Method for cleaning and frame filling of autonomous driving perception data

Abstract

The application relates to the technical field of automatic driving, and provides a cleaning and frame supplementing method for automatic driving perception data. In the cleaning step, abnormal data is identified through motion continuity detection and physical space conflict detection, in the frame supplementing step, suspected data frame segments of short-time abnormality are subjected to interpolation repair conforming to physical laws, suspected data frame segments of long-time abnormality and missing frame segments and invalid data frames are uniformly filled with preset constants, so that the physical rationality of the processed data and the consistency of multi-target interaction are ensured, and the technical problem that the prior art cannot simultaneously consider the data physical rationality and evaluation fairness is solved.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to a method for cleaning and frame interpolation of autonomous driving perception data. Background Technology

[0002] The development of autonomous driving technology relies heavily on testing and evaluation in closed environments and on public roads, with the evaluation of the perception system being particularly crucial. Currently, the Reference multi-sensor fusion system is used to collect raw data, which is then processed through target identification to obtain obstacle target perception data (DET), used to evaluate the performance of the sensor perception system and the ground truth system.

[0003] However, during the acquisition and transmission of sensor data, data quality issues may arise due to sensor failure, occlusion, and other problems. These issues mainly manifest in three categories: invalid values ​​(the target type is marked as invalid, and other core fields have no valid values), missing frames (the frame number sequence is discontinuous, and the frame interval exceeds the normal sampling period), and data anomalies (the target field has values, but their values ​​do not conform to physical laws, such as extreme anomalies in speed or size). These problems directly affect the accuracy of the evaluation results.

[0004] In existing technologies, the first method is based on isolated threshold filtering and simple mathematical interpolation. This method sets a static threshold for each field, and deletes outliers if the threshold is exceeded. Missing segments are filled using linear interpolation or nearest neighbor interpolation. Its disadvantages are that linear interpolation imposes the assumption of uniform linear motion, which cannot adapt to complex motions such as turning, acceleration and deceleration; it cannot detect physical conflicts between different target trajectories; and the static threshold requires repeated manual parameter tuning, resulting in poor generalization ability.

[0005] The second method is based on filtering and state estimation using a fixed motion model. This method establishes a fixed motion model for the target, performs prediction and weighted fusion for each frame, and fills in missing data by inferring the state using the model. Its disadvantages are that the target motion in autonomous driving scenarios is highly nonlinear, and the prediction error of the fixed model will continue to accumulate; long-term occlusion will cause the filling of missing frames to lose its reference value; and it cannot handle physical conflicts such as vehicle overlap.

[0006] The third method is extrapolation-based frame filling based on kinematic equation prediction. This method uses valid data from several frames preceding the missing segment to calculate the instantaneous motion state, assumes the motion pattern remains unchanged, and extrapolates forward to fill in the missing frames. Its drawbacks are that it relies solely on historical states for unidirectional prediction, making it unable to handle behavioral changes such as turning, braking, or evasive maneuvers during the missing period; as a post-hoc, independent method, it cannot utilize anomalies or conflict information discovered during the cleaning process; and it may cause overfilling for excessively short missing segments.

[0007] Therefore, how to provide a method that can effectively solve the data quality problem in autonomous driving perception data and ensure that the processed data is physically reasonable and the evaluation is fair is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] The purpose of this application is to overcome the shortcomings of the prior art and provide a method for cleaning and frame interpolation of autonomous driving perception data, which takes into account both the physical rationality of the data and the fairness of the evaluation.

[0009] The technical solution adopted in this application is as follows: In the first aspect, a method for cleaning and frame interpolation of autonomous driving perception data is provided, the method comprising: Acquire the target to be detected from the sensing data; Based on a preset kinematic model, motion continuity detection is performed on each of the detected targets in consecutive frames to identify suspicious data frames in the perception data. Based on the geometric relationship between different detected targets in the same frame, physical space conflict detection is performed to identify invalid data frames in the perception data. For the suspicious data frame segment, if its duration does not exceed the preset maximum window threshold, interpolation repair is performed based on the motion state information of the detected target in the frames before and after the suspicious data frame segment; if its duration exceeds the preset maximum window threshold, the suspicious data frame segment is marked as a frame segment to be supplemented. Identify missing frames in the perceived data, wherein the missing frames are continuous time intervals whose duration exceeds a preset minimum duration threshold. For each frame in the segment to be supplemented, each frame in the missing segment, and the invalid data frame, the attributes of the detected target are filled with preset constants.

[0010] Preferably, the step of performing motion continuity detection on each of the detected targets in consecutive frames based on a preset kinematic model to determine suspicious data frames in the perceived data includes: If the velocity change of the detected target between adjacent frames exceeds the velocity change threshold defined by the kinematic model, and / or the acceleration change exceeds the acceleration mutation threshold defined by the kinematic model, then the corresponding frame is determined to be a suspicious data frame segment in the perceived data.

