Vehicle control methods and devices, electronic equipment and storage media

By fusing multi-source vehicle speed information and employing a hierarchical suppression strategy, the problem of ghosting suppression failure when vehicle speed is abnormal in existing technologies has been solved. This enables accurate ghosting identification and processing even when vehicle speed is abnormal, ensuring driving safety and comfort.

CN123104674APending Publication Date: 2026-11-17CHONGQING CHANGAN AUTOMOBILE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611410616.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-11
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, static target filtering methods based on speed thresholds cannot effectively suppress ghosting when the vehicle speed is abnormal, causing the radar ghosting suppression mechanism to fail, which affects the driving experience and driving safety.

Method used

By acquiring multi-source vehicle speed information, and utilizing inherent physical properties independent of vehicle speed, such as Doppler physics, spatial geometric constraints, and kinematic continuity, combined with a hierarchical suppression strategy, ghosting can be identified and processed.

Benefits of technology

It enables accurate ghost image identification even when vehicle speed is lost, improving the accuracy and robustness of identification and ensuring driving safety and comfort.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123104674A_ABST
    Figure CN123104674A_ABST
Patent Text Reader

Abstract

This application relates to a vehicle control method, device, electronic device, and storage medium. The method includes: acquiring multi-source vehicle speed information; determining whether the vehicle speed is in a failed state based on the degree of deviation between different speed information sources; if the vehicle speed is determined to be in a failed state, acquiring each detection point detected by radar in the vehicle and extracting ghosting features of each detection point in multiple physical dimensions; calculating the deviation of the corresponding detection point in each physical dimension based on the ghosting features of each detection point to obtain the ghosting confidence score of the corresponding detection point; performing a graded suppression operation on the detection points based on the ghosting confidence score and outputting a list of suppressed detection points for subsequent vehicle control. This application solves the problem that the existing static target filtering method based on speed thresholds for ghosting suppression often fails.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of automotive parts technology, specifically to a vehicle control method and device, electronic equipment, and storage medium. Background Technology

[0002] Millimeter-wave radar, due to its all-weather operation and direct speed measurement advantages, has become one of the core sensors in Advanced Driver Assistance Systems (ADAS). However, in practical applications, radar suffers from a long-standing problem—low-speed ghosting. Low-speed ghosting refers to false targets generated by radar due to multipath reflection and strong stationary reflective objects (metal guardrails, tunnel walls, bridge piers, gantries, etc.). These false targets are particularly common in low-speed urban driving conditions, leading to false braking, abnormal acceleration, or unexplained warnings, severely impacting driving experience and safety. Current technologies suppress ghosting through static target filtering based on speed thresholds. This involves using vehicle speed and radar Doppler velocity to classify targets with near-zero absolute speed as stationary obstacles and filter them out. However, when vehicle speed becomes abnormal, the overall suppression mechanism fails. In practical applications, vehicle speed dynamic correction may fail due to wheel speed sensor slippage, CAN bus interference or delay, vehicle bumps or jumps, or changes in tire size. At this time, the vehicle speed signal cannot accurately reflect the vehicle's motion state, and the traditional vehicle speed dynamic correction method cannot correct it in time, causing the ghosting suppression logic based on vehicle speed to fail. Summary of the Invention

[0003] This application provides a vehicle control method and apparatus, electronic device and storage medium to solve the problem that ghosting suppression based on speed threshold static target filtering in the prior art fails when the vehicle speed is abnormal.

[0004] In a first aspect, this application provides a method for controlling a vehicle, including: Acquire multi-source vehicle speed information; wherein, the multi-source vehicle speed information includes speed information collected by various different types of devices in the vehicle; Based on the degree of deviation between the speed information in the multi-source vehicle speed information, it is determined whether the vehicle speed is in a failure state, wherein the failure state indicates that the vehicle speed cannot be reliably estimated through dynamic correction. When it is determined that the vehicle speed is in the failure state, each detection point detected by the radar in the vehicle is obtained, and the ghost feature quantity of each detection point in multiple physical dimensions is extracted. The ghost feature quantity is used to characterize the degree to which the detection point deviates from the real physical target in the corresponding physical dimension. Based on the ghost feature quantity of each detection point, the deviation of the corresponding detection point in each physical dimension is fused and calculated to obtain the ghost confidence of the corresponding detection point, wherein the ghost confidence is used to characterize the probability that the detection point belongs to a ghost. Based on the ghost confidence level, a graded suppression operation is performed on the detection points, and a list of suppressed detection points is output for subsequent vehicle control.

[0005] Secondly, this application provides a vehicle control device, comprising: The acquisition module is used to acquire multi-source vehicle speed information; wherein, the multi-source vehicle speed information includes speed information collected by various different types of devices in the vehicle; The first processing module is used to determine whether the vehicle speed is in a failure state based on the degree of deviation between the speed information in the multi-source vehicle speed information, wherein the failure state indicates that the vehicle speed cannot be reliably estimated through dynamic correction. The second processing module is used to acquire each detection point detected by the radar in the vehicle when it is determined that the vehicle speed is in the failure state, and extract the ghost feature quantity of each detection point in multiple physical dimensions, wherein the ghost feature quantity is used to characterize the degree to which the detection point deviates from the real physical target in the corresponding physical dimension. The third processing module is used to perform a fusion calculation on the deviation of the corresponding detection point in each physical dimension based on the ghost feature quantity of each detection point to obtain the ghost confidence of the corresponding detection point, wherein the ghost confidence is used to characterize the probability that the detection point belongs to a ghost. The fourth processing module is used to perform graded suppression operations on the detection points according to the ghost confidence level, and output a list of suppressed detection points for subsequent vehicle control.

[0006] Thirdly, this application provides an electronic device, including: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory is used to store a computer program; and the processor is used to implement the vehicle control method described in the first aspect when executing the computer program.

[0007] Fourthly, this application provides a storage medium storing a computer program thereon, characterized in that the computer program, when executed by a processor, implements the vehicle control method described in the first aspect.

[0008] In this application, autonomous identification of vehicle speed failure states is achieved through deviation verification between multi-source vehicle speed information. Problems can be detected early when wheel speed signals become abnormal, rather than reactively responding after ghosting occurs. Furthermore, after vehicle speed failure, ghosting identification no longer relies on unreliable absolute speed values, but instead utilizes inherent physical properties independent of vehicle speed, such as Doppler laws, spatial geometric constraints, and kinematic continuity, ensuring the effectiveness of processing under failure conditions. Moreover, the evaluations of multiple physical dimensions are independent and complementary; misjudgments in a single dimension can be corrected by normal performance in other dimensions, improving the accuracy and robustness of ghosting identification. Further, the hierarchical suppression strategy avoids hard, black-and-white decisions. High-confidence ghosts are completely filtered out to ensure safety, medium-confidence suspicious points are retained but their use is restricted to maintain comfort, and low-confidence points are output normally to ensure perceptual integrity, achieving refined processing. Attached Figure Description

[0009] Figure 1 A flowchart of the vehicle control method provided in the embodiments of this application; Figure 2 A flowchart of a radar low-speed ghosting suppression method provided in an embodiment of this application when vehicle speed dynamic correction fails; Figure 3 This is a schematic diagram of the structure of the vehicle control device provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0010] The embodiments of this application will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be understood that the preferred embodiments are only for illustrating this application and are not intended to limit the scope of protection of this application.

[0011] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0012] For ease of description, spatial relative terms may be used in the text to describe the relative position or movement of one element or feature relative to another element or feature, as shown in the figure. These relative terms include, for example, "inside," "outside," "middle," "outer," "below," "below," "above," "front," "back," etc. Such spatial relative terms are intended to include different orientations of the device in use or operation, other than those depicted in the figure. For example, if the device in the figure undergoes a positional flip, orientation change, or change of motion, these directional indications will change accordingly. For instance, an element described as "below other elements or features" or "below other elements or features" will subsequently be oriented "above other elements or features" or "above other elements or features." Therefore, the example term "below" can include both upper and lower orientations. The device may be otherwise oriented (rotated 90 degrees or in other directions), and the spatial relative descriptors used in the text will be interpreted accordingly.

[0013] To address the issue that existing ghosting suppression methods based on speed thresholds fail when vehicle speeds are abnormal, this application provides a vehicle control method, such as... Figure 1 As shown, the steps of this method include: Step 101: Obtain multi-source vehicle speed information; wherein, the multi-source vehicle speed information includes speed information collected by various different types of devices in the vehicle; In this embodiment, multi-source vehicle speed information refers to raw data or estimated values ​​related to the vehicle's speed that are independently measured by at least two different types of sensors or devices installed on the vehicle at the same time. These devices include, but are not limited to: wheel speed sensors (which calculate vehicle speed by measuring the number of wheel rotations), inertial measurement units (which calculate speed changes by integrating acceleration), global navigation satellite system receivers (which directly calculate absolute speed through satellite positioning), millimeter-wave radar (which infers vehicle speed by detecting the Doppler frequency shift of stationary targets), and vision sensors (which estimate their own speed by analyzing changes in optical flow in an image).

[0014] Because different types of sensors operate on completely different physical principles (wheel speed relies on mechanical rotation, IMU on inertial measurement, GNSS on satellite signals, radar on electromagnetic wave reflection, and vision on image optical flow), their respective failure modes and interference sources are independent of each other. For example, wheel speed may be distorted due to spinning on a slippery road surface, but GNSS and radar can still provide accurate speed references during this period. Utilizing this inherent physical redundancy, the system can cross-verify the authenticity of vehicle speed signals, rather than blindly trusting a single signal source.

[0015] Step 102: Based on the degree of deviation between the speed information in the multi-source vehicle speed information, determine whether the vehicle speed is in a failure state. The failure state means that the vehicle speed cannot be reliably estimated through dynamic correction. In this embodiment, the degree of deviation refers to the absolute value of the speed difference between any two different sources after spatiotemporal alignment of multi-source vehicle speed information (unifying signals from different sources to the same time and the same vehicle coordinate system). If the speed values ​​from each source are close to each other, it indicates that the information given by all sensors is consistent, and the vehicle speed signal is reliable; if the speed values ​​from some sources deviate significantly from other sources, it indicates that these sources may be abnormal, and the overall reliability of the vehicle speed decreases.

[0016] Furthermore, dynamic correction refers to a vehicle speed estimation method that uses wheel speed signals and combines them with a vehicle dynamics model for real-time compensation and correction (e.g., using Kalman filtering to smooth and correct wheel speed). When the vehicle experiences drive wheel slippage, tire size changes, CAN bus transmission delays, or severe vehicle vibrations causing instantaneous jumps in wheel speed signals, the dynamic correction algorithm, due to its physical limitations in response speed and correction range, cannot promptly correct the abnormal wheel speed signals back to their true values. At this point, an unacceptable deviation exists between the actual and true values ​​of the vehicle speed signal, indicating a failure in the vehicle speed assessment. It's important to note that this failure does not mean all vehicle speed information sources are invalid, but rather that the dynamic correction link centered on wheel speed is no longer reliable. Even with wheel slippage, non-wheel speed sources such as IMU and GNSS may still function normally. This means that multi-source verification can detect wheel speed link failures and, under these conditions, activate backup strategies, utilizing reference speeds from other sources to continue subsequent processing.

[0017] Step 103: When it is determined that the vehicle speed is in a failure state, obtain each detection point detected by the radar in the vehicle, and extract the ghost feature quantity of each detection point in multiple physical dimensions. The ghost feature quantity is used to characterize the degree to which the detection point deviates from the real physical target in the corresponding physical dimension. In this embodiment, a detection point refers to a discrete observation data unit output by the radar receiver after the electromagnetic waves emitted by the radar are reflected by an object in the environment, and after signal processing steps such as constant false alarm rate detection and peak extraction. Each detection point typically contains attribute information such as range, azimuth, elevation, Doppler velocity, and reflection intensity. It should be noted that a detection point is not equivalent to a target object. Real targets (such as vehicles in front) often generate multiple detection points, which need to be clustered and tracked to form a stable target. Ghosting also appears in the radar output in the form of detection points. Therefore, ghosting suppression must be performed at the detection point level to prevent false points from aggregating into false targets.

