A method for processing lidar return signals

Abstract

The application relates to the technical field of laser radar signal processing, and discloses a laser radar echo signal processing method. The method acquires a raw echo signal sequence of multiple scanning periods, the sequence containing signal intensity data of all detection channels at each timestamp; signal intensity change modes of each detection channel in each scanning period and all detection channels in all scanning periods are compared, and a signal fluctuation index of each detection channel in each scanning period is extracted; then, signal intensity distribution intervals and the signal fluctuation index of each detection channel in each scanning period are analyzed to generate a signal dynamic change amount; then, signal space distribution characteristics and adjacent channel signal intensity gradient differences of each detection channel in each scanning period and all other detection channels are compared to extract a spatial correlation difference; and finally, the signal dynamic change amount and the spatial correlation difference of all detection channels are fused to generate a signal quality evaluation value of each scanning period.

Description

Technical Field

[0001] This invention relates to the field of lidar signal processing technology, specifically a lidar echo signal processing method. Background Technology

[0002] In lidar applications, the quality of echo signal processing directly affects the accuracy of target detection and identification. Fields such as autonomous driving, environmental monitoring, and surveying rely heavily on echo signal processing technology. Currently, mainstream lidar echo signal processing methods focus on analyzing signal intensity data within a single scan cycle, filtering and classifying signals by setting fixed thresholds. However, this approach has significant limitations. Due to the complex and variable operating environment of lidar, factors such as changes in ambient light, atmospheric particulate interference, and differences in target object surface materials can cause significant fluctuations in the signal intensity of the same detection channel across different scan cycles. Relying solely on data from a single cycle cannot accurately reflect the true pattern of signal variation, easily misjudging valid signals as noise or missing weak, valid signals masked by noise.

[0003] Existing methods often neglect the spatial correlation between channels when processing multi-channel signals. LiDAR detection channels are typically arranged in a spatial array, with overlapping or correlated areas detected by adjacent channels. The gradient changes in signal intensity of each channel can reflect the spatial contour features of the target object. However, current methods often treat each detection channel as an independent entity, failing to analyze the spatial distribution characteristics and intensity gradient differences between channels. This results in the processed signal failing to fully retain the target's spatial information, thus affecting the accuracy of subsequent target localization and identification. Furthermore, in the signal quality assessment stage, existing methods often use single indicators such as the mean and variance of signal intensity, which are insufficient to comprehensively reflect the dynamic changes of the signal in the time dimension and the correlation characteristics in the spatial dimension. This leads to discrepancies between the assessment results and the actual signal quality. Signal classification models trained based on these assessment results have poor generalization ability and cannot adapt to different working environments and detection scenarios, limiting the application effectiveness of LiDAR in complex scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a method for processing lidar echo signals to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for processing lidar echo signals, the method comprising:

[0006] Acquire raw echo signal sequences for multiple scan cycles. Each raw echo signal sequence contains signal strength data for all detection channels at each timestamp.

[0007] By comparing the signal intensity change patterns of each detection channel in each scanning cycle with those of all detection channels in all scanning cycles, the signal fluctuation index of each detection channel in each scanning cycle is extracted.

[0008] Analyze the signal intensity distribution range and the signal fluctuation index of each detection channel in each scanning cycle to generate the dynamic change of the signal of each detection channel in each scanning cycle;

[0009] By comparing the signal spatial distribution characteristics of each detection channel with all other detection channels in each scanning cycle, as well as the signal intensity gradient differences between each detection channel and its adjacent detection channels, the spatial correlation differences of each detection channel in each scanning cycle are extracted.

[0010] By integrating the dynamic changes and spatial correlation differences of the signals from all detection channels in each scanning cycle, a signal quality assessment value for each scanning cycle is generated.

[0011] Based on the signal quality assessment value, class labels are assigned to all scanning cycles. A deep learning network is then trained using the scanning cycles with assigned class labels to obtain an echo signal classification model.

[0012] Preferably, the process of extracting the signal fluctuation index of each detection channel in each scanning cycle is as follows:

[0013] Arrange all signal strength data of each detection channel in each scanning cycle in chronological order to generate a signal strength time series curve;

[0014] Calculate the global trend characteristics of the time-series curves of each signal intensity;

[0015] Obtain the mean of the global trend characteristics of all detection channels in all scanning cycles;

[0016] The signal fluctuation index is obtained by comparing the global change trend characteristic quantity of each detection channel with the mean value in each scanning cycle.

