Search

CN-120976501-B - Prediction boundary box dynamic correction method based on multidimensional kinematic features

CN120976501BCN 120976501 BCN120976501 BCN 120976501BCN-120976501-B

Abstract

The invention provides a prediction boundary frame dynamic correction method based on multidimensional kinematic features, and belongs to the technical field of prediction boundary frame dynamic correction. The method comprises the following steps of parameter initialization and data preparation, track effectiveness judgment, border frame offset calculation based on historical displacement offset, track acceleration calculation, turning state detection, border frame offset calculation combined with displacement offset, acceleration and turning state, self-adaptive threshold calculation based on recent movement trend reliability, border frame offset adjustment combined with self-adaptive threshold, and adjustment of an original prediction border frame. The invention can dynamically adjust parameters without additional training data, can be integrated into the existing tracking or predicting system as a light module, can effectively reduce the prediction error of the boundary frame, improves the target positioning stability in complex dynamic scenes, and saves hardware cost and deployment complexity.

Inventors

  • CHENG SHAOWU
  • CHEN SIQIN
  • LIU XINHAO

Assignees

  • 哈尔滨工业大学

Dates

Publication Date
20260508
Application Date
20250811

Claims (10)

  1. 1. The prediction boundary box dynamic correction method based on the multidimensional kinematic features is characterized by comprising the following steps of: Step one, parameter initialization and data preparation, namely setting a prediction boundary box dynamic correction control parameter, and determining target input data, wherein the input data comprises historical track sequence data T and an original prediction boundary box ; Step two, judging the track effectiveness, namely judging the track as an invalid track when the track length n of the historical track sequence data T in the step one is smaller than 2, and directly executing the step nine if effective track analysis cannot be performed at the moment; Extracting the motion trend of the historical track sequence data T and giving a higher weight to recent motion in the historical track sequence data T, and multiplying the offset representing the position change of two adjacent frames of boundary frames in the historical track sequence data T by the corresponding weight to obtain a weighted average offset reflecting the overall motion trend of the target; The fourth step of track acceleration calculation, namely quantitatively describing the motion state of the target through three-level characteristics of displacement, speed and acceleration, wherein the displacement is an offset, the speed is approximately represented through the offset, and the acceleration is obtained through the calculation of the difference value of adjacent speeds; fifthly, detecting turning states, namely based on position information in the historical track sequence data, quantifying direction change by calculating cosine similarity of continuous direction vectors of the position information, and further judging whether a target is in the turning states or not; Step six, calculating the boundary frame offset combining the displacement offset, the acceleration and the turning state, namely combining the weighted average offset obtained in the step three with the acceleration obtained in the step four to generate a basic prediction offset, and enhancing the basic prediction offset to adapt to the motion characteristic of the target in turning when the target is detected to be in the turning state in the step five to obtain an enhanced prediction offset; Step seven, calculating an adaptive threshold based on the reliability of the recent motion trend, namely dynamically constructing the adaptive threshold to quantify the reliability degree of the recent motion trend of the target; step eight, adjusting the boundary frame offset combined with the self-adaptive threshold, namely further adjusting the enhanced predicted offset obtained by the calculation in the step six according to the self-adaptive threshold obtained by the calculation in the step seven to obtain a corrected offset; And step nine, adjusting an original prediction boundary frame, namely carrying out consistency check on the original prediction offset calculated according to the original prediction boundary frame and the correction offset obtained in the step eight through setting a tolerance range to obtain a final correction offset, and outputting the original prediction offset when the track length n of the historical track sequence data T in the step one is smaller than 2.
  2. 2. The method for dynamically modifying a prediction boundary box based on multi-dimensional kinematic features according to claim 1, wherein said prediction boundary box dynamic modification control parameters in step one comprise historical weight attenuation factors Predicting step number Acceleration influencing factor Threshold value of turning detection Factor of turning influence Adaptive threshold scaling factor Adaptive minimum threshold Original prediction offset adjustment threshold 。
