CN116363163B - Space target detection tracking method, system and storage medium based on event camera - Google Patents

Abstract

The invention discloses a space target detection and tracking method, system and storage medium based on an event camera, which belongs to the field of image processing technology. The method includes: step S1, characterizing the collected event signal sequence into three-dimensional voxels; step S2, converting the collected event signal sequence into three-dimensional voxels. The three-dimensional voxels are input into the trained target detection network to obtain the position sequence of the target within the current t moments; step S3, fit the position sequence within the current t moments to obtain the target motion trajectory within the current t moments. Function; Step S4, use the target trajectory function to calculate the target position at the next moment, and at the same time determine whether the target detection network has a detection result at the next moment, and correct the target position at the next moment according to the detection result and the calculation result of trajectory prediction. ; and update the target location. The invention also provides a corresponding detection and tracking system. The present invention can use event data for target detection and tracking, and improves the detection and tracking accuracy of weak and small targets at long distances in space.

Description

Space target detection and tracking method, system and storage medium based on event camera

Technical field

The invention belongs to the field of image processing technology, and more specifically, relates to an event camera-based spatial target detection and tracking method, system and storage medium.

Background technique

Biological vision systems use neuromorphic computing architecture and principles that are completely different from traditional digital signal processors to achieve the purpose of identification, detection and tracking. They are far superior to traditional digital systems in terms of intelligence, operation, quality and power consumption. The event camera is a new type of neuromorphic sensor with the characteristics of high dynamics, low data redundancy and low latency. It can solve the problems of traditional frame imaging cameras such as information redundancy and difficulty in capturing fast moving targets, and achieve low power consumption and low latency. Motion trajectory detection.

Traditional frame imaging sensors use global synchronous integral sampling to record the absolute brightness value of the external environment, while event cameras use asynchronous differential sampling between pixels to only record changes in external illumination changes. In comparison, event signals are more sparse and efficient. However, there is a huge difference in the data form between event data and image frames, and the manifestation of target-background-interference in event data is also essentially different from that in images, which makes it impossible to directly use event data in the existing technology. Perform target detection and tracking.

In long-distance space target detection and tracking tasks, such as space-based ground target detection and tracking, the target's trajectory has the characteristics of local linearity and no sudden changes; the main difficulties in current target detection and tracking tasks are: long detection distance As a result, the target signal is attenuated, and there is background clutter interference such as clouds in the detection path. The existing detection and tracking algorithm is prone to missed detection and false detection of targets, and is difficult to meet the high-precision detection and tracking requirements for weak and small targets at long distances.

Contents of the invention

In view of the shortcomings and improvement needs of the existing technology, the present invention provides a space target detection and tracking method, system and storage medium based on event cameras. The purpose is to use event data for target detection and tracking, and improve the accuracy of weak and small targets in long distances in space. Detection and tracking accuracy.

In order to achieve the above object, according to one aspect of the present invention, a space target detection and tracking method based on event cameras is provided, including:

Step S1: Characterize the collected event signal sequence into three-dimensional voxels;

Step S2: Input the three-dimensional voxels into the trained target detection network to obtain the position sequence of the target within the current t moments; wherein the target detection network is a convolutional neural network;

Step S3: Fit the position sequence within the current t moments to obtain the target motion trajectory function within the current t moments;

Step S4: Use the target trajectory function to calculate the target position V _{C (t+1)} at the next moment, and at the same time determine whether the target detection network has a detection result at the next moment:

If there is a detection result, and the Euclidean distance between the detected target position V _{S (t+1)} and the calculated target position V _{C (t+1)} is less than a given threshold, then the detected target position V _S is used. _(t+1) as the corrected target position V _(t+1) at the next moment. Otherwise, the calculated target position V _{C (t+1)} is used as the corrected target position V _(t +1) at the next moment. ₊₁₎ ;

If there is no detection result, the calculated target position V _{C (t+1)} is used as the corrected target position V _(t+1) at the next moment.

Further, it also includes steps:

Step S5: Using the first-in, first-out method, use the corrected target position V _(t+1) to update the position sequence of the target within the current t times;

Step S6: Repeat steps S3 to S5 to update the target tracking trajectory.

Further, step S2 includes:

Step S21: Input the three-dimensional voxels into the trained target detection network, and output the detection result sequence of the target within the current t times;

Step S22: Eliminate abnormal detection results in the detection result sequence to obtain a normal detection result sequence;

Step S23: Take the last n values in the normal detection result sequence as the position sequence of the target within the current t times, where n is a positive integer.

