A radar dual-branch fusion tracking method based on evidence score

Abstract

A radar dual-branch fusion tracking method based on evidence scores belongs to the field of radar target detection and tracking. The method involves: obtaining a first target candidate output by a neural network detection branch and a second target candidate output by a traditional detection branch based on radar echo data; matching the first and second target candidates according to at least one of spatial positional relationship, Doppler information, and target size information to generate associated target candidates; calculating the admission evidence score and trajectory evidence score of the associated target candidates; based on the association results between the associated target candidates and existing trajectories, using the admission evidence score to determine whether to directly establish or delay orbit establishment, and using the trajectory evidence score to update the current frame evidence state of the transient trajectory or confirmed trajectory; when the transient trajectory meets the confirmation condition, converting the transient trajectory into a confirmed trajectory; and when the confirmed trajectory meets preset output conditions, outputting the target tracking result.

Description

Technical Field

[0001] This invention belongs to the field of radar target detection and tracking technology, and relates to a radar dual-branch fusion tracking method based on evidence scoring. Background Technology

[0002] Millimeter-wave radar is commonly used for target detection and tracking in scenarios such as vehicles, robots, and unmanned platforms. Radar echo data typically requires signal processing in the range, Doppler, and angle dimensions to generate data such as range-Doppler spectra, angle estimation results, point traces, or radar feature maps. Based on this data, different detection and tracking methods can be used to obtain target candidates, detection boxes, or trajectory results.

[0003] Existing radar target detection methods generally fall into two categories. One category is neural network-based detection methods, such as target detection networks based on range-Doppler maps, range-angle maps, or radar tensor inputs. These methods have strong target recognition capabilities, but their outputs mainly depend on network confidence and are prone to lacking physically interpretable constraints such as target size, Doppler consistency, and the number of points.

[0004] Another category is traditional radar detection methods, such as those based on CFAR, point clustering, Doppler filtering, and target size estimation. These methods can provide physical observation information such as the number of target points, target width, target height, Doppler broadening, and the proportion of the Doppler peak, but they are prone to problems such as fragmented detection, mistracking, and unstable trajectories in complex scenarios.

[0005] In practical radar tracking systems, simply merging the neural network detection results with the traditional detection results can easily lead to the following problems:

[0006] First, the target candidate outputs from the neural network detection branch and the traditional detection branch have different qualities. Simple splicing can lead to duplicate targets, false targets, or low-quality targets entering the trajectory manager.

[0007] Second, some targets may have been effectively measured in the current frame, but due to the requirements of the conventional confirmation rules, such as the number of consecutive hits, the number of points, and stability, the transient trajectory cannot be converted into a confirmed trajectory in a timely manner, resulting in missed detections.

[0008] Third, some trajectories may have been confirmed, but due to the standard display threshold requirements for the number of confirmation tests, display quality, or stable display conditions, the confirmed trajectories may not be included in the final output, resulting in missed detections.

[0009] Fourth, although some target candidates output by the traditional detection branch do not meet the conditions for direct establishment of new tracks, their evidence quality is close to the entry threshold. If they are discarded directly, it may result in missed track establishment; if they are established directly, it may result in incorrect track establishment.

[0010] Therefore, there is a need for a radar dual-branch fusion tracking method that can integrate neural network detection evidence, traditional detection physical evidence, Doppler evidence, temporal continuity evidence, and motion continuity evidence, and perform track establishment, trajectory confirmation, and output control based on evidence scores, in order to improve the reliability of trajectory confirmation and output control. Summary of the Invention

[0011] The purpose of this invention is to provide a radar dual-branch fusion tracking method based on evidence scores. This method matches a first target candidate output by a neural network detection branch with a second target candidate output by a traditional detection branch. It converts the target existence information provided by the neural network detection branch, the physical observation information provided by the traditional detection branch, and the target's temporal and motion continuity information into corresponding evidence items. Based on these evidence items, it calculates an admission evidence score and a trajectory evidence score. The admission evidence score is used for direct and delayed orbit establishment judgments, and the trajectory evidence score is used for trajectory evidence status updates, transient trajectory confirmation, and confirmed trajectory output control. This reduces missed detections caused by the failure to establish orbits for valid target candidates in a timely manner, the failure to confirm transient trajectories in a timely manner, and the failure to output confirmed trajectories in a timely manner. It also improves the recall rate and output reliability of the fusion tracking results while controlling false alarms.

[0012] The objective of this invention is achieved through the following technical solution.

[0013] This invention discloses a radar dual-branch fusion tracking method based on evidence scoring, comprising the following steps:

[0014] Step 1: Obtain radar echo data, and based on the radar echo data, obtain the first target candidate output by the neural network detection branch and the second target candidate output by the traditional detection branch.

[0015] Step 2: Match the first target candidate and the second target candidate according to the spatial position relationship and at least one physical observation information to determine the mutually matching candidate pairs, and construct associated target candidates based on the mutually matching candidate pairs. The physical observation information may include at least one of Doppler information and target size information.

[0016] Step 3: Calculate the admission evidence score and trajectory evidence score of the associated target candidate respectively.

[0017] Step 4: Based on the target position and Doppler observation of the associated target candidate, and based on at least one of the predicted position, predicted velocity and historical Doppler observation of the existing trajectory, the associated target candidate is associated with the existing trajectory, which includes transient trajectory and confirmed trajectory.

[0018] If the associated target candidate is successfully associated with the transient trajectory, the associated transient trajectory is updated using the associated target candidate, and the trajectory evidence status of the transient trajectory is updated according to the trajectory evidence score of the associated target candidate; wherein, the trajectory evidence status includes the current frame trajectory evidence score, the cumulative trajectory evidence score, the average trajectory evidence score, and the number of evidence hits, and the current frame trajectory evidence score is determined by the trajectory evidence score of the associated target candidate;

[0019] If the associated target candidate is successfully associated with the confirmed trajectory, the associated target candidate is used to update the successfully associated confirmed trajectory, and the trajectory evidence score of the associated target candidate is used as the current frame trajectory evidence score of the confirmed trajectory.

