JP6873009B2 - Radar device and target detection method - Google Patents

Description

The present invention relates to a radar device and a target detection method.

Conventionally, for example, there is a radar device that is mounted on a vehicle, receives a reflected wave that is reflected by a transmitted wave transmitted from the vehicle and hits a target, and periodically detects the target by analyzing the received signal.

In such a radar device, by analyzing the reflection points detected based on the transmitted wave and the reflected wave, an observed value (also called an instantaneous value) including factors such as the distance to the target and the velocity is generated. To. Then, the radar device generates target data corresponding to each observed value by applying a filter process for each observed value obtained.

For example, radar devices are known to use particle filters in filtering. In the particle filter, the target data is estimated from the observed values by plotting the latest observed values on a state space model in which a predetermined number of particle data are distributed (see, for example, Patent Document 1).

Here, in the particle filter, prediction processing is performed to predict the distribution state of the particle data used in the latest cycle based on the distribution state of the particle data used in the previous cycle. In the prediction process, for example, the latest particle data is generated by moving the particle data of the previous cycle according to the motion model such as the velocity included in the target data of the previous cycle.

Japanese Unexamined Patent Publication No. 2013-238442

However, the conventional technique has room for improvement in improving the prediction accuracy in the prediction process. Specifically, in the prediction process, the particle data is moved on the assumption that the target detected in the previous cycle moves in almost the same way in the latest cycle, so that, for example, the target state is one previous. If there is a large change between the period and the latest period, the prediction error of the latest particle data may become relatively large.

The present invention has been made in view of the above, and an object of the present invention is to provide a radar device and a target detection method capable of improving the prediction accuracy of particle data in a particle filter.

In order to solve the above-mentioned problems and achieve the object, the radar device according to the present invention includes a detection unit and a filter processing unit. The detection unit detects an observed value corresponding to the target based on the frequency-modulated transmitted wave and the reflected wave of the transmitted wave by the target. The filter processing unit generates target data corresponding to the observed value by applying a particle filter that associates a predetermined number of particle data with the observed value detected by the detection unit. Further, the filter processing unit includes a prediction unit that applies a Kalman filter to predict the current particle data in the prediction processing from the previous particle data to the current particle data in the particle filter. The prediction unit corresponds to the previous particle data and applies the Kalman filter to a smaller number of representative particle data than the previous particle data.

According to the present invention, it is possible to improve the prediction accuracy of particle data in the particle filter.

FIG. 1A is a diagram showing an outline of a target detection method according to an embodiment. FIG. 1B is a diagram showing an outline of a target detection method according to an embodiment. FIG. 2 is a block diagram showing a configuration of a radar device according to an embodiment. FIG. 3 is an explanatory diagram of processing from the pre-stage processing of the signal processing unit to the peak extraction processing in the signal processing unit. FIG. 4A is a processing explanatory view of the angle estimation process. FIG. 4B is a process explanatory diagram (No. 1) of the pairing process. FIG. 4C is a process explanatory diagram (No. 2) of the pairing process. FIG. 5 is a diagram showing the processing contents of the prediction unit. FIG. 6 is a diagram showing the processing contents of the prediction unit. FIG. 7 is a diagram showing the processing contents of the prediction unit. FIG. 8 is a diagram showing the processing contents of the prediction unit. FIG. 9 is a flowchart showing a processing procedure of processing executed by the radar device according to the embodiment. FIG. 10 is a flowchart showing a processing procedure of processing executed by the Kalman filter according to the embodiment.

Hereinafter, embodiments of the radar device and the target detection method disclosed in the present application will be described in detail with reference to the accompanying drawings. The present invention is not limited to this embodiment. In the following, the case where the radar device 1 is of the FM-CW (Frequency Modulated Continuous Wave) method will be described as an example, but the radar device 1 will be described by another method such as the FCM (Fast-Chirp Modulation) method. There may be.

First, the outline of the target detection method according to the embodiment will be described with reference to FIGS. 1A and 1B. 1A and 1B are diagrams showing an outline of a target detection method according to an embodiment. As shown in FIG. 1A, the radar device 1 is mounted in, for example, the front grill of the vehicle MC, and detects a target (for example, the preceding vehicle LC, etc.) existing in the traveling direction of the vehicle MC. The radar device 1 may be mounted on other places such as the windshield, the rear grille, and the left and right side parts (for example, the left and right door mirrors).

As shown in FIG. 1A, the radar device 1 according to the embodiment first detects an observed value corresponding to a target. The observed value is a value indicating the state vector of the reflection point detected based on the frequency-modulated transmitted wave and the reflected wave of the transmitted wave by the target. Such a state vector includes values such as the distance to the reflection point and the relative velocity. The observed value is sometimes called an instantaneous value or the like.