[0011] Preferably, the step of performing physical space conflict detection based on the geometric relationship between different detected targets in the same frame to determine invalid data frames in the perceived data includes: If the two-dimensional bounding boxes of any two detected targets in the same frame intersect, then the frame is determined to be an invalid data frame in the perception data.

[0012] Preferably, the method for determining the intersection of the two-dimensional bounding boxes of any two detected targets in the same frame includes: Calculate the center distance between the two detected targets; If the center distance is less than the product of the average width or average length of the two detected targets and the safety factor, then it is determined that an intersection has occurred.

[0013] Preferably, the interpolation repair based on the motion state information of the detected target in the preceding and following frames of the suspicious data frame segment includes: The normal motion state information of the detected target in the previous and next frames of the suspicious data frame segment is used as boundary constraints. Based on the normal motion state information, a preset constant velocity model or a preset constant acceleration model is dynamically adapted to calculate the motion state value of the detected target in each frame of the suspicious data frame segment.

[0014] Preferably, the method further includes: For the detected target whose target type is marked as invalid in the perception data, delete the field data of the detected target except for the frame number and timestamp.

[0015] Preferably, the method further includes: The cleaned and frame-pasted sensory data is organized along a unified timeline and output as a standardized dataset.

[0016] Preferably, before outputting the standardized dataset, the method further includes: For each frame in the standardized dataset, perform timeline continuity and field integrity checks. If the checks pass, output is allowed; if the checks fail, generate a flag and re-execute the cleaning and frame interpolation operations.

[0017] Preferably, the preset constant is 0 or a set of preset tag values ​​with specific semantics.

[0018] Preferably, the preset maximum window threshold is greater than the preset minimum duration threshold, and the preset maximum window threshold is an integer multiple of the preset minimum duration threshold.

[0019] In a second aspect, a system for cleaning and frame interpolation of autonomous driving perception data is provided, the system comprising: The perception data parsing unit is used to acquire the detected target from the perception data; The cleaning unit is used to perform motion continuity detection on each of the detected targets in a continuous frame based on a preset kinematic model to identify suspicious data frame segments in the perception data, and to perform physical space conflict detection based on the geometric relationship between different detected targets in the same frame to identify invalid data frames in the perception data. The frame interpolation unit is used to perform interpolation repair based on the motion state information of the detected target in the preceding and following frames of the suspicious data frame segment if its duration does not exceed a preset maximum window threshold; and to mark the suspicious data frame segment as a frame segment to be interpolated if its duration exceeds the preset maximum window threshold. It is also used to identify missing frames in the perceived data, wherein the missing frames are continuous time intervals whose time length exceeds a preset minimum duration threshold; It is also used to fill the attributes of the detected target with preset constants for each frame in the frame segment to be supplemented, each frame in the missing frame segment, and the invalid data frame.

[0020] In a third aspect, a computer device is provided, including a memory and a processor; the memory stores a computer program that can run on the processor, and when the computer program is executed by the processor, it performs the method provided in the first aspect.

[0021] In the fourth aspect, a computer-readable storage medium is provided that stores a computer program, which, when executed, implements the method provided in the first aspect.

[0022] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application relates to the field of autonomous driving technology and provides a method for cleaning and frame interpolation of autonomous driving perception data. The method includes: acquiring the detected targets in the perception data; performing motion continuity detection on each detected target in consecutive frames based on a preset kinematic model to identify suspicious data frame segments in the perception data, and performing physical space conflict detection based on the geometric relationship between different detected targets in the same frame to identify invalid data frames in the perception data; for suspicious data frame segments, if their duration does not exceed a preset maximum window threshold, then performing interpolation repair based on the motion state information of the detected targets in the frames before and after the suspicious data frame segment; if their duration exceeds the preset maximum window threshold, then marking the suspicious data frame segment as a frame segment to be interpolated; identifying missing frame segments in the perception data, where a missing frame segment is a continuous time interval whose duration exceeds a preset minimum duration threshold; and filling each frame in the frame segment to be interpolated, each frame in the missing frame segment, and invalid data frames with the attributes of the detected targets as preset constants. In the cleaning step, this application identifies abnormal data through motion continuity detection and physical space conflict detection. In the frame interpolation step, suspicious data frames with short-term anomalies are repaired by interpolation in accordance with physical laws, and suspicious data frames with long-term anomalies, missing frames, and invalid data frames are uniformly filled with preset constants. This ensures the physical rationality of the processed data and the consistency of multi-target interaction, and solves the technical problem that the existing technology is unable to balance the physical rationality of the data and the fairness of the evaluation. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a general flowchart of the method for cleaning and frame interpolation of autonomous driving perception data provided in the embodiments of this application; Figure 2 This is a flowchart of the process for determining suspicious data frames provided in an embodiment of this application; Figure 3 This is a flowchart of determining invalid data frames provided in an embodiment of this application; Figure 4 This is a flowchart provided in an embodiment of the present application for determining whether the two-dimensional bounding boxes of two detected targets intersect; Figure 5 This is a flowchart of the interpolation repair provided in the embodiments of this application; Figure 6 This is a flowchart of the output of cleaned and frame-patched sensor data provided in an embodiment of this application; Figure 7This is a schematic diagram of the system for cleaning and frame interpolation of autonomous driving perception data provided in an embodiment of this application; Figure 8 This is a schematic diagram of the computer device provided in the embodiments of this application. Detailed Implementation

