A radar data storage method and apparatus

Abstract

The application relates to the field of radar technology, and particularly discloses a radar data storage method and device, which comprises the following steps: acquiring historical laser radar data units of each channel in a multi-channel respectively; predicting data density based on the historical laser radar data units to generate predicted data density; determining a data sampling period based on the predicted data density; and acquiring and storing laser radar data units corresponding to each channel in the multi-channel based on the data sampling period. The application adjusts the data acquisition period by dynamically adapting the data characteristics of each channel, effectively solving the problems of key data loss and storage resource waste caused by fixed period acquisition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar, in particular to a radar data storage method and device. BACKGROUND

[0002] With the rapid development of automatic driving, environmental monitoring and other fields, multi-channel laser radar systems are widely used because they can simultaneously obtain multi-dimensional environmental information. However, in the prior art, multi-channel laser radar data storage generally adopts a fixed period collection strategy, that is, all channels collect and store data according to a preset uniform period. This approach has significant drawbacks: on the one hand, the environmental characteristics of the detection areas corresponding to different channels differ greatly, for example, some channels may continuously detect dynamic target dense areas, while some channels may face open scenes for a long time. The fixed period cannot adapt to such differences, resulting in insufficient data collection in dynamic areas and the loss of key features, while in open areas, excessive redundant data occupies a large amount of storage resources; on the other hand, the data volume of multi-channel systems grows exponentially, and fixed period collection will cause uneven load on the storage system, reducing data read-write efficiency, and subsequent data processing requires filtering of a large amount of invalid data, increasing computing cost. SUMMARY

[0003] In order to overcome the above technical problems existing in the prior art, the embodiments of the present application provide a radar data storage method and device.

[0004] On the one hand, the present application provides a radar data storage method applied to a multi-channel laser radar system, which comprises:

[0005] Respectively acquiring historical laser radar data units of each channel in the multi-channel;

[0006] Predicting data density based on the historical laser radar data units to generate a predicted data density;

[0007] Determining a data sampling period based on the predicted data density;

[0008] Acquiring and storing laser radar data units corresponding to each channel in the multi-channel based on the data sampling period.

[0009] Preferably, the historical laser radar data units include first historical laser radar data units within a first time window and second historical laser radar data units within a second time window, wherein the length of the second time window is greater than the length of the first time window, and the second time window is a multiple of the first time window; predicting data density based on the historical laser radar data units to generate a predicted data density comprises:

[0010] determine a degree of echo power fluctuation based on the first historical lidar data unit;

[0011] determine a first data density based on the degree of echo power fluctuation;

[0012] divide the second historical lidar data into a plurality of data slices based on a window size of the first time window;

[0013] extract radar features corresponding to each slice to generate a radar feature sequence;

[0014] splice the radar feature sequence with the first data density;

[0015] input the spliced features into a preset LSTM model to obtain a second data density;

[0016] fuse the first data density and the second data density to obtain a predicted data density.

[0017] Preferably, fusing the first data density and the second data density to obtain a predicted data density comprises:

[0018] calculate a channel similarity based on the first historical lidar data unit;

[0019] determine a channel spatial weight based on the channel similarity;

[0020] determine a confidence degree corresponding to the first data density based on the degree of echo power fluctuation and the channel spatial weight.

[0021] Preferably, fusing the first data density and the second data density to obtain a predicted data density further comprises:

[0022] determine a confidence degree corresponding to the second data density based on a prediction error of the preset LSTM model;

[0023] fuse the first data density and the second data based on the confidence degree corresponding to the first data density and the confidence degree corresponding to the second data to obtain a predicted data density.

[0024] Preferably, based on the data sampling period, acquiring and storing the to-be-stored lidar data unit corresponding to each channel in the plurality of channels comprises:

[0025] determine the to-be-stored lidar data unit corresponding to each channel in the plurality of channels based on the data sampling period;

[0026] perform importance analysis on the to-be-stored lidar data unit to determine a data type;

[0027] determine a storage strategy according to a data type;

[0028] store the to-be-stored lidar data unit based on the storage strategy.

[0029] Preferably, the to-be-stored lidar data unit is analyzed for importance, and the data type is determined, including:

[0030] Obtain a radar data topology graph.

[0031] Abstract the to-be-stored lidar data unit as a graph node, and extract corresponding node features.

[0032] Based on a preset spatio-temporal correlation rule, establish a spatio-temporal correlation edge between each graph node and other graph nodes in the radar data topology graph, and determine the corresponding edge weight to obtain a new radar data topology graph.