[0018] Furthermore, the multiple physical dimensions in this application's embodiments refer to different physical perspectives for evaluating whether a detection point truly exists. Specifically, it can be evaluated from three dimensions: the signal domain (whether the Doppler velocity of the detection point matches the expected radial velocity caused by the vehicle's motion), the spatial domain (whether the spatial position of the detection point falls within a reasonable area where a physically possible object might exist), and the time domain (whether the motion trajectory of the detection point conforms to Newtonian kinematics). Each dimension reflects the same issue from different angles, namely, whether the detection point possesses the attributes that a real physical target should have.

[0019] Furthermore, the ghosting feature quantity is the quantitative output of the above evaluation process. For each detection point and each physical dimension, the system calculates a numerical index reflecting the degree of anomaly in that dimension. The higher the degree of anomaly, the less the detection point conforms to the expected behavior of a real target in that dimension, and the more likely it is a ghosting caused by multipath reflection or a strongly stationary object. For example, in the signal dimension, if the Doppler velocity of a detection point measured by the radar differs significantly from the theoretical value calculated based on the vehicle's reference speed and the azimuth angle of that point, then the ghosting feature quantity of that point in that dimension will be very high, because the Doppler velocity of a real physical target should follow this physical law.

[0020] In this way, when vehicle speed fails, the system no longer relies on simple filtering based on whether the absolute speed is close to zero (because vehicle speed itself is unreliable). Instead, it utilizes the Doppler physical characteristics inherent in the radar signal, the spatial distribution pattern of the target, and kinematic constraints to examine each detection point from multiple independent dimensions, providing sufficient basis for subsequent ghosting determination.

[0021] Step 104: Based on the ghost feature quantity of each detection point, the deviation of the corresponding detection point in each physical dimension is fused and calculated to obtain the ghost confidence of the corresponding detection point. The ghost confidence is used to characterize the probability that the detection point belongs to a ghost. In this embodiment, fusion computation refers to a fusion strategy combining hierarchical decision-making and fuzzy inference to synthesize deviation metrics across various physical dimensions into a continuous confidence value between 0 and 1. The processing logic is as follows: First, a hierarchical decision tree structure is used to determine an initial confidence interval based on the number of dimensions in which the detection point is marked as anomaly. Then, within this interval, continuous smoothing adjustments are made based on the specific anomaly values ​​of each dimension, ultimately outputting a precise confidence value. The ghost confidence score is the final output comprehensive score index; the higher the value, the greater the probability that the detection point belongs to a ghost. This confidence score is not a simple yes / no binary judgment but a continuous metric, which provides a refined decision-making basis for subsequent implementation of differentiated suppression strategies.

[0022] By fusing evidence from multiple physical dimensions in this way, the potential bias and uncertainty of judgment based on a single dimension can be overcome. For example, although a detection point may show an anomaly in the signal dimension, if its spatial location is in the center of the lane and the visual sensor confirms the presence of a target, the fusion calculation will comprehensively consider all evidence and output a moderate confidence value, rather than simply classifying it as a ghost, thereby effectively reducing the risk of misjudgment.

[0023] Step 105: Based on the ghost confidence level, perform graded suppression operation on the detection points and output a list of suppressed detection points for subsequent vehicle control.

[0024] In this embodiment, the graded suppression operation refers to classifying detection points into several levels based on their confidence level, with each level corresponding to a different processing action, rather than employing a single filtering or binary retention strategy. Specifically: For detection points with extremely high confidence (clearly ghosting), the most stringent processing method is adopted, i.e., they are directly removed from the radar point cloud, preventing them from participating in any subsequent fusion and perception processing. For detection points with moderate confidence (suspected but uncertain), they are not directly deleted, but a limiting label is added, allowing them to be used only for safety-related functions (such as emergency braking) and not for comfort functions (such as adaptive cruise control). For detection points with low confidence but still suspicious, a delayed confirmation strategy is adopted, i.e., the point is not used immediately, and is only allowed to be established as a target after it has appeared stably and continuously for multiple consecutive processing cycles. For detection points with extremely low confidence (basically credible), no suppression processing is performed, and they participate normally in subsequent perception fusion.

[0025] Furthermore, the list of suppressed detection points is the final output of the system, where each detection point has either been removed or has a corresponding confidence level label. This list is sent to the vehicle's control module for subsequent decision-making and control functions. It is evident that by using a tiered strategy rather than a blanket filtering approach, high-confidence ghosting is completely eliminated while excessive suppression of real targets is avoided, achieving a balance between safety and comfort.

[0026] Based on steps 101 to 105 above, autonomous identification of vehicle speed failure states is achieved through deviation verification between multi-source vehicle speed information. Problems can be detected early when wheel speed signals become abnormal, rather than passively responding after ghosting occurs. Furthermore, after vehicle speed failure, the system no longer relies on unreliable absolute speed values, but instead utilizes inherent physical properties independent of vehicle speed for ghosting identification, such as Doppler laws, spatial geometric constraints, and kinematic continuity, ensuring the effectiveness of processing under failure conditions. Moreover, the evaluations of multiple physical dimensions are independent and complementary; misjudgments in a single dimension can be corrected by normal performance in other dimensions, improving the accuracy and robustness of ghosting identification. Further, the hierarchical suppression strategy avoids hard, black-and-white decisions. High-confidence ghosts are completely filtered out to ensure safety, medium-confidence suspicious points are retained but their use is restricted to maintain comfort, and low-confidence points are output normally to ensure perceptual integrity, achieving refined processing.

[0027] In an optional embodiment of this application, the method for obtaining multi-source vehicle speed information in step 101 above may further include: Step 11: Acquire at least two of the following information: wheel speed pulse signal provided by wheel speed sensor in vehicle, acceleration and angular velocity provided by inertial measurement unit, absolute velocity provided by global navigation satellite system, inverted velocity of stationary target detected by radar, and motion flow velocity provided by visual sensor.

[0028] In this embodiment, the wheel speed sensor is a rotational speed measuring device installed near the vehicle wheel, typically operating on the principles of magnetoelectric or Hall effect. This sensor measures the wheel's angular velocity by detecting the pulse signals generated by the toothed ring rotating with the wheel, and then calculates the vehicle's forward speed by combining this with the known tire rolling radius.

[0029] Furthermore, an Inertial Measurement Unit (IMU) is a measurement device based on inertial principles, integrating a three-axis accelerometer and a three-axis gyroscope. The accelerometer measures the linear acceleration of the vehicle in three-dimensional space, while the gyroscope measures the vehicle's yaw rate, pitch rate, and roll rate. By integrating the acceleration signal along the time axis, the change in vehicle speed can be calculated; combined with the known initial speed, an absolute speed estimate at any given moment can be obtained. The IMU has unique advantages as a source of vehicle speed information: it is completely independent of external signals (such as satellite signals or electromagnetic wave reflections) and unaffected by the wheel-ground contact state. Therefore, even when wheel slippage occurs, the IMU can still provide continuous and reliable motion information. In this application, the acceleration and angular velocity provided by the IMU are important independent sources of evidence for determining vehicle speed health. Especially at the moment of a sudden change in wheel speed signal, the IMU's output, due to the independence of its physical principles, can serve as a reliable reference for determining wheel speed anomalies.

[0030] Furthermore, Global Navigation Satellite System (GNSS) refers to satellite navigation and positioning systems including GPS, BeiDou, GLONASS, and Galileo. A GNSS receiver receives navigation signals transmitted by multiple satellites, calculates its three-dimensional spatial position using triangulation, and then calculates its absolute velocity using Doppler shift measurement or position difference methods. The velocity provided by GNSS is an absolute velocity, that is, the velocity relative to a fixed Earth coordinate system, not a relative velocity relative to the ground. This characteristic makes it an absolute velocity reference independent of the vehicle's own motion. Unlike wheel speed and IMU, GNSS is not affected by wheel slippage and does not have integral drift problems; its velocity measurement accuracy can reach the order of 0.1 m / s when the signal is good. In this application, the absolute velocity provided by GNSS serves as an independent source of multi-source information, providing a high-confidence absolute velocity reference when the signal is available. Especially in situations where wheel speed and IMU are disputed, GNSS often acts as an adjudicator in determining vehicle speed health.

[0031] Furthermore, when millimeter-wave radar detects stationary targets in the environment (such as roadside utility poles, streetlights, traffic signs, stationary vehicles, etc.), it can obtain the radial relative velocity of the target relative to the vehicle by measuring the Doppler frequency shift of the target's reflected echo. Since the target is stationary and its absolute velocity is zero, the value of this radial relative velocity actually reflects the projected component of the vehicle's velocity in the radar beam direction. By jointly calculating the Doppler velocities of multiple stationary targets at different azimuth angles, the vehicle's velocity vector in the vehicle coordinate system can be deduced. This method of vehicle speed acquisition has two significant advantages. First, it does not rely on wheel speed signals and is not affected by wheel slippage. Second, its physical principles are completely different from those of IMU and GNSS, utilizing the Doppler effect of electromagnetic waves rather than mechanical rotation or satellite signals, thus creating natural physical redundancy with other sources. When there are a sufficient number of stationary targets and they are well distributed within the radar's field of view, the retrieved vehicle speed has high accuracy and confidence. In this application, the inversion speed of a stationary target detected by radar is a source of vehicle speed information with great engineering value because its physical mechanism is completely different from that of wheel speed signals, and it can provide the most direct counter-evidence when wheel speed slips.

[0032] Furthermore, the motion flow velocity provided by visual sensors (typically referring to vehicle-mounted monocular or binocular cameras) refers to the technique of estimating the vehicle's own speed by analyzing pixel motion patterns (i.e., optical flow fields) between consecutive image frames. Specifically, the camera acquires image sequences of the road scene ahead, and uses optical flow algorithms to calculate the pixel displacement of each feature point in the image (such as road texture, lane line edges, roadside building outlines, etc.) between different frames. Combining the camera's intrinsic parameter matrix and the vehicle motion model, these pixel displacements can be converted into the vehicle's three-dimensional motion velocity relative to the environment. The physical principle of this vehicle speed acquisition method is based on optical imaging and computer vision, which is different from wheel speed, IMU, GNSS, and radar. Visual sensors, similar to radar, belong to the category of environmental perception sensors, but compared to radar, visual sensors can provide rich texture and semantic information (such as "this is a pedestrian" or "that is a vehicle"), which provides an additional verification dimension for subsequent spatial plausibility determination. In this application, the motion flow velocity provided by the visual sensor, as one of the optional sources of multi-source vehicle speed information, becomes another independent witness in the multi-source redundancy chain due to its unique physical principle.

[0033] As can be seen, this application embodiment explicitly lists five vehicle speed information sources: wheel speed pulse signal, IMU acceleration and angular velocity, GNSS absolute velocity, radar stationary target inversion velocity, and visual motion flow velocity. This provides a clear and unambiguous implementation method for multi-source vehicle speed information, making the solution clearly operable in engineering. The flexible limitation of at least two sources allows the solution to adapt to vehicle platforms with different sensor configuration levels. It is suitable for high-end models equipped with a full set of perception sensors, as well as entry-level models with only basic sensor configurations. As long as the vehicle model has at least two different types of vehicle speed information sources, it can be implemented, demonstrating good engineering versatility and commercial promotion value. In addition, in specific implementation, the system can dynamically select the combination of sources participating in vehicle speed health determination based on the availability and signal quality of each information source under the current operating conditions, exhibiting adaptive capabilities. In open road sections with good GNSS signals, absolute velocity can be fully utilized as a reference, while in tunnels or urban canyons, it automatically switches to the wheel speed and IMU combination, which does not rely on external signals, ensuring the robustness of the solution under different environmental conditions.

[0034] In an optional embodiment of this application, the method of determining whether the vehicle speed is in a faulty state based on the degree of deviation between the speed information in the multi-source vehicle speed information involved in step 102 above may further include: Step 21: Perform spatiotemporal alignment processing on the speed information in the multi-source vehicle speed information to obtain multiple vehicle speed estimates with the same reference. In this embodiment of the application, the spatiotemporal alignment process includes operations at both the temporal alignment and spatial alignment levels.

