A trajectory data fusion method based on activation function

Abstract

The application discloses a trajectory data fusion method based on an activation function, which comprises the following steps: S1, setting an outlier elimination condition, and eliminating GPS displacement data that does not meet the condition as outliers; S2, establishing a motion system model for an object and filtering GPS displacement data and magnetic induction data by using a Kalman filtering formula; S3, performing alignment calculation on the basis of magnetic induction speed data, and performing interpolation alignment on GPS displacement data and magnetic induction displacement data according to different sequences; and S4, introducing an activation function based on motion speed, and calculating fused data. The application adopts the event-triggered outlier elimination strategy to eliminate outliers in GPS data, respectively performs Kalman filtering on different sensor data, adopts a bidirectional interpolation mode to perform data alignment, and introduces a speed-oriented activation function in data fusion, and the like, so that the data fusion technology is improved, and more accurate trajectory data can be obtained.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of data fusion, and particularly relates to a trajectory data fusion method based on an activation function. BACKGROUND

[0002] Data fusion technology refers to an information processing technology that utilizes a computer to automatically analyze and synthesize a plurality of observation information obtained in time sequence under certain criteria to complete a required decision and evaluation task. At present, data fusion technology is widely applied in military and non-military fields, and is widely applied in automatic target recognition, battlefield monitoring and the like in the military field, and is widely applied in environmental monitoring, medical technology and the like in the non-military field.

[0003] Data fusion technology realizes more accurate measurement and judgment than a single sensor by reasonably fusing data and related information from a plurality of sensors. Compared with single sensor measurement, a plurality of sensors can still work normally under the condition that one or more sensors are faulty, and are more reliable. Secondly, a plurality of sensor measurement can improve measurement diversity and improve the observation process, such as acceleration, speed and displacement measurement of a moving object. However, different sensors usually have different measurement properties, and when different sensor outputs are subjected to multi-sensor data fusion, the fusion technology used also needs to be designed according to the performance of the sensor.

[0004] Outdoor motion trajectory recording mainly relies on position data provided by GPS positioning, and therefore, the accuracy of GPS positioning data directly affects the quality of motion trajectory recording. However, the technology for improving the accuracy of GPS positioning data has the following defects:

[0005] (1) Direct filtering estimation or data fusion on original data is easily affected by outliers (information with large deviation).

[0006] (2) GPS positioning data is often fused with IMU data, but IMU data is easily affected by electromagnetic interference, and the installation and use environment has relatively strict requirements, and IMU data output is acceleration data, which needs to be integrated to be fused with GPS position data.

[0007] (3) Different sensors have different properties, and the data update frequency is also different, and there is a difference in the data collected at different times. If the output data of the sensors is directly fused, the measurement difference caused by the different update times will make the fused data deviate from the true value. SUMMARY

[0008] The present application aims to provide a trajectory data fusion method based on an activation function, which can obtain more accurate trajectory data.

[0009] The above-mentioned objective of this invention is achieved through the following technical solution: a trajectory data fusion method based on activation functions, characterized by comprising the following steps:

[0010] S1. Set out the outlier removal criteria:

[0011] ||S g (x g y g )-S(x,y)||<h×||V l (x l y l )×(T g -T l )|| Formula (1)

[0012] Collect GPS displacement data S g (x g y g ), and the GPS displacement data S that do not satisfy formula (1) g (x g y g ) are excluded as outliers;

[0013] S2. Establish a motion system model for the object:

[0014]

[0015] Using the Kalman filter formula:

[0016]

[0017] GPS displacement data S g (x g y g ) Perform filtering, z(k)=S g (x g y g After filtering, the GPS displacement data S is obtained. kg (x kg y kg ); Filter the magnetic induction data, z(k) = S l (x l y l After filtering, the magnetic displacement data S is obtained. km (x km y km ) and magnetic induction velocity data V km (x km y km );

[0018] S3, alignment calculation is performed with the magnetic induction speed data, and interpolation alignment is performed on the GPS displacement data and the magnetic induction displacement data according to different sequences, T g is the GPS displacement data update time, T m is the magnetic induction displacement data update time, and the data interpolation calculation is as follows:

[0019] a, aligning to the GPS displacement data, assuming that the aligned magnetic induction displacement data is S rm (x rm , y rm ), then:

[0020] S rm (x rm , y rm ) = S km (x km , y km ) + V km (x km , y km ) × (T g -T l ) Formula (4)

[0021] b, aligning to the magnetic induction displacement data, assuming that the aligned GPS displacement data is S rg (x rg , y rg ), then:

[0022] S rg (x rg , y rg ) = S kg (x kg , y kg ) + V km (x km , y km ) × (T l -T g ) Formula (5)

[0023] Select the data with smaller alignment error in the two sets of alignment data a and b as the alignment result data, and the alignment time point is recorded as T r . For set a, the alignment error is the absolute value difference between S rm (x rm , y rm ) and S kg (x kg , y kg ); for set b, the alignment error is the absolute value difference between S rg (x rg , y rg ) and S km (x km , ykm The absolute value difference;

[0024] S4. Introduce an activation function based on motion speed:

[0025]

[0026] In the formula, Δ is the offset;

[0027] The merged data is calculated using the following formula:

[0028]

[0029] This invention removes outliers from GPS data. This outlier removal strategy is simple to design, fast to calculate, and effectively eliminates outlier data, avoiding the influence of points with large errors on the trajectory estimation results. This invention applies Kalman filtering to data from different sensors separately to avoid mutual interference between data and weaken the impact of noise and observation interference on the measurement data. This invention uses bidirectional interpolation for data alignment, obtaining better alignment results based on interpolation errors, effectively reducing the interference of measurement errors caused by different sampling times of different sensors on the data fusion results. A velocity-guided activation function is introduced into the data fusion process, using motion velocity as the independent variable to adaptively adjust the fusion weights, cleverly reducing the impact of GPS measurement data drift on the estimation of the true motion trajectory. Therefore, through the above improvements, this invention can obtain more accurate trajectory data.

[0030] In formula (4) of this invention, when T g >T m When T increases the magnetic induction displacement data by a certain displacement, it is for forward interpolation alignment; when T g <T m At that time, the magnetic induction displacement data is subtracted by a certain displacement amount for backward interpolation alignment.

[0031] In formula (5) of this invention, when T g >T m At that time, the GPS displacement data is increased by a certain displacement amount for forward interpolation alignment; when T g <T m At that time, the GPS displacement data is subtracted by a certain displacement amount for backward interpolation alignment.

[0032] In this invention, in formula (1), the weight h is greater than or equal to 3 and less than or equal to 7.

[0033] Compared with the prior art, the present invention has the following significant effects:

[0034] (1) The application eliminates the outliers in GPS displacement data, and the outlier elimination strategy is simple in design, fast in calculation, and can effectively eliminate outlier data, avoiding the influence of points with large errors on the trajectory estimation results.

[0035] (2) The application respectively performs Kalman filtering on different sensor data, avoids mutual interference between data, and weakens the influence of noise interference, observation interference and the like on the measurement data.

[0036] (3) The application performs data alignment in a bidirectional interpolation manner, obtains better alignment results according to interpolation errors, and effectively reduces the interference of measurement errors caused by different sampling times of different sensors on the data fusion results.

[0037] (4) In the data fusion, a velocity-oriented activation function is introduced, taking the motion velocity as the independent variable, and adaptively adjusting the fusion weight, which skillfully reduces the influence of GPS measurement data drift on the estimation of the real motion trajectory.

[0038] (5) The application can obtain more accurate trajectory data through the above improvement measures of the data fusion technology. BRIEF DESCRIPTION OF DRAWINGS

[0039] The application will be further described in detail below in combination with the drawings and specific embodiments.

[0040] Figure 1 is a flow chart of the application. DETAILED DESCRIPTION

[0041] In order to facilitate the description, the main symbols used in the description of the application are defined as follows: world coordinate system (x, y), the current displacement output by the GPS is S(x, y), the current displacement and current velocity output by the magnetic induction module are S(x, y) and V(x, y) respectively, the current displacement and velocity of the system are S(x, y) and V(x, y) respectively, the displacement and velocity of the system at the last time are S(x, y) and V(x, y) respectively, and the last data update time of the system is T. g g g m m m m m m l l l l l l l

[0042] ​​​​​​​​​​​​​​​​This invention leverages the long-term stability of GPS positioning and the accuracy of magnetic sensor speed measurement by filtering and fusing data from multiple sensors to obtain more accurate trajectory data. For the magnetic induction module, this invention uses an electromagnetic speed sensor on a bicycle wheel as an example for illustration.