[0033] Based on the new radar data topology graph, calculate the degree centrality and betweenness centrality of each graph node.

[0034] Determine the node importance based on the degree centrality and the betweenness centrality.

[0035] Determine the data type according to the node importance.

[0036] Preferably, the data type includes at least core data, and the method further includes:

[0037] After storing the data corresponding to N data sampling periods, obtain the core data proportion.

[0038] Determine whether to update the data sampling period according to the core data proportion.

[0039] In another aspect, the present application provides a radar data storage device suitable for the method of any one of the preceding aspects, the device comprising:

[0040] A data acquisition unit for acquiring historical lidar data units of each channel in the multi-channel respectively.

[0041] A density prediction unit for predicting data density based on the historical lidar data units to generate a predicted data density.

[0042] A cycle determination unit for determining a data sampling period based on the predicted data density.

[0043] A data storage unit for acquiring and storing to-be-stored lidar data units corresponding to each channel in the multi-channel based on the data sampling period.

[0044] Through the technical solutions provided by the present application, the present application has at least the following technical effects:

[0045] By dynamically adapting the data characteristics of each channel to adjust the data acquisition period, the problem of key data loss and storage resource waste caused by fixed cycle acquisition is effectively solved.

[0046] Other features and advantages of the embodiments of the present application will be described in detail in the following detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0047] The accompanying drawings are included to provide a further understanding of the embodiments of the present application, and constitute a part of the specification, and are used to explain the embodiments of the present application together with the following detailed description, but do not constitute a limitation of the embodiments of the present application. In the drawings:

[0048] Figure 1 is a flowchart of a radar data storage method provided by the embodiments of the present application;

[0049] Figure 2 is a structural schematic diagram of a radar data storage device provided by the embodiments of the present application. DETAILED DESCRIPTION

[0050] The specific embodiments of the embodiments of the present application will be described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the embodiments of the present application, and do not limit the embodiments of the present application.

[0051] The terms "system" and "network" in the embodiments of the present application can be used interchangeably. "Multiple" means two or more, and therefore, "multiple" in the embodiments of the present application can also be understood as "at least two". "And / or", which describes the association relationship of the associated objects, means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / ", if not specially stated, generally represents an "or" relationship between the associated objects. In addition, it should be understood that in the description of the embodiments of the present application, "first", "second", etc. are only used for the purpose of distinguishing the description, and cannot be understood as indicating or implying relative importance, nor can it be understood as indicating or implying order.

[0052] With the rapid development of automatic driving, environmental monitoring and other fields, multi-channel laser radar systems are widely used because they can simultaneously obtain multi-dimensional environmental information. However, in the prior art, the multi-channel laser radar data storage generally adopts a fixed period acquisition strategy, that is, all channels perform data acquisition and storage according to a preset uniform period. This approach has significant drawbacks: on the one hand, the environmental characteristics of the detection regions corresponding to different channels differ greatly, for example, some channels may continuously detect dynamic target dense regions, while some channels may face open scenes for a long time. The fixed period cannot adapt to such differences, resulting in insufficient data acquisition in dynamic regions and missing key features, while in open regions, excessive redundant data occupies a large amount of storage resources; on the other hand, the data volume of a multi-channel system grows exponentially, and fixed period acquisition will cause the storage system to be unbalanced in load, reducing data read-write efficiency, and when subsequent data processing is performed, a large amount of invalid data needs to be filtered, increasing the computing cost.

[0053] Based on the above reasons, see Figure 1 The embodiments of the present application provide a radar data storage method applied to a multi-channel laser radar system, which comprises: acquiring historical laser radar data units of each channel in the multi-channel respectively; predicting data density based on the historical laser radar data units to generate predicted data density; determining a data sampling period based on the predicted data density; and acquiring and storing the to-be-stored laser radar data units corresponding to each channel in the multi-channel based on the data sampling period.

[0054] It should be noted that the radar frames collected by a channel in one scanning period are a laser radar data unit, the historical laser radar data unit includes multiple historical radar frames, and each laser radar data unit contains key information such as point cloud coordinates, echo intensity, time stamp, channel identifier, and signal-to-noise ratio.

