A multi-source data analysis and collection system and method for an automobile radar

Abstract

This application relates to the field of edge computing and distributed data processing technology, and discloses a multi-source data analysis and acquisition system and method for automotive radar. The system divides the shared memory pool of the radar on-chip system into a safety reserve level and a transaction quota level; an evaluation module calculates the perceptual difference between the radar and visual frames, triggering capture when the difference exceeds a threshold; an encoding module generates a standardized inventory index; an arbitration module queries a local Bloom filter, performing speculative locking if there is a miss and the number of pending memory blocks is below the quota level, otherwise performing a circuit breaker; a settlement module sends the index to the cloud and schedules the controller to transmit data based on the supply and demand saturation instructions fed back from the cloud. This invention solves the problem of resource contention between high-frequency vehicle data acquisition and driving safety functions through edge-cloud collaborative verification and hierarchical management of memory resources, improving the effectiveness and security of data acquisition.

Description

Technical Field

[0001] This invention relates to the field of edge computing and distributed data processing technology, specifically to a multi-source data analysis and acquisition system and method for automotive radar. Background Technology

[0002] With the evolution of intelligent connected vehicles (ICVs) and advanced driver assistance systems (ADAS), utilizing mass-produced vehicles as distributed terminals to collect real-world road scene data has become a core path driving the iteration of deep learning algorithms and building high-precision data assets. Existing data acquisition technologies mainly rely on static preset triggering rules (such as AEB triggering) or full-volume cyclic logging modes to transmit raw sensor data back to the cloud.

[0003] However, in practical engineering applications, existing data acquisition schemes face multiple severe challenges: First, automotive radar systems-on-a-chip (SoCs) typically employ embedded architectures, resulting in severely limited computing power and SRAM storage resources. In traditional acquisition architectures, non-real-time data transfer tasks often rely on CPU-dominated memory copying, which easily leads to CPU resource contention and bus I / O bottlenecks under high-concurrency loads. This can block the main road perception computing thread, causing frame-level latency jitter or even frame drops in real-time tasks such as automatic emergency braking, resulting in non-deterministic failures and safety hazards in the underlying computing system. Existing technologies lack a dynamic scheduling mechanism that achieves isolation and non-interference coexistence between "security services" and "transaction services" at the underlying physical resource level.

[0004] Secondly, existing data filtering logic is relatively simplistic, typically relying solely on the confidence level of a single sensor or target velocity for threshold judgment. This single-logic-based filtering method struggles to handle data consistency calculations and "cognitive conflicts" between heterogeneous sensors. For example, in scenarios like "ghost peeks" where radar detects an obstacle but visual detection misses it, these inconsistent data features from heterogeneous perceptions are precisely the long-tail samples with the highest information entropy value in machine learning model training. The lack of in-depth value assessment makes it easy for the acquisition system to miss truly crucial "high-game" data.

[0005] Finally, traditional vehicle-side data collection strategies are often executed in isolation, lacking real-time interaction with the global inventory status in the cloud. This leads to a situation in large-scale distributed data collection networks where numerous edge computing nodes (vehicles) on the same road segment may simultaneously upload highly repetitive routine road condition data, causing expensive network communication bandwidth congestion and a huge waste of cloud backend storage resources. Meanwhile, data from specific scarce scenarios may be discarded due to local resource exhaustion or rigid strategies, failing to achieve an optimal balance between data collection costs and data asset value across the entire network.

[0006] Therefore, this invention proposes a multi-source data analysis and acquisition system and method for automotive radar to address the shortcomings of existing technologies. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a multi-source data analysis and acquisition system and method for automotive radar. It solves the problem that existing vehicle edge computing nodes, under conditions of limited memory and computing power, cannot balance the resource contention between high-priority secure computing tasks and high-concurrency data acquisition requests, and also address the issues of data redundancy and network bandwidth waste caused by the lack of distributed deduplication and global management.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a multi-source data analysis and acquisition system and method for automotive radar, comprising: The configuration module is used to logically divide the security reserve level and the transaction quota level in the shared memory pool of the radar on-chip system; The evaluation module is used to perform temporal alignment of radar frames and visual frames, calculate the perceptual difference of the aligned data, and generate a capture signal when the perceptual difference exceeds a trigger threshold. The encoding module is used to generate a standardized inventory index based on the target motion attributes in response to the captured signal; The arbitration module is used to query the local Bloom filter and, when the local Bloom filter fails, to detect the number of pending memory blocks in the shared memory pool. If the number of pending memory blocks is lower than the transaction quota level, speculative locking of memory resources is performed. If the number of pending memory blocks reaches the transaction quota level, circuit breaking is performed. The settlement module is used to send the standardized inventory index to the cloud platform and, according to the instructions returned by the cloud platform, to schedule the direct memory access controller to transmit locked memory block data or release locked memory blocks.

[0009] Preferably, the configuration module determines the safety reserve level by: obtaining the frame refresh rate of the radar sensor and the maximum time consumed by the digital signal processing core to process a single frame of data; calculating the product of the frame refresh rate and the maximum time consumed, and adding a safety redundancy coefficient, and rounding the calculation result up to obtain the number of memory blocks corresponding to the safety reserve level; the configuration module subtracts the number of memory blocks corresponding to the safety reserve level from the total number of physical memory blocks in the shared memory pool to obtain the number of memory blocks corresponding to the transaction quota level.

[0010] Preferably, the evaluation module calculates the perceptual difference as follows: It maintains a visual frame buffer covering historical time periods, using the radar frame generation timestamp as an anchor point, and searches for nearest neighbor image frames with time deviations within an effective threshold in the visual frame buffer to complete temporal alignment; it obtains the radar's detection confidence of the target and the visual perception model's detection confidence in the radar projection box's neighborhood, and calculates the absolute value of the difference between the two as a first evaluation value; it obtains the intersection-union ratio (IU) of the radar projection box and the visual detection box, and calculates the difference between the first IU and the IU as a second evaluation value; it multiplies the first evaluation value and the second evaluation value by their respective weighting coefficients and then adds them together to obtain the perceptual difference.

[0011] Preferably, the encoding module generates the standardized inventory index by: calculating the absolute velocity vector magnitude of the target relative to the ground; if the absolute velocity vector magnitude is greater than or equal to the stationary determination threshold, the target is determined to be a dynamic traffic participant; performing binning quantization on the target's longitudinal relative velocity, lateral relative velocity, and estimated collision time respectively; and concatenating the quantized feature codes to generate the standardized inventory index.

[0012] Preferably, the encoding module is further configured to perform the following steps: calculate the absolute velocity vector magnitude of the target relative to the ground; if the absolute velocity vector magnitude is less than the static determination threshold, then determine that the target is a static geographic facility; map the projected metric coordinates of the target to a discretized geographic grid identification number, and generate the standardized inventory index without time attributes by combining the radar cross section quantization level.