  3. 3. The method for dynamically modifying a prediction boundary box based on a multidimensional kinematic feature according to claim 2, wherein the multidimensional kinematic feature parameter in the step one further comprises a prediction uncertainty adjustment threshold The input data further includes a prediction bounding box uncertainty value Input prediction bounding box uncertainty value When predicting bounding box uncertainty values And if the prediction uncertainty adjustment threshold is smaller than the prediction uncertainty adjustment threshold, directly adopting the original prediction boundary frame, otherwise, adjusting the original prediction boundary frame.
  4. 4. The method for dynamically correcting a prediction boundary box based on multidimensional kinematic features according to claim 2, wherein the specific process of obtaining the weighted average offset reflecting the overall motion trend of the target in the step three is as follows: Inputting a historical track sequence , wherein, Represent the first A bounding box of a step of time, Is the upper left corner coordinate of the bounding box, Is the lower right corner of the bounding box, n is the track length, here Otherwise, no correction is needed; offset amount The calculation formula of the position change of the boundary frames of two adjacent frames is as follows: Wherein, the , To highlight the effect of recent offset, the weight decays exponentially with time, historical two adjacent frame offsets Weights of (2) The calculation is as follows: Wherein, the As a result of the historical weight decay factor, Represent the first The decay weights of the individual offsets are used, The smaller the offset, the older the offset, the smaller the weight, the molecular is A kind of electronic device The denominator is the sum of the attenuation weights of all offsets, i.e Normalizing the offset weight; Multiplying the offset by the corresponding weight, and summing, wherein the calculation formula is as follows: Obtaining weighted average offset reflecting overall motion trend 。
  5. 5. The method for dynamically correcting a prediction boundary box based on multidimensional kinematic features according to claim 4, wherein the specific calculation method of the acceleration in the fourth step is as follows: In the step, the length n of the historical track is more than or equal to 3, if the length of the historical track is less than 3, the step is skipped, and the step five is executed, wherein in the continuous frames, the first is that At intervals of time, the speed of the target Approximately the offset of adjacent frames I.e. , The core of acceleration is the rate of change of velocity, i.e. the difference between two adjacent velocities: Wherein, the , The representation is from the first The first speed is to The variation of each speed is used for reducing single-frame noise interference, and the average value of all the variation of the speed is taken as the final acceleration; Wherein, the The average acceleration is 4-dimensional and contains Acceleration components of the four coordinates, Is the upper left corner coordinate of the bounding box, Is the coordinates of the lower right corner of the bounding box.
  6. 6. The method for dynamically correcting a prediction boundary box based on multidimensional kinematic features according to claim 5, wherein the specific method for determining whether the target is in a turning state in the fifth step is as follows: In the step, the length n of the historical track is more than or equal to 3, if the length n of the historical track is less than 3, the step is skipped, and the step six is executed, so that the sequence of the historical track is followed Extracting boundary frame coordinates of the last three frames: , bounding boxes representing the third last frame, the second last frame and the last frame respectively, and calculating two continuous motion direction vectors based on the coordinates of the last three frames: , wherein, Representing slave To the point of Is defined in the direction vector of (a), Representing slave To the point of Each component of the vector corresponds to a bounding box coordinate Direction change amount of (2); If the module length of the direction vector is too small, namely, is close to 0, the target is almost stationary in the time period, and no turning phenomenon exists, and the default judgment is in a non-turning state, and the module length calculation formula is as follows: Wherein, the For four components of the direction vector, if Or (b) , If the value is the minimum value, judging that the vehicle is in a non-turning state; and then normalizing the effective direction vector, wherein the normalization formula is as follows: then, calculating cosine similarity, wherein the cosine similarity can reflect the consistency of the directions; Wherein, the Is the included angle between the two direction vectors, The closer the value is to 1, the more consistent the direction, the closer the value is to-1, and the more opposite the direction; , =1, 2,3,4, respectively normalized back direction vector Corresponding to four components of (a) The components of the direction are calculated by Divided by the corresponding component of (2) In the same way as the above, the following steps, Is that Divided by the corresponding component of (2) ; After the cosine similarity value of the direction vectors of the two directions is calculated, if the cosine similarity is smaller than the turning detection threshold value Then, the turning state is determined, Is a turning state judging function.