[0035] Time alignment refers to the fact that data collected by different sensors may have different times and frequencies (for example, a wheel speed sensor may output at 100Hz, an IMU at 200Hz, and a GNSS at only 10Hz), meaning the raw data is not strictly synchronized on the timeline. Time alignment involves unifying all data sources to the same timestamp through interpolation or synchronization triggering, ensuring that the speed information compared is from the same moment. Without time alignment, comparing data collected by one sensor earlier than another will introduce errors due to time asynchrony, leading to misjudgments.

[0036] Spatial alignment refers to the fact that different sensors may define different velocity directions or coordinate systems. For example, wheel speed sensors provide longitudinal velocity along the vehicle's direction of travel, IMUs provide a three-dimensional velocity vector in the vehicle coordinate system, GNSS provides eastward and northward velocity components in the geographic coordinate system, and radar-retrieved velocity is the radial velocity component along the radar beam direction. Spatial alignment involves transforming this velocity information from different coordinate systems into a unified reference coordinate system (e.g., the vehicle coordinate system) and extracting comparable velocity components (usually the longitudinal velocity component along the vehicle's direction of travel), making velocity values ​​from different sources comparable.

[0037] After spatiotemporal alignment, the original data from each source are transformed into vehicle speed estimates based on the same benchmark, that is, multiple speed values ​​at the same time, in the same coordinate system, and under the same physical dimensions. These values ​​can be directly compared mathematically and their deviations calculated.

[0038] Step 22: Pair any two speed estimates from different sources together and calculate the deviation between each pair of speed estimates. In this embodiment, "any two different sources" means that no pre-screening of sources is performed; instead, all possible pairwise combinations are paired and calculated. If there are n available sources of vehicle speed estimates, the total number of pairwise combinations is C_n^2 = n(n-1) / 2. For example, when there are three sources: wheel speed, IMU, and GNSS, the pairwise combinations to be calculated are: wheel speed and IMU, wheel speed and GNSS, and IMU and GNSS, a total of three pairs. This approach of exhaustively enumerating all combinations rather than selecting only one pair for comparison ensures that the system can obtain complete information on the relationships between all sources, and any consistency or deviation between a pair of sources will be captured.

[0039] Furthermore, in engineering implementation, the deviation value typically refers to the absolute difference between two vehicle speed estimates (i.e., |vi - vj|), but sometimes a relative deviation (i.e., |vi - vj| / max(vi, vj)) can be used depending on actual needs. In this application, the deviation value is a positive number, and its magnitude directly reflects the degree of disagreement between the estimates of the same physical quantity (vehicle speed) from the two information sources at the same time. That is, the larger the deviation, the more serious the discrepancy between the two information sources. These deviation values ​​are the input data for subsequent threshold comparisons. The core significance of this step is that it decomposes the macroscopic problem of consistency between multiple sources into several microscopic problems of consistency between pairs of sources, making the consistency monitoring problem mathematically feasible.

[0040] Step 23: Compare each pair of deviation values ​​with their respective preset tolerance thresholds to obtain the deviation judgment result between each pair of vehicle speed estimates; In this embodiment, the preset tolerance threshold refers to the upper limit of allowable deviation pre-set and stored in the system for each specific sensor pair. It is important to emphasize that different sensor pairs correspond to different tolerance thresholds, rather than using the same uniform threshold for all pairs. This is because speed estimates from different sources naturally have different accuracy characteristics and noise levels. For example, the wheel speed and IMU pair is considered a golden combination in engineering, and their deviation is usually very small under normal circumstances, so a relatively strict tolerance threshold (e.g., 3 km / h) can be set. However, the accuracy of motion flow velocity from visual sensors is greatly affected by lighting and texture, and its normal deviation range from wheel speed is inherently wider. Therefore, the pair involving vision should be set with a relatively lenient tolerance threshold (e.g., 6 km / h). If a uniform threshold is used for all pairs, it is possible that an overly strict threshold could lead to misjudgment of deviations within the normal error range of the sensors, or an overly lenient threshold could prevent the identification of genuine anomalies. The setting of differentiated thresholds reflects the refined design of this solution at the engineering calibration level.

[0041] Furthermore, the deviation determination result is a binary (Boolean) result: when the deviation value of a pair exceeds its corresponding preset tolerance threshold, the determination result for that pair is deviation; when the deviation value does not exceed the tolerance threshold, the determination result is consistency. This operation of converting continuous deviation values ​​into discrete "deviation / consistency" determination results allows subsequent statistical steps to be quantified by event counting. It is evident that only deviations significantly exceeding the normal fluctuation range of the sensor pair are considered abnormal, avoiding frequent false judgments triggered by inherent sensor noise.

[0042] Step 24: Based on all deviation judgment results, calculate the overall degree of deviation between each vehicle speed estimate and generate a vehicle speed health index. In this embodiment, all the divergence determination results refer to all binary determination results obtained after pairwise comparisons of all sensor pairs. These results together constitute a complete information matrix describing the relationships between all vehicle speed sources at the current moment, i.e., which sensor pairs are divergent and which pairs are consistent.

[0043] Furthermore, calculating the overall degree of deviation between various vehicle speed estimates involves weighted statistical processing of these binary judgment results, rather than simply counting the number of deviations. Specifically, different sensor pairs have varying degrees of reliability in engineering, so their contributions to the overall degree of deviation should differ when they deviate. For example, the wheel speed and IMU combination, being both high-precision sources with complementary physical principles, indicates a higher severity of the problem if they deviate, thus warranting a higher weight in the statistics. Conversely, the vision sensor and wheel speed combination, due to the lower precision of vision itself, has a lower reliability in indicating anomalies even if deviations occur, thus its weight can be correspondingly reduced. The specific method for weighted statistics is as follows: for each deviation event, a preset weight coefficient is assigned based on the sensor source combination corresponding to that event; then, the weighted values ​​of all deviation events are summed to obtain the total weighted deviation score.

[0044] Furthermore, the vehicle speed health index is a normalized comprehensive quantitative indicator with a value ranging from 0 to 1. The calculation logic is as follows: divide the weighted total deviation score by the total number of combinations of all sensor pairs to obtain the normalized deviation ratio; then subtract this normalized deviation ratio from the value 1 to obtain the vehicle speed health index. The engineering implication of this formula is that the vehicle speed health index reflects the weighted proportion of consistent pairings among all vehicle speed sources. The closer the index is to 1, the more consistent the sensor pairs are, and the higher the overall reliability of the vehicle speed signal; the closer the index is to 0, the more divergent the sensor pairs are, and the lower the overall reliability of the vehicle speed signal. For vehicle speed estimates that do not diverge, their weighted contribution is zero.

[0045] As can be seen, by aggregating the discrete and scattered multiple pairs of deviation judgment results into a single continuous numerical index through the above method, the vehicle speed health status of the system can be quantified and classified, thereby providing an objective basis for subsequent threshold judgment.

[0046] Step 25: If the vehicle speed health index is lower than the preset health threshold, the vehicle speed signal is determined to be in a failed state.

[0047] In this embodiment, the preset health threshold is a constant value determined in advance through bench testing and real vehicle calibration, typically set at a relatively high level (e.g., 0.8). This threshold represents the maximum degree of deviation between vehicle speed information sources that the system can tolerate. When the vehicle speed health index is higher than or equal to this threshold, it indicates that although there may be minor deviations (due to normal errors such as sensor noise), the overall consistency between the sources is good, and the vehicle speed signal is trustworthy. When the vehicle speed health index is lower than this threshold, it indicates that abnormally deviating sensor pairs have occupied a significant proportion, and the overall reliability of the vehicle speed signal no longer meets the system requirements. Furthermore, determining that the vehicle speed signal is in the failed state is the final output of the vehicle speed health diagnosis process, i.e., a binary status flag. Once this flag is set to valid, the system switches from normal mode to failed mode, and the subsequent radar ghosting suppression process is initiated. It is worth noting that this determination result is based on the combined vote of multiple independent sensors, rather than the alarm signal of a single sensor, thus possessing higher reliability and robustness.

[0048] For steps 21 to 25 above, in a specific example, assume that at a certain moment, the three types of vehicle speed information collected by the system, after spatiotemporal alignment, yield the following baseline vehicle speed estimates: wheel speed estimate of 38 km / h (higher due to drive wheel slippage), IMU integral speed of 26.5 km / h, and radar stationary target inversion speed of 26.2 km / h. After the three sets of vehicle speed estimates are fed into the consistency monitoring model, the following processing is performed: The wheel speed pulse signal, IMU acceleration integral value, and radar stationary target Doppler inversion value are interpolated and synchronized to the same time point. The velocity vectors of the IMU and radar are then uniformly projected onto the vehicle's forward direction, resulting in three longitudinal velocity estimates based on the same reference: 38 km / h, 26.5 km / h, and 26.2 km / h. At this point, the three data streams are under the same spatiotemporal reference and can be directly compared numerically.

[0049] The three sources form a total of C_3^2=3 pairs of combinations. The system calculates the deviation between wheel speed and IMU as |38-26.5|=11.5km / h, the deviation between wheel speed and radar as |38-26.2|=11.8km / h, and the deviation between IMU and radar as |26.5-26.2|=0.3km / h.

[0050] The system reads the pre-stored tolerance thresholds for each sensor pair. The tolerance threshold for the wheel speed and IMU pair is 3 km / h; 11.5 km / h far exceeds 3 km / h, therefore this pair is determined to be "disconnected." The tolerance threshold for the wheel speed and radar pair is 3.5 km / h; 11.8 km / h also far exceeds 3.5 km / h, thus it is determined to be disconnected. The tolerance threshold for the IMU and radar pair is 2 km / h; 0.3 km / h does not exceed 2 km / h, therefore it is determined to be consistent. Thus, the system obtains the disconnection determination results for each of the three sensor pairs—two pairs are disconnected, and one pair is consistent.

[0051] Each pair of divergence events is assigned a preset weight. The wheel speed and IMU pair has a weight of 0.9 (a golden combination, severe divergence), the wheel speed and radar pair has a weight of 0.8 (radar inversion accuracy is slightly inferior), and the IMU and radar pair, although identical, has a weight of 0.6. The weighted total divergence score is 0.9×1 + 0.8×1 + 0.6×0 = 1.7. Dividing the weighted total divergence score by the total number of combinations (3) yields the normalized divergence ratio as 1.7 / 3 ≈ 0.5667. The vehicle speed health index I = 1 - 0.5667 = 0.4333.

[0052] The system's internal preset health threshold is 0.8. The currently calculated vehicle speed health index of 0.4333 is significantly lower than 0.8, and based on this, the system determines that the vehicle speed is currently in a faulty state. This status flag takes effect immediately, triggering subsequent ghost feature extraction and graded suppression processes.

[0053] In the example above, without a differentiated tolerance threshold—that is, using a uniform 3 km / h threshold for all pairs—a 0.3 km / h deviation between the IMU and the radar would be within 3 km / h and would not affect the judgment result. However, consider another scenario: on a bumpy road, the IMU's acceleration integral may experience an instantaneous fluctuation of approximately 3.5 km / h due to vibration and noise. In this case, the deviation between the IMU and the radar reaches 3.8 km / h. If a uniform threshold of 3 km / h is used, this pair would be misjudged as diverging, causing an unnecessary drop in the vehicle speed health index and potentially triggering a false alarm. In this solution, the tolerance thresholds for the IMU and radar are independently calibrated to 5 km / h. 3.8 km / h does not exceed this threshold and is still correctly judged as consistent, thus avoiding false triggers.

[0054] As can be seen, through steps 21 to 25 above, the preprocessing operation of spatiotemporal alignment eliminates the incomparability between different sensors caused by different sampling times and coordinate systems, ensuring the accuracy of subsequent deviation calculations and fundamentally guaranteeing the effectiveness of vehicle speed consistency monitoring. Furthermore, by comparing all possible sensor combinations in pairs, rather than selecting only one or a few pairs for comparison, a complete information matrix describing the interrelationships between all vehicle speed information sources is obtained, avoiding missing abnormal signals due to the omission of certain combinations. By setting independent preset tolerance thresholds for different sensor pairs, the objective differences in accuracy characteristics and noise levels of different sensors are fully considered, avoiding misjudgments or omissions that may be caused by uniform thresholds, reflecting the refinement level of engineering calibration. By transforming discrete deviation judgment results into a normalized continuous indicator of vehicle speed health index, the health status of vehicle speed can be quantitatively described, hierarchically managed, and historically tracked, rather than merely remaining at a qualitative level of good / bad. The final determination of the failure state is made by objectively comparing the vehicle speed health index with the preset health threshold, which ensures the uniformity, reproducibility and calibrability of the determination results, and makes the porting and calibration of the solution under different vehicle models and different sensor configurations have a standardized process.