[0013] Preferably, the arbitration module performs the speculative locking in the following manner: it modifies the descriptor list pointer of the direct memory access controller using atomic operations; it removes the descriptor node pointing to the current memory block from the free circular list and attaches it to the tail of the lock list to be traded, while inserting a new free memory block descriptor into the original position, so as to complete the logical locking without moving the physical memory data.

[0014] Preferably, the arbitration module executes the circuit breaker in the following manner: when the number of pending memory blocks reaches the transaction quota level, it forcibly abandons the lock on the current memory block; generates an unmet demand log containing the standardized inventory index and timestamp, and stores the unmet demand log in the circular overwrite log area.

[0015] Preferably, the settlement module is further configured to: calculate the transmission priority based on the perceived difference degree and the bandwidth weight of the current network connection type; if the transmission priority is higher than the immediate transmission threshold, then immediately call the high-priority thread to transmit the locked memory block data; if the transmission priority is lower than the immediate transmission threshold, then add the locked memory block data to the idle waiting queue until the network connection type is detected to have changed to Wi-Fi or vehicular Ethernet before uploading.

[0016] Preferably, the system further includes a cloud platform, which includes a clearing module and an analysis module; The clearing module is used to: receive the standardized inventory index, retrieve the corresponding scene feature partition in the global spatiotemporal index library; count the number of valid samples stored in the scene feature partition and the number of samples with issued tokens but in transit across the entire network, and calculate the supply and demand saturation; if the supply and demand saturation is lower than the redundancy threshold, send a confirmation instruction to the settlement module; if the supply and demand saturation is higher than or equal to the redundancy threshold, send a rejection instruction. The analysis module is used to: receive unmet demand logs uploaded by the vehicle terminal, perform spatiotemporal clustering on the unmet demand logs and calculate the regional scarcity index in combination with the inventory status of the scene feature partition, and dynamically adjust the collection strategy parameters sent to the configuration module based on the regional scarcity index.

[0017] This invention also provides a method for multi-source data analysis and acquisition from automotive radar, comprising the following steps: The configuration module logically divides the security reserve level and the transaction quota level in the shared memory pool of the radar on-chip system; The evaluation module performs temporal alignment of radar frames and visual frames, calculates the perceptual difference of the aligned data, and generates a capture signal when the perceptual difference exceeds a trigger threshold. The encoding module responds to the captured signal and generates a standardized inventory index based on the target motion attributes; The arbitration module queries the local Bloom filter and, if the local Bloom filter fails to find the target, checks the number of pending memory blocks in the shared memory pool. If the number of pending memory blocks is lower than the transaction quota level, speculative locking of memory resources is performed. If the number of pending memory blocks reaches the transaction quota level, circuit breaking is performed. The settlement module sends the standardized inventory index to the cloud platform and, based on the instructions returned by the cloud platform, schedules the direct memory access controller to transmit locked memory block data or release locked memory blocks.

[0018] This invention provides a multi-source data analysis and acquisition system and method for automotive radar. It has the following beneficial effects: 1. This invention establishes a dual resource management mechanism based on a security reserve level and a transaction quota level within the shared memory pool of the radar on-chip system, achieving non-interference coexistence of sensing and transaction services at the physical resource level. This mechanism precisely quantifies the minimum resource boundary required for real-time sensing based on the sensor's physical refresh rate and processing latency, ensuring the absolute priority of vehicle driving safety functions. Simultaneously, it utilizes remaining fragmented resources for speculative locking, effectively resolving the issues of system memory contention and main computing thread blocking that may arise from high-frequency data read / write and transmission tasks in heterogeneous computing environments. This improves the utilization rate of underlying hardware resources and the security of task scheduling without increasing additional hardware overhead.

[0019] 2. This invention introduces a value quantification assessment method based on the difference between radar and visual perception, and an end-side deduplication mechanism based on Bloom filters, optimizing the inventory structure and transmission costs of data assets. Instead of blindly collecting all raw data, the system accurately selects cognitive conflict samples with high game-theoretic value by calculating the confidence and geometric position deviation of multi-source sensors after spatiotemporal alignment. Combined with a local efficient hash deduplication verification algorithm based on Bloom filters, the system can intercept redundant low-value data from the entire network at the source with extremely low space complexity before the data enters the memory- and bandwidth-intensive locking and transmission stages, thereby significantly reducing communication traffic costs and increasing the information entropy density of the cloud-based training dataset.

[0020] 3. This invention constructs a closed-loop strategy iteration system based on unmet demand logs and supply-demand saturation analysis, enabling the data acquisition strategy to adaptively respond to changes in physical road conditions. When the vehicle-side is unable to respond to high-value acquisition requests due to resource constraints, the system retains potential market demand information by uploading lightweight logs. Based on this, the cloud platform assesses the data scarcity of specific areas using spatiotemporal clustering statistical algorithms and dynamically issues and adjusts acquisition quota instructions and parameters to edge nodes. This distributed feedback mechanism upgrades data acquisition from a static task to an adaptive scheduling process based on dynamic optimization of global storage status, achieving efficient allocation of data production factors. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram of the cloud platform processing flow of the present invention; Figure 4 This is a schematic diagram of the memory pool dynamics and circuit breaker mechanism of the present invention; Figure 5 This is a simulation comparison diagram of the cumulative amount of high-value samples under different acquisition strategies of the present invention. Figure 6 This is a schematic diagram comparing the convergence of global inventory redundancy rates under different acquisition strategies of the present invention.

[0022] Among them, 100 is the vehicle-mounted terminal; 110 is the configuration module; 120 is the evaluation module; 130 is the coding module; 140 is the arbitration module; 150 is the settlement module; 200 is the cloud platform; 210 is the clearing module; and 220 is the analysis module. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] See attached document Figure 1 This embodiment constructs a multi-source data analysis and acquisition system for automotive radar, which includes an on-board terminal 100 and a cloud platform 200.

[0025] The vehicle-mounted terminal 100 is built on a radar system-on-a-chip and integrates a digital signal processing core, a logic operation core, and a shared memory pool. The vehicle-mounted terminal 100 includes: a configuration module 110, an evaluation module 120, an encoding module 130, an arbitration module 140, and a settlement module 150.

[0026] Configuration module 110 is used to establish data asset levels and define the security level and transaction quota of the shared memory pool.

[0027] Evaluation module 120 is used to perform temporal registration and value calculation of radar and visual data.

[0028] The encoding module 130 is used to generate a standardized inventory index based on the target motion attributes.

[0029] Arbitration module 140 is used to manage the direct memory access controller and perform speculative locking or circuit breaking of memory resources.