[0055] In one possible implementation, first, the historical lidar data units of each channel in the past preset time period are extracted in real time through the multi-channel data interface of the lidar system and stored in a local cache database classified by channel ID. Second, a data density prediction module is constructed based on the historical lidar data units: for each channel, the number of point clouds in a unit time (i.e., data density) in the historical data is counted, the change rate of the data density is calculated through a sliding window algorithm (the window size is set to 5 minutes), the data density change trend in the next 1 minute is predicted in combination with a polynomial fitting method, and the predicted data density of the channel is generated. Then, a mapping relationship model of the predicted data density and the data sampling period is established: a preset data density threshold interval (such as low density: 0-50 points / second, medium density: 51-200 points / second, and high density: 201 points / second or more) is set, different data sampling periods are set correspondingly, and the exclusive data sampling period of each channel is determined according to the threshold interval into which the predicted data density of the channel falls. Finally, the to-be-stored lidar data units of each channel are acquired in real time according to the determined data sampling period, and after channel identification verification, the to-be-stored lidar data units are stored classifiedly by using a distributed storage architecture (such as HDFS), and the data sampling period and the timestamp of each piece of data are recorded, which facilitates subsequent data tracing and period adjustment.

[0056] The embodiment of the application adjusts the data acquisition period by dynamically adapting the data characteristics of each channel, effectively solving the problems of key data loss and storage resource waste caused by fixed period acquisition. On the one hand, in the data-intensive area, the data acquisition period is increased to ensure the complete recording of environmental details and dynamic target information, providing high-quality data support for subsequent target detection, trajectory prediction, and other applications; on the other hand, in the data sparse area, the data acquisition period is reduced, significantly reducing the storage amount of redundant data, and improving the space utilization rate and data read-write efficiency of the storage system. At the same time, the differentiated processing strategy for multiple channels enables the system to better adapt to complex and variable detection environments, enhancing the practicality and reliability of the multi-channel lidar system in actual scenes.

[0057] Further, the accuracy of data density prediction directly determines the rationality of the data acquisition cycle. In actual application, if only the historical lidar data unit of a single time window is relied on, for example, only short-term data or long (such as 1 hour) data is used for prediction. Although the short-term window data can reflect the recent data fluctuation, it is easily affected by instantaneous noise, resulting in oscillation of the prediction result; although the long-term window data can reflect the overall trend, it lags in response to sudden data density changes (such as sudden appearance of dense target groups), and cannot timely capture short-term dramatic fluctuations. Especially in a multi-channel lidar system, the data characteristics of each channel are affected by the environment and present a complex coupling relationship between short-term fluctuations and long-term trends, and a single window prediction method cannot take both into account, resulting in a large deviation between the predicted data density and the actual value, and further making the data acquisition cycle determined based on the prediction result unreasonable, affecting the data storage effect.

[0058] In the embodiment of the application, the historical lidar data unit includes a first historical lidar data unit in a first time window and a second historical lidar data unit in a second time window, wherein the length of the second time window is greater than the length of the first time window, and the second time window is a multiple of the first time window; the data density is predicted based on the historical lidar data unit to generate a predicted data density, including: determining the echo power fluctuation degree based on the first historical lidar data unit; determining the first data density based on the echo power fluctuation degree; dividing the second historical lidar data into multiple data slices based on the window size of the first time window; extracting the corresponding radar features of each slice to generate a radar feature sequence; splicing the radar feature sequence with the first data density; inputting the spliced features into a preset LSTM model to obtain a second data density; and fusing the first data density and the second data density to obtain the predicted data density.

[0059] In a possible implementation, first, the historical lidar data unit is divided into a first historical lidar data unit of a first time window (such as 1s, T1) and a second historical lidar data unit of a second time window (such as 60s, T2=60xT1). For the first data density: based on the first historical data, the echo power, target distance, reflectivity and other physical parameters of each data unit are extracted and substituted into a preset physical model (such as data density=kx echo power / (target distance)2x reflectivity), to obtain the first data density. ), wherein k is a coefficient related to reflectivity, determined by calibration experiment), the initial physical density is calculated; then the initial physical density is corrected in combination with the degree of echo power fluctuation (calculated by standard deviation) (the greater the fluctuation, the smaller the correction coefficient, such as the correction coefficient is 0.8 when the degree of fluctuation is greater than 0.5), and finally the first data density is obtained. For the second data density (long-term prediction): the second historical laser radar data unit is divided into 60 data segments according to T1, and the corresponding first data density, density change rate, environmental temperature (affecting laser propagation) and other features of each data segment are extracted to form a feature sequence; the sequence is input into the pre-trained LSTM model (trained by multi-channel long-term data, containing 3 layers of LSTM units, focusing on capturing periodic and trend changes), and the second data density is output. Fusion stage: dynamic weighting method is adopted, when the physical parameter stability (such as reflectivity fluctuation <5%) of the first historical data is high, the weight of the first data density is set to 0.7, and the weight of the second data density is set to 0.3; when the physical parameter is unstable, the weight of the first data density decreases to 0.4, and the weight of the second data density increases to 0.6, and the weighted predicted data density is obtained.