  7. 7. The method for dynamically modifying a prediction bounding box based on multidimensional kinematic features according to claim 6, wherein the specific method for obtaining the enhanced prediction offset in the sixth step is as follows: Firstly, combining the calculated offset based on the historical displacement offset with the acceleration to generate a basic prediction offset, wherein the calculation formula is as follows: Wherein, the The amount of offset is predicted on a basis, For the weighted average offset calculated in step three, Is a preset acceleration influence factor which is set, For the average acceleration calculated in step four, In order to predict the number of steps, When the fifth step detects that the target is in a turning state, namely The basic predicted offset needs to be enhanced to adapt to the motion characteristic of the target steering, and the enhanced predicted offset has the following calculation formula: Wherein, the In order to enhance the predicted offset amount, Is a preset turning impact factor.
  8. 8. The method for dynamically modifying a prediction bounding box based on multidimensional kinematic features according to claim 7, wherein the specific method for dynamically constructing the adaptive threshold in the seventh step is as follows: The offset sequence of the historical track is Extracting the last offset And penultimate offset Which represent the recent trend and the last trend, respectively, and are symbolized as Calculating the absolute value proportion of the recent offset and the last offset for the component of each direction vector, and taking the maximum value as the integral change proportion; Wherein, the Represents four values of (2) The offset components of the four coordinates, Is the upper left corner coordinate of the bounding box, Is the coordinates of the lower right corner of the bounding box, Is that With the offset component, the absolute ratio of the recent offset to the last offset, At the level of the minimum value of the total number of the components, Characterizing the maximum variation ratio of the last two offsets; Based on the maximum variation ratio Adaptive threshold scaling factor with prediction Generating adaptive thresholds The formula is: Wherein, the In order to adapt the minimum threshold value to be used, For the interval limiting function, the self-adaptive threshold is ensured not to be lower than 。
  9. 9. The method for dynamically correcting a prediction boundary box based on multidimensional kinematic features according to claim 8, wherein the specific method for obtaining the correction offset in the step eight is as follows: according to the self-adaptive threshold value obtained by calculation in the step seven, the enhanced offset obtained by calculation in the step six is calculated Further adjustment is carried out, and the adjustment rule is as follows: Wherein, the In order to correct the amount of offset, Is L2 norm of vector and is used for quantizing the overall size of the offset, and when the modular length of the offset after enhancement is less than or equal to the self-adaptive threshold value When the product of the recent offset modular length is multiplied, the recent movement trend is determined to be reliable, and the method adopts As an adjusted offset.
  10. 10. The method for dynamically correcting a prediction bounding box based on multidimensional kinematic features according to claim 9, wherein the specific method for obtaining the final correction offset in step nine is as follows: Consistency verification is carried out on the original prediction offset and the correction offset through setting a tolerance range, so that the advantage complementation of two prediction results is realized; first, it is necessary to calculate the original prediction offset from the original prediction bounding box The calculation formula is as follows: Wherein, the For the original prediction bounding box, The method comprises the steps that as a boundary frame of the last frame in a historical track, the two boundary frames adopt a representation mode of adding the left upper corner coordinate of the boundary frame to the right lower corner coordinate; and then, calculating the difference between the original predicted offset and the corrected offset obtained in the step eight, wherein the formula is as follows: Wherein, the The Euclidean distance of the original predicted offset and the corrected offset is used for reflecting the integral deviation degree of the original predicted and corrected results, Represents four values of (2) The direction vector of the four coordinates, Tolerance range adjusts threshold according to set original prediction offset Performing a calculation in the form of a representation of an interval If the Euclidean distance between the original predicted offset and the corrected offset is within the tolerance range or the length n <2 of the historical track, the original predicted offset is adopted, otherwise, the corrected offset is adopted, and the formula is as follows: Wherein, the The offset is corrected for the end.