[0030] The settlement module 150 is used to perform interactions with the cloud platform 200 and the final delivery of data assets.

[0031] The cloud platform 200 includes a clearing module 210 and an analysis module 220, which are used for global inventory management and supply and demand strategy optimization, respectively.

[0032] See attached document Figure 2 This embodiment constructs a multi-source data analysis and acquisition method for automotive radar, including the following steps: S1, the configuration module 110 reads the radar hardware firmware parameters to determine the data asset level, receives the policy tiles issued by the cloud platform 200, and defines the safety level reserved for real-time business and the quota level allowed for transactions in the shared memory pool. S2, the evaluation module 120 acquires radar frames and visual frames, performs timing alignment based on timestamps, calculates the perceptual difference of the aligned data, and generates a capture signal when the difference exceeds the trigger threshold. S3, the encoding module 130 responds to the capture signal, reads the target's absolute velocity relative to the ground, and generates a first type of hash index based on geographic grid coordinates when the velocity is less than a set threshold, and generates a second type of hash index based on relative motion characteristics when the velocity is greater than or equal to the set threshold. S4, the arbitration module 140 queries the local Bloom filter. When there is a miss, it checks the number of pending memory blocks in the shared memory pool. If the number is lower than the quota level, it modifies the direct memory access descriptor list to lock the current memory block. If the number reaches the quota level, it executes circuit breaking and generates logs. S5, the settlement module 150 sends the hash index or log to the cloud platform 200, and the cloud platform 200 returns a confirmation instruction or a rejection instruction based on the status of the global spatiotemporal index library; S6, the settlement module 150 schedules the direct memory access controller to transmit locked memory block data according to the confirmation instruction and the current network standard, or releases the locked memory block according to the rejection instruction.

[0033] To further clarify the implementation of each technical aspect of this embodiment, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.

[0034] See attached document Figure 1 The configuration module 110 is responsible for executing the core resource boundary definition and acquisition strategy loading tasks during system power-on initialization and subsequent operation.

[0035] In the data asset level mapping sub-step S101, the configuration module 110 does not directly rely on the application layer configuration. Instead, it accesses the register abstraction layer of the radar on-chip system through the standard interface provided by the board-level support package, reading the firmware version identifier and the data flow descriptors exposed by the digital signal processing unit. Based on the read underlying descriptor information, the configuration module 110 maps the vehicle's current data supply capability to a standardized data asset level. As an enumerated value indicating data richness and bandwidth consumption, It served as the underlying benchmark for asset pricing in subsequent transactions. The specific mapping logic is set as follows: When When the system provides a constant false alarm rate (CFAR) filter, it indicates that the system only provides the target list data. This level has the smallest data volume and is suitable for low-bandwidth network environments. When, it indicates that sparse point cloud data containing distance, azimuth, Doppler velocity, and reflection intensity is provided; when When, it indicates that unfiltered distance-Doppler spectral data is provided, which fully preserves the micro-Doppler characteristics of the environment; when At this time, it refers to the raw time-domain sampled data provided by the analog-to-digital converter output. This data contains the most complete environmental information but has the largest data volume. Once determined... The configuration module 110 encapsulates it into the capability registration message and sends it to the cloud platform through an encrypted channel, thereby completing the handshake between the end and the cloud to ensure consistency of data specifications.

[0036] After the handshake is completed, the process proceeds to the policy tile parsing sub-step S102. The configuration module 110 receives and unpacks the policy tile data packet sent from the cloud platform. The physical structure of this data packet includes a set of geofence coordinates, a Bloom filter bit array, and a hash function seed list. The configuration module 110 first parses the geofence coordinates to establish the spatial range in which the current collection policy is effective. Considering deduplication efficiency, the configuration module 110 directly loads the Bloom filter bit array into the high-speed static random access memory of the on-chip system and synchronously updates the hash function seed parameters. This technical detail ensures that the vehicle-mounted terminal and the cloud use completely consistent hash mapping logic, enabling the vehicle-mounted terminal to quickly determine whether a newly generated asset index already exists in the cloud's global inventory using its local high-speed bit array. This effectively intercepts the generation of duplicate data at the source, significantly reducing unnecessary computation and storage overhead.

[0037] In the shared memory pool level partitioning sub-step S103, the configuration module 110 performs logical address space partitioning for the shared memory pool shared by the radar DSP core and ARM core. Given the risk that data transaction tasks may encroach on real-time sensing resources in a heterogeneous computing environment, this embodiment sets two key level thresholds based on queuing theory: a safety reserve level. and trading quota level .

[0038] To accurately quantify the safety boundary, configuration module 110 constructs the following safety retention level calculation model based on the physical refresh rate of the radar sensor and the end-to-end delay characteristics of the signal processing link: ; In the formula, This indicates a floor function to ensure that the number of memory blocks allocated is an integer. This parameter represents the frame refresh rate of the radar sensor, measured in Hertz. It determines the rate at which data flows into the memory pool. This represents the maximum time, in seconds, for the DSP core to process a single frame of data and output it to the downstream controller. This parameter reflects the upper limit of the memory release rate. It is worth noting that a safety redundancy factor is introduced in the formula. Its value range is typically set to [0.2, 0.5]. Introducing... The physical purpose is to cover the nondeterministic lag caused by bus arbitration delay, interrupt response jitter or operating system scheduling delay under extreme operating conditions, and to prevent the loss of sensing frames due to exhaustion of the receive buffer caused by instantaneous processing blockage.

[0039] Based on the calculated safe reserve level, configuration module 110 further determines the transaction quota level. The calculation logic is as follows: ; In the formula, This represents the total number of physically available fixed-size memory blocks in the shared memory pool. In actual deployment, configuration module 110 performs a strict integrity check: if the calculation result shows... This indicates that current hardware resources are insufficient even to meet the most basic real-time security requirements. In this case, the system will be forced into a security failure mode, prohibiting any data transaction functions and reporting an error to upper-layer applications. If the verification passes, This will be written into the system's global resource management table. This parameter clearly defines the maximum number of data blocks that the system is allowed to speculatively lock for transactions without affecting driving safety, thereby achieving a non-interference coexistence mechanism at the physical resource level where "sensing services take priority, and transaction services utilize fragmented resources generated by sensing redundancy."

[0040] See attached document Figure 1 The evaluation module 120 continuously performs the task of spatiotemporal correlation and value determination of multi-source sensor data during system operation.