[0017] Preferably, the process of generating the dynamic changes in the signal of each detection channel in each scanning cycle is as follows:

[0018] Calculate the signal intensity range of each detection channel in each scanning cycle;

[0019] The dynamic changes in the signal are directly proportional to the signal strength range and the signal fluctuation index, respectively.

[0020] Preferably, the process of extracting the spatial correlation differences of each detection channel in each scanning cycle is as follows:

[0021] A scanning dot matrix distribution model is constructed, and the detection channels arranged according to spatial coordinates in each scanning cycle are divided into several regional units;

[0022] Calculate the average signal strength of all detection channels within each regional unit, and extract the signal strength offset between the regional unit where the target detection channel is located and the adjacent regional units.

[0023] The cumulative value of the signal strength offset of the target detection channel at all timestamps in each scanning cycle is calculated.

[0024] Based on the distribution characteristics of key inflection points in the signal strength time series curve, a time interval variation characteristic quantity is generated.

[0025] The cumulative value of the signal strength offset and the change in time interval are normalized to obtain the spatial correlation difference.

[0026] Preferably, the process of generating signal quality evaluation values ​​for each scan cycle is as follows:

[0027] Calculate the weighted fusion result of the dynamic changes of signals in each detection channel and the spatial correlation differences in each scanning cycle;

[0028] The signal quality assessment value is the sum of the weighted fusion results of all detection channels in each scanning cycle.

[0029] Preferably, the process of assigning category labels to all scan cycles based on the signal quality assessment value is as follows:

[0030] Based on the signal quality assessment values ​​of each scanning cycle, the initial classification center points are determined using a density clustering algorithm;

[0031] All scan cycles are clustered using the initial classification center point to distinguish between effective echo signal scan cycles and noise interference scan cycles, and different category labels are assigned.

[0032] Preferably, the process of training the deep learning network through the scanning cycle after assigning category labels includes:

[0033] All valid echo signal scan cycles are divided into training and validation subsets according to a preset ratio;

[0034] All noise interference scanning cycles are divided into the training subset and the validation subset according to the same preset ratio;

[0035] The deep learning network is iteratively trained and validated using the training and validation subsets.

[0036] Preferably, the method further includes a feedback optimization step:

[0037] The echo signal sequence of the new scanning cycle is acquired in real time, and the echo signal classification model is input to output the prediction result;

[0038] When the prediction result is a noise interference scanning period, extract the actual signal quality assessment value of that scanning period;

[0039] Calculate the deviation between the predicted result and the actual signal quality assessment value;

[0040] An adaptive compensation factor is generated based on the deviation.

[0041] Preferably, the feedback optimization step further includes:

[0042] Record the deviation amount for multiple consecutive scan cycles to generate a historical deviation moving average;

[0043] By combining the adaptive compensation factor with the moving average of historical deviations, a dynamic correction coefficient is generated;

[0044] The calculation weight of the signal quality assessment value is adjusted by the dynamic correction coefficient.

[0045] Preferably, the method further includes:

[0046] The dynamic correction coefficient is fed back into the fusion process of the dynamic change of the signal and the spatial correlation difference;

[0047] The parameters of the echo signal classification model are updated based on the adjusted signal quality assessment values.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] This lidar echo signal processing method effectively solves many problems existing in current processing methods through multi-dimensional signal analysis and fusion. In the signal data acquisition stage, this method acquires raw echo signal sequences from multiple scan cycles, covering signal intensity data of all detection channels at each time stamp. Compared to traditional single-cycle data processing methods, it can fully capture the temporal variation patterns of the signal, avoiding signal misjudgment caused by accidental factors within a single cycle. By comparing the signal intensity variation patterns of each detection channel in each scan cycle with those of all detection channels across all scan cycles, the extracted signal fluctuation index accurately reflects the fluctuation characteristics of different channel signals over multiple cycles, providing crucial temporal evidence for subsequent signal quality assessment and making the judgment of signal stability more comprehensive and accurate.

[0050] In the signal dynamic change generation stage, this method analyzes the signal intensity distribution range and signal fluctuation index of each detection channel in each scanning cycle. It considers not only the numerical range of signal intensity but also the signal fluctuation, enabling more accurate quantification of the dynamic change trend of the signal in different cycles. This helps distinguish between temporary signal fluctuations caused by environmental interference and continuous signal changes caused by changes in the target object, reducing the interference of environmental factors on the signal analysis results. By comparing the signal spatial distribution characteristics of each detection channel with all other detection channels in each scanning cycle, and the signal intensity gradient differences between each detection channel and its adjacent detection channels, the extracted spatial correlation differences effectively uncover the spatial connections between the signals of each channel, completely preserving the spatial contour information of the target object. This overcomes the shortcomings of traditional methods that neglect the spatial correlation between channels, allowing the processed signal to more realistically reflect the spatial characteristics of the target, providing richer spatial information for subsequent target localization and identification.