[0043] like Figure 1 The diagram illustrates a trajectory data fusion method based on an activation function according to the present invention, comprising the following steps:

[0044] S1. Outlier Removal. The change in an object's velocity is continuous; abrupt changes in velocity within a short time interval are unlikely, meaning the object's displacement increment within a short period won't be significant. Therefore, estimating the current displacement based on historical data and comparing it to the current displacement can effectively remove outliers. This is done based on data from the system's previous update and the current GPS update time T. g Based on the current displacement information, the following event-triggered outlier culling strategy can be constructed:

[0045] ||S g (x g y g )-S(x,y)||<h×||V l (x l y l )×(T g -T l )|| Formula (1)

[0046] The weight h ranges from [3, 7] and can be set according to requirements. For objects with rapidly changing speeds, a larger weight can be set. Taking bicycle speed as an example, h = 3.5 can be used. When the current GPS displacement data S g (x g y g If formula (1) is not satisfied, it indicates that the current GPS displacement data S g (x g y g The displacement data S has deviated significantly from the possible range of the current position. g (x g y g Points with large errors are identified as outliers and discarded. This outlier removal strategy is simple to design, fast to compute, and can effectively remove outlier data, avoiding the influence of points with large errors on the trajectory estimation results.

[0047] S2. Kalman Filtering. To obtain more accurate data fusion results, the raw output data of the sensors needs to be filtered to reduce the impact of noise interference, observation interference, and other factors on the measurement data. First, a motion system model of the object is established:

[0048]

[0049] wherein,

[0050]

[0051] Kalman filter formula:

[0052]

[0053] wherein,

[0054]

[0055]

[0056] For a variable w, w(k) represents the value of w at time k, w(k|k) is equivalent to w(k), and w(k|k-1) represents the estimated w(k) obtained from w(k-1).

[0057] The above motion system model and Kalman filter formula are in a unified framework, when filtering GPS data, z(k) = S g (x g , y g ), and the filtered GPS displacement data is S kg (x kg , y kg ); when filtering magnetic induction data, z(k) = S l (x l , y l ), and the filtered displacement data and velocity data are S km (x km , y km ) and V km (x km , y km ) respectively.

[0058] S3, data interpolation. Different sensors have different properties, and the data update frequency is also different, which means that it is difficult to ensure that different sensors update data at the same time. The data collected at different times will inevitably have differences, and if the output data of the sensor is directly fused, the measurement difference caused by the different update times will make the fused data deviate from the true value. The application adopts an interpolation method to time-align the data output by different sensors. Since the velocity data obtained by the magnetic induction module is more real-time and more accurate, when performing data alignment, the velocity data of the magnetic induction module is used for alignment calculation, and the GPS displacement data and the magnetic induction displacement data are interpolated and aligned according to different sequences. The GPS data update time and the magnetic induction data update time are T g , T mData interpolation is calculated as follows:

[0059] a. Aligning to GPS data, assuming that the aligned magnetic induction data is S rm (x rm , y rm ), then:

[0060] S rm (x rm , y rm ) = S km (x km , y km ) + V km (x km , y km ) x (T g - T l ) Formula (4)

[0061] b. Aligning to magnetic induction data, assuming that the aligned GPS data is S rg (x rg , y rg ), then:

[0062] S rg (x rg , y rg ) = S kg (x kg , y kg ) + V km (x km , y km ) x (T l - T g ) Formula (5)

[0063] In formula (4), when T g > T m , the magnetic induction data needs to increase a certain displacement amount to compensate for the measurement error caused by the earlier data update time. This alignment method is forward interpolation alignment from the time scale. Conversely, when T g < T m , the magnetic induction data needs to be subtracted by a certain displacement amount to remove the displacement variation amount added due to the late data update. This is backward interpolation alignment.