[0041] When executing S201, the evaluation module 120 primarily addresses the time phase discrepancy issue caused by the asynchronous sampling frequencies and varying data processing delays between the millimeter-wave radar and the vehicle-mounted camera. Internally, the system maintains a circular visual frame buffer based on a first-in-first-out (FIFO) strategy, with a depth covering at least 200 milliseconds of historical data. When the evaluation module 120 receives a frame with a generation timestamp... When processing radar target list data, the system uses this timestamp as the time sequence anchor point and iterates through the visual frame buffer to search for the image frame with the closest temporal distance. To ensure the physical effectiveness of multi-source data fusion, the system executes the following strict nearest neighbor temporal alignment search logic: ; ; In the formula, This represents the timestamp of the generation of the best-matching image frame found in the search. This represents the set of timestamps for all candidate image frames in the current visual frame buffer; This represents the maximum allowable timing deviation threshold of the system. The formula introduces the absolute value symbol. This is to address the potential two-way time discrepancy between radar frames and visual frames, which may be ahead or behind. The value is not arbitrarily set, but is determined by the camera's sampling period. The decision is usually set as For example, for a 30fps camera, this value is approximately 16.6ms. The physical significance of this constraint is to ensure that the electromagnetic reflection state detected by the radar and the optical state recorded in the image belong to the same transient physical scene, preventing spatial position "ghosting" or significant drift caused by the high-speed movement of the vehicle, thereby ensuring the geometric accuracy of subsequent spatial projection. If no frame satisfies the constraint after traversal, the system will automatically discard the current radar frame to avoid introducing invalid noise.

[0042] After time alignment is completed, the process proceeds to S202. Evaluation module 120, based on the selection principle of "cognitive conflict equals high-value samples," quantitatively scores the aligned data pairs. The inputs to this step include the radar's output three-dimensional target state vector (containing range, azimuth, and elevation angles) and the visual perception module's output two-dimensional target detection box vector (containing pixel coordinates). Width and height (and confidence level).

[0043] In this embodiment, the visual perception module embeds a lightweight convolutional neural network model (such as YOLOv5-Nano or MobileNet-SSD), which includes a backbone network, a neck network, and a detection head. The model input is YUV420 format image data preprocessed by an image signal processor, and the output is a sequence of data containing target category probabilities. Bounding box coordinates ( and detection confidence level The model training process uses a mixture of open-source datasets (such as nuScenes) and self-collected datasets containing various weather and lighting conditions. The loss function uses a weighted combination of CIoU-Loss and Focal-Loss to optimize bounding box regression accuracy and address class imbalance issues.

[0044] To unify the spatial benchmark, the evaluation module 120 uses pre-calibrated extrinsic rotation and translation matrices and camera intrinsic matrices to project the 3D target bounding box detected by the radar onto the 2D image plane, generating a radar projection box. Subsequently, the evaluation module 120 calculates the value score based on the following perception difference model. : ; In the formula, This represents the radar's confidence level in detecting the target, with a value range of [0,1]. This represents the detection confidence of the visual neural network model within the neighborhood of the radar projection box. It should be noted that if the visual algorithm fails to detect the target in the corresponding area (i.e., a missed detection occurs), this confidence level is forcibly set. and This is to ensure that the formula can accurately reflect the extreme difference between "a target on radar but no target on visual radar"; This represents the intersection-over-union ratio between the radar projection frame and the visual inspection frame, used to quantify the degree of overlap in geometric positions; and These are the confidence level difference weighting coefficient and the geometric position difference weighting coefficient, respectively, and they satisfy the following conditions: The values ​​of these two coefficients are dynamically assigned by the cloud based on the iterative needs of the current algorithm model. For example, when tackling algorithmic challenges in weak visual texture scenes such as "ghost peeking" (where someone suddenly appears from behind a screen), the cloud will increase the coefficient. Values ​​(such as setting) This increases the capture weight of existence conflict samples, making the system more inclined to collect high-risk edge cases that are "confirmed by radar but ignored by vision." It should be noted that the absolute value operation is introduced into the formula to capture perceptual conflicts in both directions: including cases where radar detects the target but vision misses it, and cases where vision detects the target but radar misses it, thus ensuring the comprehensiveness of the collected data.

[0045] In S203, the evaluation module 120 will calculate the value score in real time. With the preset capture threshold Compare the capture threshold. Derived from the regional scarcity definition table in the strategy tiles, its value is set relatively high for conventional scenarios such as highways (e.g. To reduce redundant data; and for intersections or accident-prone areas, the value is set lower (e.g., ...). To capture as many potentially complex game scenarios as possible. When When this occurs, it indicates that there is a significant perception conflict or abnormal operating condition in the current scene. The evaluation module 120 then generates an asset capture trigger signal and locks the signal together with the registered radar raw data pointer and image data pointer, and enters the subsequent standardized coding process.

[0046] See attached document Figure 1 After receiving the capture signal triggered by the evaluation module 120, the encoding module 130 initiates the asset standardization index generation process.

[0047] In S301, the encoding module 130 does not directly process the raw data, but first performs a physical classification determination of the target attributes. Given the fundamental differences in data value between static environmental facilities (such as guardrails and road signs) and dynamic traffic participants (such as vehicles and pedestrians), the former relies on the long-term existence of geographical location for indexing, while the latter relies on the physical regularity of movement behavior for indexing. Therefore, the system introduces a classifier mechanism based on absolute ground velocity. The encoding module 130 calculates the target's absolute ground velocity vector through vector synthesis. ,in The radial relative velocity vector measured by radar. This is the vehicle's velocity vector obtained from the vehicle chassis bus. Subsequently, the system calculates the magnitude of the absolute velocity. and compared with the preset static determination threshold Comparison. When When the target is identified as a "static geographic facility," it is routed to S302 for processing; otherwise, it is identified as a "dynamic traffic participant" and routed to S303 for processing. The value is usually set to 0.5 m / s. This threshold is based on the speed resolution limit of Doppler radar and the slight vibration noise of the vehicle itself, and is intended to prevent stationary objects from being misjudged as dynamic objects due to signal jitter.

[0048] Upon entering S302, for targets determined to be static, the encoding module 130 aims to generate a hash index that uniquely identifies the geographic facility. Considering the drift error of GPS coordinates and the random jitter of radar measurement points, directly using latitude and longitude for hashing would result in different indices for the same object. Therefore, this embodiment employs a spatial discretization strategy. Since latitude and longitude (in angle units) and grid step size (in metric units) have different dimensions, the encoding module 130 first uses Mercator projection or UTM projection algorithms to convert the target's original geodetic coordinates (… Convert to local metric plane coordinates ( Subsequently, the discretized grid ID is calculated according to the following grid mapping logic: ; In the formula, This indicates the floor function; and These represent the metric x and y coordinates after projection; This represents the step size of the discretized mesh, which is set in this embodiment. Meters. This value was not chosen arbitrarily, but was based on a combination of the standard urban lane width (approximately 3.5 meters) and the typical positioning error of civilian GPS (approximately 1-2 meters), to ensure that the same facility in the same lane could stably fall into the same grid ID during different data collections; and This is a preset coordinate offset used to calibrate the grid origin. This is used to obtain a stable... Subsequently, the encoding module 130 combines the target's radar cross-section (RCS) quantification level and generates a static asset index without time attributes through a hash function, thereby achieving cross-temporal deduplication of inventory.

