A method for dynamically correcting a prediction bounding box based on multi-dimensional kinematic features
By integrating multi-dimensional kinematic features and adaptive threshold adjustment to dynamically correct the predicted bounding box, the problem of error accumulation in existing technologies is solved, achieving high-precision target localization in complex dynamic scenes and possessing the advantage of lightweight design.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2025-08-11
- Publication Date
- 2026-05-08
AI Technical Summary
Existing model-driven prediction bounding box correction methods cannot adapt to diverse motion modes, leading to error accumulation, especially insufficient positioning accuracy in complex dynamic scenes.
By integrating multi-dimensional kinematic features of historical trajectory displacement, acceleration, and turning state, and combining adaptive threshold adjustment and prediction uncertainty perception mechanism, the predicted bounding box is dynamically corrected to adapt to complex dynamic behaviors such as sudden changes in target speed and turning.
It significantly improves the robustness and continuity of bounding box prediction, effectively suppresses the accumulation of errors caused by noise and instantaneous disturbances, enhances the stability and accuracy of target localization, and has the advantage of being lightweight without the need for additional training data.
Smart Images
Figure CN120976501B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for dynamic correction of predicted bounding boxes based on multidimensional kinematic features, belonging to the technical field of dynamic correction of predicted bounding boxes. Background Technology
[0002] As the direct carrier of the target's spatial position, the accuracy and stability of the predicted bounding box directly affect the target detection, tracking and trajectory prediction results. Dynamic optimization of the predicted bounding box can correct the bounding box position in real time when the target's motion state changes and scene interference occurs, so as to avoid the accumulation of positioning errors and provide reliable guarantee for downstream tasks.
[0003] Currently, improving the localization accuracy of corrected predicted bounding boxes remains a critical challenge. Existing methods for correcting predicted bounding boxes mainly fall into two categories: physical model-driven and data-driven. Physical model-driven methods offer significant advantages in terms of lightweight design compared to data-driven methods. However, existing model-driven methods optimize predicted bounding boxes using single models such as uniform motion models or linear extrapolation models. These methods rely on only a single motion assumption, making it difficult to adapt to sudden changes in motion or nonlinear motion such as turning in real-world scenarios, which can easily lead to error accumulation. Summary of the Invention
[0004] The purpose of this invention is to solve the problem that the existing model-driven prediction bounding box correction method cannot adapt to diverse motion modes and will lead to cumulative errors. Therefore, this invention provides a dynamic prediction bounding box correction method based on multi-dimensional kinematic features.
[0005] This invention integrates multi-dimensional kinematic features of historical trajectory displacement, acceleration, and turning states with an adaptive threshold adjustment based on the reliability of recent motion trends and a prediction uncertainty perception mechanism to dynamically correct the original predicted bounding box. This enables real-time robust correction of the predicted bounding box to complex dynamic behaviors such as sudden changes in target speed and turning, thereby solving the problems of prediction error accumulation and inability to adapt to diverse scenarios in existing technologies in highly dynamic scenes. It breaks through the limitations of a single motion model in terms of positioning accuracy, effectively suppresses noise accumulation, and has the advantages of being lightweight and requiring no additional training data for engineering practicality. It can effectively reduce bounding box prediction errors and improve the stability of target positioning in complex dynamic scenes.
[0006] The objective of this invention is achieved through the following technical solution:
[0007] A method for dynamic correction of predicted bounding boxes based on multidimensional kinematic features includes the following steps:
[0008] Step 1: Parameter Initialization and Data Preparation: Set the dynamic correction control parameters for the predicted bounding box and determine the target input data, which includes: historical trajectory sequence data T and the original predicted bounding box. ;
[0009] Step 2, Trajectory Validity Determination: When the trajectory length n of the historical trajectory sequence data T in Step 1 is less than 2, it will be determined as an invalid trajectory. At this time, effective trajectory analysis cannot be performed, and Step 9 will be executed directly. When the trajectory length n of the historical trajectory sequence data T is greater than or equal to 2, the historical trajectory sequence data T is valid and Step 3 will be executed.
[0010] Step 3: Calculation of bounding box offset based on historical displacement: Extract the motion trend of historical trajectory sequence data T and give higher weight to recent motion in historical trajectory sequence data T. Multiply the offset representing the change in bounding box position between two adjacent frames in historical trajectory sequence data T with the corresponding weight to obtain the weighted average offset reflecting the overall motion trend of the target.