Further, in step S22, the random sampling consensus method, Kalman filtering method, or correlation filtering method is used to eliminate abnormal detection results in the detection result sequence.

Further, step S1 includes:

Step S11: Construct a three-dimensional space-time body composed of two-dimensional space and one-dimensional time, and divide the three-dimensional space-time body into evenly distributed cubic grids;

Step S12: Fill the event signal sequence into the corresponding cubic grid according to the trigger position and trigger time of the event signal sequence to obtain the three-dimensional voxels of the event signal sequence.

Further, the three-dimensional voxel V(x,y,t) is:

In the formula, 0≤x≤W, 0≤y≤H, H and W represent the number of pixels in the vertical and horizontal directions of the event camera respectively, t represents the current moment, 0≤t≤B, and B is the time within the specified time. The number of units whose dimensions are quantified; i represents the sequence number of the event, the value is [1, N], N represents the total number of events, x _i , y _i represent the pixel position triggered by event i; p _i ∈ [1,-1 ] is the polarity of event i. When the light intensity at the pixel (x _i , y _i ) changes from dark to bright, p _i =1, otherwise p _i =-1; k _b is the interpolation function, k _b (a) = max(0,1-|a|), a represents the interpolation of a certain dimension in the three-dimensional voxel, Indicates the relative position of the current event triggering moment in the three-dimensional voxel time dimension, and the expression is:

In the formula, t _i represents the timestamp triggered by the i-th event.

Further, in step S3, a linear fitting method is used to fit the position sequence within the current t times.

According to another aspect of the present invention, an event camera-based spatial target detection and tracking system is provided, including:

The event signal representation module is used to characterize the collected event signal sequence into three-dimensional voxels;

A position sequence acquisition module, used to input the three-dimensional voxels into the trained target detection network to obtain the position sequence of the target within the current t moments; wherein the target detection network is a convolutional neural network;

The target motion trajectory function fitting module is used to fit the position sequence within the current t moments to obtain the target motion trajectory function within the current t moments;

The target position correction module is used to calculate the target position V _C(t+1) at the next moment using the target trajectory function, and at the same time determine whether the target detection network has a detection result at the next moment:

If there is a detection result, and the Euclidean distance between the detected target position V _{S (t+1)} and the calculated target position V _{C (t+1)} is less than a given threshold, then the detected target position V _S is used. _(t+1) as the corrected target position V _(t+1) at the next moment. Otherwise, the calculated target position V _{C (t+1)} is used as the corrected target position V _(t +1) at the next moment. ₊₁₎ ;

If there is no detection result, the calculated target position V _{C (t+1)} is used as the corrected target position V _(t+1) at the next moment.

Furthermore, it also includes:

The position sequence update module is used to update the position sequence of the target within the current t moments using the corrected target position V _(t+1) in a first-in, first-out manner;

The target tracking module is used to repeatedly execute the target motion trajectory function fitting module and the target position correction module to update the target tracking trajectory.

According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the method according to any one of the first aspects is implemented.

Generally speaking, through the above technical solutions conceived by the present invention, the following beneficial effects can be achieved:

(1) The detection and tracking method of the present invention represents the event data of the space target as a three-dimensional voxel grid, and uses the three-dimensional voxel grid as the input of the target detection network to predict the initial trajectory and target the space target's operating trajectory. Characteristics of using the target motion trajectory function of the current t moments to predict the target position at the next moment. When the space target is obscured by clouds or occlusion caused by the external environment and the target disappears, the predicted target position is directly used as the target at the next moment. position to solve the problem that the target detection network cannot detect the target or misses the target; if the detection result output by the target detection network is significantly different from the predicted target position, the predicted target position will be used as the target position at the next moment to avoid unnecessary errors. Clutter interference, etc. cause false detection problems in the target detection network. The method of the present invention that combines target detection with local fitting trajectories can ensure a relatively accurate target position at every moment, avoid undetectable, missed and false detection of targets, and achieve high-speed and high-precision detection. Target Tracking.

(2) Based on the corrected target position obtained, by updating the position sequence of the target within the current t moments, and predicting the next round of target trajectory, the tracking trajectory update can be achieved.

(3) As an option, by removing abnormal detection results from the sequence of detection results within the current t moments of the target detection network, the convergence speed of the detection and tracking method can be accelerated, and the accuracy of detection and tracking can be improved at the same time; and a method for eliminating abnormal detection results is provided. Corresponding method.