[0049] For assets identified as dynamic targets, the system performs kinematic feature-based quantization behavior hash generation in S303. At this point, the target's absolute geographical location is no longer used as an index feature; instead, its relative motion behavior "fingerprint" feature is used. To achieve generalized matching for similar driving scenarios (such as "close-range high-speed cut-in"), the encoding module 130 performs bin quantization on continuous floating-point motion parameters. The key feature dimensions selected by the system include longitudinal relative velocity. Lateral relative velocity And the estimated time of collision (TTC). To address the mathematical singularity in TTC calculations where a relative velocity of zero leads to a division by zero, the system pre-sets that if the relative velocity is less than the minimum approximation threshold (e.g., 0.1 m / s), the TTC is directly assigned the maximum saturation value. The system performs quantization mapping for each feature dimension, including longitudinal relative velocity, lateral relative velocity, and TTC. The quantization process uniformly follows the piecewise linear model below: ; In the formula, The input continuous physical quantity representing the current processing dimension; and The effective statistical interval for this feature is defined; To quantize the step size, for example, for relative velocity, It can be set to 2 m / s, which means that all targets with relative speeds between 20 m / s and 22 m / s will be classified into the same speed class; The output is a discrete feature code. Finally, the encoding module 130 converts the discrete feature codes of each dimension into binary bit strings and concatenates them according to preset bit width weights to generate a behavioral hash index containing only kinematic semantics. Through this quantization mechanism, the system can ignore minute numerical differences and group all targets with the characteristics of "high-speed approach and rapid lateral movement" into the same type of standard inventory unit, thereby supporting efficient cloud-based transactions and model training sample retrieval based on scene semantics.

[0050] See attached document Figure 1 After receiving the asset hash index generated by the encoding module, the arbitration module 140 initiates the core storage resource scheduling and concurrent security control process.

[0051] In S401, the arbitration module 140 utilizes a lightweight probabilistic data structure to perform existence verification on assets about to be added to the inventory, aiming to avoid redundant data storage and bandwidth consumption. The system reads a Bloom filter bit array preloaded into the on-chip SRAM by the configuration module. This array, as a compact bitmap, can represent massive historical inventory with minimal space overhead. The system uses... Each independent hash function performs a mapping operation on the input SKU index. The specific mapping position calculation follows the modulo operation logic as follows: ; In the formula, Indicates the first The index of the bit array calculated by the hash function; Index of assets currently pending processing; For the first A single independent hash mapping function; This represents the total physical number of bits in the Bloom filter bit array. The value is typically set between 3 and 5 to balance computation time and false positive rate. The system checks all bits in the bit array. The system determines that an asset is not recorded in either the cloud or local historical inventory only if any corresponding bit is a logical "0", thus allowing it to enter the subsequent locking process. The physical purpose of this mechanism is to leverage the extremely high execution efficiency of bit operations to intercept duplicate data collection requests at the source, before the data enters the memory-intensive locking process, thereby significantly reducing the invalid occupancy rate of the system's internal bus.

[0052] If the asset passes the deduplication check, proceed to step S402. Arbitration module 140 needs to determine if the remaining memory resources of the current system are sufficient to support this transaction, and simultaneously execute a security circuit breaker. The system reads the current locked frame counter from the global resource management table in real time. And compare it with the transaction quota level calculated by the configuration module. Comparison. When When the system's memory resources reserved for non-real-time transactions are exhausted, continued locking will inevitably encroach on the circular buffer required for sensing operations, posing a security risk of radar target loss. At this point, the system immediately triggers a hard circuit breaker, forcibly relinquishing the lock on the current data. However, to prevent the cloud from misjudging the collection strategy due to a lack of negative samples, the arbitration module 140 generates a lightweight "unmet requirement log." This log contains only the SKU hash value and timestamp and is stored in a circularly overwritten, low-speed log area. This design ensures that the cloud can perceive "high-value data that met the strategy but was not collected due to resource constraints," thus providing crucial data support for subsequent strategy optimization or regional pricing adjustments.

[0053] Only when At this point, the system executes S403. To address the high CPU load and transmission latency issues caused by traditional memory copy operations, this embodiment employs an atomic zero-copy locking mechanism based on Direct Memory Access (DMA) descriptor linked list operations. In the conventional radar signal processing flow, the DMA controller automatically moves the ADC sampling data to the circular buffer according to the linked list descriptors. The arbitration module 140 does not move any data bytes in physical memory, but instead achieves logical locking by modifying the pointers of the DMA descriptor linked list. Specifically, the system removes the memory block descriptor currently containing high-value data from the "free circular linked list" and attaches it to the end of the "pending transaction lock list," while simultaneously inserting a new free memory block descriptor into its original position to maintain the continuity of the sensing link.

[0054] To prevent hardware interrupts or linked list breaks caused by concurrent access from multiple cores during the microsecond-level window of pointer swapping, this step is defined as a system-level critical area. The arbitration module 140 utilizes processor-provided hardware mutex primitives (such as LDREX / STREX instructions or global interrupt masking instructions under the ARM architecture) and memory barrier instructions to ensure the atomicity and timing determinism of pointer swapping operations. Its core logic can be described by the following pseudocode: ; In this process This represents the address pointer variable in the DMA descriptor linked list that points to the next node. Indicates the physical address of the memory block that needs to be locked; This indicates the physical address of the free memory block used as a substitute. Through this hardware-enforced atomicity guarantee, the system achieves the "ownership transfer" of megabytes of data within nanosecond clock cycles, logically locking the data while physically preventing any data movement, thus ensuring zero interference between the data transaction process and the main sensing services.

[0055] See attached document Figure 1 After receiving the locked memory pointer from the arbitration module, the settlement module 150 immediately initiates the end-to-cloud interaction and resource reclamation process.

[0056] When executing S501, the settlement module 150 does not directly upload the massive amount of raw radar data. Instead, it uses an asynchronous handshake protocol of "inquiry-confirmation" to avoid invalid transmissions. The system reads the header information of the locked memory block through a direct memory access (DMA) mechanism, extracts and encapsulates it to generate a lightweight transaction metadata header. This metadata header contains an asset hash index generated by the encoding module and an asset capability level representing the data quality. (e.g., point cloud density, signal-to-noise ratio), physical data packet size and effective data collection duration Subsequently, the system sends the metadata header to the cloud-based trading platform via a lightweight MQTT protocol or an HTTP / 2 long-lived connection. The cloud server compares the data against a global spatiotemporal index based on the hash index. If the asset is found to be a new sample or belongs to a high-priority data entry task, a "confirm upload" command and the corresponding upload token are issued; if the data already exists, a "reject transaction" command is issued. This metadata-based pre-verification mechanism replaces the blind transmission of megabytes of data with 100-byte-level signaling interaction, reducing the power consumption and bandwidth costs of the vehicle-side communication module from the source.