[0011] Step 4: Trajectory Acceleration Calculation: The target's motion state can be quantitatively described by three levels of features: displacement, velocity, and acceleration. Displacement is the offset; velocity is approximately represented by the offset; and acceleration is calculated by the difference between adjacent velocities.
[0012] Step 5, Turning State Detection: Based on the position information in the historical trajectory sequence data, the cosine similarity of the continuous direction vector of the position information is used to quantify the direction change, thereby determining whether the target is in a turning state;
[0013] Step 6: Calculate the bounding box offset by combining displacement, acceleration, and turning state: Combine the weighted average offset obtained in Step 3 with the acceleration obtained in Step 4 to generate the basic predicted offset. When the target is detected to be turning in Step 5, the basic predicted offset is enhanced to adapt to the motion characteristics of the target when turning, and the enhanced predicted offset is obtained.
[0014] Step 7: Adaptive threshold calculation based on recent motion trend credibility: Quantify the credibility of the target's recent motion trend by dynamically constructing an adaptive threshold;
[0015] Step 8: Adjust the bounding box offset based on the adaptive threshold: Based on the adaptive threshold calculated in Step 7, further adjust the enhanced prediction offset calculated in Step 6 to obtain the corrected offset.
[0016] Step 9: Adjustment of the original predicted bounding box: By setting a tolerance range, the consistency between the original predicted offset calculated based on the original predicted bounding box and the corrected offset obtained in Step 8 is checked to obtain the final corrected offset. When the trajectory length n of the historical trajectory sequence data T in Step 1 is less than 2, the original predicted offset is output.
[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0018] This invention fully leverages the complementarity of kinematic features in three dimensions—displacement offset, acceleration change, and turning state—from historical trajectories with an adaptive threshold adjustment mechanism based on the credibility of recent motion trends. It employs a rule-driven dynamic correction strategy to adapt to complex motion states such as speed changes and turns without requiring additional training data. This significantly improves the robustness and continuity of bounding box prediction, effectively suppresses error accumulation caused by dynamic noise and instantaneous disturbances, and maintains high stability even in challenging scenarios such as occlusion.
[0019] This invention includes optional operations, which are bounding box correction and determination methods based on the uncertainty value of the predicted bounding box. By dynamically adapting the prediction reliability to accurately control the correction timing, the correction operations in low-confidence scenarios can be proactively reduced, thereby improving real-time performance.
[0020] This invention can dynamically adjust parameters without additional training data. It can be integrated into existing tracking or prediction systems as a lightweight module, significantly saving hardware costs and deployment complexity while improving positioning accuracy. Attached Figure Description
[0021] Figure 1 This is a schematic diagram illustrating the basic principle of a predictive bounding box dynamic correction method based on multidimensional kinematic features according to the present invention.
[0022] Figure 2 The flowchart illustrates the application of this invention. Detailed Implementation
[0023] The present invention will be further described in detail below with reference to the accompanying drawings: This embodiment is implemented under the premise of the technical solution of the present invention, and detailed implementation methods are given, but the protection scope of the present invention is not limited to the following embodiments.
[0024] like Figure 1 and Figure 2 As shown in the figure, the method for dynamic correction of predicted bounding boxes based on multidimensional kinematic features involved in this embodiment includes the following steps:
[0025] Step 1: Parameter Initialization and Data Preparation
[0026] Set the dynamic correction control parameters for the predicted bounding box and determine the input data;
[0027] This method requires the inclusion of nine dynamic correction control parameters for the predicted bounding box, including: historical weight decay factor. Predicting steps Acceleration Influence Factor Turn detection threshold Adjusting the threshold for prediction uncertainty Turning Influence Factors Adaptive threshold scaling factor Adaptive minimum threshold and the original prediction offset adjustment threshold
[0028] The input data for this method consists of three types: first, historical trajectory sequence data T, which is composed of bounding boxes at each time step; second, the original predicted bounding boxes. Third, predicting uncertainty values. The confidence level of this uncertainty ranges from [0-1]. It is worth noting that the closer the value is to 1, the higher the reliability of the corresponding predicted bounding box result.