[0060] The embodiment of the application fuses the historical data features of short-term and long-term time windows, which not only captures the instantaneous fluctuation and detail change of the data by the first time window data, but also mines the long-term change trend and periodicity of the data through the second time window data and the LSTM model, effectively overcoming the limitations of single window prediction. The predicted data density can more accurately reflect the actual data distribution characteristics, especially in the environment mutation scene (such as the sudden appearance of dense vehicle group), which can quickly respond to the surge of short-term data density, and can also avoid over-prediction by combining long-term trends, so that the subsequent determination of data acquisition period is more in line with actual needs, further improving the adaptability and reliability of multi-channel laser radar data storage.

[0061] Further, the above-mentioned embodiments are also limited to predicting data itself, i.e., pure data prediction, and cannot combine the physical law of the laser radar, i.e., cannot combine the decay law of the echo power with the detection distance in the process from the emission to the reception of the laser pulse, which reflects the comprehensive influence of the three factors of geometric decay, atmospheric decay, and scattering characteristics; in another possible embodiment, since the short-term window data can reflect the recent data fluctuation, but is easily affected by instantaneous noise, leading to the oscillation of the prediction result, therefore, the embodiment of the present application does not predict the short-term window data, but directly determines the theoretical echo power corresponding to each radar frame in the short-term window. Further, a physical model can be constructed according to the decay law, the theoretical echo power is determined based on the physical model, and then the theoretical echo power fluctuation degree is determined, the corresponding first data density is determined from the preset correlation table based on the theoretical echo power fluctuation degree, and then the first data density and the second data density determined by the long-term window data are fused. It should be noted that at this time, the method of determining the second data density based on the long-term window data can not be further optimized to extract SNR, Doppler shift, power residual, and channel corresponding space-time coordinates (x, y, z, t) and other characteristics based on the long-term window data, and predict based on the extracted characteristics.

[0062] The embodiment of the present application directly calculates the theoretical echo power of the radar frame in the short-term window based on the physical decay law of the laser radar to determine the first data density, which not only avoids the interference of instantaneous noise on the short-term data from the root, ensures that the first data density conforms to the physical nature and has stronger stability, but also realizes the deep fusion of the physical law and the long-term data trend.

[0063] After determining the first data density and the second data density, it is also necessary to fuse the first data density and the second data density, and one possible way is to fuse them with fixed weights, but this way does not consider the influence of the data correlation between multiple channels on the reliability of the first data density. In fact, adjacent channels of a multi-channel laser radar often detect overlapping or associated regions, and the data characteristics thereof have strong correlation. If the first data density of a certain channel and the contemporaneous data density of the adjacent channel have large deviation, it means that the first data density may be less reliable due to noise interference; on the contrary, if the consistency with the data density of the adjacent channel is high, the reliability is higher.

[0064] In the embodiment of the present application, the first data density and the second data density are fused to obtain the predicted data density, comprising: calculating a channel similarity based on the first historical laser radar data unit; determining a channel spatial weight based on the channel similarity; determining a confidence degree corresponding to the first data density based on the echo power fluctuation degree and the channel spatial weight.

[0065] In a possible implementation, first, a channel similarity is calculated based on a first historical lidar data unit: features such as a signal-to-noise ratio, a power residual, and the like are extracted from the first historical lidar data unit, where the power residual is determined according to a difference between a theoretical value of echo power and an actual value of echo power; based on the extracted features, a cosine similarity between channels is calculated, for each channel, a number of channels (denoted as N) with a similarity greater than 0.7 to the channel is counted, if N is greater than or equal to 3, a spatial weight of the channel is set to 1.0; if 1 is less than N and N is less than 3, the spatial weight is set to 0.8; if N is equal to 0, the spatial weight is set to 0.5 (indicating that the channel data is highly independent and has low reference value). Then, a confidence of the first data density is calculated in combination with a fluctuation degree of echo power: the confidence = spatial weight x (1-fluctuation coefficient), where the fluctuation coefficient is a normalized value of the fluctuation degree of echo power, for example, if the spatial weight of a certain channel is 1.0 and the fluctuation coefficient is 0.2, the confidence of the first data density of the channel is 0.8, indicating that the data is highly reliable.