[0025] The technical solutions in the embodiments of this application will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art are within the scope of protection of this application.

[0026] As described in the background section, during the acquisition and transmission of sensor data, data quality issues may arise due to sensor failure, occlusion, or other problems. These issues mainly manifest in three ways: invalid values ​​(the target type is marked as invalid, and other core fields have no valid values), missing frames (the frame number sequence is discontinuous, and the frame interval exceeds the normal sampling period), and data anomalies (the target field has values, but their values ​​do not conform to physical laws, such as extreme anomalies in speed or size). These problems directly affect the accuracy of the evaluation results. The three existing solutions all have their own drawbacks, making it difficult to balance the physical validity of the data with the fairness of the evaluation, thus failing to effectively solve the aforementioned problems.

[0027] Based on this, this application provides a method for cleaning and frame interpolation of autonomous driving perception data, aiming to solve the data quality problems, as well as the problems of data physical rationality and evaluation fairness in the prior art.

[0028] The embodiments of this application are described in detail below with reference to the accompanying drawings.

[0029] Example 1 This embodiment provides a method for cleaning and frame interpolation of autonomous driving perception data, referencing... Figure 1 The methods include: S1: Acquire the target to be detected from the sensing data; S2: Based on the preset kinematic model, motion continuity detection is performed on each detected target in the continuous frame to identify suspicious data frame segments in the perception data, and physical space conflict detection is performed based on the geometric relationship between different detected targets in the same frame to identify invalid data frames in the perception data. S3: For suspicious data frame segments, if their duration does not exceed the preset maximum window threshold, interpolation repair is performed based on the motion state information of the detected target in the frames before and after the suspicious data frame segment; if their duration exceeds the preset maximum window threshold, the suspicious data frame segment is marked as a frame segment to be supplemented. S4: Identify missing frames in the perceived data. A missing frame is a continuous time interval whose duration exceeds a preset minimum duration threshold. S5: For each frame in the supplementary frame segment, each frame in the missing frame segment, and invalid data frames, fill the attributes of the detected target with preset constants.

[0030] Step S1 is the data acquisition step. The system reads the perception data collected by the autonomous driving test vehicle during the evaluation process from the storage device or data bus. The perception data includes at least the camera video frame sequence, the LiDAR point cloud sequence, the state information of the detected target, and the field data inherent in each data set. Each obstacle target is parsed from this perception data as a detected target. Each detected target is associated with state fields such as a unique target identifier (track_id), target type (type), 3D position (center_x, center_y, center_z), size (length, height, width), velocity (velocity_x, velocity_y), heading angle (yaw), and acceleration (acceleration_x, acceleration_y). This completes the acquisition of the detected targets.

[0031] Step S2 is the data cleaning step. The preset kinematic model can be a uniformly accelerated motion model, used to describe the evolution of the motion state of the detected target between consecutive frames, such as the mathematical constraints between position, velocity, and acceleration. Motion continuity detection is based on this kinematic model to determine whether the actual state changes of the detected target between adjacent frames, such as velocity differences or acceleration differences, fall within the reasonable deviation range allowed by the model. If they exceed this range, the corresponding consecutive frame segments are marked as suspicious data frame segments. Suspicious data frame segments refer to multiple frames of abnormal data that are consecutive in time. Physical space conflict detection targets any two detected targets in the same frame. Based on their geometric dimensions and spatial positions, it is determined whether their two-dimensional bounding boxes intersect or have too small a spacing, thereby identifying invalid data frames that cannot coexist physically. Invalid data frames refer to single frames that have spatial conflicts.

[0032] The two detection methods described above constitute the verification mechanism in the cleaning process. Motion continuity detection ensures the temporal and physical authenticity of the trajectory of a single detected target, avoiding outliers introduced by sensor noise or transmission errors. Physical space conflict detection removes frames that violate basic spatial constraints at the multi-target interaction level, ensuring the logical consistency of the positional relationships among the detected targets within the scene. Through these two verification steps, suspicious and invalid data frames in the perceived data can be comprehensively located, providing a reliable basis for subsequent interpolation repair or standardized filling.