[0029] Step 2: Determining the validity of the trajectory
[0030] Considering that subsequent analysis of historical trajectory sequence data is required, when the trajectory length n of historical trajectory sequence data T is less than 2, it will be judged as an invalid trajectory, and effective trajectory analysis cannot be performed at this time. Step 10 will be executed directly. When the trajectory length n of historical trajectory sequence data T is greater than 2, historical trajectory sequence data T is valid and step 3 will be executed.
[0031] Step 3: Boundary box adjustment decision based on predicted boundary box uncertainty
[0032] This step is optional. If no prediction bounding box uncertainty value is entered, skip this step and proceed to step four. Here, the prediction bounding box uncertainty value is in the range of [0,1]. The closer it is to 1, the higher the prediction uncertainty, indicating a worse prediction effect. If a prediction bounding box uncertainty value has been entered, the threshold will be adjusted according to the preset prediction uncertainty. Make a judgment: when the uncertainty value of the predicted bounding box is... If the value is less than the prediction uncertainty adjustment threshold, the original prediction bounding box is used directly; otherwise, proceed to step four to adjust the original prediction bounding box.
[0033] Step 4: Calculation of bounding box offset based on historical displacement offset
[0034] This step extracts the motion trends of historical trajectories and assigns higher weight to recent motion, making the prediction more closely match the current motion state of the target. The historical trajectory sequence is then input. ,in, Indicates the first The bounding box at each time step. The coordinates of the top-left corner of the bounding box. Here, is the coordinate of the bottom right corner of the bounding box, and n is the trajectory length. Otherwise, no correction is needed;
[0035] offset It refers to the positional change of the bounding boxes between two adjacent frames, and its calculation formula is:
[0036]
[0037] in, , To highlight the impact of recent offsets, the weights decay exponentially over time, and the offsets between two adjacent historical frames are considered. weight The calculation is as follows:
[0038]
[0039] in, Historical weight decay factor This represents the decay weight of the k-th offset. The smaller k is, the older the offset, and the smaller the weight. The numerator is... of The power is the sum of the attenuation weights of all offsets, i.e. , used to normalize the offset weights;
[0040] Multiplying the offset by its corresponding weight and summing the results yields a weighted average offset that reflects the overall trend of the movement. The calculation formula is:
[0041]
[0042] Step 5: Trajectory Acceleration Calculation
[0043] The motion state of a target can be quantitatively described using three levels of features: displacement, velocity, and acceleration. Displacement is represented by the change in position between adjacent frames, i.e., the offset. Speed reflects the change in displacement per unit time. In continuous frame scenes, since the frame interval is a fixed value, it can be measured by the offset. Approximate representation; acceleration reflects the rate of change of velocity per unit time, specifically calculated by the difference between adjacent velocities. By quantifying acceleration, the dynamic change trend of the target during motion can be effectively captured, thereby correcting the cumulative error caused by relying solely on average velocity for prediction in multi-step prediction scenarios and improving prediction accuracy.
[0044] Acceleration calculation requires at least 3 frames of valid historical trajectory, i.e., historical trajectory length n≥3. If the historical trajectory length is less than 3, skip this step and proceed to step six, which calculates the target's velocity in the k-th time interval within consecutive frames. It can be approximated as the offset between adjacent frames, i.e. , The core of acceleration is the rate of change of velocity, that is, the difference between two adjacent velocities:
[0045]
[0046] in, , This indicates the speed from the k-th speed to the... To reduce noise interference in a single frame, the average value of all velocity changes is taken as the final acceleration.
[0047]
[0048] in, The average acceleration has 4 dimensions, including... Acceleration components of the four coordinates, It is the coordinate of the top left corner of the bounding box. These are the coordinates of the bottom right corner of the bounding box;
[0049] Step Six: Turning Status Detection
[0050] The core of turning state detection is to analyze the degree of change in the target's recent motion direction to determine whether its trajectory has changed significantly. This step is based on the position information of the three most recent frames in the historical trajectory. It quantifies the change in direction by calculating the cosine similarity of continuous direction vectors, and then determines whether it is in a turning state. The essential difference between the target's straight-line motion and turning motion lies in the continuity of the motion direction vectors: in the straight-line motion scenario, the direction vectors in adjacent time intervals are highly consistent, and their angle is close to 0°, with the corresponding cosine similarity close to 1; while in the turning motion scenario, the angle between adjacent direction vectors will increase significantly, causing the cosine similarity to decrease significantly and be less than the preset threshold. Based on this, by calculating the cosine similarity of the two continuous direction vectors formed by the three most recent frames in the target's historical trajectory, the magnitude of the change in direction can be quantified, thereby achieving accurate detection of the target's turning state.