[0066] In the embodiments of the present application, the spatial weight is determined by introducing channel similarity analysis, and the confidence of the first data density is calculated in combination with the fluctuation degree of echo power, so that the reliability evaluation of the first data density is more in line with the actual correlation characteristics of multi-channel data. When the channel data has high consistency with adjacent channels and stable fluctuation, the first data density is given a higher confidence, otherwise the weight is reduced, effectively avoiding the interference of unreliable data on the fusion result.

[0067] Of course, in another possible implementation, if the first data density is determined based on physical laws, the confidence thereof should be determined according to the power residual, and the smaller the power residual, the closer the confidence to 1. In actual application, the confidence corresponding to the first data density is at most 0.95.

[0068] Further, the confidence of the second data density also needs to be considered. In actual application, the prediction performance of the LSTM model is affected by factors such as input data distribution, and there may be a large prediction error (for example, when a special environment scene not included in the training sample is encountered). If the reliability of the second data density is not judged and a fixed or only experience-based weight is directly given, the second data density may still have a significant impact on the final result when the model prediction deviation is large, and thus the predicted data density after fusion deviates from the actual value, affecting the rationality of the data sampling period.

[0069] In the embodiments of the present application, the first data density and the second data density are fused to obtain a predicted data density, and the method further includes: determining a confidence corresponding to the second data density based on a prediction error of the preset LSTM model; and fusing the first data density and the second data based on the confidence corresponding to the first data density and the confidence corresponding to the second data density to obtain a predicted data density.

[0070] In a possible implementation, for the second data density, the confidence thereof is determined based on the prediction error of the preset LSTM model: after the LSTM model is trained, the test set data is used for verification, the absolute error between the second data density output by the model and the actual data density in each group of data is recorded, the mean (denoted as μ) and the standard deviation (denoted as σ) of the error are calculated; in actual prediction, for the current input radar feature sequence, the real-time error (denoted as e) between the second data density output by the model and the recent actual data density (such as the last 30 seconds) is calculated, if e≤μ-σ, it is indicated that the prediction accuracy of the second data density is extremely high, and the confidence thereof is set to 1.0; if μ-σ<e≤μ, the confidence is set to 0.8; if μ<e≤μ+σ, the confidence is set to 0.5; if e>μ+σ, it is indicated that the prediction deviation of the second data density is large, and the confidence is set to 0.2. Secondly, the confidence corresponding to the first data density (C1) determined in the foregoing and the confidence corresponding to the second data density (C2) obtained in this step are combined, the first data density and the second data density are fused by using a weighted fusion formula, that is, the predicted data density=(C1x the first data density+C2x the second data density) / (C1+C2), and in this way, the fusion result is more suitable for the actual reliability of the two kinds of data densities, and the accuracy of the predicted data density is improved.

[0071] The embodiments of the application evaluate the confidence of the second data density by introducing the prediction error of the LSTM model, so that the fusion process can dynamically adapt to the real-time prediction performance of the model. When the model prediction is accurate, the second data density is given a higher weight, so as to fully exert the advantage of capturing long-term trends; when the model prediction deviation is large, the weight thereof is reduced, so as to avoid misleading the final result.

[0072] After determining the predicted data density, the corresponding data sampling period can be determined according to the preset correlation, and data storage processing is performed based on the data sampling period, thereby realizing on-demand allocation of storage resources and avoiding the problems of data redundancy or missing of critical information under a fixed period. However, there are still obvious limitations in practical application: a unified storage strategy (such as the same storage medium, redundancy, compression algorithm, etc.) is used to process the to-be-stored data obtained by all channels according to the data acquisition period, which ignores the significant difference in actual information value of different data units. For example, data units containing dynamic targets, complex terrain and other critical information are crucial for core tasks such as environmental perception and target identification, and require high reliability and low latency storage guarantee; while the static background data in open areas has low information value and does not need to occupy too much high-quality storage resource. This "one-size-fits-all" storage method not only may lead to the risk of loss or damage of critical data due to insufficient allocation of storage resources, but also may cause low storage efficiency and rising costs due to the excessive occupation of high-cost storage resources by low-value data. At the same time, when processing data later, the key information needs to be screened from a large amount of low-value data, increasing the computing overhead and affecting the overall performance of the system. Therefore, how to further develop a differentiated storage strategy based on the information value of the data unit on the basis of the dynamic data acquisition period has become the key to improving the comprehensive performance of the multi-channel laser radar data storage system.

[0073] In the embodiment of the application, the to-be-stored laser radar data units corresponding to each channel in the multi-channel are obtained and stored based on the data sampling period, including: determining the to-be-stored laser radar data units corresponding to each channel in the multi-channel based on the data sampling period; performing importance analysis on the to-be-stored laser radar data units to determine the data type; determining the storage strategy according to the data type; and storing the to-be-stored laser radar data units based on the storage strategy.