[0057] Once the "confirm upload" command is received from the cloud, the system enters S502. Given the highly dynamic changes in the communication environment during vehicle operation, the settlement module 150 dynamically calculates the transmission priority based on the current network service quality (QoS) and the asset's intrinsic value. This determines whether to transmit immediately or with a delay. The calculation of transmission priority follows a weighted logic: ; In the formula, The perceived value score calculated by the evaluation module in S202 has a value range of [0,1]. The higher the value, the more complex the game scenario contained in the data, and the higher the requirement for timeliness. As a preferred approach, a bandwidth weighting function for the current network connection type is set for Wi-Fi connections or in-vehicle Ethernet wired debugging interfaces. For 5G standalone (SA) network environments, set For 4G / LTE environments, set The physical basis for setting this weight is to reflect the differences in unit bit transmission cost and bandwidth stability under different network standards; and These are the value weight and the network weight, respectively, and satisfy the following conditions: In this embodiment, it is set With a slight emphasis on data value, the system will calculate the... With the preset immediate transmission threshold Comparison (usually) ).like This indicates that the data has extremely high timeliness value (such as accident scene data) or that the current network environment is excellent, and the system immediately calls a high-priority thread to execute the fragmented upload; if The system will add the data pointer to the "idle waiting queue" until the vehicle connects to a Wi-Fi hotspot or detects an improvement in the network environment before uploading.

[0058] Once the data upload is complete and an integrity verification receipt is received from the cloud, or after receiving a "transaction rejected" instruction from the cloud in S501, the process proceeds to S503. At this point, the memory block has completed its transaction lifecycle, and the settlement module 150 performs a resource reclamation operation. To maintain the continuity of the circular buffer, this step performs the inverse operation of step S403 in the arbitration module. The system first clears the "locked" status bit and the "pending transaction" flag bit in the DMA descriptor, resetting the ownership identifier of the memory block to "free". Subsequently, atomic operations are used to relink the descriptor node to the head of the free list, allowing it to be requested and used again by the interrupt service routine of the radar front end. Through this reference-based memory recycling mechanism, the system achieves efficient reuse of limited SRAM resources without physical memory allocation and release, completely eliminating the risk of memory fragmentation that may occur during long-term operation.

[0059] See attached document Figure 3 The cloud platform 200 serves as the control center of the entire distributed data trading system, responsible for executing inventory clearing and strategy iteration from a global perspective.

[0060] In step S601, after receiving the transaction request initiated by the vehicle terminal, the clearing module 210 does not perform simple key-value matching, but instead performs a supply and demand saturation check based on statistics. The system maintains a dynamically updated "global spatiotemporal index library," which is essentially a sparse matrix in a high-dimensional feature space. It divides continuous driving scenarios into several discrete scenario feature partitions, each uniquely indexed by the Cartesian product of multi-dimensional features such as "weather-road type-traffic density-vehicle behavior." The clearing module 210 locates the corresponding scenario feature partition based on the received hash index and calculates the supply and demand saturation of the current scenario according to the following formula. : ; In the formula, This indicates the number of valid samples of this type of scenario that have actually been stored in the global database; This represents the number of "in-transit" samples that have been issued upload tokens but have not yet completed transmission across the entire network. The physical significance of introducing this variable is to eliminate the state delay of the distributed system and prevent multiple vehicles from redundantly collecting data on the same saturated scene in the same short period of time due to network transmission time. Indicates the current autonomous driving model The theoretical target sample size required to achieve a specific performance metric is typically preset by the model training team based on the convergence curve of the loss function. The system will calculate the resulting... With the preset redundancy threshold (In this embodiment, it is preferably set to 1.1, that is, allowing 10% redundancy to cope with the risk of data corruption.) The comparison is performed. If... The system determines that the data is scarce and sends a "confirm transaction" command to the vehicle terminal; if If the system determines that the sample size for that scenario is approaching saturation, it will issue a "reject transaction" instruction to the vehicle terminal. This dynamic clearing mechanism based on global supply and demand effectively prevents the omission of samples in long-tail scenarios and the ineffective accumulation of samples in high-frequency scenarios, thereby significantly improving the information entropy density of the training dataset.

[0061] As massive transactions continue, the system enters step S602, where the analysis module 220 initiates adaptive optimization of the data collection strategy. The core input of this analysis module 220 includes not only the data distribution of successfully completed transactions, but more importantly, the "unmet demand logs" uploaded by the vehicle-mounted terminal in step S402. These logs record potential data points that "meet the current strategy but are forced to be abandoned due to limited vehicle-mounted resources," constituting a "potential data distribution state" reflecting the real-world road condition distribution. The analysis module 220 then analyzes the specified geographical area... All unmet logs within the region are spatiotemporally clustered, and combined with the current inventory status, the scarcity index for various scenarios in that region is calculated. : ; In the formula, and The definition is consistent with that in S601, with the first item representing the degree of current inventory shortage; This represents the number of unmet requirements logs reported in this area per unit of time; This represents the total number of vehicles passing through the area per unit of time. To prevent extremely small positive numbers with a denominator of zero (take 10 as an example) -5 The constant 1 was introduced to avoid when The mathematical singularity of the undefined logarithm occurs; and As the weighting coefficient, it is set in this embodiment. This is to emphasize the potential value reflected in unmet needs. The physical meaning of this is to quantify the "scarcity of data" in a specific region and scenario. The role of the logarithmic term is to smooth out the data and avoid excessive fluctuations in the metrics caused by a surge in the number of logs due to sudden congestion.

[0062] Based on the calculated scarcity index, the analysis module 220 dynamically adjusts the configuration parameters sent to vehicles in the region. The specific parameter adjustment logic follows the negative feedback control model: ; In the formula, Update the vehicle-side data collection quota (such as the maximum number of megabytes allowed to be uploaded per day or the number of transactions); Basic quota; This is the scarcity balance threshold; To adjust the gain coefficient (value range 0.1 to 0.5); The hyperbolic tangent function serves to provide a soft saturation limit, ensuring that even if the scarcity exponent rises abnormally, the system's quota adjustment will not exceed the base value. This formula ensures that when the scarcity index of a region exceeds a threshold, the system automatically increases the data collection quota and acquisition price for vehicles in that region, incentivizing more vehicles to participate in capturing high-value data; conversely, it reduces the quota to save costs. Through this closed-loop feedback mechanism, the cloud platform 200 can automatically sense changes in road conditions in the physical world and schedule various vehicle terminals in real time to perform targeted high-density data collection or resource-saving strategy adjustments, achieving automated and efficient allocation of data production factors.