[0051] Here, the historical trajectory length n ≥ 3 is also required. If the historical trajectory length is less than 3, skip this step and proceed to step seven, starting from the historical trajectory sequence. Extract the bounding box coordinates of the last three frames: , These represent the bounding boxes of the third-to-last frame, the second-to-last frame, and the last frame, respectively, in the format of... , It is the coordinate of the top left corner of the bounding box. These are the coordinates of the bottom right corner of the bounding box. Based on the coordinates of the most recent three frames, two consecutive motion direction vectors are calculated: ,in, Indicates from arrive directional vector, Indicates from arrive The direction vector, where each component of the vector corresponds to the bounding box coordinates. The change in direction;
[0052] If the magnitude of the direction vector is too small, i.e., close to 0, it indicates that the target is almost stationary during that time period and there is no turning phenomenon. In this case, it is judged by default to a non-turning state. The formula for calculating the magnitude is:
[0053]
[0054] in, Let these be the four components of the direction vector, if or , If the value is a minimum, it is determined to be a non-turning state;
[0055] The effective direction vectors are then normalized using the following formula:
[0056]
[0057] Next, cosine similarity is calculated, which can reflect the consistency of direction;
[0058]
[0059] in, Let be the angle between two direction vectors. The closer the value is to 1, the more consistent the direction; the closer the value is to -1, the more opposite the direction. (i=1,2,3,4) are the normalized direction vectors. The four components correspond to The directional component is calculated as follows: The corresponding component divided by Similarly, for The corresponding component divided by .
[0060] After calculating the cosine similarity values of the direction vectors in two directions, if the cosine similarity is less than the turning detection threshold... If so, it is determined to be a turning state.
[0061]
[0062] Step 7: Calculate the bounding box offset by combining displacement, acceleration, and turning states.
[0063] First, the calculated offset based on historical displacement is combined with the acceleration to generate the basic predicted offset, and the calculation formula is as follows:
[0064]
[0065] in, Based on the predicted offset, This refers to the weighted average offset calculated in step four. The preset acceleration influence factor, The average acceleration calculated in step five, To predict the number of steps, for example, when predicting the first frame in the future, =1;
[0066] When step six detects that the target is in a turning state, that is... The base prediction offset needs to be enhanced to adapt to the motion characteristics of the target during turning. The enhanced prediction offset is calculated using the following formula:
[0067]
[0068] in, For the enhanced prediction offset, The preset turning influence factor;
[0069] Step 8: Adaptive threshold calculation based on the credibility of recent movement trends
[0070] The core logic of dynamically constructing an adaptive threshold to quantify the reliability of recent motion trends is as follows: when the target is in a stable motion state, that is, when the change ratio of the two most recent offsets is small, the recent motion trend has high reliability and is therefore given priority; when the target undergoes a sudden change in motion, that is, when the change ratio of the two most recent offsets is large, the recent motion trend may be affected by noise or instantaneous disturbances. In this case, it is necessary to use the basic offset obtained by fusing historical information for correction. By calculating the adaptive threshold based on the ratio of the two most recent offsets, the reliability of the recent trend and the dependence on historical patterns can be dynamically balanced, thereby effectively improving the robustness of the predicted offset.
[0071] offset sequence of historical trajectory Extract the last offset and the second-to-last offset These represent the recent trend and the previous trend, respectively, and are represented by the symbols: For each component of the directional vector, calculate the ratio of the absolute value of the recent offset to the previous offset, and take the maximum value as the overall change ratio.