Subsequently, as shown in FIG. 1A, the radar device 1 according to the embodiment performs a filter process on the detected observed value to generate target data which is a filter value corresponding to the observed value. As shown in FIG. 1A, for example, a particle filter is used for the filtering process.

The particle filter estimates target data by plotting a predetermined number of particle data and observed values in a predetermined state space and analyzing the positional relationship in the state space. As shown in FIG. 1A, the particle filter generates target data from observed values by performing "prediction", "allocation", "weighting", "resampling", and "data generation" processing. ..

Here, each process of the particle filter will be briefly described. "Prediction" is a process of predicting the distribution state of particle data used in the latest cycle in the state space. Specifically, "prediction" is a prediction process for predicting the distribution state in the current particle data, which is the latest cycle, from the distribution state in the previous particle data, which is the previous cycle.

"Assignment" is a process of allocating the observed value of this time, which is the latest cycle, to the predicted particle data of this time. In "assignment", for example, observation values existing in a predetermined allocation range from the previous target data are assigned to the current particle data.

The "weighting" is a process of weighting each of the current particle data based on the assigned current observation value. "Resampling" is a process of rearranging (resampling) each of the current particle data based on the weight of each of the current particle data by "weighting". "Data generation" is a process of generating the target data of this time from the observed value of this time based on the resampled particle data of this time. The details of each process of the particle filter will be described later.

Here, a conventional target detection method will be described. In the conventional target detection method, "prediction" in the particle filter is performed according to a motion model. Specifically, "prediction" generates the current particle data by moving the previous particle data from one point to another based on the information such as the position and velocity included in the previous target data. It was. In other words, the particle data this time was predicted on the assumption that the target moves in almost the same way between the previous cycle and the latest cycle.

However, for example, when the state such as the speed and the moving direction of the target changes significantly between the previous cycle and the latest cycle (for example, sudden acceleration / deceleration, sharp turn, etc.), the motion model is followed as in the conventional case. When "prediction" is performed, there is a possibility that the prediction error in the particle data this time becomes relatively large.

Therefore, in the target detection method according to the embodiment, in the "prediction" process, a Kalman filter is applied to predict the current particle data from the previous particle data. As the Kalman filter, for example, an unscented Kalman filter can be used, but the present invention is not limited to this, and for example, an extended Kalman filter or the like may be used, and a non-linear Kalman filter may be used. Hereinafter, a case where an unscented Kalman filter is applied in the target detection method according to the embodiment will be described.

In general, the Kalman filter calculates the optimum Kalman gain that minimizes both errors based on the latest observed value including the error and the previous state data including the error, and uses the optimum Kalman gain to calculate the optimum Kalman gain. The latest state data is predicted from the state data of.

That is, the Kalman filter KF according to the embodiment assumes that both the latest observed value obtained and the previous particle data contain an error, and when the previous particle data is moved according to the motion model, such an error occurs. Calculate the minimum optimum Kalman gain. Then, the Kalman filter KF generates the current particle data Pb from the previous particle data Pa by using the motion model and the optimum Kalman gain.

Specifically, as shown in FIG. 1B, the radar device 1 according to the embodiment applies the Kalman filter KF to each of the previous particle data Pa1 to Pa8 (may be described as the previous particle data Pa). ..

For example, the Kalman filter KF calculates the optimum Kalman gain based on the previous particle data Pa1 and the current observed value, and predicts the current particle data Pb1. That is, the Kalman filter KF calculates the optimum Kalman gain for each of the previous particle data Pa1 to Pa8, and predicts the current particle data Pb1 to Pb8. Specifically, the difference between the predicted current particle data Pb1 to Pb8 and the true value is generated by generating the current particle data Pb1 to Pb8 from the previous particle data Pa1 to Pa8 using the calculated optimum Kalman gain. (That is, the error) can be reduced. Note that FIG. 1B shows eight previous particle data Pa1 to Pa8 and current particle data Pb1 to Pb8, but the number is not limited to eight and any number can be set.

As a result, the error of the particle data Pb1 to Pb8 of this time, which is the output result of the Kalman filter KF, becomes the minimum. That is, according to the target detection method according to the embodiment, it is possible to improve the prediction accuracy of the particle data Pb1 to Pb8 in the particle filter.

The radar device 1 according to the embodiment does not need to apply the Kalman filter KF to all the previous particle data Pa1 to Pa8. For example, the radar device 1 corresponds to one representative particle data corresponding to the average of the previous particle data Pa1 to Pa8. The Kalman filter KF may be applied. Details of this point will be described later.

Further, the radar device 1 according to the embodiment can further improve the prediction accuracy by further inputting the previous observation value into the Kalman filter KF in addition to the current observation value, and the details of this point will be described later. To do.