[0020] If the associated target candidate fails to be associated with an existing trajectory, and the admission evidence score of the associated target candidate is not lower than the threshold for directly creating a new trajectory, then a new transient trajectory is created based on the associated target candidate, and the trajectory evidence state of the transient trajectory is initialized according to the trajectory evidence score of the associated target candidate.

[0021] If the associated target candidate fails to associate with an existing trajectory, and the admission evidence score of the associated target candidate is lower than the direct new trajectory threshold but not lower than the candidate retention threshold, then the associated target candidate is added to the delayed trajectory establishment pool. Cross-frame association is performed on the target candidates in the delayed trajectory establishment pool. When the cumulative hit count and admission evidence score of the same delayed trajectory establishment object meet the release conditions, a transient trajectory is created based on the target candidate corresponding to that delayed trajectory establishment object, and the trajectory evidence state of the transient trajectory is initialized according to the trajectory evidence score of that target candidate.

[0022] If the associated target candidate fails to be successfully associated with an existing trajectory, and the admission evidence score of the associated target candidate is lower than the candidate retention threshold, then no new transient trajectory will be created based on the associated target candidate, and the associated target candidate will not be added to the delayed trajectory establishment pool.

[0023] Step 5: When the number of evidence hits for the same transient trajectory, the current frame trajectory evidence score of the transient trajectory, and the average trajectory evidence score of the transient trajectory reach the corresponding thresholds, the transient trajectory is converted into a confirmed trajectory.

[0024] Step 6: When the confirmed trajectory meets the preset output conditions in the current frame, output the target tracking result corresponding to the confirmed trajectory.

[0025] Furthermore, when matching the first target candidate and the second target candidate, the normalized spatial distance and the normalized Doppler difference are calculated, and the matching cost is obtained by weighting the normalized spatial distance and the normalized Doppler difference.

[0026] Calculate the first target candidate With the second target candidate Normalized spatial distance :

[0027]

[0028] in, This indicates the horizontal matching threshold. Indicates the vertical matching threshold. , Indicates the first target candidate The center coordinates, , Indicates the second target candidate The center coordinates.

[0029] Calculate the first target candidate With the second target candidate Normalized Doppler differences :

[0030]

[0031] in, Indicates the first target candidate Doppler observations, Indicates the second target candidate Doppler observations, This indicates the Doppler difference after considering the Doppler cycle numbering. This represents the Doppler matching threshold.

[0032] Matching cost between the first target candidate and the second target candidate Represented as:

[0033]

[0034] in, and , where are weighting coefficients, used to adjust the proportion of normalized spatial distance and normalized Doppler difference in the matching cost, respectively.

[0035] when Not greater than the preset normalized space threshold, and When the value is not greater than the preset normalized Doppler threshold, and A candidate pair is identified as a matchable pair. If multiple matchable pairs exist for the same candidate, the matching cost is selected. The smallest candidate pair is selected as the matching result.

[0036] For successfully matched candidate pairs, the neural network confidence, target location, and Doppler observations in the first target candidate are combined with the number of points, target size, Doppler broadening, Doppler peak ratio, and traditional detection quality in the second target candidate to construct associated target candidates.

[0037] Furthermore, the admission evidence score and the trajectory evidence score are obtained based on at least one of the following evidence items:

[0038] Neural network confidence evidence is used to characterize the degree of confidence of the detection branch of a neural network in the existence of a target.

[0039] Physical size evidence is used to characterize at least one of the following: the number of points output by a conventional detection branch, the target width, the target height, or the conventional detection quality.

[0040] Doppler consistency evidence is used to characterize at least one of the following: target Doppler broadening, proportion of the main Doppler peak, and absolute Doppler intensity.

[0041] Evidence of temporal continuity is used to characterize the continuity of target appearance across frames.

[0042] Evidence of motion continuity is used to characterize the dynamic scale of a target or the consistency of motion across frames.

[0043] Furthermore, the admission evidence score is:

[0044]

[0045] The baseline score for the trajectory evidence is:

[0046]

[0047] in, This represents the confidence level evidence of the neural network. This indicates evidence of physical dimensions. This indicates evidence of Doppler consistency. Evidence indicating temporal continuity. This indicates evidence of the continuity of motion.

[0048] The trajectory evidence score is determined as follows:

[0049]

[0050] in, Score the trajectory evidence. For trajectory evidence-based score, Score for evidence of admission.

[0051] Furthermore, the neural network detection branch is implemented by a pre-trained target detection network.

[0052] The training samples for the target detection network include range-Doppler maps corresponding to radar echo data, as well as corresponding target annotation information.

[0053] The target annotation information includes at least one of the following: target center position, target distance, target azimuth angle, target presence annotation, or Doppler annotation information.

[0054] The neural network detection branch is used to input the range-Doppler image corresponding to the radar echo data of the current frame into the target detection network and output the first target candidate.

[0055] Furthermore, the traditional detection branches include spectral peak detection, angle estimation, spot generation, and target candidate aggregation.

[0056] Among them, spectral peak detection is used to obtain the range-Doppler detection unit based on range-Doppler spectral data.

[0057] Angle estimation is used to generate an azimuth response based on the multi-channel range-Doppler data, MIMO Doppler index, and preset angle calibration matrix corresponding to the range-Doppler detection unit, and to determine the target azimuth angle based on the peak value of the azimuth response.

[0058] The spot generation is used to generate radar spots based on range, Doppler, and azimuth.

[0059] Target candidate aggregation is used to generate the second target candidate based on the radar spot.

[0060] Furthermore, the first target candidate includes the target center position, target distance, target azimuth angle, Doppler observation value, and neural network confidence.