(4) By characterizing the one-dimensional event sequence as a three-dimensional voxel grid, the present invention makes up for the local characteristics of the spatial dimension that do not exist in the one-dimensional event sequence, and establishes a connection between the event data and the image data, so that subsequent Use event data for target detection and tracking; using event data has the advantages of high dynamic range, high frame rate, and low data rate.

(5) The operating trajectory based on the space target has the characteristics of local linearity and no sudden changes. As an option, the linear fitting method is used to fit the position sequence within the current t moments. The method is reasonable and can improve the detection and tracking accuracy.

All in all, the present invention can reasonably characterize the event sequence and improve the accuracy of detection and tracking of weak and small targets at long distances in space through the combination of trajectory prediction and detection network.

Description of the drawings

Figure 1 is a schematic diagram of the event camera-based spatial target detection and tracking method of the present invention.

Figure 2 is a flow chart of the event camera-based spatial target detection and tracking method of the present invention.

Figure 3 is a schematic diagram of the present invention using a random sampling consistent method to exclude and detect outliers and linear fitting.

Figure 4 is a schematic diagram of the invention's mutual correction of target detection results and trajectory prediction results.

Figure 5 is a schematic flow chart of trajectory correction and updating according to the present invention.

Figure 6(a) shows the trajectory output by the target detection network in the embodiment of the present invention; Figure 6(b) is a schematic diagram of trajectory fitting using the detection and tracking method of the present invention; Figure 6(c) shows the target trajectory during the event Illustration of the visualization effect on the frame.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

As shown in Figures 1 and 2, the event camera-based spatial target detection and tracking method of the present invention mainly includes the following steps:

Step S1: Characterize the event signal sequence collected by the event camera into three-dimensional voxels;

Step S2: Input the three-dimensional voxels into the trained target detection network to obtain the position sequence of the target within the current t moments; where the target detection network is a convolutional neural network;

Step S3: Fit the position sequence within the current t moments to obtain the target motion trajectory function within the current t moments;

Step S4, the detection results and the calculation results of trajectory prediction are mutually corrected: use the target trajectory function of the current t moments to calculate the position of the target at t+1 (next moment) V _C(t+1) , and at the same time determine whether the target detection network is next Whether there is a detection result at a moment: If the target detection network has no detection result at the next moment t+1 (that is, the target detection network has no output at t+1), the calculated position of the target at the next moment V _{C(t+ 1)} As the corrected target position V _(t+1) at the next moment;

If the target detection network has a detection result at the next moment, the detected target position is V _S(t+1) , and the detected target position V _S(t+1) is the same as the calculated target position at the next moment V _{C(t +1)} is less than the given threshold, then the detected target position V _{S (t+1)} is used as the corrected target position V _(t+1) at the next moment, otherwise it is considered a false detection generated by the target detection network , at this time, the calculated target position V _C(t+1) at the next time t+1 is used as the corrected target position V _(t+1) at the next time.

Target trajectory prediction is performed through the above steps, and based on the corrected target position V _(t+1) obtained in step S4, the next round of target trajectory prediction is performed to achieve target tracking trajectory update, which includes the following steps:

Step S5, target position sequence update: using the first-in, first-out method, use the corrected target position V _(t+1) to update the target position sequence within the current t moments;

Step S6: Repeat steps S3 to S5 to update the target tracking trajectory.

Specifically, step S1 includes the steps:

Step S11: Construct a three-dimensional space-time body composed of two-dimensional space and one-dimensional time, and divide the three-dimensional space-time body into evenly distributed cubic grids;

Step S12: According to the trigger position and trigger time of the event signal sequence collected by the event camera, fill the one-dimensional event signal sequence into the corresponding cubic grid to obtain the event signal sequence {(xi,y _{i ,} t _i , p _i )} _i∈[1,N] three-dimensional voxels V(x,y,t) to achieve three-dimensional grid representation of event signals;

Among them, 0≤x≤W, 0≤y≤H, H and W represent the number of pixels in the vertical and horizontal directions of the event camera respectively, t represents the current moment; i represents the sequence number of the event, and the value is [1, N], N represents the total number of events, x _i and y _i represent the pixel position triggered by event i, t _i represents the timestamp triggered by event i, p _i ∈ [1,-1] is the polarity of event i, when the pixel is ( _xi , y _i ) When the light intensity changes from dark to bright, p _i =1, otherwise p _i =-1.