[0033] Steps S3 to S5 are data frame completion steps. In step S3, the preset maximum window threshold is a time threshold used to distinguish between short-term and long-term anomalies. The duration of a suspicious data frame segment refers to the time span from the start frame to the end frame. When the duration of a suspicious data frame segment does not exceed the preset maximum window threshold, it indicates that the anomaly duration is short, and there are still reliable normal state frames before and after the segment as boundaries. In this case, interpolation repair can be used to calculate the reasonable state values ​​of each intermediate frame using the motion state information of the preceding and following frames, thereby restoring the data to its completeness while maintaining physical continuity. If the duration of a suspicious data frame segment exceeds the preset maximum window threshold, it indicates that the anomaly duration is too long, and the time interval between the preceding and following normal frames is too far. The result of interpolation repair is no longer of reference value, and forced interpolation may introduce greater deviation. Therefore, repair is no longer attempted, and the segment is marked as a frame segment to be completed, proceeding to the subsequent standardized filling process.

[0034] The logic behind this differentiated processing is that short-term anomalies are usually caused by instantaneous sensor noise or brief occlusion. During this period, the target's movement can still be constrained by the preceding and following normal states, and the true trajectory can be accurately reconstructed through interpolation that conforms to kinematic laws. Long-term anomalies, on the other hand, often correspond to prolonged target occlusion or continuous sensor failure. In this case, it is impossible to accurately predict the target's actual motion state, and continuing to use interpolation will produce unreliable data. By setting a preset maximum window threshold, this first embodiment achieves a balance between repair reliability and data integrity, maximizing the preservation of valid information while avoiding the introduction of false data due to over-repair, providing a clear boundary for subsequent standardized data filling.

[0035] As an example, the preset maximum window threshold can be set to 167ms, corresponding to a time span of approximately 5 frames of data. If the duration of a suspicious data frame segment does not exceed 167ms, it is considered a short-term anomaly and interpolation repair is performed; if it exceeds 167ms, it is considered a long-term anomaly and marked as a frame segment to be supplemented.

[0036] In step S4, a missing frame segment refers to a continuous time interval in the sensed data that is naturally missing due to storage or transmission reasons. The method for identifying missing frame segments is to scan the trajectory time series of each detected target and check the difference in timestamps between adjacent frames. If this difference exceeds a preset minimum duration threshold, it is determined that there are missing frames in between. Consecutive missing frames constitute a missing frame segment. The preset minimum duration threshold is a time threshold used to determine whether data loss has occurred. It is usually set slightly larger than the normal sampling period. For example, if the normal sampling period is 33.33ms, the preset minimum duration threshold can be set to 33.40ms. The purpose of this threshold is to distinguish between normal frame intervals and abnormal data loss. Since the sampling frequency may have slight fluctuations, the timestamps of adjacent frames will not be strictly equal to the sampling period. Therefore, a tolerance deviation needs to be set; only when this deviation is exceeded is it considered a missing frame.

[0037] The missing frame segments and the frames to be filled marked in step S3 differ in origin. The frames to be filled originate from long-term anomalies detected by motion continuity detection, meaning that although the data exists, it does not conform to physical laws; while the missing frame segments originate from natural gaps in the data itself, meaning that they are not recorded at all on the timeline. Both are subjected to the same standardized filling process in the subsequent step S5, but their identification logic is independent of each other. By setting a preset minimum duration threshold, this embodiment can accurately capture the location and range of data loss, avoiding misjudgment as missing due to normal frame interval fluctuations.

[0038] In a preferred embodiment, the preset maximum window threshold is greater than the preset minimum duration threshold, and the preset maximum window threshold is an integer multiple of the preset minimum duration threshold.

[0039] The preset minimum duration threshold is used to determine the starting point of natural missing data, while the preset maximum window threshold is used to define the upper limit of repairable short-term anomalies. The latter typically corresponds to several times the standard sampling period, such as 5 times. This integer multiple relationship simplifies parameter configuration, ensures that the two thresholds work together on the same time reference, and avoids judgment conflicts caused by inconsistencies in scale.

[0040] In step S5, the attributes of the detected target, namely the state fields mentioned above, include numerical parameters such as three-dimensional position, size, velocity, heading angle, and acceleration. The preset constant is a predetermined fixed value used to replace the original values ​​in the aforementioned attributes. The purpose of uniformly filling each frame in the to-be-filled frame segment, each frame in the missing frame segment, and invalid data frames with this preset constant is to prevent attempts to infer unreliable specific states for long-term abnormal segments, naturally missing segments, and invalid frames with physical conflicts that cannot be repaired through interpolation. Instead, a standardized and consistent constant is used to fill these segments, thus forming a complete and continuous data sequence on the timeline. This also ensures that all evaluated systems receive completely consistent input under the same missing scenario, fundamentally guaranteeing the fairness of the evaluation.