[0061] The second target candidate includes the target center location, target width, target height, number of points, Doppler observation value, Doppler broadening, proportion of Doppler main peak, and traditional detection quality.

[0062] Furthermore, the processing method for the delayed track establishment pool described in step 4 is as follows:

[0063] When the associated target candidate fails to be associated with an existing trajectory, and its admission evidence score is lower than the direct new trajectory threshold but not lower than the candidate retention threshold, the associated target candidate is added to the delayed trajectory establishment pool.

[0064] In subsequent frames, target candidates in the delayed orbit establishment pool are associated across frames based on region identifier, lateral distance, longitudinal distance, and Doppler difference;

[0065] When the cumulative hit count and admission evidence score of the same delayed track establishment object in the delayed track establishment pool meet the release conditions, a new transient trajectory is created based on the target candidate corresponding to the delayed track establishment object, and the trajectory evidence state of the transient trajectory is initialized according to the trajectory evidence score of the target candidate.

[0066] When a delayed track establishment object in the delayed track establishment pool exceeds a preset lifespan or no longer meets the cross-frame association conditions, it is deleted.

[0067] Furthermore, the preset output conditions described in step 6 include:

[0068] The confirmed trajectory is supported by both neural network detection branches and traditional detection branches in the current frame.

[0069] The number of confirmation checks for the confirmed trajectory is not less than the preset minimum number of confirmation checks.

[0070] The number of points corresponding to the target candidate associated with the current frame of the confirmed trajectory is not less than the preset output point threshold.

[0071] The trajectory evidence score of the current frame of the confirmed trajectory is not less than the preset display evidence threshold.

[0072] Beneficial effects:

[0073] 1. The present invention discloses a radar dual-branch fusion tracking method based on evidence scores. By matching the first target candidate output by the neural network detection branch and the second target candidate output by the traditional detection branch, and comprehensively utilizing the target existence information provided by the neural network detection branch, the physical observation information provided by the traditional detection branch, and the target's temporal continuity and motion continuity information to calculate the evidence score, the target judgment basis can be expanded from a single branch result to a dual-branch multi-source evidence joint judgment, thereby improving the interpretability and usability of associated target candidates.

[0074] 2. This invention uses admission evidence scores to determine whether target candidates that have not been successfully associated with existing trajectories should be directly created as transient trajectories, enter the delayed trajectories pool, or not retained. Therefore, it can avoid the direct trajectories of low-quality target candidates and reduce the risk of missed trajectories caused by the direct discarding of associated target candidates whose evidence is close to the admission threshold.

[0075] 3. This invention retains target candidates with admission evidence scores lower than the direct new track creation threshold but not lower than the candidate retention threshold by delaying track creation pool, and determines whether to release the new transient trajectory based on cross-frame association, cumulative hit count and admission evidence score. Therefore, it can reduce the risk of missed or incorrect track creation caused by one-time admission judgment.

[0076] 4. This invention uses trajectory evidence scores to update the trajectory evidence status of transient trajectories, and determines whether to convert transient trajectories into confirmed trajectories based on the number of evidence hits, the trajectory evidence score of the current frame, and the average trajectory evidence score. Therefore, it can enable timely confirmation of transient trajectories with sufficient evidence and reduce missed detections caused by the failure of valid target candidates to be converted into confirmed trajectories in a timely manner.

[0077] 5. This invention uses trajectory evidence scores to confirm the output control of the trajectory. When there is sufficient evidence in the current frame of the confirmed trajectory, it is allowed to enter the final output. Therefore, it can reduce the missed detection caused by the confirmed trajectory not entering the final output in time and improve the output reliability of the target tracking results.

[0078] 6. Under the same data segment and evaluation criteria, compared with the original fusion scheme that does not enable evidence score confirmation and output control, this invention can reduce the number of missed detections and improve recall and F1 score while keeping the number of false positives unchanged. Here, F1 represents the harmonic mean of precision and recall. Attached Figure Description

[0079] Figure 1 A flowchart illustrating the overall process of the radar dual-branch fusion tracking method based on evidence scores provided in this embodiment of the invention;

[0080] Figure 2 A schematic diagram of the dual-branch candidate matching, evidence score calculation, and trajectory processing flow provided in an embodiment of the present invention;

[0081] Figure 3 This is a schematic diagram illustrating the operational results of radar point cloud, detected target, and tracking trajectory provided in an embodiment of the present invention. Detailed Implementation

[0082] To better illustrate the purpose, technical solution, and beneficial effects of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0083] like Figure 1 As shown, this embodiment discloses a radar dual-branch fusion tracking method based on evidence scores, including radar echo data acquisition, neural network detection branch processing, traditional detection branch processing, dual-branch candidate matching, associated target candidate generation, evidence score calculation, association of target candidates with existing trajectories, orbit establishment judgment for unassociated target candidates, transient trajectory confirmation, and confirmed trajectory output. For target candidates that have not been successfully associated with existing trajectories and whose admission evidence score is lower than the direct orbit establishment threshold but not lower than the candidate retention threshold, delayed orbit establishment processing can be further performed. The specific process can be combined with... Figure 2 And the following explanation.

[0084] like Figure 1As shown in the figure, the radar dual-branch fusion tracking method based on evidence scores disclosed in this embodiment is specifically implemented as follows:

[0085] Acquire radar echo data. The radar echo data can be raw radar ADC data or radar data cubes. Preprocess the radar echo data, including at least range-dimensional Fast Fourier Transform and Doppler-dimensional Fast Fourier Transform, to obtain range-Doppler data. The range-Doppler data can include at least one of the following: range-Doppler complex spectrum, range-Doppler amplitude spectrum, range-Doppler power spectrum, or a range-Doppler plot obtained by normalizing these spectra.