[0063] Based on the aforementioned technical solution, in order to enable those skilled in the art to understand more deeply and intuitively the actual operating logic and advantages of the present invention under complex working conditions, the following will be explained in detail with a typical scenario of "forcibly cutting in against the light on a highway", and supplemented with simulation verification analysis based on actual data.

[0064] Specific application example: Data acquisition process under emergency lane change conditions on highways In this embodiment, an intelligent vehicle equipped with the system of the present invention is used as an example. The vehicle is cruising at a speed of 100 km / h in the middle lane of a highway. The vehicle perception system is equipped with a long-range millimeter-wave radar with a refresh rate of 20Hz and a forward-looking camera with 30fps. Both share the same physical memory pool based on a heterogeneous SoC architecture.

[0065] Scene description: The scenario is set in a backlit evening environment. A target vehicle cuts into the lane to the left from a distance of 30 meters, with a relatively low relative speed but a significant lateral speed. This scenario is a typical long-tail high-risk scenario.

[0066] System response process: When the system powers on, the configuration module 110 dynamically allocates a "safety reserve level" in the shared memory pool based on the radar processing delay and a preset safety redundancy coefficient. This level ensures that sufficient physical memory is reserved for real-time safety functions such as Automatic Emergency Braking (AEB) even when the system is operating at full load. The remaining memory space is marked as a "transaction quota" for non-real-time data transaction operations.

[0067] When the intrusion event occurs, the evaluation module 120 performs data alignment and verification of the heterogeneous sensors. At this time, the millimeter-wave radar accurately detects the lateral movement component of the target and outputs a high-confidence state; however, the visual sensor is affected by backlight and glare, resulting in an extremely unstable target detection box, causing the visual confidence level to be much lower than the radar confidence level, and the intersection-over-union (IoU) ratio of their spatial projections to decrease. The system calculates the value score of the current scene based on the perceptual difference formula. Exceeding the preset capture threshold This triggers the data collection process.

[0068] The encoding module 130 reads the radar's underlying data, identifies the target's motion characteristics of "close range, high relative speed, and large lateral angular velocity," and quantifies them to generate a corresponding dynamic behavior hash index.

[0069] Arbitration module 140 first verifies that the hash index has not been repeatedly collected locally using a Bloom filter. Then, the system checks if the number of currently locked frames is below the "transaction quota." After confirming sufficient resources, the arbitration module performs a "zero-copy locking" operation: modifying the descriptor list pointer of the Direct Memory Access (DMA) controller using atomic instructions. This operation removes the current memory block from the free queue and adds it to the transaction queue within microseconds. This process does not involve CPU movement of physical data, thus avoiding blocking the main path awareness algorithm.

[0070] Settlement module 150 uploads the hash index to the cloud. Clearing module 210 queries the global spatiotemporal index library, confirms that the sample inventory for this type of backlighting scenario has not reached the saturation threshold, and issues a "confirm" command. The vehicle terminal then schedules high-priority threads to complete the data upload based on the current 5G network status.

[0071] To verify the beneficial effects of this invention in resource scheduling and value capture, this embodiment simulated a mixed test set including highway and urban road conditions in a simulation environment, and compared the proposed solution (Strategy A), traditional timed random sampling (Strategy B), and sampling relying solely on local thresholds (Strategy C). To ensure the single-variable principle of the comparative experiment, the specific control logic of each strategy is defined as follows: Strategy A (this invention): It has a dual mechanism of "spatiotemporal value triggering" and "safety level circuit breaker"; Strategy B (Traditional timing): The system does not perform perceptual difference calculation, but only forcibly triggers data locking once at a fixed time interval (such as every 60 seconds), and does not have a safety level detection mechanism until physical memory is exhausted; Strategy C (Local Threshold): The system executes the exact same value assessment algorithm as Strategy A (i.e., (Triggered at any time), but without distinguishing between "safety reserve level" and "transaction quota level", meaning that as long as there are physically free blocks in the shared memory pool, they can be locked for data collection.

[0072] See attached document Figure 4 This sequence diagram visually reproduces the memory game process of the system under the extreme condition of "data burst". In order to clearly define the safety boundary of the system, two horizontal threshold lines are preset in the diagram: the lower dashed line represents the preset "transaction quota level"; the upper solid line represents the physical limit "safety retention level".

[0073] Experimental data reveals the starkly different response logics of the two strategies when a high-value event occurs between the 8th and 12th seconds: The curve representing Strategy A of this invention exhibits extremely high security constraints when faced with a surge in memory requests. The moment the curve climbs to the lower "transaction quota level," the system immediately triggers the circuit breaker mechanism, causing the memory usage curve to show a flat-top "peak-shaping" characteristic. This characteristic indicates that the system actively rejects subsequent unnecessary locking requests, thereby ensuring that the total memory usage is always strongly suppressed below the upper "safety reserve level," without any encroachment on real-time security services within the reserved area.

[0074] In stark contrast, the curve representing comparison strategy C, lacking this quota-based circuit breaker mechanism, exhibits an uncontrolled, linear upward trend in its peak, directly breaching the physical safety level above. In the simulation records, this physical boundary violation directly corresponds to multiple instances of frame dropping in the perception task due to memory exhaustion, powerfully verifying the decisive role of this invention in ensuring system security under extreme data surges.

[0075] See attached document Figure 5 This figure illustrates the cumulative number of high-value "long-tail samples" collected under different strategies over simulation time. The slope of the curves intuitively reflects the signal-to-noise ratio and sustainability of data acquisition. Statistical results show that the curve representing strategy A exhibits a continuous, steep, and stable linear growth trend, indicating that the vast majority of data captured are high-value samples, with an effective sample ratio exceeding 85%, and the system operates stably. Conversely, the curve representing strategy B has an extremely low and flat slope, indicating that it collects a large amount of low-value redundant data (such as smooth following), resulting in low efficiency. While the curve representing strategy C has a high growth rate in the initial stage, a clear horizontal line appears in the latter half. This is due to frequent memory overflows leading to system resets and task interruptions, proving that the high-value acquisition strategy lacking water level protection is unsustainable.