[0072]
[0073]
[0074] in, The four possible values represent Offset components of the four coordinates, It is the coordinate of the top left corner of the bounding box. These are the coordinates of the bottom right corner of the bounding box. for Under the offset component, the ratio of the absolute value of the most recent offset to the previous offset. It is the minimum value. Characterizes the maximum percentage change between the two most recent offsets;
[0075] Based on the change ratio With predictive adaptive threshold scaling factor Generate adaptive threshold The formula is:
[0076]
[0077] in, The minimum threshold of the adaptive threshold. The function is an interval-limited function that ensures the adaptive threshold is not lower than [a certain value]. ;
[0078] Step 9: Adjust the bounding box offset based on the adaptive threshold
[0079] Based on the adaptive threshold calculated in step eight, the enhanced offset calculated in step seven is adjusted accordingly. Further adjustments will be made, and the adjustment rules are as follows:
[0080]
[0081] in, To correct the offset, The L2 norm of the vector is used to quantize the overall size of the offset; when the magnitude of the enhanced offset is less than or equal to the adaptive threshold... When multiplying by the recent offset magnitude, the reliability of the recent motion trend is determined. As the adjusted offset; when the magnitude of the enhanced offset is greater than the adaptive threshold When multiplying with the recent offset magnitude, the enhanced offset is used directly. .
[0082] Step 10: Adjusting the original predicted bounding box
[0083] By setting a tolerance range, the consistency between the original prediction offset and the corrected offset is verified, so as to achieve the complementary advantages of the two prediction results.
[0084] First, the original predicted offset needs to be calculated based on the original predicted bounding box. The calculation formula is as follows:
[0085]
[0086] in, The original predicted bounding box, The bounding box is the last frame in the historical trajectory. Both bounding boxes are represented by the coordinates of the top left corner plus the coordinates of the bottom right corner.
[0087] Next, the difference between the original predicted offset and the corrected offset obtained in step nine is calculated using the following formula:
[0088]
[0089] in, The Euclidean distance between the original prediction offset and the corrected offset reflects the overall deviation between the original prediction and the corrected result. The four values of i represent... The direction vectors of the four coordinates,
[0090] The tolerance range adjusts the threshold based on the set original prediction offset. The calculation is performed, and it is represented as an interval. If the original predicted offset is within the tolerance range or the historical trajectory length n < 2, then the original predicted offset is used; otherwise, the corrected offset is used. The formula is as follows:
[0091]
[0092] in, This is the final corrected offset.
[0093] The above description is merely a preferred embodiment of the present invention. These specific embodiments are different implementations based on the overall concept of the present invention, and the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for dynamic correction of predicted bounding boxes based on multidimensional kinematic features, characterized in that, Includes the following steps: Step 1: Parameter Initialization and Data Preparation: Set the dynamic correction control parameters for the predicted bounding box and determine the target input data, which includes: historical trajectory sequence data T and the original predicted bounding box. ; Step 2, Trajectory Validity Determination: When the trajectory length n of the historical trajectory sequence data T in Step 1 is less than 2, it will be determined as an invalid trajectory. At this time, effective trajectory analysis cannot be performed, and Step 9 will be executed directly. When the trajectory length n of the historical trajectory sequence data T is greater than or equal to 2, the historical trajectory sequence data T is valid and Step 3 will be executed. Step 3: Calculation of bounding box offset based on historical displacement: Extract the motion trend of historical trajectory sequence data T and give higher weight to recent motion in historical trajectory sequence data T. Multiply the offset representing the change in bounding box position between two adjacent frames in historical trajectory sequence data T with the corresponding weight to obtain the weighted average offset reflecting the overall motion trend of the target. Step 4: Trajectory Acceleration Calculation: The target's motion state can be quantitatively described by three levels of features: displacement, velocity, and acceleration. Displacement is the offset; velocity is approximately represented by the offset; and acceleration is calculated by the difference between adjacent velocities. Step 5, Turning State Detection: Based on the position information in the historical trajectory sequence data, the cosine similarity of the continuous direction vector of the position information is used to quantify the direction change, thereby determining whether the target is in a turning state; Step 6: Calculate the bounding box offset by combining displacement, acceleration, and turning state: Combine the weighted average offset obtained in Step 3 with the acceleration obtained in Step 4 to generate the basic predicted offset. When the target is detected to be turning in Step 5, the basic predicted offset is enhanced to adapt to the motion characteristics of the target when turning, and the enhanced predicted offset is obtained. Step 7: Adaptive threshold calculation based on recent motion trend credibility: Quantify the credibility of the target's recent motion trend by dynamically constructing an adaptive threshold; Step 8: Adjust the bounding box offset based on the adaptive threshold: Based on the adaptive threshold calculated in Step 7, further adjust the enhanced prediction offset calculated in Step 6 to obtain the corrected offset. Step 9: Adjustment of the original predicted bounding box: By setting a tolerance range, the consistency between the original predicted offset calculated based on the original predicted bounding box and the corrected offset obtained in Step 8 is checked to obtain the final corrected offset. When the trajectory length n of the historical trajectory sequence data T in Step 1 is less than 2, the original predicted offset is output.