Next, the configuration of the radar device 1 according to the embodiment will be described in detail with reference to FIG. FIG. 2 is a block diagram showing the configuration of the radar device 1 according to the embodiment. Note that, in FIG. 2, only the components necessary for explaining the features of the present embodiment are represented by functional blocks, and the description of general components is omitted.

In other words, each component shown in FIG. 2 is a functional concept and does not necessarily have to be physically configured as shown in the figure. For example, the specific form of distribution / integration of each functional block is not limited to the one shown in the figure, and all or part of the functional blocks are functionally or physically distributed in arbitrary units according to various loads and usage conditions. -It is possible to integrate and configure.

As shown in FIG. 2, the radar device 1 includes a transmitting unit 10, a receiving unit 20, and a processing unit 30. The radar device 1 is connected to a vehicle control device 2 that controls the behavior of the vehicle MC.

The vehicle control device 2 controls a vehicle such as a PCS (Pre-crash Safety System) or an AEB (Advanced Emergency Braking System) based on the detection result of a target by the radar device 1.

The transmission unit 10 includes a signal generation unit 11, an oscillator 12, and a transmission antenna 13. The signal generation unit 11 generates a modulated signal for transmitting a millimeter wave frequency-modulated with a triangular wave under the control of a transmission / reception control unit 31 described later. The oscillator 12 generates a transmission signal based on the modulation signal generated by the signal generation unit 11, and outputs the transmission signal to the transmission antenna 13. As shown in FIG. 2, the transmission signal generated by the oscillator 12 is also distributed to the mixer 22 described later.

The transmission antenna 13 converts the transmission signal from the oscillator 12 into a transmission wave, and outputs the transmission wave to the outside of the vehicle MC. The transmitted wave output by the transmitting antenna 13 is a continuous wave frequency-modulated with a triangular wave. The transmitted wave transmitted from the transmitting antenna 13 to the outside of the vehicle MC, for example, forward, is reflected by a target such as the preceding vehicle LC and becomes a reflected wave.

The receiving unit 20 includes a plurality of receiving antennas 21 forming an array antenna, a plurality of mixers 22, and a plurality of A / D conversion units 23. The mixer 22 and the A / D conversion unit 23 are provided for each receiving antenna 21.

Each receiving antenna 21 receives the reflected wave from the target as a received wave, converts the received wave into a received signal, and outputs the received wave to the mixer 22. Although the number of receiving antennas 21 shown in FIG. 2 is 4, it may be 3 or less or 5 or more.

The received signal output from the receiving antenna 21 is amplified by an amplifier (for example, a low noise amplifier) (not shown) and then input to the mixer 22. The mixer 22 mixes a part of the distributed transmission signal and the reception signal input from the reception antenna 21 to remove unnecessary signal components to generate a beat signal, and outputs the beat signal to the A / D conversion unit 23. ..

The beat signal is a difference wave between the transmitted wave and the reflected wave, and is the frequency of the transmitted signal (hereinafter referred to as “transmission frequency”) and the frequency of the received signal (hereinafter referred to as “reception frequency”). It has a different beat frequency. The beat signal generated by the mixer 22 is converted into a digital signal by the A / D conversion unit 23, and then output to the processing unit 30.

The processing unit 30 includes a transmission / reception control unit 31, a signal processing unit 32, and a storage unit 33. The signal processing unit 32 includes a detection unit 32a and a filter processing unit 32b.

The storage unit 33 stores the history data 33a. The history data 33a is information including a history of target data in a series of signal processing executed by the signal processing unit 32.

The processing unit 30 is, for example, a microcomputer including a CPU (Central Processing Unit), a ROM (Read Only Memory) and a RAM (Random Access Memory) corresponding to the storage unit 33, a register, and other input / output ports, and is a radar. Controls the entire device 1.

When the CPU of the microcomputer reads and executes the program stored in the ROM, it functions as the transmission / reception control unit 31 and the signal processing unit 32. The transmission / reception control unit 31 and the signal processing unit 32 can all be configured by hardware such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

The transmission / reception control unit 31 controls the transmission unit 10 including the signal generation unit 11 and the reception unit 20. The signal processing unit 32 periodically executes a series of signal processing. Next, each component of the signal processing unit 32 will be described.

The detection unit 32a includes a frequency analysis unit 321a, a peak extraction unit 322a, and an observation value generation unit 323a, and corresponds to the target based on the frequency-modulated transmission wave and the reflected wave of the transmission wave by the target. Detect observed values.

The frequency analysis unit 321a performs a fast Fourier transform (FFT) process (hereinafter, referred to as “FFT process”) on the beat signal input from each A / D transform unit 23, and peaks the result. Output to the extraction unit 322a. The result of such FFT processing is the frequency spectrum of the beat signal, and is the power value (signal level) for each frequency of the beat signal (for each frequency bin set at a frequency interval according to the frequency resolution).