[0076] See attached document Figure 6This figure reveals the redundancy rate trend of similar scenario samples in the cloud database during long-term operation, reflecting the system's ability to suppress duplicate data. Experiments show that the curve representing strategy A, after a brief rise at the beginning of system startup, quickly converges downwards and stabilizes at a low level of around 5% as the cloud-based global spatiotemporal index is established and the feedback mechanism takes effect. This indicates that the system successfully identified and suppressed duplicate collection of data from the central saturated region, achieving "on-demand data supply." Conversely, the curves representing strategies B and C, lacking a dynamic feedback loop between cloud supply and demand, consistently maintain a high redundancy rate oscillating above 60%. This means that valuable bandwidth resources are largely wasted on saturated conventional scenario transmissions, failing to maximize the marginal benefits of data assets.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-source data analysis and acquisition system for automotive radar, characterized in that, include: The configuration module is used to logically divide the security reserve level and the transaction quota level in the shared memory pool of the radar on-chip system; An evaluation module is used to perform temporal alignment of radar frames and visual frames, calculate the perceptual difference of the aligned data, and generate a capture signal when the perceptual difference exceeds a trigger threshold. The evaluation module calculates the perceptual difference as follows: maintaining a visual frame buffer covering historical time periods; using the generation timestamp of the radar frame as an anchor point; searching for nearest neighbor image frames with temporal deviations within an effective threshold in the visual frame buffer to complete temporal alignment; obtaining the radar's detection confidence of the target and the visual perception model's detection confidence in the radar projection box's neighborhood; calculating the absolute value of the difference between the two as a first evaluation value; obtaining the intersection-union ratio (IUGR) of the radar projection box and the visual detection box; calculating the difference between the IUGR and the IUGR as a second evaluation value; and multiplying the first evaluation value and the second evaluation value by their respective weighting coefficients and then adding them together to obtain the perceptual difference. An encoding module is used to respond to the captured signal and generate a standardized inventory index based on the target's motion attributes. The encoding module generates the standardized inventory index by: calculating the target's absolute velocity vector magnitude relative to the ground; if the absolute velocity vector magnitude relative to the ground is greater than or equal to a stationary determination threshold, the target is determined to be a dynamic traffic participant; binning quantization is performed on the target's longitudinal relative velocity, lateral relative velocity, and estimated collision time, which are the target's motion attributes; and the quantized feature codes are concatenated to generate the standardized inventory index. The arbitration module is used to query the local Bloom filter and, when the local Bloom filter fails, to detect the number of pending memory blocks in the shared memory pool. If the number of pending memory blocks is lower than the transaction quota level, speculative locking of memory resources is performed. If the number of pending memory blocks reaches the transaction quota level, circuit breaking is performed. The settlement module is used to send the standardized inventory index to the cloud platform and, according to the instructions returned by the cloud platform, to schedule the direct memory access controller to transmit locked memory block data or release locked memory blocks.

2. The multi-source data analysis and acquisition system for automotive radar according to claim 1, characterized in that, The configuration module determines the safe retention water level in the following way: Obtain the frame refresh rate of the radar sensor and the maximum time consumed by the digital signal processing core to process a single frame of data; Calculate the product of the frame refresh rate and the maximum time consumption, and add a safety redundancy coefficient. Round the result up to get the number of memory blocks corresponding to the safety reserve level. The configuration module subtracts the number of memory blocks corresponding to the security reserve level from the total number of physical memory blocks in the shared memory pool to obtain the number of memory blocks corresponding to the transaction quota level.

3. The multi-source data analysis and acquisition system for automotive radar according to claim 1, characterized in that, The encoding module is also used to perform the following steps: Calculate the target's absolute velocity vector magnitude relative to the ground; If the magnitude of the absolute velocity vector relative to the ground is less than the static determination threshold, the target is determined to be a static geographical facility. The projected metric coordinates of the target are mapped to a discretized geographic grid identification number, and the standardized inventory index without time attributes is generated by combining the radar cross section quantization level.

4. The multi-source data analysis and acquisition system for automotive radar according to claim 1, characterized in that, The specific method by which the arbitration module performs the speculative locking is as follows: The descriptor list pointer of the direct memory access controller is modified using atomic operations; The descriptor node pointing to the current memory block is removed from the free circular linked list and attached to the end of the lock list to be traded. At the same time, a new free memory block descriptor is inserted into the original position to complete the logical lock without moving the physical memory data.

5. The multi-source data analysis and acquisition system for automotive radar according to claim 1, characterized in that, The arbitration module executes the circuit breaker in the following specific manner: When the number of pending memory blocks reaches the transaction quota level, the lock on the current memory block is forcibly abandoned; Generate an unmet demand log containing the standardized inventory index and timestamp, and store the unmet demand log in the circular overwrite log area.

6. The multi-source data analysis and acquisition system for automotive radar according to claim 1, characterized in that, The settlement module is also used for: Based on the perceived difference and the bandwidth weight of the current network connection type, the transmission priority is calculated using a weighted average. If the transmission priority is higher than the immediate transmission threshold, then the high-priority thread is immediately invoked to transmit the locked memory block data; If the transmission priority is lower than the immediate transmission threshold, the locked memory block data is placed in the idle waiting queue until the network connection type is detected to have changed to Wi-Fi or vehicular Ethernet before it is uploaded.

7. The multi-source data analysis and acquisition system for automotive radar according to claim 1, characterized in that, The system also includes a cloud platform, which includes a clearing module and an analysis module. The clearing module is used to: receive the standardized inventory index, retrieve the corresponding scene feature partition in the global spatiotemporal index library; count the number of valid samples stored in the scene feature partition and the number of samples with issued tokens but in transit across the entire network, and calculate the supply and demand saturation; if the supply and demand saturation is lower than the redundancy threshold, send a confirmation instruction to the settlement module; if the supply and demand saturation is higher than or equal to the redundancy threshold, send a rejection instruction. The analysis module is used to: receive unmet demand logs uploaded by the vehicle terminal, perform spatiotemporal clustering on the unmet demand logs and calculate the regional scarcity index in combination with the inventory status of the scene feature partition, and dynamically adjust the collection strategy parameters sent to the configuration module based on the regional scarcity index.

8. A method for multi-source data analysis and acquisition of automotive radar, applied to the system described in any one of claims 1-7, characterized in that, Includes the following steps: The configuration module logically divides the security reserve level and the transaction quota level in the shared memory pool of the radar on-chip system; The evaluation module performs temporal alignment of radar frames and visual frames, calculates the perceptual difference of the aligned data, and generates a capture signal when the perceptual difference exceeds a trigger threshold. The encoding module responds to the captured signal and generates a standardized inventory index based on the target motion attributes; The arbitration module queries the local Bloom filter and, if the local Bloom filter fails to find the target, checks the number of pending memory blocks in the shared memory pool. If the number of pending memory blocks is lower than the transaction quota level, speculative locking of memory resources is performed. If the number of pending memory blocks reaches the transaction quota level, circuit breaking is performed. The settlement module sends the standardized inventory index to the cloud platform and, based on the instructions returned by the cloud platform, schedules the direct memory access controller to transmit locked memory block data or release locked memory blocks.

Images