2. The method for dynamic correction of predicted bounding boxes based on multidimensional kinematic features according to claim 1, characterized in that, The dynamic correction control parameters for the predicted bounding box mentioned in step one include: historical weight decay factor. Predicting steps Acceleration Influence Factor Turn detection threshold Turning Influence Factors Adaptive threshold scaling factor Adaptive minimum threshold and the original prediction offset adjustment threshold .
3. The method for dynamic correction of predicted bounding boxes based on multidimensional kinematic features according to claim 2, characterized in that, The multidimensional kinematic feature parameters in step one also include a threshold for adjusting prediction uncertainty. The input data also includes the predicted bounding box uncertainty value. Input the uncertainty value of the predicted bounding box. When predicting the bounding box uncertainty value If the value is less than the prediction uncertainty adjustment threshold, the original prediction bounding box is used directly; otherwise, the original prediction bounding box is adjusted.
4. The method for dynamic correction of predicted bounding boxes based on multidimensional kinematic features according to claim 2, characterized in that, The specific process for obtaining the weighted average offset reflecting the overall movement trend of the target in step three is as follows: Input historical trajectory sequence ,in, Indicates the first The bounding box at each time step, The coordinates of the top-left corner of the bounding box. Here, is the coordinate of the bottom right corner of the bounding box, and n is the trajectory length. Otherwise, no correction is needed; offset It refers to the positional change of the bounding boxes between two adjacent frames, and its calculation formula is: in, , To highlight the impact of recent offsets, the weights decay exponentially over time, and the offsets between two adjacent historical frames are considered. weight The calculation is as follows: in, Historical weight decay factor Indicates the first The attenuation weight of each offset The smaller the value, the older the offset, the smaller the weight, and the numerator is... of The power is the sum of the attenuation weights of all offsets, i.e. , used to normalize the offset weights; The formula is to multiply the offset by its corresponding weight and then sum them up: Obtain the weighted average offset that reflects the overall movement trend. .
5. The method for dynamic correction of predicted bounding boxes based on multidimensional kinematic features according to claim 4, characterized in that, The specific calculation method for acceleration mentioned in step four is as follows: In this step, the length of the historical trajectory n ≥ 3. If the length of the historical trajectory is less than 3, skip this step and proceed to step five; in consecutive frames, the first... The target's speed at each time interval Approximately the offset between adjacent frames ,Right now , The core of acceleration is the rate of change of velocity, that is, the difference between two adjacent velocities: in, , Indicates from the first The speed to the first To reduce noise interference in a single frame, the average value of all velocity changes is taken as the final acceleration. in, The average acceleration has 4 dimensions, including... Acceleration components of the four coordinates, It is the coordinate of the top left corner of the bounding box. These are the coordinates of the bottom right corner of the bounding box.
6. The method for dynamic correction of predicted bounding boxes based on multidimensional kinematic features according to claim 5, characterized in that, The specific method for determining whether the target is in a turning state in step five is as follows: In this step, the length of the historical trajectory n ≥ 3. If the length of the historical trajectory n < 3, skip this step and proceed to step six, starting from the historical trajectory sequence. Extract the bounding box coordinates of the last three frames: , Let these represent the bounding boxes of the third-to-last frame, the second-to-last frame, and the last frame, respectively. Based on the coordinates of the three most recent frames, calculate two consecutive motion direction vectors: ,in, Indicates from arrive directional vector, Indicates from arrive The direction vector, where each component of the vector corresponds to the bounding box coordinates. The change in direction; If the magnitude of the direction vector is too small, i.e., close to 0, it indicates that the target is almost stationary during that time period and there is no turning phenomenon. In this case, it is judged by default to a non-turning state. The formula for calculating the magnitude is: in, Let these be the four components of the direction vector, if or , If the value is a minimum, it is determined to be a non-turning state; The effective direction vectors are then normalized using the following formula: Next, cosine similarity is calculated, which reflects the consistency of direction. in, Let be the angle between two direction vectors. The closer the value is to 1, the more consistent the direction; the closer the value is to -1, the more opposite the direction. , =1,2,3,4, which are the normalized direction vectors. The four components correspond to The directional component is calculated as follows: The corresponding component divided by Similarly, for The corresponding component divided by ; After calculating the cosine similarity values of the direction vectors in two directions, if the cosine similarity is less than the turning detection threshold... If so, it is determined to be a turning state. This is a function for determining the turning state.