[0086] The neural network detection branch is used to output the first target candidate based on the range-Doppler data, while the traditional detection branch is used to perform spectral peak detection, angle estimation, spot generation, and target candidate aggregation based on the range-Doppler spectral data, and output the second target candidate.

[0087] The neural network detection branch can be implemented by a pre-trained target detection network. Training samples include the range-Doppler image corresponding to the radar echo data and the corresponding target annotation information. The target annotation information may include at least one of the following: target center position, target range, target azimuth, target presence annotation, or Doppler annotation information. After training, the range-Doppler image corresponding to the current frame's radar echo data is input into the target detection network, and a first target candidate is output. The first target candidate may include the target center position, target range, target azimuth, Doppler observation value, and neural network confidence score.

[0088] In one implementation, the i-th first target candidate output by the neural network detection branch is represented as:

[0089]

[0090] in, , Indicates the center coordinates of the first target candidate. Indicates the distance to the target. Indicates the target azimuth. Represents Doppler observations, This represents the confidence level of the neural network.

[0091] Traditional detection branches may include spectral peak detection, angle estimation, spot generation, and target candidate aggregation. Spectral peak detection is used to obtain range-Doppler detection units based on range-Doppler spectral data; angle estimation is used to generate an azimuth response based on the multi-channel range-Doppler data corresponding to the range-Doppler detection units, the MIMO Doppler index, and a preset angle calibration matrix, and to determine the target azimuth angle based on the peak value of the azimuth response; spot generation is used to generate radar spots based on range, Doppler, and azimuth angle; target candidate aggregation is used to generate second target candidates based on the radar spots. The second target candidate may include the target center position, target width, target height, number of spots, Doppler observation value, Doppler broadening, Doppler main peak ratio, and traditional detection quality.

[0092] In one implementation, the j-th second target candidate output by the conventional detection branch is represented as:

[0093]

[0094] in, , Indicates the center coordinates of the second target candidate. , Indicates the target width and target height. This indicates the number of points corresponding to the second target candidate. Represents Doppler observations, This indicates Doppler broadening. Indicates the percentage of the Doppler peak. This indicates the quality of traditional testing.

[0095] In one implementation, It can be obtained statistically from the Doppler observations of the corresponding points of the second target candidate, such as taking the mean, median or mode of the Doppler observations of the points.

[0096] The value can be determined based on the range or variance of the Doppler distribution of the corresponding points of the second target candidate, for example, by taking the difference between the largest and smallest Doppler units:

[0097]

[0098] The ratio of the number of points corresponding to the main Doppler element to the total number of candidate points for the second target can be determined:

[0099]

[0100] in, This indicates the number of points corresponding to the main Doppler element. A preset constant is used to prevent the denominator from being zero.

[0101] It can be calculated based on at least one of the following: number of points, dynamic point ratio, Doppler main peak ratio, or local signal-to-noise ratio. For example:

[0102]

[0103] in, This represents the result of point normalization. Indicates the proportion of dynamic points or the quality of motion observation. Indicates the percentage of the Doppler peak. , , , where is the weighting coefficient, and clamp indicates that the calculation results will be limited to a preset range.

[0104] In one implementation, Based on the number of points corresponding to the second target candidate Compared with the preset point reference value The calculation yielded:

[0105]

[0106] This can be calculated based on the ratio of the number of dynamic points in the second target candidate to the total number of points:

[0107]

[0108] in, This indicates the number of points in the second target candidate whose absolute Doppler value is greater than the preset dynamic threshold. A preset constant is used to prevent the denominator from being zero.

[0109] This can be calculated based on the ratio of the number of points corresponding to the main Doppler elements in the second target candidate to the total number of points:

[0110]

[0111] in, Can be with Using the same calculation results, it is also possible to calculate separately according to the preset normalization rules used in traditional quality testing calculations.

[0112] like Figure 2 As shown, after obtaining the first and second target candidates, they are converted into a unified candidate format for matching. The unified candidate format includes at least the candidate source identifier, the horizontal coordinates of the target center, the vertical coordinates of the target center, the Doppler observations, and the candidate confidence information.

[0113] For the first target candidate, the candidate source identifier is recorded as the neural network detection branch, and the candidate confidence information includes the neural network confidence score. For the second target candidate, the candidate source identifier is recorded as the traditional detection branch, and the candidate confidence information includes the traditional detection quality, and further records the number of points, target width, target height, Doppler broadening, and the proportion of the Doppler main peak.

[0114] Under a unified candidate format, the first target candidate and the second target candidate are matched based on their spatial location relationship and at least one physical observation information to determine mutually matching candidate pairs. The physical observation information may include at least one of Doppler information and target size information.

[0115] For successfully matched candidate pairs, the neural network confidence, target location, and Doppler observations in the first target candidate are combined with the number of points, target size, Doppler broadening, Doppler peak ratio, and traditional detection quality in the second target candidate to construct associated target candidates.

[0116] In one implementation, a first target candidate is calculated. With the second target candidate Normalized spatial distance:

[0117]

[0118] in, This indicates the horizontal matching threshold. This indicates the vertical matching threshold.

[0119] Calculate the first target candidate With the second target candidate Normalized Doppler differences:

[0120]

[0121] in, This indicates the Doppler difference after considering the Doppler cycle numbering. This represents the Doppler matching threshold.

[0122] In one implementation, the matching cost between the first target candidate and the second target candidate is expressed as:

[0123]

[0124] in, and These are the weighting coefficients.

[0125] when Not greater than the preset normalized space threshold, and When the value is not greater than the preset normalized Doppler threshold, and A candidate pair is identified as a matchable pair. If multiple matchable pairs exist for the same candidate, the matching cost is selected. The smallest candidate pair is selected as the matching result.

[0126] For successfully matched candidate pairs, the neural network confidence, target location, and Doppler observations from the first target candidate are combined with the number of points, target size, Doppler broadening, Doppler peak proportion, and traditional detection quality from the second target candidate to construct associated target candidates. Thus, the associated target candidates simultaneously contain target existence information provided by the neural network detection branch and physical observation information provided by the traditional detection branch.