Among them, in step S12, the voxel grid (three-dimensional voxels) V(x, y, t) generated by the event sequence is:

In the formula, k _b (a) is the interpolation function, k _b (a) = max (0,1-|a|), a represents the interpolation of a certain dimension in the three-dimensional voxel, Represents the relative position of the current event triggering moment in the time dimension of the voxel grid, and its expression is:

In the formula, t ₁ represents the timestamp triggered by the first event, t _N represents the timestamp triggered by the Nth event, B is the number of units whose time dimension is quantified within the specified time, t represents the current moment, 0≤t ≤B.

After the above three-dimensional voxel representation of the event signal sequence, the event sequence can be characterized as a W×H×B tensor, that is, V(x, y, t) represents a W×H×B tensor. The amount is input into the target detection network in step S2 to detect the initial position of the target.

By characterizing the one-dimensional event sequence as a three-dimensional voxel grid, the present invention makes up for the local characteristics of the spatial dimension that do not exist in the one-dimensional event sequence, and establishes a connection between the event data and the image data, so that the event data can be used subsequently. Perform target detection and tracking. Specifically, step S2 includes:

Step S21: Input the three-dimensional voxels into the trained target detection network, and output the detection result sequence of the target within the current t moments;

Step S22: Use the detection result sequence as the initialization point for trajectory prediction, eliminate abnormal detection results in the detection result sequence, and obtain a normal detection result sequence;

Step S23: Take the last n values in the normal detection result sequence as the position sequence of the target within the current t times, where n is a positive integer. In this embodiment, the length of the detection result sequence output by the target detection network within the current t moments is 10, and the value of n is 5; in practical applications, the value of n is obtained based on experience.

Preferably, in step S22, abnormal detection results in the detection result sequence are eliminated through a random sampling consensus method, a Kalman filter method, or a correlation filter method.

As shown in Figure 3, the random sampling consistent method is used to eliminate abnormal inspection results in the detection result sequence, which mainly includes: randomly selecting any two detection results in the detection result sequence multiple times for linear fitting; among which the outlier points (abnormality There is a big difference in the slope of the straight line fitted by the test result) and the straight line fitted by the normal point (normal test result). This difference is used to exclude abnormal detection points. Since the number of normal detection points in the detection result sequence is greater than the number of abnormal detection points, after multiple random samplings, a set of normal detection points and abnormal detection points can be obtained respectively, thereby eliminating abnormal detection points.

By excluding abnormal detection results, the convergence speed of the detection and tracking method can be accelerated and the accuracy of detection and tracking can be improved.

As a preferred option, the target detection network is a convolutional neural network such as Faster-RCNN and YOLO.

In step S3, the position sequence within the current t moments is fitted through linear fitting or other data fitting methods to obtain the target motion trajectory function of the current t moments.

In step S4, in this embodiment, the given threshold of the Euclidean distance is 5 (calculated in pixel units). In other embodiments, it is selected according to the detection and tracking accuracy.

Specifically, as shown in Figures 4 and 5, the predicted value in the figure refers to the calculated position V _C(t+1) of the target at the next moment, and the detection value refers to the detection result output by the target detection network at the next moment. , that is, the detected target position is V _S(t+1) . In Figure 4, the left picture shows that the Euclidean distance between the predicted value and the detected value at time t+1 is less than the set threshold, and the detected value is used as the target position at the next moment; the right picture shows the difference between the predicted value and the detected value at time t+1. If the Euclidean distance is greater than the set threshold, the predicted value is used as the target position at the next moment.

The detection and tracking method of the present invention uses the target motion trajectory function of the current t moments to predict the target position at the next moment. When the space target is obscured by clouds or the occlusion caused by the external environment causes the target to disappear, the predicted target position is directly used as the next target. The target position at one moment to solve the problem that the target detection network cannot detect the target or misses the target; if the detection result output by the target detection network is significantly different from the predicted target position, the predicted target position will be used as the target at the next moment. location to avoid false detection problems in the target detection network due to clutter interference. In view of the characteristics of the space target's running trajectory, the method of the present invention that combines target detection and local linear fitting trajectory has a relatively accurate target position at every moment, which can avoid undetectable, missed detection and false detection of the target. , which can achieve high-speed and high-precision target tracking.