[0041] It should be noted that if the same frame is marked as both a suspicious data frame segment and an invalid data frame, the invalid data frame will be used as the standard, and the process will directly proceed to the constant filling step without interpolation repair.

[0042] As a preferred implementation, the preset constant is 0 or a set of preset tag values ​​with specific semantics.

[0043] For the marked missing, invalid, and unfilled frame segments, a standardized zero-state data frame is generated at each time point where filling is required. All numerical state variables (such as position, size, velocity, heading angle, acceleration, etc.) are filled with 0, while the frame number, timestamp, and target type information are retained to indicate the target and type to which the zero-value frame belongs. Besides all zeros, preset constants can also be a set of preset marker values ​​with specific semantics, such as filling the velocity field with -999. As long as this set of constants is consistently defined across all datasets and significantly distinguishable from normal values, the technical effect of marking missing data and ensuring evaluation fairness can still be achieved.

[0044] For missing frames, invalid frames, and other data frames that cannot be repaired through interpolation, their true motion state cannot be accurately determined. Using speculative filling would introduce unreliable information. A unified pre-set constant for filling ensures that all evaluated systems receive completely consistent input under the same missing data scenario, thus eliminating the incomparability of evaluation results caused by differences in preprocessing algorithms. Zero values ​​are preferred due to their computational friendliness and clear semantics, while pre-set marker values ​​with specific semantics can clearly separate missing data from real data while retaining the data missing identifier, facilitating identification by subsequent analysis tools.

[0045] In summary, this embodiment identifies abnormal data through motion continuity detection and physical space conflict detection during the cleaning step. In the frame completion step, suspicious data frames with short-term anomalies are repaired using physically consistent interpolation, while suspicious data frames with long-term anomalies, missing frames, and invalid data frames are uniformly filled with preset constants. This ensures the physical rationality of the processed data and the consistency of multi-target interactions, solving the technical problem of existing technologies struggling to balance data physical rationality and evaluation fairness. Furthermore, compared to the three solutions in the background technology, this embodiment eliminates the need to set static thresholds for each field, avoiding repeated manual parameter tuning. Physical space conflict detection identifies and handles geometric contradictions such as vehicle overlap, compensating for the shortcomings of existing methods in handling multi-target interaction anomalies. Short-term anomalies are interpolated using a bidirectional constrained motion model, while long-term anomalies and missing segments are uniformly filled with preset constants, ensuring both data physical rationality and fundamentally solving the evaluation fairness problem.

[0046] As a preferred embodiment, refer to Figure 2 Based on a pre-defined kinematic model, motion continuity detection is performed on each detected target in consecutive frames to identify suspicious data frames in the perceived data, including: S20: If the velocity change of the detected target between adjacent frames exceeds the velocity change threshold defined by the kinematic model, and / or the acceleration change exceeds the acceleration mutation threshold defined by the kinematic model, then the corresponding frame is determined to be a suspicious data frame segment in the perception data.

[0047] In this step, a kinematic model is used to constrain the state changes between adjacent frames. When the change in velocity exceeds the velocity change threshold, it indicates that the velocity of the detected target has changed in a way that does not conform to the laws of physics. When the change in acceleration exceeds the acceleration change threshold, it indicates that the detected target has undergone a physically impossible velocity change in a very short time. For example, an ordinary passenger car cannot increase its speed from 0 m / s to 30 m / s in 33 ms.

[0048] As an example, suppose the speed of the detected target in four consecutive frames is 10m / s, 10m / s, 50m / s, and 10m / s respectively. The speed change threshold defined by the kinematic model is 5m / s. Then the speed change of the third frame relative to the second frame is 40m / s, which far exceeds the speed change threshold. The continuous frame segment consisting of this frame and the adjacent frames before and after it will be marked as a suspicious data frame segment.

[0049] As another example, suppose the acceleration of the detected target in three consecutive frames is 0 m / s², 0 m / s², and 20 m / s² respectively, and the acceleration change threshold defined by the kinematic model is 5 m / s². Then the acceleration change in the third frame relative to the second frame is 20 m / s², which far exceeds the acceleration change threshold. This frame and the consecutive frame segments formed by it and its preceding and following frames will be marked as suspicious data frames.

[0050] As a preferred embodiment, refer to Figure 3 Physical spatial conflict detection is performed based on the geometric relationships between different detected targets within the same frame to identify invalid data frames in the perceived data, including: S21: If the two-dimensional bounding boxes of any two detected targets in the same frame intersect, then the frame is determined to be an invalid data frame in the perception data.