[0127] For the aforementioned candidate targets, neural network confidence information, traditional detection physical observation information, Doppler observation information, and cross-frame continuity information are converted into multiple evidence items. Based on these multiple evidence items, an admission evidence score, a trajectory evidence basis score, and a trajectory evidence score are calculated. The multiple evidence items may include neural network confidence evidence, physical size evidence, Doppler consistency evidence, temporal continuity evidence, and motion continuity evidence. In one embodiment, the admission evidence score, trajectory evidence basis score, and trajectory evidence score are calculated as follows:

[0128]

[0129]

[0130]

[0131] in, To score the admission evidence, For trajectory evidence-based score, Score the trajectory evidence; This serves as confidence evidence for neural networks, characterizing the degree of confidence that a branch of the neural network has in the existence of a target. As physical size evidence, it is used to characterize the number of points output by the traditional detection branch, the target width, the target height, or the degree to which the traditional detection quality supports the existence of the target; This serves as evidence of Doppler consistency, used to characterize the degree to which the target's Doppler broadening, the proportion of the Doppler main peak, or the absolute intensity of the Doppler beam support the consistency of the target's motion. This serves as evidence of temporal continuity, used to characterize the continuity of target appearance across frames; It serves as evidence of motion continuity, used to characterize the dynamic scale of a target or the consistency of motion across frames.

[0132] Among them, the admission evidence score Used to determine whether candidate targets that have not been successfully associated with existing trajectories should be directly created as transient trajectories, added to the delayed trajectory establishment pool, or not retained; Trajectory Evidence Base Score Used to evaluate the degree to which the associated target candidates support the continuity of the trajectory from the perspectives of trajectory confirmation and trajectory output; trajectory evidence score. Used to update or initialize the trajectory evidence state, and for transient trajectory confirmation judgment and confirmed trajectory output judgment.

[0133] In one implementation, neural network confidence evidence The confidence level is determined by the output of the neural network detection branch. If a target candidate does not have support from a neural network detection branch, then... It can be set to 0.

[0134] Physical size evidence It can be calculated based on the number of points output by the traditional detection branch, the target width, the target height, or the traditional detection quality. For example:

[0135]

[0136]

[0137] in, The number of points corresponding to the target candidate. This is a reference value for point normalization. This is for traditional testing quality.

[0138] Doppler consistency evidence It can be calculated based on Doppler broadening, the proportion of the Doppler main peak, and the absolute intensity of the Doppler. For example:

[0139]

[0140]

[0141]

[0142] in, This indicates Doppler broadening. This represents the Doppler broadening normalized reference value. Indicates signed Doppler observations. This represents the normalized reference value for the absolute intensity of Doppler. This indicates the percentage of the Doppler peak.

[0143] Evidence of temporal continuity The target candidate can be determined based on the number of times it appears across frames, the cumulative number of hits in the delayed orbit establishment pool, or the number of consecutive observation frames.

[0144] Evidence of motion continuity It can be determined based on the target dynamic scale or cross-frame motion consistency. For example:

[0145]

[0146] in, This represents the ratio of the number of dynamic points among the target candidates to the total number of points.

[0147] It should be noted that the above weights are only one implementation method. In other implementation methods, the weights of each evidence item can be adjusted according to the radar type, detection network type, target category, application scenario, or calibration results.

[0148] Then, based on the target location and Doppler observations of the associated target candidate, and based on at least one of the predicted location, predicted velocity, and historical Doppler observations of the existing trajectory, the associated target candidate is associated with the existing trajectory. The existing trajectory includes transient trajectories and confirmed trajectories.

[0149] In one implementation, for the k-th associated target candidate, its target position is denoted as... Doppler observations For the l-th existing trajectory, predict the trajectory position in the current frame based on the state of the existing trajectory in the previous frame, and obtain the predicted position. And obtain the predicted velocity or historical Doppler observations of the existing trajectory. The prediction can employ at least one of a uniform velocity model, a uniform acceleration model, a Kalman filter model, or an α-β filter model. For example, in one implementation using a uniform velocity model, the predicted position can be expressed as:

[0150]

[0151]

[0152] in, , This indicates the target position of the existing trajectory in the previous frame. , This indicates the velocity component of the existing trajectory in the previous frame. This indicates the time interval between adjacent frames.

[0153] For Doppler observations, the most recently successfully associated Doppler observation, the moving average of historical Doppler observations, or the radial velocity obtained from the trajectory velocity projection can be used as the Doppler reference value for the existing trajectory. For example, the moving average of historical Doppler observations can be expressed as:

[0154]

[0155] in, This represents the Doppler observation value recorded when the l-th existing trajectory was successfully associated M times recently, where M is the preset historical window length.

[0156] Calculate the normalized positional difference between the associated target candidate and the existing trajectory:

[0157]

[0158] in, Indicates the horizontal threshold for trajectory association. This indicates the longitudinal threshold for trajectory association.

[0159] Calculate the Doppler difference between the candidate associated target and the existing trajectory:

[0160]

[0161] in, This indicates the Doppler difference after considering the Doppler cycle numbering. This indicates the trajectory-associated Doppler threshold.

[0162] In one implementation, the trajectory association cost between the associated target candidate and the existing trajectory is expressed as:

[0163]

[0164] in, and These are the weighting coefficients.

[0165] when Not greater than the preset normalized trajectory space threshold, and If the value is not greater than the preset normalized trajectory Doppler threshold, the k-th candidate target and the l-th existing trajectory are identified as associatable objects. If multiple associatable existing trajectories exist for the same candidate target, the trajectory association cost is selected. The existing trajectory with the lowest cost is selected as the association result; if the same existing trajectory corresponds to multiple candidate association targets, the trajectory association cost is selected. The smallest associated target candidate updates the existing trajectory.