The peak extraction unit 322a extracts the peak frequency that becomes the peak in the result of the FFT processing by the frequency analysis unit 321a, and outputs the extraction result to the observation value generation unit 323a. The peak extraction unit 322a extracts peak frequencies for each of the "UP section" and "DN section" of the beat signal described later.

The observation value generation unit 323a executes an angle estimation process for calculating the arrival angle of the reflected wave corresponding to each of the peak frequencies extracted by the peak extraction unit 322a and the power value thereof. Since the arrival angle is the angle at which the target is estimated to exist at the time of executing the angle estimation process, it may be described as "estimated angle" below.

Further, the observation value generation unit 323a executes a pairing process for determining the correct combination of the peak frequencies of the “UP section” and the “DN section” based on the calculation result of the calculated estimated angle and the power value.

Further, the observation value generation unit 323a calculates the distance and the relative speed of each target with respect to the vehicle MC from the determined combination result. Further, the observation value generation unit 323a outputs the calculated estimated angle, distance, and relative velocity of each target to the filter processing unit 32b as observation values for the latest period (latest scan).

In order to make the explanation easy to understand, the flow of processing from the pre-stage processing of the signal processing unit 32 to this point in the signal processing unit 32 is shown in FIGS. 3 to 4C. FIG. 3 is an explanatory diagram of processing from the pre-stage processing of the signal processing unit 32 to the peak extraction processing in the signal processing unit 32.

Further, FIG. 4A is a processing explanatory view of the angle estimation process. Further, FIGS. 4B and 4C are processing explanatory views (No. 1) and (No. 2) of the pairing process. Note that FIG. 3 is divided into three regions by two thick downward white arrows. In the following, each of these areas will be referred to as an upper row, a middle row, and a lower row in this order.

As shown in the upper part of FIG. 3, the transmission signal fs (t) is transmitted from the transmission antenna 13 as a transmission wave, then reflected by the target and arrives as a reflected wave, and the reception signal fr (t) is transmitted by the reception antenna 21. ) Is received.

At this time, as shown in the upper part of FIG. 3, the received signal fr (t) is delayed by a time difference T with respect to the transmitted signal fs (t) according to the distance between the vehicle MC and the target. Due to this time difference T and the Doppler effect based on the relative speed of the vehicle MC and the target, the beat signal has a frequency up in the "UP section" where the frequency rises and a frequency fdn in the "DN section" where the frequency falls. Obtained as a repeating signal (see middle section of FIG. 3).

The lower part of FIG. 3 schematically shows the result of FFT processing of the beat signal in the frequency analysis unit 321a for each of the "UP section" side and the "DN section" side.

As shown in the lower part of FIG. 3, after the FFT processing, waveforms in the respective frequency regions on the “UP section” side and the “DN section” side are obtained. The peak extraction unit 322a extracts the peak frequency that becomes the peak in such a waveform.

For example, in the case of the example shown in the lower part of FIG. 3, the peak extraction threshold value is used, and on the “UP section” side, peaks Pu1 to Pu3 are determined as peaks, and peak frequencies fu1 to fu3 are extracted, respectively.

Further, on the "DN section" side, peaks Pd1 to Pd3 are determined as peaks, respectively, and peak frequencies fd1 to fd3 are extracted, respectively, based on the peak extraction threshold value.

Here, the frequency components of each peak frequency extracted by the peak extraction unit 322a may be a mixture of reflected waves from a plurality of targets. Therefore, the observation value generation unit 323a performs an angle estimation process for calculating the direction of each peak frequency, and analyzes the existence of the corresponding target for each peak frequency.

The directional calculation in the observation value generation unit 323a can be performed by using a known arrival direction estimation method such as ESPRIT (Estimation of Signal Parameters via Rotational Invariance Techniques).

FIG. 4A schematically shows the orientation calculation result of the observed value generation unit 323a. The observation value generation unit 323a calculates the estimated angle of each target (each reflection point) corresponding to each of the peaks Pu1 to Pu3 from the peaks Pu1 to Pu3 of the directional calculation result. Further, the magnitude of each peak Pu1 to Pu3 is the power value. As shown in FIG. 4B, the observation value generation unit 323a performs such an angle estimation process on each of the “UP section” side and the “DN section” side.

Then, the observed value generation unit 323a performs a pairing process in which the peaks having similar estimated angles and power values are combined in the directional calculation result. Further, from the combination result, the observation value generation unit 323a calculates the distance and the relative velocity of each target (each reflection point) corresponding to the combination of each peak.

The distance can be calculated based on the relationship of "distance ∝ (fup + fdn)". The relative velocity can be calculated based on the relationship of "velocity ∝ (fup-fdn)". As a result, as shown in FIG. 4C, a pairing processing result showing the instantaneous values of the estimated angle, distance, and relative velocity of each reflection point RP with respect to the vehicle MC is obtained.