[0074] In a possible implementation, according to the data sampling period, the to-be-stored laser radar data units are acquired, which contain basic information such as point cloud coordinates, echo intensity, time stamp and channel identification. Secondly, the importance analysis is performed on the to-be-stored laser radar data units to determine the data type: the features are extracted from the to-be-stored laser radar data units, the extracted features are input into a preset importance evaluation model to obtain a score result; and the to-be-stored laser radar data units are divided into core data (8-10 points, such as multi-channel overlapping region data containing near-distance fast-moving targets), important data (5-7 points, such as single-channel key region data containing far-distance static targets) and general data (0-4 points, such as empty region data without significant targets) according to the score result, that is, the determination of the data type is completed. Then, the corresponding storage strategy is formulated according to the determined data type: the core data adopts multi-copy redundant storage (3 copies), the storage medium is selected as a high-speed solid state storage (SSD), the compression mode adopts lossless compression to ensure data integrity, and the storage period is set to 1 year; the important data adopts double-copy storage, the storage medium is selected as a hybrid storage (SSD cache+HDD), the compression mode adopts medium-loss compression, and the storage period is set to 3 months; the general data adopts single-copy storage, the storage medium is selected as a large-capacity mechanical hard disk (HDD), the compression mode adopts high-loss compression, and the storage period is set to 1 month. Finally, based on the above storage strategy, the to-be-stored laser radar data units of each channel are classified and stored in the corresponding storage medium according to the data type, and the data type, storage location and strategy parameters are recorded through the metadata management module, so that the differential storage management of data is realized.

[0075] The embodiment of the application realizes the accurate allocation of storage resources by performing importance analysis on the to-be-stored laser radar data units and adopting a differential storage strategy. The core data is stored in a high-redundancy, high-speed and long-term manner to ensure its safety and accessibility, thereby providing reliable data support for key applications such as automatic driving decision and environmental event backtracking. The important data and the general data are stored in an adaptive storage scheme according to their value, thereby significantly reducing the overall storage cost and improving the space utilization of the storage system. At the same time, the classified storage enables the subsequent data retrieval and analysis to directly focus on the core data, thereby reducing the processing amount of invalid data and improving the data application efficiency.

[0076] The above embodiment relies on the local features (such as the number of targets, distance) of the to-be-stored laser radar data unit when performing the importance analysis process, and ignores the spatio-temporal correlation between the to-be-stored laser radar data unit and other data. In fact, multi-channel laser radar data has strong spatio-temporal correlation: for example, a data unit itself has fewer targets (local features show low importance), but it is a key node connecting multiple core data units, and its actual importance is high. Only by evaluating the importance through local features, it may lead to the misjudgment of such data with important correlation value as low importance, and then the unreasonable storage strategy is adopted, which affects the integrity and correlation of the overall data, and reduces the accuracy of subsequent multi-channel data fusion and scene reconstruction

[0077] In the embodiment of the application, the importance of the to-be-stored laser radar data unit is analyzed, and the data type is determined, including: obtaining a radar data topology graph; abstracting the to-be-stored laser radar data unit as a graph node and extracting the corresponding node features; establishing a spatio-temporal correlation edge between each graph node and other graph nodes in the radar data topology graph based on a preset spatio-temporal correlation rule, and determining the corresponding edge weight to obtain a new radar data topology graph; calculating the degree centrality and betweenness centrality of each graph node based on the new radar data topology graph; determining the node importance based on the degree centrality and the betweenness centrality; and determining the data type according to the node importance.

[0078] In a possible implementation, first, a radar data topology graph is constructed, and then the to-be-stored laser radar data unit is processed: each to-be-stored laser radar data unit is abstracted as a new graph node, and its timestamp, spatial coordinates, channel ID, SNR, echo power and other features are extracted as node features. Based on a preset spatio-temporal correlation rule, for example, if two nodes satisfy the condition that the time difference is less than or equal to 1 second and the spatial distance is less than or equal to 10 meters (in the same channel) or less than or equal to 5 meters (in adjacent channels), a spatio-temporal correlation edge is established, and the edge weight is set to the normalized inverse ratio of the time difference and the spatial distance, and the higher the value, the stronger the correlation. The spatio-temporal distance between the new node and all nodes in the radar data topology graph is calculated, the correlation edge that meets the condition is established and the weight is assigned, and an extended radar data topology graph containing the new node is formed, that is, the new radar data topology graph. Then, the node importance is calculated: the degree centrality and betweenness centrality of each new node are calculated by using a graph theory algorithm, the two are normalized, and the degree centrality and betweenness centrality are weighted and summed according to the weight of 40% and 60% respectively, to obtain the node comprehensive importance. Finally, the data type is determined according to the comprehensive importance threshold (≥0.8 for core data, 0.5-0.79 for important data, and <0.5 for general data).