[0166] If the associated target candidate is successfully associated with the transient trajectory, the associated transient trajectory is updated using the associated target candidate, and the trajectory evidence status of the transient trajectory is updated according to the trajectory evidence score of the associated target candidate; wherein, the trajectory evidence status includes the current frame trajectory evidence score, the cumulative trajectory evidence score, the average trajectory evidence score, and the number of evidence hits, and the current frame trajectory evidence score is determined by the trajectory evidence score of the associated target candidate;

[0167] In one implementation, when an associated target candidate is supported by both a neural network detection branch and a traditional detection branch, has target size information, has a number of points not less than a preset point threshold, and has a trajectory evidence score not less than a preset evidence hit threshold, it is recorded as an evidence hit, and the cumulative trajectory evidence score, the number of evidence hits, and the average trajectory evidence score of the transient trajectory are updated as follows:

[0168]

[0169]

[0170]

[0171] in, The cumulative trajectory evidence score for the same transient trajectory. The number of evidence hits for this transient trajectory. The average trajectory evidence score for this transient trajectory.

[0172] If the associated target candidate is successfully associated with the confirmed trajectory, the confirmed trajectory is updated using the associated target candidate, and the trajectory evidence score of the associated target candidate is used as the current frame trajectory evidence score of the confirmed trajectory. At the same time, at least one of the following is updated for the confirmed trajectory: the number of confirmed detections, the most recent matching frame number, or the number of unmatched frames.

[0173] For existing trajectories that have not been successfully associated with any associated target candidate in the current frame, their number of unmatched frames can be updated; when the number of unmatched frames of an existing trajectory exceeds a preset deletion threshold, or when the trajectory quality is lower than a preset deletion condition, the existing trajectory can be deleted.

[0174] If the associated target candidate is not successfully associated with an existing trajectory, and the admission evidence score of the associated target candidate is not lower than the threshold for directly creating a new trajectory, then a new transient trajectory is created based on the associated target candidate, and the trajectory evidence score of the associated target candidate is used as the current frame trajectory evidence score and the initial average trajectory evidence score of the transient trajectory. The number of evidence hits is initialized according to whether the associated target candidate meets the evidence hit conditions.

[0175] If a candidate target is not successfully associated with an existing trajectory, and the admission evidence score of the candidate target is lower than the direct new trajectory threshold but not lower than the candidate retention threshold, then the candidate target is added to the delayed trajectory establishment pool.

[0176] In one implementation, when a target candidate is added to the delayed trajectory establishment pool, a delayed trajectory establishment object is created for that target candidate. The delayed trajectory establishment object records at least the target candidate's region identifier, lateral coordinates, longitudinal coordinates, Doppler observations, admission evidence score, trajectory evidence score, cumulative admission evidence score, cumulative number of hits, most recent hit frame number, number of consecutive unmatched frames, and number of live frames.

[0177] In subsequent frames, for new candidate targets for association, the cross-frame association relationship between them and existing delayed track establishment objects in the delayed track establishment pool is calculated. If the region identifier of the candidate target for association in the current frame is the same as that of the delayed track establishment object, and the difference in their lateral distance, longitudinal distance, and Doppler difference is not greater than the corresponding threshold, then the candidate target for association is considered to have been successfully associated with the delayed track establishment object across frames.

[0178] In one embodiment, the lateral distance difference, longitudinal distance difference, and Doppler difference are respectively expressed as:

[0179]

[0180]

[0181]

[0182] in, , and These represent the horizontal coordinates, vertical coordinates, and Doppler observations of the target candidate in the current frame, respectively. , and These represent the horizontal coordinates, vertical coordinates, and Doppler observations of the delayed orbit establishment object, respectively.

[0183] When the following conditions are met:

[0184]

[0185]

[0186]

[0187] When it is determined that the current frame target candidate and the delayed track establishment object are successfully associated across frames, wherein, This indicates the threshold for the horizontal correlation of the delayed track-building pool. This indicates the threshold for the longitudinal correlation of the delayed track-building pool. This indicates the Doppler correlation threshold for delayed orbit establishment pooling.

[0188] After successful cross-frame association, update the horizontal coordinates, vertical coordinates, Doppler observations, admission evidence score, and trajectory evidence score of the delayed orbit establishment object, and update its cumulative admission evidence score, cumulative hit count, and most recent hit frame number. Simultaneously, set the number of consecutive unmatched frames to zero.

[0189]

[0190]

[0191]

[0192] in, The cumulative admission evidence score for the same delayed track establishment object. This represents the cumulative number of hits for the delayed orbit establishment target. The average admission evidence score for this delayed track establishment object.

[0193] If the cumulative hit count of the same delayed trajectory establishment object is not less than the preset release hit count, and at least one of the following conditions is met, the delayed trajectory establishment object is released to create a new transient trajectory: the admission evidence score of the target candidate successfully associated with the delayed trajectory establishment object across frames in the current frame is not less than the preset release evidence threshold; the average admission evidence score of the delayed trajectory establishment object is not less than the preset release evidence threshold. When releasing the delayed trajectory establishment object, a new transient trajectory is created based on the target candidate successfully associated with the delayed trajectory establishment object across frames in the current frame, and the trajectory evidence score of the target candidate is used as the current frame trajectory evidence score and the initial average trajectory evidence score of the transient trajectory. The evidence hit count is initialized according to whether the target candidate meets the evidence hit condition.

[0194] The number of live frames for a delayed track establishment object is updated with each frame. If a delayed track establishment object fails to be successfully associated with any target candidate across frames in the current frame, the number of consecutive unmatched frames for that delayed track establishment object is increased. If the number of live frames for a delayed track establishment object exceeds a preset live duration, or the number of consecutive unmatched frames exceeds a preset number of unmatched frames, the delayed track establishment object is removed from the delayed track establishment pool.