[0055] In an optional embodiment of this application, the method of generating a vehicle speed health index by calculating the overall degree of deviation between each vehicle speed estimate based on all deviation judgment results in step 24 above may further include: Step 31: Record the comparison result where the deviation of each pair of vehicle speed estimates exceeds its corresponding preset tolerance threshold as a deviation event. In this embodiment, a deviation event is a term with a specific technical meaning. It does not refer to a specific event in the physical world, but rather to an anomaly marker recorded internally by the system when the deviation value exceeds a preset tolerance threshold during a pairwise comparison mathematical operation. Each deviation event corresponds to a specific pair of vehicle speed information sources, indicating a non-negligible discrepancy between the speed estimates at the current moment between those sources. This operation transforms continuous numerical values ​​(deviation values) into discrete event records (deviation or no deviation). This event-based processing allows subsequent statistical operations to be performed at the integer level (counting and weighting) without requiring complex mathematical calculations on the deviation values ​​themselves, thereby reducing computational complexity and the computing power requirements of the embedded platform.

[0056] It should be noted that the recording of divergence events is binary, meaning it only records whether a divergence occurred or not, without recording the specific degree of divergence (e.g., whether the deviation is 11.5 km / h or 8 km / h). Although the specific degree of divergence may contain useful information in some scenarios, this design chooses to use whether it exceeds the tolerance threshold as the judgment criterion. This is because the tolerance threshold itself is pre-calibrated based on the normal error range of each sensor pair. That is, exceeding this threshold already indicates that there is a significant divergence between the sources of the pair, while the specific extent to which it exceeds the threshold provides limited incremental information at the statistical level.

[0057] Step 32: For each deviation event, assign a corresponding weight coefficient based on the combination of vehicle speed estimation sources corresponding to the deviation event; In this embodiment, the weighting coefficient is a positive number pre-determined through engineering calibration, typically ranging from 0 to 1, representing the credibility weight or importance of the sensor combination in determining vehicle speed health. Different sensor combinations correspond to different weighting coefficients, and these coefficients are constants determined and stored in the system during the system development phase, rather than being calculated online in real time. The assignment logic of the weighting coefficients is based on the engineering confidence assessment of each sensor pair. Specifically, if two sensors have high measurement accuracy and complementary physical principles, then when they deviate from each other, it indicates a higher severity of the problem, and the evidence of vehicle speed failure is stronger, thus requiring a higher weight. Conversely, if one sensor in a pair has low accuracy or is susceptible to environmental interference (e.g., the motion flow velocity of a visual sensor), then even if they deviate from the other sensor, the deviation may partly stem from the normal error range of the low-accuracy sensor. That is, although the error is still within the tolerance threshold, because the lower limit of the sensor's accuracy is high, the evidence of its deviation is relatively weak, and therefore a lower weight should be assigned.

[0058] Let's illustrate the logic behind setting the weighting coefficients with a specific calibration example. Assume the system has three sources of vehicle speed information: wheel speed sensors (high accuracy, but susceptible to slippage), IMUs (high accuracy, unaffected by slippage, but subject to integral drift), and vision sensors (accuracy affected by lighting and texture, relatively low stability). For the wheel speed and IMU combination, since both are high-precision sources based on completely different physical principles (mechanical rotation and inertial navigation), if they deviate from each other, it's almost certain that at least one source is faulty. Therefore, the weighting coefficient for this combination can be set to a relatively high value of 0.9. For the wheel speed and vision combination, although a deviation between them might also indicate abnormal vehicle speed, the accuracy and stability of the vision sensor itself are lower than that of the IMU, so the evidence of its deviation is relatively weaker. The weighting coefficient can be set to 0.6. For the IMU and vision combination, the weight can be set somewhere in between.

[0059] Step 33: Sum the weight coefficients corresponding to all divergence events to obtain the weighted total divergence score; In this embodiment, the weighted deviation score is a numerical scalar, calculated as follows: Iterate through all sensor pairs. For each pair, if it is determined to be deviating at the current moment (i.e., its deviation exceeds a preset tolerance threshold), the corresponding weight coefficient is added to the total score; if it is not determined to be deviating, it does not contribute any value (i.e., contributes 0). Mathematically, this is expressed as: Weighted Deviation Score = Σ(ω_ij × Indication Function of Deviation), where the indicator function is 1 when a pair deviates, and 0 otherwise.

[0060] The lower limit of the total score is 0 (all sensor pairs are consistent, and no divergence events occur), and the upper limit is the sum of the weight coefficients of all sensor pairs (all sensor pairs diverge). This score directly reflects the weighted scale of the conflicting sensor pairs among all vehicle speed information sources at the current moment. In other words, the higher the score, the more irreconcilable disagreements exist between important sensor combinations.

[0061] Step 34: Divide the weighted total divergence score by the total number of combinations of vehicle speed estimates to obtain the normalized divergence ratio. In this embodiment, the total number of combinations of vehicle speed estimates refers to the total number of pairwise combinations that can be formed when there are n vehicle speed information sources. Its mathematical expression is the number of combinations C_n^2 = n(n-1) / 2. This is a fixed value determined entirely by the number of vehicle speed information sources and is independent of the actual deviation. For example, when there are 3 sources, the total number of combinations is 3; when there are 5 sources, the total number of combinations is 10. Furthermore, the normalized deviation ratio is the quotient obtained by dividing the weighted total deviation score by the total number of combinations. The purpose of this division is to eliminate the influence of different information source numbers on the magnitude of the statistical results, making the vehicle speed health index comparable across different vehicle configurations. Without this normalization, the weighted total deviation score of a vehicle equipped with 5 sensors (total combination of 10) would naturally be much larger than that of a vehicle equipped with 3 sensors (total combination of 3), requiring drastically different failure thresholds for the two vehicle types, increasing the complexity and inconsistency of system calibration.

[0062] However, more precisely, in the weighted case, the weighted total divergence score is not directly equal to the number of divergences, but rather the sum of the weights of each divergence event. The sum of the weight coefficients is related to the number of sensor combinations. The normalized divergence ratio = weighted total divergence score / total number of combinations. This value represents the average weighted divergence score contributed by each sensor, which can also be understood as the divergence density after considering the weights. After normalization, the numerical range of this ratio is compressed to a uniform, comparable scale, allowing vehicles with different configurations to share the same set of health threshold calibration parameters.

[0063] Step 35: Subtract the normalized deviation ratio from the value 1 to obtain the vehicle speed health index.

[0064] In this embodiment, the vehicle speed health index is the final output indicator of this process, with a value ranging from 0 to 1. This index describes the health of vehicle speed information from the perspective of consistency, the opposite of divergence. When the normalized divergence ratio is 0 (all sensor pairs are consistent, with no divergence events), the vehicle speed health index = 1 - 0 = 1, indicating that the vehicle speed information has reached the theoretically healthiest state, i.e., all sources are completely consistent, and the vehicle speed signal is extremely reliable. When the normalized divergence ratio is 1 (all sensor pairs diverge, and the average contribution reaches its maximum), the vehicle speed health index = 1 - 1 = 0, indicating that the vehicle speed information has reached the theoretically unhealthiest state, i.e., there are irreconcilable discrepancies between all sources, and the vehicle speed signal is completely unreliable.

[0065] For steps 31 to 35 above, in a specific example, assume that the three vehicle speed sources collected by the system at the current moment, after spatiotemporal alignment, yield the following results: wheel speed estimate 38 km / h, IMU integral speed 26.5 km / h, and radar stationary target inversion speed 26.2 km / h. The pairwise deviation calculation and threshold comparison have been completed above, and the resulting deviation judgment is: deviation occurs between wheel speed and IMU, deviation occurs between wheel speed and radar, and consistency exists between IMU and radar. A specific example is as follows: The deviation between wheel speed and IMU (11.5 km / h) exceeding its corresponding tolerance threshold (3 km / h) is recorded as one deviation event; the deviation between wheel speed and radar (11.8 km / h) exceeding its corresponding tolerance threshold (3.5 km / h) is recorded as another deviation event; the deviation between IMU and radar (0.3 km / h) does not exceed its tolerance threshold (2 km / h) and is therefore not recorded as a deviation event. Thus, the system records a total of two deviation events. Then, based on the sensor combination corresponding to each deviation event, pre-stored weighting coefficients are read. The wheel speed and IMU combination was assigned a weighting coefficient of 0.9 during the calibration phase, therefore the deviation event between wheel speed and IMU receives a weight of 0.9. The wheel speed and radar combination was assigned a weighting coefficient of 0.8, therefore the deviation event between wheel speed and radar receives a weight of 0.8. No deviation event occurred between IMU and radar, therefore no weight is assigned.

[0066] Then, the weighted coefficients of the two divergence events are added together, resulting in a weighted total divergence score of 0.9 + 0.8 = 1.7. This score reflects the weighted sum of all divergence events among the current vehicle speed information sources. Specifically, the divergence between the wheel speed and IMU (the golden combination) contributes 0.9, and the divergence between the wheel speed and radar (an important combination) contributes 0.8, totaling 1.7. Dividing the weighted total divergence score of 1.7 by the total number of combinations (3 pairs of combinations from three sources) yields a normalized divergence ratio of 1.7 / 3 ≈ 0.5667. This value indicates that each sensor pair contributes approximately 0.5667 to the weighted divergence score on average, or that approximately 56.7% of sensor pairs diverged when weights were considered. Subtracting the normalized divergence ratio of 0.5667 from the value of 1 yields the vehicle speed health index: 1 - 0.5667 = 0.4333. The index indicates that at the current moment, the weighted proportion of all vehicle speed information sources that maintain consistency is approximately 43.33%, meaning that the overall reliability of the vehicle speed signals is more than 40%.

[0067] To further illustrate the role of the weighting coefficients, consider another comparative scenario: Suppose the deviation between the IMU and radar is exactly 2.5 km / h. Although this does not exceed the tolerance threshold of 5 km / h between the two pairs, and therefore is not counted as a deviation event, it will not affect the total deviation score. However, because the weighting coefficient of the vision sensor is set relatively low (assuming the weight of all vision-related sensor pairs is 0.5), its contribution when deviating from the radar is less than its contribution when deviating from the wheel speed and IMU. This allows the system to automatically reduce the severity of deviations from low-reliability sensor combinations when calculating vehicle speed health, thereby preventing the entire vehicle speed signal from being misjudged and rendered invalid due to the anomaly of a single low-quality sensor.

[0068] Through steps 31 to 35, the discretization of deviation events transforms continuous deviation comparison results into countable binary events, enabling subsequent statistical operations to be performed efficiently at the integer level, reducing computational complexity and facilitating real-time implementation on embedded platforms. By assigning differentiated weighting coefficients to different sensor combinations, the differences in measurement accuracy, physical principles, and engineering reliability among various vehicle speed information sources are fully considered. This allows the vehicle speed health index to more realistically and precisely reflect the actual reliability of the vehicle speed signal; that is, high-confidence sensor combinations have a greater impact on the health index when deviations occur, while low-confidence combinations have a smaller impact, avoiding excessive interference from abnormal signals from low-quality sensors in failure determination. Furthermore, through normalization, the weighted total deviation score is divided by the total number of combinations, eliminating the influence of different numbers of vehicle speed information sources on the magnitude of the statistical results. This allows vehicles with different sensor configurations to share the same set of health threshold calibration parameters, improving the engineering adaptability and calibration efficiency of the solution. Furthermore, by mapping the normalized deviation ratio to the vehicle speed health index (1 minus the normalized deviation ratio), the final output index semantically transforms from how severe the deviation is to how high the credibility is. That is, the higher the value, the healthier the vehicle, and the lower the value, the more suspicious the vehicle. This positive measurement method is more intuitive in engineering applications and is naturally connected and semantically consistent with the decision logic of determining failure when the health index is lower than the preset threshold.