[0064] In formula (5), when T g > T m , the GPS displacement data increases a certain displacement amount to compensate for the measurement error caused by the earlier data update time. This alignment method is forward interpolation alignment from the time scale. When T g < T mAt this time, the GPS displacement data is subtracted by a certain displacement amount to remove the displacement variation amount added due to late data updating, for alignment of the backward interpolation.

[0065] Finally, the data with smaller error in the alignment data of the two groups a and b is selected as the alignment result data, and the alignment time point is recorded as T r For the group a, the alignment error is S rm (x rm , y rm ) and the absolute value difference of S kg (x kg , y kg ). For the group b, the alignment error is S rg (x rg , y rg ) and the absolute value difference of S km (x km , y km ).

[0066] S4, data fusion and data updating. After data alignment of the filtering result, the next step is to fuse the data obtained by different sensors. The common fusion method is to set the fusion weight according to the observation error variance of the sensor, but considering the drift problem of GPS at low speed, the application introduces an activation function based on the motion speed for data fusion:

[0067]

[0068] Wherein, e is a natural constant, and the value is e≈2.72, and Δ is a displacement amount, which can be used to control when the speed is discarded GPS data, and the walking speed of a person is taken as the minimum speed, and Δ can be set to 4mk / h. The purpose of introducing the activation function is to reduce the reliability of GPS data at low speed. The fused data can be calculated by the following formula:

[0069]

[0070] The embodiments of the application are not limited to this, according to the above content of the application, according to the ordinary technical knowledge and common means in the art, other various forms of modification, replacement or change can be made without departing from the above technical idea of the application, and all fall within the protection scope of the application.

Claims

1. A trajectory data fusion method based on activation function, characterized in that The method comprises the following steps: S1, establishing an external point rejection condition: ||S g (x g ,y g ) - S(x, y) || <h×||V l (x l ,y l ) × (T g -T l ) || Formula (1) The weight h is greater than or equal to 3 and less than or equal to 7; Collect GPS displacement data S g (x g ,y g GPS displacement data S that do not satisfy formula (1) g (x g ,y g ) are excluded as outliers; S2, establishing a motion system model for the object: The Kalman filtering formula is adopted: filtering the GPS displacement data S g (x g ,y g ), z(k) = S g (x g ,y g ), to obtain filtered GPS displacement data S kg (x kg ,y kg ); filtering the magnetic induction data, z(k) = S l (x l ,y l ), to obtain filtered magnetic induction displacement data S km (x km ,y km ) and filtered magnetic induction velocity data V km (x km ,y km ); S3, alignment calculation is performed with the magnetic induction speed data, and interpolation alignment is performed on the GPS displacement data and the magnetic induction displacement data according to different sequences, T g is the GPS displacement data update time, T m is the magnetic induction displacement data update time, and the data interpolation calculation is as follows: a. Aligning to GPS displacement data, let the aligned magnetic induction displacement data be S rm (x rm ,y rm ), then: S rm (x rm ,y rm ) = S km (x km ,y km ) + V km (x km ,y km ) × (T g - T l ) Equation (4) When T g >T m When T increases the magnetic induction displacement data by a certain displacement, it is for forward interpolation alignment; when T g <T m At that time, the magnetic induction displacement data is subtracted by a certain displacement amount for backward interpolation alignment; b. Align the magnetic induction displacement data, and let the aligned GPS displacement data be S rg (x rg ,y rg ), then: S rg (x rg ,y rg ) = S kg (x kg ,y kg ) + V km (x km ,y km ) × (T l - T g ) Equation (5) When T g >T m At that time, the GPS displacement data is increased by a certain displacement amount for forward interpolation alignment; when T g <T m At that time, the GPS displacement data is subtracted by a certain displacement amount for backward interpolation alignment; Select a, b two groups of alignment data in the alignment error smaller data as the alignment result data, alignment time point is recorded as T r , for a group, alignment error is S rm (x rm ,y rm ) and S kg (x kg ,y kg ) absolute value difference; for b group, alignment error is S rg (x rg ,y rg ) and S km (x km ,y km ) absolute value difference; S4, introducing a motion speed-based activation function: In the formula, Δ is an offset; The fused data is calculated by the following formula:

Images