[0195] For a transient trajectory, when the number of evidence hits for the same transient trajectory is not less than a preset hit count threshold, the current frame trajectory evidence score of the transient trajectory is not less than the current frame confirmation threshold, and the average trajectory evidence score of the transient trajectory is not less than the average confirmation threshold, the transient trajectory is converted into a confirmed trajectory.

[0196] For a confirmed trajectory, when the confirmed trajectory meets preset output conditions in the current frame, the target tracking result corresponding to the confirmed trajectory is output. The preset output conditions may include: the confirmed trajectory is supported by both neural network detection branches and traditional detection branches in the current frame; the number of confirmed detections of the confirmed trajectory is not less than the preset minimum number of confirmed detections; the number of target candidates associated with the confirmed trajectory in the current frame is not less than the preset output point threshold; and the trajectory evidence score of the confirmed trajectory in the current frame is not less than the preset display evidence threshold.

[0197] In one specific embodiment, the preset point threshold, preset hit count threshold, and preset output point threshold can all be 2, and the current frame confirmation threshold, average confirmation threshold, and preset display evidence threshold can all be 0.70. The above values ​​are only used to illustrate one embodiment of the present invention and do not constitute a limitation on the scope of protection.

[0198] Figure 3 The results of the operation of an embodiment of the present invention on a segment of radar data are shown, including radar point cloud, detected target and tracking trajectory. Figure 3 This is only used to intuitively illustrate the output effect of the embodiments of the present invention, and does not constitute a limitation on specific scenarios, number of targets, coordinate range or radar model.

[0199] To verify the effectiveness of the embodiments of the present invention, different methods were compared under the same radar data segment, the same tag mapping relationship, and the same matching threshold. Evaluation metrics included TP, FP, FN, Precision, Recall, and F1. TP represents the number of predicted targets successfully matched with the tag target, FP represents the number of predicted targets that did not successfully match the tag target, FN represents the number of tag targets that were not matched by the predicted target, Precision represents precision, Recall represents recall, and F1 represents the harmonic mean of precision and recall.

[0200] In one experimental case, the target corresponding to the output confirmed trajectory was used as the predicted target, and it was matched with the label target according to the 4m matching threshold.

[0201] The experimental results are as follows:

[0202] The original fusion scheme without evidence score confirmation and output control: TP=189, FP=1, FN=30, Precision=0.9947, Recall=0.8630, F1=0.9242.

[0203] An embodiment of the present invention with enabled evidence score confirmation and output control: TP=197, FP=1, FN=22, Precision=0.9949, Recall=0.8995, F1=0.9448.

[0204] Simple candidate merge scheme: TP=191, FP=3, FN=28, Precision=0.9845, Recall=0.8721, F1=0.9249.

[0205] The above results demonstrate that, under the experimental data range and evaluation criteria, compared to the original fusion scheme without evidence score confirmation and output control, the embodiments of the present invention reduce the number of missed detections while maintaining the same number of false positives, and improve the recall rate and F1 score. These results indicate that trajectory establishment judgment, trajectory confirmation, and trajectory output control based on evidence scores can reduce missed detections caused by the failure to establish trajectories for valid target candidates in a timely manner, the failure to confirm transient trajectories in a timely manner, and the failure of confirmed trajectories to enter the final output in a timely manner.

[0206] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. All modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A radar dual-branch fusion tracking method based on evidence scoring, characterized in that, Includes the following steps: Step 1: Acquire radar echo data, and based on the radar echo data, obtain the first target candidate output by the neural network detection branch and the second target candidate output by the traditional detection branch, respectively; Step 2: Match the first target candidate and the second target candidate according to the spatial position relationship and at least one physical observation information to determine the mutually matching candidate pairs, and construct associated target candidates based on the mutually matching candidate pairs. The physical observation information may include at least one of Doppler information and target size information. Step 3: Calculate the admission evidence score and trajectory evidence score of the associated target candidate respectively; Step 4: Based on the target position and Doppler observation of the associated target candidate, and based on at least one of the predicted position, predicted velocity and historical Doppler observation of the existing trajectory, associate the associated target candidate with the existing trajectory, wherein the existing trajectory includes transient trajectory and confirmed trajectory; If the associated target candidate is successfully associated with the transient trajectory, the associated transient trajectory is updated using the associated target candidate, and the trajectory evidence status of the transient trajectory is updated according to the trajectory evidence score of the associated target candidate; wherein, the trajectory evidence status includes the current frame trajectory evidence score, the cumulative trajectory evidence score, the average trajectory evidence score, and the number of evidence hits, and the current frame trajectory evidence score is determined by the trajectory evidence score of the associated target candidate; If the associated target candidate is successfully associated with the confirmed trajectory, the associated target candidate is used to update the successfully associated confirmed trajectory, and the trajectory evidence score of the associated target candidate is used as the current frame trajectory evidence score of the confirmed trajectory. If the associated target candidate fails to be associated with an existing trajectory, and the admission evidence score of the associated target candidate is not lower than the threshold for directly creating a new trajectory, then a new transient trajectory is created based on the associated target candidate, and the trajectory evidence state of the transient trajectory is initialized according to the trajectory evidence score of the associated target candidate. If the associated target candidate fails to be associated with an existing trajectory, and the admission evidence score of the associated target candidate is lower than the direct new trajectory threshold but not lower than the candidate retention threshold, then the associated target candidate is added to the delayed trajectory establishment pool; cross-frame association is performed on the target candidates in the delayed trajectory establishment pool, and when the cumulative hit count and admission evidence score of the same delayed trajectory establishment object meet the release conditions, a transient trajectory is created based on the target candidate corresponding to the delayed trajectory establishment object, and the trajectory evidence state of the transient trajectory is initialized according to the trajectory evidence score of the target candidate; If the associated target candidate is not successfully associated with an existing trajectory, and the admission evidence score of the associated target candidate is lower than the candidate retention threshold, then no new transient trajectory is created based on the associated target candidate, and the associated target candidate is not added to the delayed trajectory establishment pool; Step 5: When the number of evidence hits of the same transient trajectory, the current frame trajectory evidence score of the transient trajectory, and the average trajectory evidence score of the transient trajectory reach the corresponding thresholds, the transient trajectory is converted into a confirmed trajectory; Step 6: When the confirmed trajectory meets the preset output conditions in the current frame, output the target tracking result corresponding to the confirmed trajectory.