[0069] In an optional embodiment of this application, the method for extracting ghost features of each detection point in multiple physical dimensions in step 103 above may further include: Step 41: Obtain the reference speed of the vehicle based on the non-wheel speed source, and calculate the expected Doppler speed of the detection point by combining it with the spatial azimuth angle of the detection point. Then, compare the measured Doppler speed of the detection point with the expected Doppler speed to obtain the first deviation metric of the detection point in the signal dimension. The first deviation metric is used to characterize the degree of abnormality of the Doppler speed of the detection point relative to the expected value of the real physical target. In this application embodiment, non-wheel speed sources refer to vehicle speed information sources other than wheel speed sensors, including but not limited to the integral acceleration value of the inertial measurement unit, the absolute velocity value of the global navigation satellite system, and the inverted velocity value of a stationary target detected by radar. The emphasis on non-wheel speed sources is because the overall triggering condition of this application is the failure of dynamic vehicle speed correction, meaning that the vehicle speed signal centered on wheel speed is no longer reliable. In this situation, if wheel speed continues to be used as the benchmark for calculating the expected Doppler velocity, the calculation result itself is based on an unreliable foundation, and subsequent comparisons will be meaningless. Therefore, when the system determines that the vehicle speed has failed, it must switch to using the vehicle reference speed obtained from non-wheel speed sources. These sources do not depend on the contact state between the wheels and the ground and can still provide relatively reliable reference values ​​when wheel speed slips. Specifically, the system can determine the reference speed based on at least one of the integral acceleration value of the inertial measurement unit, the absolute velocity value of the global navigation satellite system, and the inverted velocity value of a stationary target detected by radar, rather than being limited to a single source.

[0070] Furthermore, the spatial azimuth angle refers to the angle between the radar detection point and the radar normal or the vehicle's direction of travel on the horizontal plane. When detecting targets, millimeter-wave radar, in addition to measuring distance and relative velocity, can also measure the target's azimuth angle through the phase difference of the antenna array. The azimuth angle is a key geometric parameter for calculating the expected Doppler velocity because the radar can only measure the radial velocity component of the target along the radar beam direction, not the target's absolute velocity vector. Additionally, the expected Doppler velocity refers to the radial relative velocity that the detection point should exhibit if it were a stationary real physical target (such as a roadside guardrail, utility pole, or streetlight pole), calculated using the Doppler effect formula in physics, based on the vehicle's current reference speed and the spatial azimuth angle of the detection point. The formula is v_rel_exp = v_ego ×cosθ, where v_ego is the vehicle's reference speed and θ is the azimuth angle of the detection point. The physical meaning of this formula is: the projection component of the velocity of a stationary target relative to a moving vehicle in the vehicle's direction of travel is equal to the projection of the vehicle's velocity in the target's direction. In other words, if the detection point is a real, stationary object, its radar echo should exhibit a radial relative velocity proportional to the cosine of the azimuth angle—targets directly in front (θ=0°) should exhibit the greatest approach velocity, while targets to the side (θ increases) should exhibit a gradually decreasing approach velocity.

[0071] Furthermore, the measured Doppler velocity is the radial relative velocity of the detection point directly measured by the radar. Millimeter-wave radar utilizes the Doppler effect of electromagnetic waves; that is, when an electromagnetic wave illuminates a moving target, the frequency of the reflected wave shifts relative to the emitted wave. The magnitude of this shift is proportional to the radial velocity of the target relative to the radar. By measuring this frequency shift, the radar can directly calculate the radial relative velocity of the target. This measurement is entirely derived from physical measurements and does not rely on any vehicle speed estimation or model assumptions. The first deviation metric is a quantitative result obtained by comparing the measured Doppler velocity with the expected Doppler velocity. It characterizes the degree of anomaly of the Doppler velocity at the detection point relative to the expected value of a real physical target. The specific calculation method is as follows: calculate the difference between the theoretical relative velocity and the measured relative velocity, and then divide this difference by a preset standard deviation factor to obtain the Doppler anomaly factor for the detection point. The larger the factor value, the further the measured Doppler velocity deviates from the theoretical expectation, and the lower the credibility of the detection point as a real physical target in the signal dimension. The deviation metric is a continuous value, rather than a discrete yes / no judgment, which allows subsequent confidence fusion calculations to integrate information from various dimensions at a fine scale.

[0072] The Doppler velocity is determined by both the vehicle's speed and the target's azimuth. If the measured Doppler velocity at a certain detection point deviates significantly from this physical law, then that detection point is highly unlikely to be a real physical target, but rather a ghost image generated by non-physical sources such as multipath reflection or electromagnetic interference. This detection method based on physical laws does not require a large amount of training data, nor does it rely on prior environmental knowledge, thus possessing universality and interpretability.

[0073] Step 42: Obtain the spatial position of the detection point in the vehicle coordinate system, and determine whether the spatial position falls within the blacklist area of ​​a pre-calibrated or real-time constructed strongly reflective static object. If so, combine the target detection results of the visual sensor in the corresponding spatial area for comprehensive judgment to obtain the second deviation measure of the detection point in the spatial dimension. The second deviation measure is used to characterize the degree of abnormality of the spatial position of the detection point relative to the area where the real physical target may exist. In this embodiment, the spatial position in the vehicle coordinate system refers to the coordinate values ​​of the detection point distance and azimuth information obtained by radar measurement, transformed into a Cartesian coordinate system with the vehicle as the origin after coordinate transformation. This coordinate transformation process converts polar coordinates (distance, angle) into planar rectangular coordinates (x, y), enabling the position of the detection point to be spatially matched and compared with the detection results of map data or visual sensors in the same coordinate system. Furthermore, the pre-calibrated or real-time constructed blacklist area of ​​strongly reflective static objects refers to a group of areas in physical space marked as having a high probability of generating ghosting. These areas include, but are not limited to, the locations of known strong radar reflectors such as metal guardrails, sound barriers, tunnel walls, bridge piers, gantry cranes, road manhole covers, speed bumps, and railway tracks. These objects share two common characteristics: first, they have a strong radar cross-section, capable of generating strong radar echoes; second, they are stationary, repeatedly appearing in the radar field of view, and easily generating ghosting under multipath reflection conditions. Pre-calibration refers to recording the locations of these highly reflective objects into the system database in advance using high-precision map data or on-site surveying during the system development phase; real-time construction refers to identifying and recording highly reflective points in the environment online and dynamically updating the blacklist area during vehicle operation using synchronous positioning and mapping technology.

[0074] Furthermore, the target detection result of the visual sensor in the corresponding spatial region refers to the output result of the visual target detection algorithm within that image region after projecting the spatial position of the detection point onto the image plane of the visual sensor. If the visual sensor detects a valid target (such as a vehicle, pedestrian, or bicycle) in the corresponding region, it indicates that a real object does exist at that spatial position. In this case, even if the point falls into the blacklist region, it cannot be simply determined as a ghost image, because there may indeed be an object at that position, and that object happens to be near the blacklist region (e.g., a real car parked next to a bridge pier). Conversely, if the visual sensor does not detect any valid target in the corresponding region, and the position falls into the blacklist region, it indicates that the detection point is highly likely to be a false echo generated by a strong reflector. Further, the second deviation metric is a quantitative index obtained by combining the spatial position of the detection point and the visual detection result, used to characterize the degree of anomaly of the spatial position of the detection point relative to the area where a real physical target might exist. The deviation metric is calculated as follows: when the detection point falls within the blacklist area and the visual system does not detect the target, the deviation metric is high, indicating that the point is highly suspicious in the spatial dimension; when the detection point falls within the blacklist area but the visual system detects the target, the deviation metric is medium, indicating that although the location of the point is suspicious, the possibility of the presence of a real object cannot be ruled out; when the detection point does not fall within the blacklist area, the deviation metric is low, indicating that the point has no obvious abnormalities in the spatial dimension.

[0075] It should be noted that radar itself cannot distinguish whether an echo originates from a real target or from multipath reflection; that is, the two may be equivalent in terms of distance and angle information. However, by spatially matching the location of the detection point with the known locations of strong reflectors and combining this with the independent detection results from the visual sensor, the system can provide additional evidence for assessing the authenticity of the detection point from a spatial dimension.

[0076] Step 43: Obtain the position records of the detection point in multiple consecutive historical frames, calculate the motion acceleration or change in motion direction of the detection point based on the position records, and compare the motion acceleration or change in motion direction with the physically achievable motion range to obtain the third deviation metric of the detection point in the time dimension. The third deviation metric is used to characterize the degree of abnormality of the motion trajectory of the detection point relative to the motion law of the real physical target.

[0077] In this embodiment, the position record in multiple consecutive historical frames refers to the position history of the detection point in several consecutive processing cycles prior to the current frame. In the radar perception system, data is processed in frames, with each frame corresponding to a radar scan cycle (typically 50 milliseconds or 100 milliseconds). The system maintains a short-term historical trajectory buffer for each persistent detection point, recording the spatial position of the detection point in the past several frames (e.g., the previous 5 or 10 frames). These historical records constitute the data basis for evaluating the motion law of the detection point. Furthermore, motion acceleration or change in motion direction are two key kinematic parameters calculated from the historical position records. Motion acceleration refers to the rate of change of the detection point's velocity over time, which can be obtained by performing a second-order difference calculation on the position sequence; change in motion direction refers to the angle of change of the detection point's motion direction between adjacent frames. These two parameters together describe the smoothness or abruptness of the detection point's motion, i.e., the motion of a real physical target usually follows Newton's laws of kinematics, its acceleration is limited by the upper limit of the physical driving force, and its change in direction is limited by the capability of the steering mechanism.

[0078] Furthermore, the physically achievable range of motion refers to the reasonable upper limit of acceleration and directional changes of a real physical target during motion, set according to the laws of physics and common sense about the vehicle environment. For example, the maximum acceleration of a real vehicle is limited by its engine power and tire adhesion, typically not exceeding a certain number of m / s²; similarly, the acceleration and directional changes of a real pedestrian are limited by human movement capabilities. If the calculated acceleration or directional change of a detection point exceeds the physically achievable range of motion—for example, if an impossible, huge speed jump or directional change occurs between two adjacent frames—it indicates that the motion of that detection point does not conform to the motion laws of any real physical target, significantly increasing its suspicion as a ghost. Further, the third deviation metric is a quantitative indicator obtained by comparing the motion acceleration or directional change of a detection point with the physically achievable range of motion, used to characterize the degree of abnormality of the detection point's trajectory relative to the motion laws of real physical targets. The specific calculation method is as follows: calculate the second-order difference of the position of the detection point in the current frame relative to the previous two frames (i.e., the discrete approximation of acceleration), and compare the difference with the upper limit of physically feasible acceleration; the more the difference exceeds the upper limit, the higher the value of the third deviation metric, indicating that the motion continuity of the detection point in the time dimension is worse and it is less likely to be a real physical target.

[0079] It's important to note that the motion of real physical targets is continuous and smooth; their velocity and direction do not change abruptly in a very short time because any real object has mass, and acceleration requires force and time. Ghosting, however, is not subject to this constraint. The position of multipath reflections can randomly change between adjacent frames because the generation and disappearance of ghosting depend on changes in the geometric relationship of the electromagnetic wave reflection path, rather than physical motion. By detecting these non-physical motion abrupt changes, the system can effectively identify ghosting in the time dimension.

[0080] As can be seen, through steps 41 to 43 above, the three dimensions are based on completely different physical principles and data sources, and their respective failure modes and interference sources are independent of each other. That is, the signal dimension may be affected by electromagnetic interference, the spatial dimension may be affected by map data accuracy, and the time dimension may be affected by tracking algorithms. This independence ensures that when the judgment of one dimension is biased due to specific interference, the other dimensions can still provide independent evidence, thereby improving the robustness of the comprehensive judgment. Therefore, under the special condition of vehicle speed failure, the traditional ghost filtering method based on the absolute value of vehicle speed completely fails. However, the feature extraction method of the three dimensions proposed in this solution does not directly rely on the absolute value of vehicle speed. That is, the signal dimension uses a reference speed that is not from wheel speed (which can still be obtained under failure state through multi-source redundancy), the spatial dimension relies on map and visual information, and the time dimension relies on kinematic laws. Therefore, it can continue to work effectively under the condition of vehicle speed failure, filling the gap in the existing technology.