[0079] It should be noted that the abstraction of each to-be-stored laser radar data unit into a new graph node refers to the abstraction of the laser radar data acquisition unit of each channel in the data acquisition period into an independent graph node, that is, the number of new graph nodes is consistent with the number of channels.

[0080] The embodiment of the application breaks through the limitation of single local feature evaluation by introducing radar data topology graph to analyze the space-time correlation relationship of the to-be-stored laser radar data unit, so that the importance judgment is more in line with the actual value of the data in the whole system. For those laser radar data units that are not outstanding in local features but have key correlation, they can be accurately identified and given a higher importance, ensuring that they adopt appropriate storage strategies, thereby ensuring the overall correlation and integrity of the multi-channel laser radar data.

[0081] In the process of constructing the radar data topology graph and generating a new radar data topology graph, the corresponding space-time correlation edges and the edge weights corresponding to the space-time correlation edges are constructed, and the semantic correlation, that is, the correlation between node features, is ignored. In another possible implementation, the similarity between the graph node and other graph nodes is calculated based on the node features of the graph node and the node features of other graph nodes in the radar data topology graph, and the semantic correlation edges and the corresponding edge weights are generated based on the similarity.

[0082] Further, if the corresponding data acquisition period is calculated every time the data storage is performed, the computing overhead will increase. In order to reduce the computing overhead, it is set that the data acquisition period is updated only once after the data is acquired multiple times according to each data acquisition period and the corresponding storage is performed.

[0083] In the embodiment of the application, the data type includes at least core data, and the method further includes: after storing the data corresponding to N data sampling periods, acquiring the core data proportion; determining whether to update the data sampling period according to the core data proportion.

[0084] In a possible implementation, the preset N is the number of data acquisition periods, and a core data proportion statistical process is triggered after the acquisition and storage operation of N periods is completed. All data units stored in N periods are retrieved, the ratio of the total number of core data to the total amount of data is counted, and the core data proportion (denoted as R) is obtained. Next, a reasonable threshold range (such as 30%-60%) of the core data proportion is set, and a feedback adjustment rule is established: if R<30% or R>60%, it indicates that the current data acquisition period cannot meet the actual demand, and the current data acquisition period needs to be updated, and the update method is the same as the determination method of the current data acquisition period, which is not described in detail in the embodiment of the application; if 30%≤R≤60%, the current data acquisition period remains unchanged. Finally, the updated data acquisition period is used as the data acquisition basis for the next round (N period data), and the update time, the data acquisition period before and after the update, and the core data proportion are recorded to generate log data for subsequent optimization adjustment rules.

[0085] The embodiment of the application dynamically monitors the core data proportion and feeds back the adjustment of the data acquisition period, so that the system can continuously adapt to the changes of the detection environment and ensure that the core data proportion is always within a reasonable range. The adaptive ability of the multi-channel laser radar system is enhanced, so that the data acquisition and storage strategy can remain efficient and reasonable for a long time, while meeting the data application requirements in different scenarios, and balancing the storage resource consumption and data quality to the maximum extent.

[0086] Please refer to Figure 2 Based on the same inventive concept, the application provides a radar data storage device, which comprises: a data acquisition unit configured to acquire historical laser radar data units of each channel in the multi-channel respectively; a density prediction unit configured to predict data density based on the historical laser radar data units to generate predicted data density; a period determination unit configured to determine a data sampling period based on the predicted data density; and a data storage unit configured to acquire and store to-be-stored laser radar data units corresponding to each channel in the multi-channel based on the data sampling period.

[0087] It should be understood that the radar data storage device provided by the embodiment of the application and the radar data storage method provided by the above embodiment are based on the same inventive concept, and the more specific working principles of each unit in the embodiment of the application can be referred to the above embodiment, which is not described in detail in the embodiment of the application.

[0088] The optional implementation of the embodiment of the application is described in detail above with reference to the drawings, but the embodiment of the application is not limited to the specific details in the above implementation, and various simple modifications can be made to the technical solutions of the embodiment of the application within the technical concept range of the embodiment of the application, and these simple modifications all belong to the protection range of the embodiment of the application.