2. The method as described in claim 1, characterized in that: When matching the first target candidate and the second target candidate, the normalized spatial distance and the normalized Doppler difference are calculated, and the matching cost is obtained by weighting the normalized spatial distance and the normalized Doppler difference. Calculate the first target candidate With the second target candidate Normalized spatial distance : in, This indicates the horizontal matching threshold. Indicates the vertical matching threshold. , Indicates the first target candidate The center coordinates, , Indicates the second target candidate The center coordinates; Calculate the first target candidate With the second target candidate Normalized Doppler differences : in, Indicates the first target candidate Doppler observations, Indicates the second target candidate Doppler observations, This indicates the Doppler difference after considering the Doppler cycle numbering. Indicates the Doppler matching threshold; Matching cost between the first target candidate and the second target candidate Represented as: in, and These are weighting coefficients, used to adjust the proportions of normalized spatial distance and normalized Doppler difference in the matching cost, respectively. when Not greater than the preset normalized space threshold, and When the value is not greater than the preset normalized Doppler threshold, and A candidate pair is identified as a matchable pair; if multiple matchable objects exist for the same candidate, the matching cost is selected. The smallest candidate pair is taken as the matching result; For successfully matched candidate pairs, the neural network confidence, target location, and Doppler observations in the first target candidate are combined with the number of points, target size, Doppler broadening, Doppler peak ratio, and traditional detection quality in the second target candidate to construct associated target candidates.

3. The method as described in claim 1, characterized in that: The admission evidence score and the trajectory evidence score are obtained based on at least one of the following evidence items: Neural network confidence evidence is used to characterize the degree of confidence of a neural network detection branch in the existence of a target; Physical size evidence, used to characterize at least one of the following: the number of points output by the conventional detection branch, the target width, the target height, or the conventional detection quality; Doppler consistency evidence is used to characterize at least one of the following: target Doppler broadening, proportion of the Doppler main peak, and absolute Doppler intensity; Evidence of temporal continuity is used to characterize the continuity of target appearance across frames; Evidence of motion continuity is used to characterize the dynamic scale of a target or the consistency of motion across frames.

4. The method as described in claim 3, characterized in that: The score for the admission evidence is: The baseline score for the trajectory evidence is: in, This represents the confidence level evidence from the neural network; Indicates evidence of physical dimensions; This indicates evidence of Doppler consistency; Evidence indicating temporal continuity; Evidence of continuity of motion; The trajectory evidence score is determined as follows: in, Score the trajectory evidence. For trajectory evidence-based score, Score for evidence of admission.

5. The method as described in claim 1, characterized in that: The neural network detection branch is implemented by a pre-trained target detection network; The training samples of the target detection network include range-Doppler maps corresponding to radar echo data, and corresponding target annotation information; The target annotation information includes at least one of the following: target center position, target distance, target azimuth, target presence annotation, or Doppler annotation information; The neural network detection branch is used to input the range-Doppler image corresponding to the radar echo data of the current frame into the target detection network and output the first target candidate.

6. The method as described in claim 1, characterized in that: The traditional detection branches include spectral peak detection, angle estimation, spot generation, and target candidate aggregation; Among them, peak detection is used to obtain the range-Doppler detection unit based on range-Doppler spectral data; Angle estimation is used to generate an azimuth response based on the multi-channel range-Doppler data, MIMO Doppler index, and preset angle calibration matrix corresponding to the range-Doppler detection unit, and to determine the target azimuth based on the peak value of the azimuth response; The spot generation function is used to generate radar spots based on range, Doppler, and azimuth. Target candidate aggregation is used to generate the second target candidate based on the radar spot.

7. The method as described in claim 1, characterized in that: The first target candidate includes the target center location, target distance, target azimuth angle, Doppler observation value, and neural network confidence level; The second target candidate includes the target center location, target width, target height, number of points, Doppler observation value, Doppler broadening, proportion of Doppler main peak, and traditional detection quality.

8. The method as described in claim 1, characterized in that: The method for processing the delayed orbit establishment pool in step 4 includes: When the associated target candidate fails to be associated with an existing trajectory, and its admission evidence score is lower than the direct new trajectory threshold but not lower than the candidate retention threshold, the associated target candidate is added to the delayed trajectory establishment pool. In subsequent frames, target candidates in the delayed orbit establishment pool are associated across frames based on region identifier, lateral distance, longitudinal distance, and Doppler difference; When the cumulative hit count and admission evidence score of the same delayed track establishment object in the delayed track establishment pool meet the release conditions, a new transient trajectory is created based on the target candidate corresponding to the delayed track establishment object, and the trajectory evidence state of the transient trajectory is initialized according to the trajectory evidence score of the target candidate. When a delayed track establishment object in the delayed track establishment pool exceeds a preset lifespan or no longer meets the cross-frame association conditions, it is deleted.

9. The method as described in claim 1, characterized in that: The preset output conditions mentioned in step 6 include: The confirmed trajectory is supported by both neural network detection branches and traditional detection branches in the current frame; The number of confirmation checks for the confirmed trajectory is not less than the preset minimum number of confirmation checks; The number of points corresponding to the target candidate associated with the current frame of the confirmed trajectory is not less than the preset output point threshold; The trajectory evidence score of the current frame of the confirmed trajectory is not less than the preset display evidence threshold.

Images