[0081] In an optional embodiment of this application, the method of fusing and calculating the deviation of the corresponding detection point in each physical dimension based on the ghost feature quantity of each detection point to obtain the ghost confidence of the corresponding detection point in step 104 above may further include: Step 51: Using a hierarchical decision tree structure, determine the initial confidence interval corresponding to the detection point based on whether the detection point has one or more combinations of the first deviation measure, the second deviation measure, and the third deviation measure. In this embodiment, the hierarchical decision tree structure is a tree-like decision model composed of multiple decision nodes and branch paths. Starting from the root node, each layer divides the detection points into different branches according to a decision condition, and finally outputs the corresponding decision results at the leaf nodes. In this scheme, the decision condition of the decision tree examines whether the detection point is marked as abnormal in each physical dimension (i.e., whether each deviation metric exceeds its corresponding preset abnormality threshold). It starts from the first layer to determine whether the first deviation metric is present, then the second layer to determine whether the second deviation metric is present, and then the third layer to determine whether the third deviation metric is present. Finally, it determines different output results based on eight possible abnormal combinations (normal in all three dimensions, abnormal only in signal, abnormal only in space, abnormal only in time, abnormal in both signal and space, abnormal in both signal and time, abnormal in both space and time, and abnormal in all three dimensions).

[0082] Furthermore, the initial confidence interval is a range of confidence values ​​output by the decision tree, rather than a single value. For example, when a detection point is marked as anomaly in all three dimensions, the initial confidence interval output by the decision tree is 0.9 to 1.0, indicating a very high probability that the detection point belongs to the ghost category; when a detection point is marked as anomaly only in the signal dimension and normal in the spatial and temporal dimensions, the initial confidence interval is 0.5 to 0.7, indicating a moderate probability that the detection point belongs to the ghost category; when a detection point is normal in all three dimensions, the initial confidence interval is 0 to 0.3, indicating a very low probability that the detection point belongs to the ghost category. The reason for outputting an interval rather than a single value is that there are still differences in the degree of anomalousness within the same combination of anomalies. That is, even if there is only anomaly in the signal dimension, if the anomaly factor is 6.5, the probability of ghosting is high, and if the anomaly factor is only 3.2, the probability of ghosting is low. In other words, the interval reserves space for fine-tuning in the second stage.

[0083] As can be seen, using a hierarchical decision tree to quickly classify detection points according to the discrete feature of anomaly dimension combination makes the judgment logic clear and highly interpretable. Each layer of decision corresponds to the inspection result of a physical dimension, and the final output initial confidence interval directly reflects the anomaly pattern of the detection point in all physical dimensions. This pattern recognition method can effectively distinguish between pure ghosting (multidimensional anomaly) and real targets that are mistakenly damaged due to vehicle speed failure (single-dimensional anomaly).

[0084] Step 52: Continuously adjust the initial confidence level within the range of each deviation metric based on its specific value, and output the ghost confidence level of the detection point.

[0085] In this embodiment, the specific values ​​of each deviation metric refer to the specific values ​​of the first deviation metric (Doppler anomaly factor in the signal dimension), the second deviation metric (positional suspicion factor in the spatial dimension), and the third deviation metric (motion anomaly factor in the time dimension) calculated above. These values ​​are not merely discrete labels used to mark whether something is abnormal, but rather contain detailed information about the degree of anomaly in each dimension. For example, in the calculation of Doppler anomaly, although both anomaly factors 6.5 and 4.0 exceed the anomaly threshold of 3.0 and are marked as signal dimension anomalies, the severity of the anomaly reflected by 6.5 is much higher than that reflected by 4.0. This difference should be reflected in the final confidence level. Furthermore, continuous adjustment refers to performing smooth interpolation or offset operations based on the specific values ​​of each deviation metric within the initial confidence level interval, so that the output confidence level changes continuously within the interval, rather than simply taking a few discrete values ​​within the interval. The specific implementation can employ fuzzy logic: define membership functions for each deviation metric, mapping specific abnormal values ​​to corresponding membership values; then, based on a pre-defined fuzzy rule base, continuously adjust within the initial confidence interval under the combined effect of the membership values ​​of each dimension. For example, given that the initial confidence interval for the combination of only signal dimension anomalies is 0.5 to 0.7, if the first deviation metric of the signal dimension is high (e.g., 5.96), adjust towards the higher end within this interval, outputting a confidence value close to 0.7; if the deviation metric is only slightly higher than the threshold (e.g., 3.2), adjust towards the lower end, outputting a confidence value close to 0.5.

[0086] It is evident that if only discrete confidence levels (such as high, medium, and low) are output from the decision tree, even if there are significant differences in the degree of anomaly between different detection points, they will be assigned the same confidence level as long as they fall into the same combination pattern, leading to insufficient resolution in the judgment. By continuously adjusting the system within the initial confidence interval using specific values ​​of the deviation metric, the system can further distinguish between mild and severe anomalies within the same combination pattern, thereby achieving a more refined ghosting confidence assessment. In other words, the decision tree structure makes the system's judgment logic highly transparent and traceable; engineers can clearly understand why a detection point that is anomalous in all three dimensions will obtain a confidence interval of 0.9 or higher, rather than relying on a "mysterious number" output by an unexplainable black-box model. At the same time, the continuous adjustment mechanism ensures that the final output confidence value reflects the actual degree of anomaly of the detection point in each dimension, avoiding the accuracy loss caused by discrete grading.

[0087] For steps 51 and 52 above, in a specific example, we will take the fusion calculation process of two detection points in the previous example: point A (guardrail ghost) and point B (real vehicle in front) as an example for comparison and explanation.

[0088] Point A yielded the following results in the feature extraction across the three dimensions: the first deviation metric (Doppler anomaly) Af≈6.53, exceeding the anomaly threshold of 3.0, therefore the signal dimension was marked as anomaly; the second deviation metric indicated that its spatial location fell within the guardrail blacklist area and no target was detected visually, therefore the spatial dimension was marked as anomaly; the third deviation metric indicated that its historical trajectory exhibited random fluctuations, therefore the temporal dimension was marked as anomaly. All three anomaly markings were valid, resulting in a combination of all three dimensions being anomaly.

[0089] After the combination is fed into the hierarchical decision tree, the decision tree outputs the corresponding initial confidence interval according to the preset rules. According to the system's preset decision rules, the initial confidence interval for the combination with anomalies in all three dimensions is 0.9 to 1.0. The physical meaning of this interval is: the detection point is judged to have seriously deviated from the expected attributes of the real physical target in three completely independent physical dimensions: the signal domain, the spatial domain, and the time domain. That is, three independent pieces of evidence point to the same conclusion from three different angles. Therefore, the basic probability that the detection point belongs to ghosting is extremely high. Even considering the slight difference in the degree of anomaly, its confidence should fall within the high confidence interval of 0.9 or above. In addition, based on the initial confidence interval of 0.9 to 1.0 output in the first stage, the system enters the second stage—continuous adjustment. The system reads the specific values ​​of the three deviation measures of point A: Af≈6.53. The second deviation measure is also very high (the spatial location falls into the blacklist and there is no visual target), and the third deviation measure is also very high (the trajectory is obviously abnormal). Since the anomaly of point A in each dimension far exceeds its respective anomaly threshold, the system adjusts upward within the range of 0.9 to 1.0, and the final output ghost confidence score is approximately 0.95. This value accurately reflects the extreme anomaly achieved by point A within the high-risk combination of anomalies in all three dimensions, meaning that the deviations in all dimensions are severe, and its ghost confidence score is close to the upper limit.

[0090] The fusion calculation for point B presents a distinctly different trajectory. Point B's first deviation metric, Af ≈ 5.96, in the signal dimension exceeds the anomaly threshold of 3.0, therefore the signal dimension is flagged as anomaly. However, in the spatial dimension, point B is located precisely in the center of the lane, and a clear vehicle outline is visually detected; therefore, the spatial dimension is not flagged as anomaly. In the temporal dimension, point B's historical trajectory is continuous and smooth; therefore, the temporal dimension is also not flagged as anomaly. The combination of anomaly flags across the three dimensions results in an anomaly only in the signal dimension.

[0091] Furthermore, after feeding this combination into a hierarchical decision tree, the decision tree outputs the corresponding initial confidence interval. According to the system's preset decision rules, the initial confidence interval for the combination with only signal dimension anomalies is 0.5 to 0.7. The physical meaning of this interval is fundamentally different from that of point A: although point B also exhibits Doppler anomalies in the signal dimension, since this anomaly only appears in the signal domain, and both the spatial and temporal domains show that the detection point conforms to the attributes of a real physical target, this detection point is more likely to be a real target misjudged due to the symptom of the special condition of vehicle speed failure, rather than being a ghost itself. The interval of 0.5 to 0.7 reflects the system's cautious attitude towards such single-dimensional anomaly targets, i.e., moderate confidence, not easily judged as a ghost. Furthermore, based on the initial confidence interval of 0.5 to 0.7 output in the first stage, the system enters the second stage—continuous adjustment. The system reads the specific values ​​of each deviation metric for point B: Although the deviation metric Af≈5.96 in the signal dimension also exceeds the threshold, the system also considers that the spatial and temporal dimensions of point B are normal (i.e., the second and third deviation metrics are very low). Therefore, it adjusts downward within this range, and the final output ghost confidence score is approximately 0.65. This value is in the lower-middle range of 0.5 to 0.7, reflecting the system's judgment on point B: somewhat suspicious (abnormal signal dimension), but due to the normal performance in the spatial and temporal dimensions, the possibility of ghosting is not high.

[0092] The above comparison clearly shows that although both point A (the ghost) and point B (the real target) exhibit abnormalities in the signal dimension, their performance in the other two dimensions is drastically different. The hierarchical decision tree assigns them to completely different initial confidence intervals: the former falls into the 0.9 to 1.0 interval, while the latter falls into the 0.5 to 0.7 interval. This results in a fundamental difference in the final output confidence values. This demonstrates that the fusion method employed in this application, which first grades abnormal combinations and then fine-tunes them according to the degree of abnormality, can effectively distinguish between pure ghosts and implicated real targets in the complex condition of vehicle speed failure. Its judgment is not based on the magnitude of anomalies in a single dimension, but rather on the complete evidentiary pattern constituted by combinations of anomalies across multiple dimensions.

[0093] In an optional embodiment of this application, the method of performing graded suppression operation on the detection point based on the ghost confidence level in step 105 above may further include: Step 61: If the ghost confidence of the detection point is higher than or equal to the first preset threshold, perform first-level suppression on the detection point. First-level suppression means directly removing the detection point from the radar point cloud. In this embodiment, the first preset threshold is the highest value among the pre-calibrated confidence thresholds, typically set to around 0.8 in a typical implementation. The physical meaning of this threshold is that when the ghost confidence of a detection point reaches or exceeds this value, the system has a sufficiently high degree of certainty that the detection point is a ghost, allowing for the most thorough handling measures. Level 1 suppression is the strongest suppression level in this application, specifically removing the detection point directly from the radar point cloud. This means that the detection point is not only ignored in the fusion processing of the current frame but also completely excluded from subsequent target clustering, target tracking, and functional applications. Essentially, the detection point is completely filtered out in the early stages of the perception processing pipeline, making it completely undetectable by subsequent control modules.

[0094] It is evident that for such unmistakably identified ghost images, merely downgrading their use or delaying confirmation instead of completely removing them would waste computational resources of subsequent tracking and fusion modules. Furthermore, the prediction mechanism of the tracking algorithm might propagate the ghost image into later frames, causing it to continuously pollute the perception results. Direct removal is the cleanest and most efficient approach, fundamentally eliminating the impact of high-confidence ghost images on any subsequent functional modules. Simultaneously, since this operation only applies to detection points with extremely high confidence levels (e.g., ≥0.8), the risk of mistakenly deleting real targets is extremely low. This is because such a high confidence level only occurs when all three physical dimensions exhibit abnormalities, and such multi-dimensional anomaly patterns are rarely seen in real targets.