7. The method for dynamic correction of predicted bounding boxes based on multidimensional kinematic features according to claim 6, characterized in that, The specific method for obtaining the enhanced predicted offset in step six is as follows: First, the calculated offset based on historical displacement is combined with the acceleration to generate the basic predicted offset, and the calculation formula is as follows: in, Based on the predicted offset, This refers to the weighted average offset calculated in step three. The preset acceleration influence factor, The average acceleration calculated in step four, To predict the number of steps, When step five detects that the target is in a turning state, that is... The base prediction offset needs to be enhanced to adapt to the motion characteristics of the target during turning. The enhanced prediction offset is calculated using the following formula: in, For the enhanced prediction offset, This is the preset turning influence factor.
8. The method for dynamic correction of predicted bounding boxes based on multidimensional kinematic features according to claim 7, characterized in that, The specific method for dynamically constructing the adaptive threshold in step seven is as follows: The offset sequence of historical trajectories is Extract the last offset and the second to last offset These represent the recent trend and the previous trend, respectively, and are represented by the symbols: For each component of the directional vector, calculate the ratio of the absolute value of the recent offset to the previous offset, and take the maximum value as the overall change ratio. in, The four possible values represent Offset components of the four coordinates, It is the coordinate of the top left corner of the bounding box. These are the coordinates of the bottom right corner of the bounding box. for Under the offset component, the ratio of the absolute value of the most recent offset to the previous offset. It is the minimum value. Characterizes the maximum percentage change between the two most recent offsets; Based on the maximum change ratio With predictive adaptive threshold scaling factor Generate adaptive threshold The formula is: in, To achieve an adaptive minimum threshold, The function is an interval-limited function that ensures the adaptive threshold is not lower than [a certain value]. .
9. The method for dynamic correction of predicted bounding boxes based on multidimensional kinematic features according to claim 8, characterized in that, The specific method for obtaining the corrected offset in step eight is as follows: Based on the adaptive threshold calculated in step seven, the enhanced offset calculated in step six is adjusted. Further adjustments will be made, and the adjustment rules are as follows: in, To correct the offset, The L2 norm of the vector is used to quantize the overall size of the offset; when the magnitude of the enhanced offset is less than or equal to the adaptive threshold... When multiplying by the recent offset magnitude, the reliability of the recent motion trend is determined. As the adjusted offset.
10. The method for dynamic correction of predicted bounding boxes based on multidimensional kinematic features according to claim 9, characterized in that, The specific method for obtaining the final corrected offset in step nine is as follows: By setting a tolerance range, the consistency between the original prediction offset and the corrected offset is verified, so as to achieve the complementary advantages of the two prediction results. First, the original predicted offset needs to be calculated based on the original predicted bounding box. The calculation formula is as follows: in, The original predicted bounding box, The bounding box is the last frame in the historical trajectory. Both bounding boxes are represented by the coordinates of the top left corner plus the coordinates of the bottom right corner. Next, the difference between the original predicted offset and the corrected offset obtained in step eight is calculated using the following formula: in, The Euclidean distance between the original prediction offset and the corrected offset reflects the overall degree of deviation between the original prediction and the corrected result. The four possible values represent The direction vectors of the four coordinates, The tolerance range adjusts the threshold based on the set original prediction offset. The calculation is performed, and it is represented as an interval. If the Euclidean distance between the original predicted offset and the corrected offset is within the tolerance range or the historical trajectory length n < 2, then the original predicted offset is used; otherwise, the corrected offset is used. The formula is as follows: in, This is the final corrected offset.