[0051] In this context, a two-dimensional bounding box refers to a rectangular frame formed by projecting the dimensions of the detected target onto a plane. It is typically determined by the coordinates of its center point, width, and length (or height). In real-world autonomous driving scenarios, two independent road users should not simultaneously occupy the same physical space. Therefore, once the intersection of two two-dimensional bounding boxes is detected, it can be determined that there is a physically impossible conflict in the frame data, and the frame is marked as an invalid data frame.

[0052] As a preferred embodiment, refer to Figure 4 Methods for determining the intersection of the two-dimensional bounding boxes of any two detected targets in the same frame include: S210: Calculate the center distance between two detected targets; S211: If the center distance is less than the product of the average width or average length of the two detected targets and the safety factor, then it is determined that an intersection has occurred.

[0053] As an example, at the same time, any two detected targets and The formula for calculating the center distance is: If the center distance is less than the average width of both and the safety factor The product of is considered to have an intersection. Where, ( , () represents the center location of the target being detected. The width of the target being detected, and the safety factor. A preset value less than 1, used to allow the minimum safe distance.

[0054] As another example, at the same time, the formula for calculating the center distance between any two detected targets i and j is: If the center distance is less than the average length of both and the safety factor The product of is considered to have an intersection. Where, ( , () represents the center location of the target being detected. The length of the target being detected, and the safety factor. A preset value less than 1, used to allow the minimum safe distance.

[0055] Among them, the safety factor in two examples The value can range from 0.6 to 0.8, and can be configured according to the target type, sensor accuracy, and specific evaluation scenario requirements.

[0056] In the two examples above, the center distance between the two detected targets is less than the reasonable range allowed by their respective widths or lengths, indicating a risk of collision or overlap. The corresponding frames should be marked as invalid data frames.

[0057] It should be noted that this judgment only applies to false boundary box intersections in non-real physical space caused by perception detection errors, labeling errors, and algorithm prediction anomalies, which are judged as invalid data frames. If the target physical overlap and box intersection are generated by real vehicle collisions on real roads (such as car accidents), they are not considered invalid data frames and will not be cleaned or removed. Such accident scene frames are retained in their entirety as extreme safety scene data.

[0058] As a preferred embodiment, refer to Figure 5 Interpolation repair is performed based on the motion state information of the detected target in the preceding and following frames of the suspicious data frame segment, including: S30: Use the normal motion state information of the detected target in the previous and next frames of the suspicious data frame segment as boundary constraints. S31: Dynamically adapt a preset constant velocity model or a preset constant acceleration model based on normal motion state information to calculate the motion state value of the detected target in each frame of the suspicious data frame segment.

[0059] Using the normal motion state information of the frame preceding and following the suspicious data segment as boundaries ensures that the repair result smoothly connects with the real data on the timeline, avoiding the accumulation of errors caused by unidirectional prediction. Dynamic adaptation refers to automatically selecting an appropriate model based on the motion state of the preceding and following boundary frames. If the velocities of the preceding and following frames are approximately equal, a constant velocity model is adapted; if there is a significant difference in velocity, a constant acceleration model is adapted, making the interpolated state values ​​more closely match the actual motion trend of the target.

[0060] It should be noted that this embodiment emphasizes that the motion state of the previous and next frames must be normal because only when the data of the boundary frames is true and reliable can the interpolation result have reference value. If the boundary frames themselves are also abnormal data, the interpolation repair will further amplify the error and produce a physically less reliable trajectory. In order to ensure the normality of the boundary frames, this embodiment has verified all data in step S2 through motion continuity detection and physical space conflict detection. Only frames that are not marked as suspicious data frames and not marked as invalid data frames are identified as normal frames and can be used as reliable boundaries for interpolation repair.

[0061] After adaptation, the intermediate state value of each frame within the segment is calculated by combining the selected model with the state parameters of the boundary frames (such as position, velocity, and acceleration) and the time length of the suspicious data frame segments, thereby completing the physical consistency repair of short-term anomalies.

[0062] As an example, suppose the suspected data segment is missing three frames. The velocity of the previous frame is 10 m / s, the velocity of the next frame is 11 m / s, and the time interval between the two boundary frames is 200 ms. Since the velocity exhibits a linear trend, the system automatically adapts to a constant acceleration model, calculating an acceleration of 5 m / s². This leads to the deduced velocities of the missing frames: 10.25 m / s, 10.5 m / s, and 10.75 m / s, respectively. The positions are calculated using the uniform acceleration motion formula. If the velocities of both frames are 15 m / s, then a constant velocity model is applied, filling in the missing frame velocities with 15 m / s, and the positions are calculated using uniform motion.

[0063] As a preferred embodiment, the method further includes: For detected targets whose target type is marked as invalid in the perception data, delete the field data of the detected target except for the frame number and timestamp.