Description

Prediction boundary box dynamic correction method based on multidimensional kinematic features Technical Field The invention relates to a prediction boundary frame dynamic correction method based on multidimensional kinematic features, and belongs to the technical field of prediction boundary frame dynamic correction. Background The prediction boundary box is used as a direct carrier of the target space position, the accuracy and the stability of the prediction boundary box can directly influence the target detection, tracking and track prediction effects, and the dynamic optimization of the prediction boundary box can correct the boundary box position in real time to avoid the accumulation of positioning errors when the target motion state changes and the scene is interfered, so that reliable guarantee is provided for downstream tasks. Currently, improving the positioning accuracy of the modified prediction bounding box is still a challenge to be solved. In the prediction boundary box correction method in the prior art, the method is mainly divided into two types, namely physical model driving and data driving. Compared with a data driving method, the physical model driving method has the advantage of remarkable light weight. However, the existing model driving method optimizes the prediction boundary box through a uniform motion model, a linear extrapolation model and other single models, and the method only depends on a single motion assumption, so that the method is difficult to adapt to motion mutation, turning and other nonlinear motions in a real scene, and error accumulation is easy to cause. Disclosure of Invention The invention aims to solve the problems that the prediction boundary frame correction method based on model driving in the prior art cannot adapt to various motion modes and can cause accumulated errors, and further provides a prediction boundary frame dynamic correction method based on multidimensional kinematic characteristics. According to the invention, through combining the multidimensional kinematic characteristics of displacement offset, acceleration and turning states of the historical tracks and the self-adaptive threshold adjustment and prediction uncertainty sensing mechanism based on the reliability of recent motion trend, the original prediction boundary frame is dynamically corrected, so that the real-time robust correction of the prediction boundary frame on complex dynamic behaviors such as sudden speed change and turning of a target is realized, the problems that the prediction error is accumulated and multiple scenes cannot be adapted in a high dynamic scene in the prior art are solved, the limitation of a single motion model is broken through in positioning precision, the noise accumulation can be effectively restrained, and meanwhile, the method has the engineering practical advantages of light weight and no need of additional training data, the prediction error of the boundary frame can be effectively reduced, and the target positioning stability in the complex dynamic scene is improved. The invention aims at realizing the following technical scheme: a prediction boundary box dynamic correction method based on multidimensional kinematic features comprises the following steps: Step one, parameter initialization and data preparation, namely setting a prediction boundary box dynamic correction control parameter, and determining target input data, wherein the input data comprises historical track sequence data T and an original prediction boundary box ; Step two, judging the track effectiveness, namely judging the track as an invalid track when the track length n of the historical track sequence data T in the step one is smaller than 2, and directly executing the step nine if effective track analysis cannot be performed at the moment; Extracting the motion trend of the historical track sequence data T and giving a higher weight to recent motion in the historical track sequence data T, and multiplying the offset representing the position change of two adjacent frames of boundary frames in the historical track sequence data T by the corresponding weight to obtain a weighted average offset reflecting the overall motion trend of the target; The fourth step of track acceleration calculation, namely quantitatively describing the motion state of the target through three-level characteristics of displacement, speed and acceleration, wherein the displacement is an offset, the speed is approximately represented through the offset, and the acceleration is obtained through the calculation of the difference value of adjacent speeds; fifthly, detecting turning states, namely based on position information in the historical track sequence data, quantifying direction change by calculating cosine similarity of continuous direction vectors of the position information, and further judging whether a target is in the turning states or not; Step six, calculating the boundary frame offset comb