In the embodiment of the present invention, as shown in Figure 6(a), the target detection network is directly used for target trajectory detection. It can be seen that there are missed detection positions in the middle, and false detection points are also generated; as shown in Figure 6(b) is shown as the result of trajectory fitting using the detection and tracking method of the present invention. Figure 6(c) is the visualization effect of the target trajectory on the event frame. The horizontal and vertical coordinates in Figures 6(a)-6(b) represent pixels. Position. The straight line in the figure is the target trajectory after linear fitting using the Random Sampling Consistency Algorithm (RANSAC). The points in Figure 6(a) represent the target position obtained by using the target detection network. The points in Figure 6(b) The points represent the target positions obtained by using the combined prediction and detection method of the present invention. It can be seen that the present invention can use event cameras to accurately exclude falsely detected targets in an environment with complex backgrounds, complete missed targets, and pass target detection. Combined with local linear fitting trajectories, high-speed and high-precision target tracking is achieved.

According to another aspect of the present invention, a space target detection and tracking system based on event cameras is provided, which mainly includes:

The event signal representation module is used to characterize the collected event signal sequence into three-dimensional voxels;

The position sequence acquisition module is used to input three-dimensional voxels into the trained target detection network to obtain the position sequence of the target within the current t moments; where the target detection network is a convolutional neural network;

The target motion trajectory function fitting module is used to fit the position sequence within the current t moments to obtain the target motion trajectory function within the current t moments;

The target position correction module is used to calculate the target position V _C(t+1) at the next moment using the target trajectory function, and at the same time determine whether the target detection network has a detection result at the next moment:

If there is a detection result, and the Euclidean distance between the detected target position V _S(t+1) and the calculated target position V _C(t+1) is less than the given threshold, then the detected target position V _S(t+1) is used ₎ as the corrected target position V _(t+1) at the next moment, otherwise, use the calculated target position V _{C (t+1)} as the corrected target position V _(t+1) at the next moment;

If there is no detection result, the calculated target position V _{C (t+1)} is used as the corrected target position V _(t+1) at the next moment.

Also includes:

The position sequence update module is used to update the position sequence of the target within the current t times with the corrected target position V _(t+1) in a first-in, first-out manner;

The target tracking module is used to repeatedly execute the target motion trajectory function fitting module, target position correction module and position sequence update module to achieve target tracking trajectory update.

Each module corresponds to each step of implementing the event camera-based spatial target detection and tracking method in the above embodiment. According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, each of the event camera-based spatial target detection and tracking methods in the above embodiments is implemented. step.

It is easy for those skilled in the art to understand that the above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements, etc., made within the spirit and principles of the present invention, All should be included in the protection scope of the present invention.

Claims (5)

1. The method for detecting and tracking the spatial target based on the event camera is characterized by comprising the following steps:

step S1, representing an acquired event signal sequence as a three-dimensional voxel;

s2, inputting the three-dimensional voxels into a trained target detection network to obtain a position sequence of the target in the current t moments; wherein the target detection network is a convolutional neural network;

step S3, fitting a position sequence in the current t moments to obtain a target motion track function in the current t moments;

s4, calculating a target position V at the next moment by using the target motion trail function _C(t+1) And judging whether the target detection network has a detection result at the next moment:

if there is a detection result, and the detected target position V _S(t+1) And the calculated target position V _C(t+1) If the Euclidean distance of (2) is smaller than a given threshold, the detected target position V is adopted _S(t+1) As the target position V after correction at the next time _(t+1) Otherwise, adopting the calculated target position V _C(t+1) As the target position V after correction at the next time _(t+1) ；

If no detection result is obtained, the calculated target position V is adopted _C(t+1) As the target position V after correction at the next time _(t+1) ；

The step S1 comprises the following steps:

s11, constructing a three-dimensional space-time body consisting of a two-dimensional space and a one-dimensional time, and dividing the three-dimensional space-time body into evenly distributed cube grids;

step S12, filling the event signal sequence into a corresponding cube grid according to the triggering position and the triggering time of the event signal sequence to obtain a three-dimensional voxel of the event signal sequence;

the three-dimensional voxels V (x, y, t) are:

wherein x is more than or equal to 0 and less than or equal to W, y is more than or equal to 0 and less than or equal to H, H and W respectively represent the number of pixels in the vertical direction and the horizontal direction of the event camera, t represents the current moment, t is more than or equal to 0 and less than or equal to B, and B is the number of units with quantized time dimension in a specified time; i represents the sequence number of the event, and the value is [1, N]N represents the total number of events, x _i 、y _i Representing the pixel location triggered by event i; p is p _i ∈[1,-1]For the polarity of event i, when the pixel (x _i ，y _i ) P when the light intensity is changed from dark to bright _i =1, otherwise p _i ＝-1；k _b As an interpolation function, k _b (a) =max (0, 1- |a|), a represents the interpolation of a certain dimension in the three-dimensional voxel,and representing the relative position of the current event trigger moment in the three-dimensional voxel time dimension, wherein the expression is as follows:

wherein t is _i A time stamp representing the i-th event trigger;

the step S2 comprises the following steps:

s21, inputting the three-dimensional voxels into a trained target detection network, and outputting a detection result sequence of the target in the current t moments;

s22, eliminating abnormal detection results in the detection result sequence to obtain a normal detection result sequence;

s23, taking the last n values in the normal detection result sequence as a position sequence of the target in the current t moments, wherein n is a positive integer;

the method also comprises the steps of:

step S5, adopting a first-in first-out mode to use the corrected target position V _(t+1) Updating the position sequence of the target in the current t moments;

and S6, repeating the steps S3-S5 to update the target tracking track.

2. The method according to claim 1, wherein in step S22, an anomaly detection result in the detection result sequence is removed by a random sampling coincidence method, a kalman filtering method, or a correlation filtering method.

3. The method according to claim 1, characterized in that in step S3, a linear fitting method is used to fit the sequence of positions within the current t moments.

4. A spatial target detection tracking system based on an event camera, comprising:

the event signal characterization module is used for characterizing the acquired event signal sequence as a three-dimensional voxel; the method specifically comprises the following steps: constructing a three-dimensional space-time body consisting of a two-dimensional space and a one-dimensional time, and dividing the three-dimensional space-time body into evenly distributed cube grids; filling the event signal sequence into a corresponding cube grid according to the triggering position and the triggering time of the event signal sequence to obtain a three-dimensional voxel of the event signal sequence; the three-dimensional voxels V (x, y, t) are:

wherein x is more than or equal to 0 and less than or equal to W, y is more than or equal to 0 and less than or equal to H, H and W respectively represent the number of pixels in the vertical direction and the horizontal direction of the event camera, t represents the current moment, t is more than or equal to 0 and less than or equal to B, and B is the number of units with quantized time dimension in a specified time; i represents the sequence number of the event, and the value is [1, N]N represents the total number of events, x _i 、y _i Representing thingsPixel location triggered by element i; p is p _i ∈[1,-1]For the polarity of event i, when the pixel (x _i ，y _i ) P when the light intensity is changed from dark to bright _i =1, otherwise p _i ＝-1；k _b As an interpolation function, k _b (a) =max (0, 1- |a|), a represents the interpolation of a certain dimension in the three-dimensional voxel,and representing the relative position of the current event trigger moment in the three-dimensional voxel time dimension, wherein the expression is as follows:

wherein t is _i A time stamp representing the i-th event trigger;

the position sequence acquisition module is used for inputting the three-dimensional voxels into a trained target detection network to obtain a position sequence of the target in the current t moments; the method specifically comprises the following steps: inputting the three-dimensional voxels into a trained target detection network, and outputting a detection result sequence of the target in the current t moments; removing abnormal detection results in the detection result sequence to obtain a normal detection result sequence; taking the last n values in the normal detection result sequence as a position sequence of the target in the current t moments, wherein n is a positive integer, and the target detection network is a convolutional neural network;

the target motion track function fitting module is used for fitting the position sequences in the current t moments to obtain the target motion track functions in the current t moments;

a target position correction module for calculating a target position V at the next moment by using the target motion track function _C(t+1) And judging whether the target detection network has a detection result at the next moment:

if there is a detection result, and the detected target position V _S(t+1) And the calculated target position V _C(t+1) If the Euclidean distance of (2) is less than a given thresholdUsing the detected target position V _S(t+1) As the target position V after correction at the next time _(t+1) Otherwise, adopting the calculated target position V _C(t+1) As the target position V after correction at the next time _(t+1) ；

If no detection result is obtained, the calculated target position V is adopted _C(t+1) As the target position V after correction at the next time _(t+1) ；

Further comprises:

a position sequence updating module for using the corrected target position V in a first-in first-out manner _(t+1) Updating the position sequence of the target in the current t moments;

and the target tracking module is used for repeatedly executing the target motion track function fitting module, the target position correction module and the position sequence updating module to update the target tracking track.

5. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method according to any of claims 1-3.