[0064] These invalid detected targets typically indicate that the perception module did not output a valid target in the current frame, or that the data is filler data reserved for fixed data entries. Their core fields themselves have no physical meaning, and retaining these fields would consume storage space and interfere with subsequent evaluations. By deleting field data other than frame number and timestamp, the integrity of the timeline is preserved, while avoiding the potential impact of meaningless data on statistical analysis or algorithm performance evaluation.

[0065] As a preferred embodiment, refer to Figure 6 The methods also include: S6: Organize the cleaned and frame-padding sensory data along a unified time axis and output a standardized dataset.

[0066] The cleaned and padded sensor data is organized along a unified timeline, aligning the trajectory frames of all detected targets according to their timestamps to form a complete and continuous time series. Since the original data may contain missing frames or time shifts for different targets, this unified organization ensures that the data structure of the output standardized dataset remains consistent at every moment, facilitating direct use by subsequent evaluation algorithms.

[0067] As a preferred implementation, the method further includes the following steps before outputting the standardized dataset: Perform timeline continuity and field integrity checks on each frame in the standardized dataset. If the checks pass, output is allowed; if the checks fail, a flag is generated and the cleaning and frame interpolation operations are re-executed.

[0068] The verification process includes checking whether the timestamps are strictly incremental, whether there are intervals exceeding the preset tolerance range, and whether each frame contains the required fields with correct numerical types. If the verification passes, the dataset meets the evaluation requirements and output is allowed; if the verification fails, a marker containing the location and type of the problematic frames is generated, and the cleaning and frame-padding operations are re-executed.

[0069] As an example, the perception data collected from the autonomous driving test section, after cleaning and frame interpolation, yielded a set of trajectory data containing three vehicles. Then, all frames from the three vehicles were aligned along a unified timeline to generate a standardized dataset from 0 to 10 seconds, with each frame approximately 33ms apart. This dataset was then validated, revealing that at the 5th second, the timestamp of one frame significantly exceeded the preset minimum duration threshold, and some fields were missing. The system automatically generated a "missing frame segment needs interpolation" marker, re-identified the missing segment, and performed preset constant interpolation again. Finally, the validation passed, and a complete standardized dataset was output.

[0070] Example 2 Embodiment 2 of this application provides a system for cleaning and frame interpolation of autonomous driving perception data, referring to... Figure 7 The system includes: The perception data parsing unit is used to acquire the detected target from the perception data; The cleaning unit is used to perform motion continuity detection on each detected target in a continuous frame based on a preset kinematic model to identify suspicious data frame segments in the perception data, and to perform physical space conflict detection based on the geometric relationship between different detected targets in the same frame to identify invalid data frames in the perception data. The frame interpolation unit is used to perform interpolation repair based on the motion state information of the detected target in the frames before and after the suspicious data frame segment if its time length does not exceed the preset maximum window threshold; if its time length exceeds the preset maximum window threshold, the suspicious data frame segment is marked as a frame segment to be interpolated. It is also used to identify missing frames in perceived data. A missing frame is a continuous time interval whose duration exceeds a preset minimum duration threshold. It is also used to fill the attributes of the detected target with preset constants for each frame in the frame to be supplemented segment, each frame in the missing frame segment, and invalid data frames.

[0071] The technical effects of the system in Embodiment 2 are similar to those of the method in Embodiment 1. Please refer to the description of Embodiment 1 for details, which will not be repeated here.

[0072] As a preferred embodiment, refer to Figure 8 The system also includes a dataset output unit, which organizes the cleaned and frame-padding sensory data along a unified time axis and outputs it as a standardized dataset.

[0073] Example 3 This embodiment three provides a computer device, including a memory and a processor; the memory stores a computer program that can run on the processor, and when the computer program is executed by the processor, it executes the method for cleaning and interpolating autonomous driving perception data provided in embodiment one above.

[0074] Among them, such as Figure 8 As shown, an exemplary computer device of this embodiment is illustrated, which may specifically include a processor 1510, a video display adapter 1511, a disk drive 1512, an input / output interface 1513, a network interface 1514, and a memory 1520. The processor 1510, video display adapter 1511, disk drive 1512, input / output interface 1513, network interface 1514, and memory 1520 can be communicatively connected via a communication bus 1530.

[0075] The processor 1510 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solution provided in this application.

[0076] The memory 1520 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1520 can store the operating system 1521 for controlling the operation of the computer device, and the basic input / output system 1522 for controlling the low-level operations of the computer device. Additionally, it can store a web browser 1523, a data storage management system 1524, and a device identification information processing system 1525, etc. The aforementioned device identification information processing system 1525 can be the application program that specifically implements the aforementioned steps in this embodiment. In summary, when implementing the technical solution provided in this application through software or firmware, the relevant program code is stored in the memory 1520 and is called and executed by the processor 1510.