Returning to FIG. 2, the filter processing unit 32b will be described. As shown in FIG. 2, the filter processing unit 32b includes a prediction unit 321b, an allocation unit 322b, a weighting unit 323b, a resampling unit 324b, and a data generation unit 325b.

The filter processing unit 32b generates target data corresponding to the observed value by applying a particle filter that associates a predetermined number of particle data Pb with the observed value detected by the detection unit 32a.

The prediction unit 321b performs prediction processing for predicting the current particle data Pb from the previous particle data Pa in the particle filter. Specifically, the prediction unit 321b is based on _{the distribution state X t-1} of the time t-1 of the previous cycle, where the latest period is the time t and the distribution state X _{t of each particle data at the time t.} N pieces of particle data are arranged (sampled) based on the probability density function.

At this time, the prediction unit 321b, an error predicts the distribution X _t of this particle data so as to minimize on the basis of the output of the Kalman filter KF. The detailed processing contents of the prediction unit 321b will be described later in FIGS. 5 to 8.

The allocation unit 322b performs a process of allocating the observed value in the latest cycle to the current particle data Pb, which is the prediction result of the prediction unit 321b. Specifically, the allocation unit 322b predicts the current target data from the previous target data, and allocates the observed values existing within the predetermined allocation range based on the prediction result.

If there is an observed value that does not exist within the allocated range of any target data, the allocation unit 322b treats the observed value as a candidate for a new target. In such a case, the particle data in the initial state is assigned to the observed value corresponding to the new target. The initial state means, for example, a state in which a plurality of particle data are arranged in a predetermined positional relationship with respect to a new target.

The weighting unit 323b weights each of the current particle data Pb based on the assigned current observation value. Specifically, the weighting unit 323b increases the weight of the particles of the current particle data Pb that are close to the observed value, and decreases the weight of the particles that are far from the observed value. The terms "near" and "far" here mean that the Mahalanobis distance is "close" and "far". The Mahalanobis distance is a distance expressed using a plurality of parameters of the target data, and is a multidimensional distance considering the correlation between the previous target and the current target.

The resampling unit 324b rearranges (resampling) the particle data based on the weight of each of the particle data Pb this time. Specifically, the resampling unit 324b moves the particle data having a small weight closer to the observed value.

The data generation unit 325b recalculates the probability density function based on the current particle data Pb rearranged by the resampling unit 324b, and generates target data based on the center of gravity of the recalculated probability density function. The data generation unit 325b generated the target data based on the center of gravity of the probability density function, but for example, the target data may be generated based on the average of the probability density function.

Next, the specific processing contents of the prediction unit 321b will be described with reference to FIGS. 5 to 8. 5 and 8 are diagrams showing the processing contents of the prediction unit 321b. As shown in FIG. 5, the prediction unit 321b applies the Kalman filter KF to one representative particle data PA corresponding to the previous particle data Pa1 to Pa8.

Specifically, the prediction unit 321b generates representative particle data PA corresponding to the average of the previous particle data Pa1 to Pa8. For example, the prediction unit 321b generates the representative particle data PA by calculating the average value of the previous particle data Pa1 to Pa8 in the state space.

Then, the prediction unit 321b inputs the representative particle data PA to the Kalman filter KF. That is, the prediction unit 321b does not input all the previous particle data Pa1 to Pa8 to the Kalman filter KF, but inputs only the representative particle data PA to the Kalman filter KF. As a result, in the Kalman filter KF, the optimum Kalman gain corresponding to the representative particle data PA may be calculated, so that the processing load can be reduced.

Further, the prediction unit 321b generates the representative particle data PA corresponding to the average of the previous particle data Pa1 to Pa8, so that the output result of the Kalman filter KF, that is, the prediction accuracy of the current particle data Pb is lowered. Can be prevented.

The prediction unit 321b is not limited to generating a simple average value as the representative particle data PA, and calculates the weighted average in which predetermined weights are given to each of the previous particle data Pa1 to Pa8 to calculate the representative particle data. PA may be generated.

Alternatively, when the distribution state of the previous particle data Pa1 to Pa8 is in a relatively narrow range, the prediction unit 321b selects any one of the previous particle data Pa1 to Pa8 as the representative particle data PA. You may.

Further, in FIG. 5, one representative particle data PA is generated, but the number is not limited to one, and the number may be smaller than that of the previous particle data Pa1 to Pa8. In such a case, the previous particle data Pa1 to Pa8 may be divided into groups for each predetermined number, and representative particle data PA may be generated from each group. That is, by using a plurality of representative particle data PAs, it is possible to suppress a decrease in prediction accuracy in the particle data Pb this time.