[0051] In the signal quality assessment and model training stages, this method integrates the dynamic changes and spatial correlation differences of signals from all detection channels in each scanning cycle. The generated signal quality assessment value combines signal characteristics in both temporal and spatial dimensions, comprehensively and objectively reflecting signal quality and avoiding the one-sidedness of traditional single-index assessments. Based on this signal quality assessment value, category labels are assigned to all scanning cycles. A deep learning network is then trained using these category-labeled scanning cycles. The resulting echo signal classification model can fully learn the complex temporal and spatial characteristics of the signal, possessing stronger generalization ability and adapting to different working environments and detection scenarios. It maintains high signal classification accuracy in both outdoor environments with significant lighting changes and industrial scenarios with abundant particulate matter. Furthermore, the entire processing flow of this method is logically clear, with each step closely integrated, forming a complete technical chain from data acquisition to model training. It requires no additional complex hardware equipment; signal processing performance can be improved solely through algorithm optimization. This demonstrates strong practicality and scalability, enabling its widespread application in multiple LiDAR application fields such as autonomous driving, environmental monitoring, and surveying and exploration, further expanding the application scope and value of LiDAR. Attached Figure Description

[0052] Figure 1 This is a schematic diagram illustrating the working principle of the lidar echo signal processing method described in this invention.

[0053] Figure 2 A flowchart for extracting the signal fluctuation index of each detection channel in each scanning cycle;

[0054] Figure 3 A flowchart for extracting the spatial correlation differences between each detection channel in each scanning cycle;

[0055] Figure 4 A flowchart for assigning category labels to all scan cycles based on signal quality assessment values. Detailed Implementation

[0056] The technical solutions of 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.

[0057] Please see Figure 1 This invention provides a method for processing lidar echo signals, the method comprising the following steps:

[0058] This method acquires raw echo signal sequences from multiple scan cycles, each containing signal intensity data for all detection channels at each timestamp. By comparing the signal intensity variation patterns of each detection channel in each scan cycle with those of all detection channels across all scan cycles, the signal fluctuation index for each detection channel in each scan cycle is extracted. The signal intensity distribution range and signal fluctuation index of each detection channel in each scan cycle are analyzed to generate the dynamic signal change of each detection channel in each scan cycle. By comparing the spatial distribution characteristics of each detection channel with all other detection channels in each scan cycle, and the signal intensity gradient differences between each detection channel and its adjacent detection channels, the spatial correlation differences of each detection channel in each scan cycle are extracted. The dynamic signal change and spatial correlation differences of all detection channels in each scan cycle are fused to generate a signal quality assessment value for each scan cycle. Based on the signal quality assessment values, category labels are assigned to all scan cycles. A deep learning network is trained using the scan cycles with assigned category labels to obtain an echo signal classification model. This method achieves accurate classification of echo signals by integrating signal temporal and spatial features, improving the accuracy of lidar data processing.

[0059] Example 1: See Figure 2The process of extracting the signal fluctuation index of each detection channel in each scanning cycle is essentially about quantifying the signal's variability through time-series analysis. This process begins with the structured reassembly of the original echo signal sequence. Each scanning cycle contains multiple detection channels, and the signal strength data captured by each channel at different timestamps constitutes an independent data stream. These data streams are first arranged in chronological order to form a continuous signal strength time-series curve. The arrangement process strictly follows the timestamp sequence of data acquisition to ensure the integrity of the temporal relationship. The generation of the time-series curve involves not only simple data sorting but also consideration of the uniformity of data point intervals. For cases with missing data or non-uniformly distributed acquisition timestamps, interpolation methods are used for preprocessing to ensure the continuity and analyzability of the curve.

[0060] After generating the signal strength time-series curve, the next step is to calculate its global trend characteristic. This characteristic aims to capture the overall direction and pattern of signal strength changes over time, rather than local instantaneous fluctuations. The calculation process employs trend fitting algorithms, such as linear fitting based on least squares or higher-order polynomial fitting. Linear fitting provides a slope parameter, reflecting the monotonically increasing or decreasing trend of the signal and its intensity. Polynomial fitting can capture more complex nonlinear trend patterns, such as curvature changes. The fitting process requires a holistic analysis of the time-series curve to eliminate outliers caused by instantaneous noise, ensuring that the extracted trend characteristics are representative. The global trend characteristic is a comprehensive numerical indicator that abstractly expresses the macroscopic characteristics of the signal behavior of the probe channel throughout the entire scanning period.