[0095] Step 62: If the ghost confidence of the detection point is lower than the first preset threshold but higher than or equal to the second preset threshold, perform secondary suppression on the detection point. The secondary suppression refers to adding a limiting label to the detection point to restrict the detection point from being used for security-related functions. In this embodiment, the second preset threshold is a confidence threshold preset by the system that is less than the first threshold, and is typically set to around 0.5 in a typical implementation of this application. Detection points with a confidence level between the second and first thresholds have a ghosting confidence level that is moderately high, meaning that there is some certainty that the detection point is a ghost, but not enough to completely delete it.

[0096] Furthermore, the limiting label in this application embodiment is a metadata marker attached to the detection point data, used to indicate the reliability level and usability limitations of the detection point. This label, along with physical measurements such as the location and speed of the detection point, is transmitted to the downstream fusion and control module, enabling the downstream module to make differentiated usage decisions based on the label. Restricting the detection point to only be used for safety-related functions is the core operation of secondary suppression. In the driving function system of mass-produced vehicles, different functions have significantly different requirements for the reliability of perception data. Taking Adaptive Cruise Control (ACC) as an example, this function maintains a safe distance from the vehicle in front by controlling the vehicle's acceleration and braking. If ACC receives a ghost target (displaying an obstacle ahead when it does not actually exist), the vehicle will perform unnecessary deceleration or even emergency braking, severely affecting driving comfort. However, it has extremely high requirements for the accuracy and stability of perception data; false detections will directly lead to abnormal acceleration and deceleration, severely affecting driving comfort. Automatic Emergency Braking (AEB), on the other hand, has the opposite requirements for perception data. This function only triggers emergency braking to avoid or mitigate a collision when the risk of collision is extremely high. It requires extremely high recall of perception data (not missing any real targets), but has a relatively high tolerance for accuracy (not minding detecting a few more suspicious targets). This is because even if the system detects several more suspected targets, as long as AEB is not triggered (which requires a very short collision time), it will not cause any adverse effects. Based on the above differences in functional characteristics, the specific meaning of the restriction on its use in safety-related functions is: detection points marked as secondary suppression can participate in the decision-making logic of safety redundancy functions such as AEB (as backup warning signals), but are not allowed to participate in the control logic of comfort functions such as Adaptive Cruise Control (ACC) and lane keeping assist.

[0097] Step 63: If the ghost confidence of the detection point is lower than the second preset threshold but higher than or equal to the third preset threshold, perform three-level suppression on the detection point. The three-level suppression means that the detection point is temporarily stored in the delay queue and is only allowed to be established as a target after it has appeared continuously for a preset number of times in multiple consecutive processing cycles. In this embodiment, the third preset threshold is the lowest of the three confidence thresholds preset by the system, typically set to around 0.3 in a typical implementation. Detection points with confidence levels between the third and second thresholds have a low to medium ghost confidence level, meaning there is significant uncertainty in determining whether the detection point is a ghost; there is neither sufficient evidence to determine it as a ghost nor sufficient evidence to confirm it as a real target. The delay queue is a temporary data buffer in system memory used to temporarily store detection points that are currently deemed suspicious but not immediately usable. Detection points entering the delay queue are not immediately output to the fusion module, nor are they established as stable tracking targets; instead, they are temporarily isolated for observation, awaiting the accumulation of more information in subsequent frames. The core judgment condition for level three suppression is that a detection point is only allowed to be established as a target after it has appeared a preset number of times in multiple consecutive processing cycles. This means that a detection point must meet two conditions to be released as a normal target. The first condition is persistence: the detection point must appear repeatedly in multiple consecutive radar processing cycles (frames), rather than appearing only occasionally and then disappearing. Real physical targets have temporal persistence, while ghost images often appear and disappear randomly. The second condition is stability: the position and velocity of the detection point must remain relatively stable during its continuous appearance, rather than undergoing drastic jumps between adjacent frames. Only when both conditions are met does the system consider the detection point to have passed a sufficiently long observation period, possessing sufficient confidence, and can release it from the delay queue to participate in normal target establishment and subsequent fusion processing. The preset number of occurrences is usually determined by system calibration; a typical value is two or more occurrences in three consecutive frames. This parameter needs to strike a balance between detection delay and confidence. Too few occurrences will not effectively filter out occasional ghost images, while too many occurrences will delay the establishment time of real targets, affecting the system's response speed.

[0098] Step 64: If the ghost confidence of the detection point is lower than the third preset threshold, no suppression operation is performed on the detection point.

[0099] As can be seen, in this embodiment of the application, by replacing the traditional binary filtering with hierarchical suppression, the limitation of the binary thinking of either deleting or keeping is broken. By introducing downgraded use and delayed confirmation as intermediate disposal options, the ghosting suppression strategy has evolved from a black-and-white approach to a fine-grained hierarchical approach, which can adapt to the differentiated processing needs under different confidence levels.

[0100] The present application will be further explained below with reference to specific embodiments, which provide a method for suppressing low-speed radar ghosting when vehicle speed dynamic correction fails, such as... Figure 2 As shown, the steps of this method include: Step 201: Monitor and acquire multi-source vehicle speed information in real time; Specifically, at least two of the following data are acquired via the vehicle bus: wheel speed pulse signals provided by wheel speed sensors, acceleration and angular velocity provided by inertial measurement unit (IMU), absolute speed provided by Global Navigation Satellite System (GNSS), speed of stationary targets detected by radar, and speed of moving flow provided by visual sensors.

[0101] Step 202: Construct a multi-source vehicle speed consistency monitoring model; This involves spatiotemporal alignment of the acquired multi-source vehicle speed information and calculation of the pairwise deviations. A vehicle speed health index I is defined as follows:

[0102] in: Vehicle speed estimates from different sources; This is the deviation threshold for the corresponding sensor pair; These are the weighting coefficients; For indicator functions; This is the number of sensor logs used for normalization.

[0103] When I is lower than the preset threshold, the vehicle speed dynamic correction is determined to be in a failure state.

[0104] Step 203, Radar ghost feature extraction under vehicle speed failure state; Once the vehicle speed determination fails, a targeted ghost feature extraction module is activated to extract the following features from the original radar point cloud: Feature F1: Doppler velocity anomaly Calculate the relative velocity of the probe point Estimated speed of this vehicle Theoretical relationship between them:

[0105] in, Let be the azimuth angle of the detection point. Define the anomaly factor:

[0106] Among them, when When the value exceeds a set threshold, the point is marked as a "Doppler anomaly".

[0107] Feature F2: Spatial geometric rationality The radar point cloud is projected onto the vehicle coordinate system and combined with high-precision maps or real-time mapping to construct a blacklist of reflective areas. These areas include: vertical reflective surfaces such as guardrails and sound barriers; road surface metal structures (manhole covers, speed bumps, railway tracks); and strong reflectors such as tunnel walls and bridge piers. When a radar point falls into a blacklisted area and the visual sensor does not detect a valid target in that area, the point is marked as a spatially suspicious point.

[0108] Feature F3: Temporal continuity anomaly For each detection point, a short-term historical trajectory is established, and its kinematic rationality is calculated:

[0109] Among them, when When the value exceeds the set threshold, it indicates that the target motion has a physically impossible sudden change in acceleration or direction, and the point is marked as a trajectory anomaly point.

[0110] Step 204: Construct a hierarchical ghost confidence evaluation network; Based on the features extracted above, a lightweight ghost confidence evaluation network is constructed. This network adopts a hierarchical decision tree + fuzzy logic structure, and the specific decision rules are shown in Table 1. Table 1

[0111] Step 205: Implement a graded suppression strategy; Specifically, based on the ghost confidence scores output above, hierarchical suppression is performed: Level 1 suppression (direct filtering): Points with a confidence level ≥ 0.8 are directly removed from the point cloud and do not participate in any subsequent fusion; Secondary suppression (downgrade processing): Points with a confidence level of 0.5 ≤ confidence level < 0.8 are marked as suspicious targets and are only used for high-safety-level functions such as collision warning, not for comfort functions (such as ACC following). Level 3 Suppression (Delayed Confirmation): For points with a confidence level of 0.3 ≤ confidence level < 0.5, a delay of N frames is applied, and a new target is only allowed after it occurs M times consecutively. Normal processing: Points with a confidence level < 0.3 are included in the fusion process normally.

[0112] Step 206: Output a list of suppressed clean targets; The processed target list is sent to the planning and control module for subsequent decision-making and control.

[0113] Steps 201 to 206 above will be explained with a specific example. The current scenario is as follows: a vehicle is following another vehicle on a slippery road. The vehicle in front accelerates suddenly, and the vehicle in front accelerates suddenly as well. At this time, the drive wheels slip slightly, and the vehicle speed output by the wheel speed sensor jumps from 25 km / h to 35 km / h in a short period of time (the actual speed of the vehicle is about 27 km / h). It can be seen that the dynamic speed correction fails.

[0114] Through this specific embodiment, vehicle speed health diagnosis is performed first: wheel speed signal: 35km / h (abnormally high); IMU integrated speed: 26.8km / h; speed derived from radar stationary target: 26.5km / h; calculate the I index: the pairwise deviations all exceed the threshold, I=0.23<T_VH=0.8, and it is determined that the vehicle enters a vehicle speed dynamic correction failure state.

[0115] Then, ghost feature extraction is performed, and at this time, the radar detects two suspicious points: Point A: appears at the right guardrail position, with a relative speed of -2m / s (close to stationary) and an azimuth angle of 30°; F1 calculation: the theoretical relative speed should be 27km / h·cos30°≈6.5m / s, the actual value is -2m / s, the abnormality degree A_f=8.5>3, which is marked as Doppler abnormality; F2: Point A falls into the guardrail area, no target is detected by vision, which is marked as spatially suspicious; F3: Point A is a newly appeared point with no historical trajectory; Point B: appears directly ahead, with a relative speed of -2.5m / s, an azimuth angle of 5°, and a distance of 20m; F1 calculation: the theoretical relative speed is ≈7.5m / s, the actual value is -2.5m / s, the abnormality degree A_f=10, which is marked as Doppler abnormality; F2: Point B is within the lane and does not belong to the blacklist area; F3: Point B has a target at the same position in the previous frame, but the speed is abnormal.

[0116] Hierarchical suppression is performed based on the above extraction results: Point A: F1 high abnormality + F2 suspicious → ghost confidence 0.95 → primary suppression (direct filtering); Point B: F1 high abnormality, no F2, with F3 → ghost confidence 0.65 → secondary suppression (downgraded to a tertiary target, only used for AEB, not for ACC).

[0117] Finally, the output target list does not include point A, and point B is output marked with low confidence. The ACC module ignores point B, the vehicle follows the preceding vehicle normally, and no false braking occurs.

[0118] Corresponding to the above Figure 1 , the embodiment of the present application provides a vehicle control apparatus, as shown in Figure 3 the apparatus includes: An acquisition module 302, configured to acquire multi-source vehicle speed information of the vehicle; wherein the multi-source vehicle speed information includes speed information collected by a plurality of different types of devices in the vehicle; The first processing module 304 is used to determine whether the vehicle speed is in a failure state based on the degree of deviation between the speed information in the multi-source vehicle speed information. The failure state means that the vehicle speed cannot be reliably estimated through dynamic correction. The second processing module 306 is used to acquire each detection point detected by the radar in the vehicle when the vehicle speed is determined to be in a failure state, and extract the ghost feature quantity of each detection point in multiple physical dimensions. The ghost feature quantity is used to characterize the degree to which the detection point deviates from the real physical target in the corresponding physical dimension. The third processing module 308 is used to perform fusion calculation on the deviation of the corresponding detection point in each physical dimension based on the ghost feature quantity of each detection point, and obtain the ghost confidence of the corresponding detection point. The ghost confidence is used to characterize the probability that the detection point belongs to a ghost. The fourth processing module 310 is used to perform graded suppression operations on the detection points according to the ghost confidence level, and output a list of suppressed detection points for subsequent vehicle control.

[0119] In an optional embodiment of this application, the first processing module includes: a first processing unit, configured to perform spatiotemporal alignment processing on the speed information in the multi-source vehicle speed information to obtain multiple vehicle speed estimates with the same reference; a second processing unit, configured to pair any two vehicle speed estimates from different sources and calculate the deviation value between each pair of vehicle speed estimates; a third processing unit, configured to compare each pair of deviation values ​​with their respective preset tolerance thresholds to obtain a deviation judgment result between each pair of vehicle speed estimates; a fourth processing unit, configured to calculate the overall degree of deviation between each vehicle speed estimate based on all deviation judgment results and generate a vehicle speed health index; and a fifth processing unit, configured to determine that the vehicle speed signal is in a failed state when the vehicle speed health index is lower than a preset health threshold.