[0077] Input / output interface 1513 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0078] Network interface 1514 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0079] The communication bus 1530 includes a pathway for transmitting information between various components of the device, such as processor 1510, video display adapter 1511, disk drive 1512, input / output interface 1513, network interface 1514, and memory 1520.

[0080] In addition, the device can also obtain information on specific claim conditions from the virtual resource object claim condition information database for condition judgment, and so on.

[0081] It should be noted that although the above-described device only shows the processor 1510, video display adapter 1511, disk drive 1512, input / output interface 1513, network interface 1514, memory 1520, communication bus 1530, etc., in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the solution of this application, and does not necessarily include all the components shown in the figures.

[0082] Example 4 This fourth embodiment provides a computer-readable storage medium storing a computer program. When the computer program is executed, it implements the method for cleaning and frame interpolation of autonomous driving perception data provided in the first embodiment.

[0083] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of the embodiments of this application.

[0084] The method for cleaning and frame interpolation of autonomous driving perception data provided in this application has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and its core ideas. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for cleaning and frame interpolation of perception data for autonomous driving, characterized in that, The method includes: Acquire the target to be detected from the perceived data; Based on a preset kinematic model, motion continuity detection is performed on each of the detected targets in consecutive frames to identify suspicious data frames in the perception data. Based on the geometric relationship between different detected targets in the same frame, physical space conflict detection is performed to identify invalid data frames in the perception data. For the suspicious data frame segment, if its duration does not exceed the preset maximum window threshold, interpolation repair is performed based on the motion state information of the detected target in the frames before and after the suspicious data frame segment; if its duration exceeds the preset maximum window threshold, the suspicious data frame segment is marked as a frame segment to be supplemented. Identify missing frames in the perceived data, wherein the missing frames are continuous time intervals whose duration exceeds a preset minimum duration threshold. For each frame in the segment to be supplemented, each frame in the missing segment, and the invalid data frame, the attributes of the detected target are filled with preset constants.

2. The method for cleaning and frame interpolation of autonomous driving perception data according to claim 1, characterized in that, The step of performing motion continuity detection on each of the detected targets in consecutive frames based on a preset kinematic model to determine suspicious data frames in the perceived data includes: If the velocity change of the detected target between adjacent frames exceeds the velocity change threshold defined by the kinematic model, and / or the acceleration change exceeds the acceleration mutation threshold defined by the kinematic model, then the corresponding frame is determined to be a suspicious data frame segment in the perceived data.

3. The method for cleaning and frame interpolation of autonomous driving perception data according to claim 1, characterized in that, The step of performing physical space conflict detection based on the geometric relationships between different detected targets in the same frame to determine invalid data frames in the perceived data includes: If the two-dimensional bounding boxes of any two detected targets in the same frame intersect, then the frame is determined to be an invalid data frame in the perception data.

4. The method for cleaning and frame interpolation of autonomous driving perception data according to claim 3, characterized in that, Methods for determining the intersection of the two-dimensional bounding boxes of any two detected targets in the same frame include: Calculate the center distance between the two detected targets; If the center distance is less than the product of the average width or average length of the two detected targets and the safety factor, then it is determined that an intersection has occurred.

5. The method for cleaning and frame interpolation of autonomous driving perception data according to claim 1, characterized in that, The interpolation and repair based on the motion state information of the detected target in the preceding and following frames of the suspicious data frame segment includes: The normal motion state information of the detected target in the previous and next frames of the suspicious data frame segment is used as boundary constraints. Based on the normal motion state information, a preset constant velocity model or a preset constant acceleration model is dynamically adapted to calculate the motion state value of the detected target in each frame of the suspicious data frame segment.

6. The method for cleaning and frame interpolation of autonomous driving perception data according to claim 1, characterized in that, The method further includes: For the detected target whose target type is marked as invalid in the perception data, delete the field data of the detected target except for the frame number and timestamp.

7. The method for cleaning and frame interpolation of autonomous driving perception data according to any one of claims 1 to 6, characterized in that, The method further includes: The cleaned and frame-pasted sensory data is organized along a unified timeline and output as a standardized dataset.

8. The method for cleaning and frame interpolation of autonomous driving perception data according to claim 7, characterized in that, Before outputting the standardized dataset, the method further includes: For each frame in the standardized dataset, perform timeline continuity and field integrity checks. If the checks pass, output is allowed; if the checks fail, generate a flag and re-execute the cleaning and frame interpolation operations.

9. The method for cleaning and frame interpolation of autonomous driving perception data according to any one of claims 1 to 6, characterized in that, The preset constant is 0 or a set of preset tag values ​​with specific semantics.

10. The method for cleaning and frame interpolation of autonomous driving perception data according to any one of claims 1 to 6, characterized in that, The preset maximum window threshold is greater than the preset minimum duration threshold, and the preset maximum window threshold is an integer multiple of the preset minimum duration threshold.

Images