Next, the processing of the Kalman filter KF will be specifically described with reference to FIGS. 6 and 7. As shown in FIG. 6, the representative particle data PA corresponding to the previous particle data Pa, the current observation value, and the previous observation value are input to the Kalman filter KF. Further, the Kalman filter KF includes a conversion unit KF1, a calculation unit KF2, and an update unit KF3, and predicts the current representative particle data PAa (see FIG. 8) from the previous representative particle data PA.

The conversion unit KF1 performs unscented conversion (hereinafter referred to as U conversion) the sigma point based on the previous representative particle data PA. The sigma point is a sample point that increases or decreases according to the non-linear motion model of the target (quick acceleration / deceleration / rapid turning motion model), and is a point set around the average value of the observed value or the particle data. Specifically, the sigma point is a point provided along the ellipse of the covariance matrix calculated from the average value. The calculation unit KF2 calculates the average value of the sigma points converted by the conversion unit KF1 and the covariance matrix, and calculates the error of the representative particle data PAa of this time with respect to the prediction process based on the covariance matrix. The update unit KF3 calculates the optimum Kalman gain based on the error calculated by the calculation unit KF2, and updates the previous representative particle data PA. Specifically, the update unit KF3 calculates the weighting value that minimizes the error calculated by the calculation unit KF2 as the optimum Kalman gain, and applies the optimum Kalman gain to represent the current representative particle data PA from the previous representative particle data PA. Generate particle data. Here, the processing contents of each part of the Kalman filter KF will be described with reference to FIG. 7.

FIG. 7 is a diagram showing the processing contents of the Kalman filter KF. The upper part of FIG. 7 shows a case where the current observed value OV1, the previous observed value OV2, and the representative particle data PA are plotted in a predetermined state space. First, the conversion unit KF1 calculates a point PO indicating the average value of the previously observed value OV2 and the representative particle data PA.

Subsequently, the conversion unit KF1 calculates a point A indicating the average value of the calculated point PO and the observed value OV1 this time. Subsequently, the conversion unit KF1 calculates the covariance matrix C based on the point A. Specifically, the conversion unit KF1 calculates the degree of variance indicating how far the point PO and the observed value OV1 this time are from the point A as the covariance matrix C. Then, the conversion unit KF1 generates a plurality of sigma points sg based on the covariance matrix C. Specifically, the sigma point sg is a point located around an ellipse corresponding to the covariance matrix C.

Subsequently, as shown in the middle part of FIG. 7, the conversion unit KF1 generates the UT sigma point usg by unscented conversion (U conversion) of the generated sigma point sg. Specifically, the conversion unit KF1 assumes that the sigma point sg is the state data of the previous period, and causes the sigma point sg to time-shift to the state data of the latest period by a nonlinear function, thereby UT. Generate sigma point usg.

As the nonlinear function, for example, a function corresponding to a motion model for predicting the current particle data Pb from the previous particle data Pa is used. That is, it can be said that the UT sigma point usg is a point in which the sigma point sg is moved according to the motion model.

Subsequently, as shown in the lower part of FIG. 7, the calculation unit KF2 calculates a predetermined weighted average value (weighted average) for each UT sigma point usg converted by the conversion unit KF1 to generate a point uA. ..

Subsequently, the calculation unit KF2 calculates the covariance matrix uC based on the point uA indicating the calculated average value. The error when the covariance matrix uC is converted from the sigma point sg to the UT sigma point usg is shown. In other words, the calculation unit KF2 calculates the covariance matrix uC as an error of the representative particle data PAa (see FIG. 8) this time. That is, such an error means the current representative particle data PAa in which the previous representative particle data PA is transitioned via the Kalman filter KF and the current representative particle in which the previous representative particle data PA is transitioned without passing through the Kalman filter KF. Refers to the difference between the true value and the data PAa.

Subsequently, the update unit KF3 calculates the optimum Kalman gain that minimizes the error of the representative particle data PAa (see FIG. 8) this time based on the covariance matrix uC. Subsequently, the update unit KF3 predicts and updates the current representative particle data PAa from the previous representative particle data PA based on the motion model and the optimum Kalman gain. As a result, the error of the representative particle data PAa in the latest period can be updated to the minimum.

That is, in the Kalman filter KF, the previous particle data Pa is treated as a sigma point sg which is a virtual point, and the error caused by the time transition of the sigma point sg by the uncented transformation is estimated by the covariance matrix uC. Then, the representative particle data PAa this time is updated so that the estimated error is minimized.

That is, the covariance matrix uC showing the error transitions the current representative particle data PAa in which the previous representative particle data PA is transitioned via the Kalman filter KF and the previous representative particle data PA without passing through the Kalman filter KF. It refers to the difference between the true value and the representative particle data PAa this time.