[0120] In an optional embodiment of this application, the fourth processing unit in this application is used to perform the following steps: Recording the comparison result where the deviation value of each pair of vehicle speed estimates exceeds its corresponding preset tolerance threshold as a deviation event; assigning a corresponding weight coefficient to each deviation event based on the combination of vehicle speed estimates corresponding to the deviation event; summing the weight coefficients corresponding to all deviation events to obtain a weighted total deviation score; dividing the weighted total deviation score by the total number of combinations of vehicle speed estimates to obtain a normalized deviation ratio; and subtracting the normalized deviation ratio from the value 1 to obtain the vehicle speed health index.

[0121] In an optional embodiment of this application, the second processing module includes: The sixth processing unit is used to obtain the reference speed of the vehicle based on the non-wheel speed source, and calculate the expected Doppler speed of the detection point by combining the spatial azimuth angle of the detection point. Then, it compares the measured Doppler speed of the detection point with the expected Doppler speed to obtain the first deviation metric of the detection point in the signal dimension. The first deviation metric is used to characterize the degree of abnormality of the Doppler speed of the detection point relative to the expected value of the real physical target. The seventh processing unit is used to obtain the spatial position of the detection point in the vehicle coordinate system, determine whether the spatial position falls within the blacklist area of ​​a pre-calibrated or real-time constructed highly reflective static object, and if so, combine the target detection results of the visual sensor in the corresponding spatial area to make a comprehensive judgment and obtain the second deviation measure of the detection point in the spatial dimension. The second deviation measure is used to characterize the degree of abnormality of the spatial position of the detection point relative to the area where the real physical target may exist. The eighth processing unit is used to acquire the position records of the detection point in multiple consecutive historical frames, calculate the motion acceleration or change in motion direction of the detection point based on the position records, and compare the motion acceleration or change in motion direction with the physically achievable motion range to obtain the third deviation metric of the detection point in the time dimension. The third deviation metric is used to characterize the degree of abnormality of the motion trajectory of the detection point relative to the motion law of the real physical target.

[0122] In an optional embodiment of this application, the third processing module in this application includes: a determining unit, used to determine the initial confidence interval corresponding to the detection point by adopting a hierarchical decision tree structure based on whether the detection point has one or more combinations of a first deviation metric, a second deviation metric, and a third deviation metric; and a ninth processing unit, used to continuously adjust within the initial confidence interval according to the specific values ​​of each deviation metric, and output the ghost confidence of the detection point.

[0123] In an optional embodiment of this application, the fourth processing module includes: a first execution unit, configured to perform first-level suppression on a detection point when the ghost confidence of the detection point is higher than or equal to a first preset threshold, wherein first-level suppression refers to directly removing the detection point from the radar point cloud; a second execution unit, configured to perform second-level suppression on a detection point when the ghost confidence of the detection point is lower than the first preset threshold but higher than or equal to a second preset threshold, wherein second-level suppression refers to adding a limiting label to the detection point to restrict its use for security-related functions; a third execution unit, configured to perform third-level suppression on a detection point when the ghost confidence of the detection point is lower than the second preset threshold but higher than or equal to a third preset threshold, wherein third-level suppression refers to temporarily storing the detection point in a delay queue so that it is allowed to be established as a target only after it has appeared continuously for a preset number of times in multiple consecutive processing cycles; and a fourth execution unit, configured not to perform suppression operation on the detection point when the ghost confidence of the detection point is lower than the third preset threshold.

[0124] In an optional embodiment of this application, the acquisition module includes an acquisition unit, configured to acquire at least two of the following: wheel speed pulse signals provided by wheel speed sensors in the vehicle, acceleration and angular velocity provided by inertial measurement units, absolute velocity provided by global navigation satellite systems, inverted velocity of stationary targets detected by radar, and motion flow velocity provided by visual sensors.

[0125] like Figure 4 As shown in the figure, this application provides an electronic device, including a processor 411, a communication interface 412, a memory 413, and a communication bus 414, wherein the processor 411, the communication interface 412, and the memory 413 communicate with each other through the communication bus 414. Memory 413 is used to store computer programs; In one embodiment of this application, when the processor 411 executes the program stored in the memory 413, it implements the vehicle control method provided in any of the aforementioned method embodiments, and its function is similar, so it will not be described again here.

[0126] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the vehicle control method provided in any of the foregoing method embodiments.

[0127] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0128] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable 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 described in the various embodiments or some parts of the embodiments.

[0129] It should be understood that the terminology used herein is for the purpose of describing particular exemplary embodiments only and is not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “described” as used herein may also include the plural forms. The terms “comprising,” “including,” “containing,” and “having” are inclusive and therefore indicate the presence of the stated features, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, elements, components, and / or combinations thereof. The method steps, processes, and operations described herein are not construed as requiring them to be performed in a particular order described or illustrated unless the order of performance is explicitly indicated. It should also be understood that additional or alternative steps may be used.

[0130] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for controlling a vehicle, characterized in that, include: Acquire multi-source vehicle speed information; wherein, the multi-source vehicle speed information includes speed information collected by various different types of devices in the vehicle; Based on the degree of deviation between the speed information in the multi-source vehicle speed information, it is determined whether the vehicle speed is in a failure state, wherein the failure state indicates that the vehicle speed cannot be reliably estimated through dynamic correction. When it is determined that the vehicle speed is in the failure state, each detection point detected by the radar in the vehicle is obtained, and the ghost feature quantity of each detection point in multiple physical dimensions is extracted. The ghost feature quantity is used to characterize the degree to which the detection point deviates from the real physical target in the corresponding physical dimension. Based on the ghost feature quantity of each detection point, the deviation of the corresponding detection point in each physical dimension is fused and calculated to obtain the ghost confidence of the corresponding detection point, wherein the ghost confidence is used to characterize the probability that the detection point belongs to a ghost. Based on the ghost confidence level, a graded suppression operation is performed on the detection points, and a list of suppressed detection points is output for subsequent vehicle control.

2. The method according to claim 1, characterized in that, Determining whether the vehicle speed is in a faulty state based on the degree of deviation between the various speed information in the multi-source vehicle speed information includes: Spatiotemporal alignment processing is performed on the speed information in the multi-source vehicle speed information to obtain multiple vehicle speed estimates with the same reference. Pair any two speed estimates from different sources together and calculate the deviation between each pair of speed estimates. Each pair of deviation values ​​is compared with its corresponding preset tolerance threshold to obtain the deviation judgment result between each pair of vehicle speed estimates. Based on all the deviation judgment results, the overall degree of deviation between the estimated vehicle speeds is calculated to generate a vehicle speed health index. If the vehicle speed health index is lower than a preset health threshold, the vehicle speed signal is determined to be in the failure state.

3. The method according to claim 2, characterized in that, Based on all the deviation judgment results, the overall degree of deviation between the estimated vehicle speeds is calculated to generate a vehicle speed health index, including: The comparison result of each pair of vehicle speed estimates exceeding the corresponding preset tolerance threshold is recorded as a deviation event. For each deviation event, a corresponding weighting coefficient is assigned based on the combination of vehicle speed estimation sources corresponding to the deviation event; The weighted total divergence score is obtained by summing the weight coefficients corresponding to all divergence events. Divide the weighted total deviation score by the total number of combinations of vehicle speed estimates to obtain the normalized deviation ratio. The vehicle speed health index is obtained by subtracting the normalized deviation ratio from the value 1.

4. The method according to claim 1, characterized in that, Extract ghosting features of each detection point across multiple physical dimensions, including: The reference speed of the vehicle is obtained from the non-wheel speed source, and the expected Doppler speed of the detection point is calculated by combining the spatial azimuth angle of the detection point. The measured Doppler speed of the detection point is then compared with the expected Doppler speed to obtain the first deviation metric of the detection point in the signal dimension. The first deviation metric is used to characterize the degree of abnormality of the Doppler speed of the detection point relative to the expected value of the real physical target. The spatial position of the detection point in the vehicle coordinate system is obtained, and it is determined whether the spatial position falls within the blacklist area of ​​a pre-calibrated or real-time constructed strongly reflective static object. If so, the target detection results of the visual sensor in the corresponding spatial area are combined for comprehensive judgment to obtain the second deviation metric of the detection point in the spatial dimension. The second deviation metric is used to characterize the degree of abnormality of the spatial position of the detection point relative to the area where the real physical target may exist. The position records of the detection point in multiple consecutive historical frames are obtained. The motion acceleration or change in motion direction of the detection point is calculated based on the position records. The motion acceleration or change in motion direction is compared with the physically achievable motion range to obtain the third deviation metric of the detection point in the time dimension. The third deviation metric is used to characterize the degree of abnormality of the motion trajectory of the detection point relative to the motion law of the real physical target.

5. The method according to claim 4, characterized in that, Based on the ghost feature quantity of each detection point, the deviation of the corresponding detection point in each physical dimension is fused and calculated to obtain the ghost confidence of the corresponding detection point, including: A hierarchical decision tree structure is adopted to determine the initial confidence interval corresponding to the detection point based on whether the detection point has one or more combinations of the first deviation metric, the second deviation metric, and the third deviation metric. Based on the specific values ​​of each deviation metric, the initial confidence interval is continuously adjusted, and the ghost confidence of the detection point is output.

6. The method according to claim 1, characterized in that, Based on the ghost confidence level, a graded suppression operation is performed on the detection point, including: If the ghost confidence level of the detection point is higher than or equal to a first preset threshold, first-level suppression is performed on the detection point, wherein the first-level suppression refers to directly removing the detection point from the radar point cloud; If the confidence level of the ghost at the detection point is lower than the first preset threshold but higher than or equal to the second preset threshold, a second-level suppression is performed on the detection point. The second-level suppression refers to adding a limiting label to the detection point to restrict the detection point from being used for security-related functions. If the confidence level of the ghost at the detection point is lower than the second preset threshold but higher than or equal to the third preset threshold, three-level suppression is performed on the detection point. The three-level suppression means that the detection point is temporarily stored in a delay queue, and it is only allowed to be established as a target after it has appeared continuously for a preset number of times in multiple consecutive processing cycles. If the ghost confidence level at the detection point is lower than the third preset threshold, no suppression operation is performed on the detection point.

7. The method according to claim 1, characterized in that, Obtain multi-source vehicle speed information, including: The system acquires at least two of the following information: wheel speed pulse signals provided by wheel speed sensors in the vehicle, acceleration and angular velocity provided by inertial measurement units, absolute velocity provided by global navigation satellite systems, inverted velocity of stationary targets detected by radar, and velocity of moving flow provided by visual sensors.

8. A vehicle control device, characterized in that, include: The acquisition module is used to acquire multi-source vehicle speed information; wherein, the multi-source vehicle speed information includes speed information collected by various different types of devices in the vehicle; The first processing module is used to determine whether the vehicle speed is in a failure state based on the degree of deviation between the speed information in the multi-source vehicle speed information, wherein the failure state indicates that the vehicle speed cannot be reliably estimated through dynamic correction. The second processing module is used to acquire each detection point detected by the radar in the vehicle when it is determined that the vehicle speed is in the failure state, and extract the ghost feature quantity of each detection point in multiple physical dimensions, wherein the ghost feature quantity is used to characterize the degree to which the detection point deviates from the real physical target in the corresponding physical dimension. The third processing module is used to perform a fusion calculation on the deviation of the corresponding detection point in each physical dimension based on the ghost feature quantity of each detection point to obtain the ghost confidence of the corresponding detection point, wherein the ghost confidence is used to characterize the probability that the detection point belongs to a ghost. The fourth processing module is used to perform graded suppression operations on the detection points according to the ghost confidence level, and output a list of suppressed detection points for subsequent vehicle control.

9. An electronic device, characterized in that, include: The processor, communication interface, memory, and communication bus are connected, with the processor, communication interface, and memory communicating with each other via the communication bus. The memory is used to store computer programs; the processor is used to execute the computer programs to implement the vehicle control method according to any one of claims 1-7.

10. A storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the vehicle control method according to any one of claims 1-7.