[0089] It should be further noted that the various technical features described in the above embodiments can be combined in any suitable manner, as long as there is no contradiction. In order to avoid unnecessary repetition, the various possible combinations are not described again in the embodiments of the present application.

[0090] Those skilled in the art can understand that all or part of the steps of the methods in the above embodiments can be completed by a program instructing related hardware. The program is stored in a storage medium, and includes a plurality of instructions for enabling a single-chip microcomputer, a chip or a processor to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media capable of storing program codes.

[0091] In addition, the various different embodiments of the embodiments of the present application can also be combined in any suitable manner, as long as they do not deviate from the idea of the embodiments of the present application, and they should also be considered as disclosed in the embodiments of the present application.

Claims

1. A method for storing radar data, applied to a multi-channel lidar system, characterized by, The method comprises the following steps: respectively acquiring historical lidar data units of each channel in the multi-channel; predicting data density based on the historical lidar data units to generate predicted data density; determining a data sampling period based on the predicted data density; acquiring and storing the to-be-stored lidar data units corresponding to each channel in the multi-channel based on the data sampling period; wherein the historical lidar data units include first historical lidar data units in a first time window and second historical lidar data units in a second time window, the time length of the second time window is greater than the time length of the first time window, and the second time window is a multiple of the first time window; predicting data density based on the historical lidar data units to generate predicted data density comprises: determining a first data density based on the first historical lidar data units; obtaining a second data density based on the second historical lidar data units and the first data density; fusing the first data density and the second data density to obtain the predicted data density.

2. The radar data storage method of claim 1, wherein, predicting data density based on the historical lidar data units to generate predicted data density comprises: determining a fluctuation degree of echo power based on the first historical lidar data units; determining a first data density based on the fluctuation degree of echo power; dividing the second historical lidar data into a plurality of data slices based on the window size of the first time window; extracting radar features corresponding to each slice to generate a radar feature sequence; concatenating the radar feature sequence with the first data density; inputting the concatenated features into a preset LSTM model to obtain a second data density; fusing the first data density and the second data density to obtain the predicted data density.

3. The radar data storage method of claim 2, wherein, fusing the first data density and the second data density to obtain the predicted data density comprises: calculating a channel similarity based on the first historical lidar data units; determining a channel spatial weight based on the channel similarity; determining a confidence degree corresponding to the first data density based on the fluctuation degree of echo power and the channel spatial weight.

4. The radar data storage method of claim 2, wherein, fusing the first data density and the second data density to obtain the predicted data density further comprises: determining a confidence degree corresponding to the second data density based on the prediction error of the preset LSTM model; fusing the first data density and the second data based on the confidence degree corresponding to the first data density and the confidence degree corresponding to the second data to obtain the predicted data density.

5. The method of claim 1, wherein, acquiring and storing the to-be-stored lidar data units corresponding to each channel in the multi-channel based on the data sampling period comprises: determining the to-be-stored lidar data units corresponding to each channel in the multi-channel based on the data sampling period; performing importance analysis on the to-be-stored lidar data units to determine data types; determining a storage strategy according to the data types; storing the to-be-stored lidar data units based on the storage strategy.

6. The radar data storage method of claim 5, wherein, Performing importance analysis on the to-be-stored laser radar data unit, determining a data type, comprising: Obtaining a radar data topology graph; Abstracting the to-be-stored laser radar data unit as a graph node and extracting corresponding node features; Based on the preset spatio-temporal correlation rule, establish the spatio-temporal correlation edge between each graph node and other graph nodes in the radar data topology graph, and determine the corresponding edge weight, to obtain a new radar data topology graph; Based on the new radar data topology graph, calculate the degree centrality and betweenness centrality of each graph node; Determine the node importance based on the degree centrality and the betweenness centrality; Determine the data type according to the node importance.

7. The method of claim 5, wherein, The data type includes at least core data, and the method further comprises: After storing the data corresponding to N data sampling periods, obtain the core data proportion; Determine whether to update the data sampling period according to the core data proportion.

8. A radar data storage device adapted for use in the method of any one of claims 1-7, characterized in that, Comprise: Data acquisition unit, for respectively acquiring historical laser radar data unit of each channel in the multi-channel; Density prediction unit, for predicting data density based on the historical laser radar data unit, generating predicted data density; Period determination unit, for determining data sampling period based on the predicted data density; Data storage unit, based on the data sampling period, acquires and stores the to-be-stored laser radar data unit corresponding to each channel in the multi-channel.

Images