherein S representes state variable, and i representes pedestrian's mark, and Ω representes the state model error;

$S_k^i=\left(\begin{array}{c}{p}_{R\_x,k}^i\\ p_{R\_y,k}^i\\ v_{R\_x,k}^i\\ v_{R\_y,k}^i\\ p_{L\_x,k}^i\\ p_{L\_y,k}^i\\ v_{L\_x,k}^i\\ v_{L\_y,k}^i\end{array}\right)$ $A_k^i=\left(\begin{array}{c}{a}_{R\_x,k}^i\\ a_{R\_y,k}^i\\ a_{L\_x,k}^i\\ a_{L\_y,k}^i\end{array}\right)---\left(7\right)$

Represent state-transition matrix with Φ, Ψ representes the transition matrix of acceleration to speed and displacement conversion, and when the radar scanning frequency was 37.5Hz, T was 0.026 second; Then have:

$\mathrm{\Φ}=\left(\begin{array}{cccccccc}1& 0& T& 0& 0& 0& 0& 0\\ 0& 1& 0& T& 0& 0& 0& 0\\ 0& 0& 1& 0& 0& 0& 0& 0\\ 0& 0& 0& 1& 0& 0& 0& 0\\ 0& 0& 0& 0& 1& 0& T& 0\\ 0& 0& 0& 0& 0& 1& 0& T\\ 0& 0& 0& 0& 0& 0& 1& 0\\ 0& 0& 0& 0& 0& 0& 0& 1\end{array}\right)---\left(8\right)$

$\mathrm{\Ψ}=\left(\begin{array}{cccc}0.5T^2& 0& 0& 0\\ 0& 0.5T^2& 0& 0\\ T& 0& 0& 0\\ 0& T& 0& 0\\ 0& 0& 0.5T^2& 0\\ 0& 0& 0& 0.5T^2\\ 0& 0& T& 0\\ 0& 0& 0& T\end{array}\right)---\left(9\right)$

The systematic observation equation is following:

$M_k^i={\mathrm{HS}}_k^i+\mathrm{\ζ}---\left(10\right)$

Wherein M is an observation vector, and H is the transition matrix of S to M, and ζ is an observational error;

$H=\left(\begin{array}{cccccccc}1& 0& 0& 0& 0& 0& 0& 0\\ 0& 1& 0& 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 1& 0& 0& 0\\ 0& 0& 0& 0& 0& 1& 0& 0\end{array}\right)---\left(11\right)$

$M_k^i=\left(\begin{array}{c}{p}_{R\_x,k}^i\\ p_{R\_y,k}^i\\ p_{L\_x,k}^i\\ p_{L\_y,k}^i\end{array}\right)---\left(12\right)$

Equation (10) and equation (11) are the state equation and the observation equation of Kalman filtering; Employing standard Kalman Filter Technology comes state vector is carried out the continuous real-time optimal estimation, obtains the estimated position and the variance of each pedestrian's target under each sampling instant;

5.3) renewal of speed and acceleration

To convert feet into be one-period to definition swing pin from beginning to swing to; In one-period; The speed of swing pin and the speed of feet are upgraded by Kalman filtering automatically; The then every one-period of crossing of the acceleration of swing pin upgrades once; Upgrade with the average acceleration value in the last cycle, as the acceleration input quantity in next cycle, direction is all the time by swinging toe to feet; The distance of swing pin motion is s in the assumption period, and the time is t, and then more new formula is following for acceleration:

$a=\frac{4s-4\mathrm{vt}}{t^2}---\left(13\right).$

Images