Furthermore, as described above, the previous observation value is input to the Kalman filter KF, in other words, the updated representative particle data PAa of this time also takes into account the previous observation value, so that the prediction accuracy should be improved. Can be done. Further, as the prediction accuracy is improved, the number of the particle data Pb1 to Pb8 this time can be reduced from the number of the previous particle data Pa1 to Pa8 in the processing after the weighting unit 323b in the subsequent stage, so that the radar device 1 Processing load can be reduced. Although the previous observation value is input to the Kalman filter KF, the present invention is not limited to this, and the target data which is the previous filter value may be input.

Next, a method of reflecting the output result of the Kalman filter KF on the particle data Pb this time will be described with reference to FIG. FIG. 8 shows the previous representative particle data PA before the application of the Kalman filter KF and the current representative particle data PAa after the application of the Kalman filter KF. When the Kalman filter KF is applied to all the previous particle data Pas without generating the representative particle data PAs, the reflection method shown in FIG. 8 is omitted.

Further, as shown in FIG. 8, the prediction unit 321b performs a process of predicting the current particle data Pb1 to Pb8 from the previous particle data Pa1 to Pa8 according to the motion model, separately from the process of the Kalman filter KF.

Then, the prediction unit 321b removes an error from each of the current particle data Pb1 to Pb8 based on the current representative particle data PAa updated by the update unit KF3 in the Kalman filter KF, so that the final current particle data Pb1 ~ Pb8 is predicted.

Specifically, as shown in FIG. 8, the prediction unit 321b generates representative particle data PB corresponding to the average of the current particle data Pb1 to Pb8 generated separately from the Kalman filter KF. In other words, the representative particle data PB this time is the predicted particle data that has not passed through the Kalman filter KF, and the representative particle data PAa this time is the predicted particle data that has passed through the Kalman filter KF. That is, the representative particle data PAa this time is the predicted particle data having the minimum error from the true value, and the representative particle data PB this time is the predicted particle data having a relatively large error from the true value.

Then, the prediction unit 321b compares the representative particle data PB with the representative particle data PAa. Specifically, the prediction unit 321b calculates the difference between the representative particle data PB and the representative particle data PAa. Such a difference becomes an error included in the representative particle data PB.

Then, the prediction unit 321b reflects the calculated error in all of the particle data Pb1 to Pb8 this time. That is, the prediction unit 321b assumes that all of the current particle data Pb1 to Pb8 contain the same error, and removes the error from each of the current particle data Pb1 to Pb8. In other words, the prediction unit 321b predicts the current particle data Pb1 to Pb8 in which the difference from the true value is minimized by reflecting the error in each of the current particle data Pb1 to Pb8.

In this way, by comparing the representative particle data PB and the representative particle data PAa and reflecting them in all the particle data Pb1 to Pb8, it is possible to improve the prediction accuracy while suppressing the processing load of the Kalman filter KF.

Next, the processing procedure of the processing executed by the radar device 1 according to the embodiment will be described with reference to FIGS. 9 to 10. FIG. 9 is a flowchart showing a processing procedure of processing executed by the radar device 1 according to the embodiment.

As shown in FIG. 9, first, the detection unit 32a detects the observed value corresponding to the target based on the frequency-modulated transmitted wave and the reflected wave of the transmitted wave by the target (step S101).

Subsequently, the prediction unit 321b of the filter processing unit 32b performs particle value calculation which is particle data in the particle filter (step S102). Specifically, the prediction unit 321b acquires the particle data Pa used last time (the previous cycle).

Subsequently, the prediction unit 321b performs processing by the Kalman filter KF in the prediction processing from the previous particle data Pa to the current particle data Pb in the particle filter (step S103).

Subsequently, the prediction unit 321b predicts the current particle data Pb based on the current representative particle data PAa updated by the Kalman filter KF (step S104). Subsequently, the allocation unit 322b allocates the observed value of this time to the particle data Pb of this time (step S105).

Subsequently, the weighting unit 323b weights each of the current particle data Pb based on the current observed value (step S106). Subsequently, the resampling unit 324b resamples the particle data Pb this time based on the weighting by the weighting unit 323b (step S107).

Subsequently, the data generation unit 325b updates the probability density function of the resampled particle data Pb this time, generates target data based on the probability density function (step S108), and ends the process.

Next, the processing procedure of the processing executed by the Kalman filter KF according to the embodiment will be described with reference to FIG. FIG. 10 is a flowchart showing a processing procedure of processing executed by the Kalman filter KF according to the embodiment. Further, FIG. 10 shows a specific processing procedure of the process of “processing by Kalman filter” in step S103 of FIG. 9 and the process of “predicting the particle data of this time” in step S104.

As shown in FIG. 10, the conversion unit KF1 first acquires the previous state data (step S201). Specifically, the conversion unit KF1 acquires the previous particle data Pa and the previous observed value. At this time, the observed value of this time is also acquired.

Subsequently, the conversion unit KF1 generates a sigma point sg by performing a sigma point calculation based on the previous particle data Pa, the previous observed value, and the current observed value (step S202).

Subsequently, the conversion unit KF1 generates a UT sigma point usg by performing an unscented conversion (U conversion) (step S203). Subsequently, the conversion unit KF1 calculates the average value uA of the UT sigma points usg and the covariance matrix uC (step S204).

Subsequently, the update unit KF3 updates the state data including the current particle data Pb based on the covariance matrix uC (step S205), and ends the process. That is, steps S201 to S204 correspond to step S103, and step S205 corresponds to step S104.

As described above, the radar device 1 according to the embodiment includes a detection unit 32a and a filter processing unit 32b. The detection unit 32a detects the observed value corresponding to the target based on the frequency-modulated transmitted wave and the reflected wave of the transmitted wave by the target. The filter processing unit 32b generates target data corresponding to the observed value by applying a particle filter that associates a predetermined number of particle data Pb with the observed value detected by the detection unit 32a. Further, the filter processing unit 32b includes a prediction unit 321b that applies the Kalman filter KF to predict the current particle data Pb in the prediction processing from the previous particle data Pa to the current particle data Pb in the particle filter. Further, the prediction unit 321b applies the Kalman filter KF to the number of representative particle data PAs corresponding to the previous particle data Pa and smaller than the previous particle data Pa. Thereby, the prediction accuracy of the particle data Pb in the particle filter can be improved.

In the above-described embodiment, the radar device 1 is provided in the vehicle MC, but of course, it may be provided in a moving body other than the vehicle, for example, a ship or an aircraft.

Further, in each of the above-described embodiments, ESPRIT is mentioned as an example of the arrival direction estimation method used by the radar device 1, but the present invention is not limited to this. For example, DBF (Digital Beam Forming), PRISM (Propagator method based on an Improved Spatial-smoothing Matrix), MUSIC (Multiple Signal Classification), or the like may be used.

Further effects and variations can be easily derived by those skilled in the art. For this reason, the broader aspects of the invention are not limited to the particular details and representative embodiments expressed and described as described above. Therefore, various modifications can be made without departing from the spirit or scope of the general concept of the invention as defined by the appended claims and their equivalents.

1 Radar device 32 Signal processing unit 32a Detection unit 32b Filter processing unit 321a Frequency analysis unit 322a Peak extraction unit 323a Observation value generation unit 321b Prediction unit 322b Allocation unit 323b Weighting unit 324b Resampling unit 325b Data generation unit 33 Storage unit 33a History data Pa, Pb particle data PA, PB representative particle data KF Kalman filter KF1 conversion unit KF2 calculation unit KF3 update unit

Claims (7)

A detector that detects the observed value corresponding to the target based on the frequency-modulated transmitted wave and the reflected wave of the transmitted wave by the target.
A filter processing unit that generates target data corresponding to the observed value by applying a particle filter that associates a predetermined number of particle data with the observed value detected by the detection unit is provided.
The filter processing unit
In the prediction processing from the previous particle data to the particle data this time in the particle filter, a prediction unit for predicting the particle data this time by applying a Kalman filter is provided.
The prediction unit
A radar device that corresponds to the previous particle data and applies the Kalman filter to a smaller number of representative particle data than the previous particle data. The prediction unit
The radar device according to claim 1, wherein the Kalman filter is applied to one representative particle data corresponding to the previous particle data. The prediction unit
The radar device according to claim 2, wherein the one representative particle data corresponds to the average of the previous particle data. The prediction unit
The radar device according to any one of claims 1 to 3, wherein the Kalman filter is applied to the representative particle data obtained by further adding the previous observation value. The prediction unit
A conversion unit that uncented-converts sigma points based on the representative particle data, and
A calculation unit that calculates the average value and covariance matrix of the sigma points converted by the conversion unit, and calculates the error of the representative particle data this time with respect to the prediction process based on the covariance matrix.
The radar device according to any one of claims 1 to 4, wherein the Kalman filter having the update unit for updating the representative particle data based on the error calculated by the calculation unit is used. The prediction unit
The radar device according to claim 5, wherein the current particle data is predicted by removing the error from each of the current particle data based on the representative particle data updated by the update unit. .. A detection step of detecting an observed value corresponding to the target based on a frequency-modulated transmitted wave and a reflected wave of the transmitted wave by the target, and a detection step.
A filtering step of generating target data corresponding to the observed value by applying a particle filter for associating a predetermined number of particle data with the observed value detected by the detection step is included.
The filtering step is
In the prediction process from the previous particle data to the particle data of the present time in the particle filter, a prediction step of applying the Kalman filter to predict the particle data of the present time is included.
The prediction process is
A target detection method, which corresponds to the previous particle data and applies the Kalman filter to a smaller number of representative particle data than the previous particle data.

Images