[0061] After obtaining the global trend characteristics of all probe channels across all scan cycles, it is necessary to calculate the central tendency of these characteristics. Calculating the mean is the most direct method, as it represents the overall average level of trend characteristics across all channels across all cycles. The calculation process must consider potential biases in the data distribution. If the distribution of characteristics exhibits significant skewness, a trimmed mean or median is used as an alternative to enhance robustness. This mean serves as a benchmark reference value, measuring the degree to which the trend characteristics of individual channels deviate from the overall level.

[0062] The signal fluctuation index is ultimately obtained by comparing the global trend characteristics of each probe channel in each scan cycle with the mean value mentioned above. The comparison is usually done by calculating the relative difference or standardized distance. For example, the absolute deviation can be obtained directly using interpolation, or the relative deviation can be calculated using a ratio. To eliminate dimensional differences and facilitate subsequent fusion processing, the results are often normalized, mapping them to a uniform numerical range. The magnitude of this index directly reflects the degree of anomalousness of the channel's signal behavior relative to the overall average pattern. A higher positive index value may indicate an unusually strong positive trend in the channel signal, while a higher negative value may suggest an unusually strong negative trend or a behavior pattern significantly different from other channels. The calculation of this index provides a quantitative indicator for subsequent analysis, used to identify and focus on probe channels exhibiting unusual patterns of change.

[0063] The entire extraction process relies on batch processing of large amounts of time-series data. Calculating global trend features requires handling potentially millions of data points, making computational efficiency a critical consideration. The algorithm employs a sliding window or segmented processing strategy to decompose long-term time-series data into manageable blocks while ensuring continuity between blocks. For mean calculation, incremental updates or distributed computing are used to avoid the memory pressure of loading all data at once.

[0064] The generation of the signal fluctuation index is not merely a mathematical calculation; it involves a deeper interpretation of signal behavior. A high or low index does not directly equate to poor signal quality, but rather identifies the unique characteristics of the channel's signal behavior. In some applications, a high fluctuation index may indicate external interference or system noise, while in others, it may precisely reflect the dynamic characteristics of the actual target. Therefore, the interpretation of this index requires consideration of the specific application context and subsequent processing steps.

[0065] From a data flow perspective, this process transforms raw signal strength values ​​into abstract index indicators. The raw data is high-dimensional and rich in detail, while the signal fluctuation index is a highly condensed feature representation. It retains key information related to the change patterns in the raw data while significantly reducing the data dimensionality, thus facilitating subsequent processing steps.

[0066] The advantages of this implementation method lie in its systematic nature and scalability. By treating each detection channel as an independent time-series data source and employing a unified analysis method, consistency and comparability of the processing are ensured. The introduction of a benchmark mean makes comparisons between different periods and different channels possible. The entire process is designed for batch processing, enabling efficient handling of the large-scale data streams generated by the lidar system.

[0067] Example 2: See Figure 3 The process of generating the signal dynamics of each detection channel in each scan cycle begins with the calculation of the signal strength range. This calculation is performed on a single detection channel within a single scan cycle. For a given detection channel, its signal strength range is defined as the arithmetic difference between the maximum and minimum signal strength values ​​acquired by that channel at all timestamps within that scan cycle. This calculation directly reflects the absolute fluctuation range of the channel signal within that time period. Subsequently, the generation of the signal dynamics depends on this range and the signal fluctuation index of the channel calculated in the previous step. There is a direct proportional relationship between the two, which can be described by the following expression:

[0068]

[0069] in: This represents the dynamic change of the signal in the c-th detection channel during the p-th scan cycle. This represents the signal strength range of this channel during that scan cycle. This represents the signal fluctuation index of the channel during that scan cycle. It is a preset scaling factor used to adjust the dimensions and numerical range of the calculation result. Its value is usually set based on historical data or experience.

[0070] The process of extracting the spatial correlation differences among the detection channels in each scanning cycle involves modeling and analyzing the spatial distribution relationship of the detection channels. This process begins with the construction of a scanning lattice distribution model. This model reorganizes all detection channels in each scanning cycle according to their inherent spatial coordinates, forming a two-dimensional or three-dimensional spatial lattice. Based on this, a grid partitioning algorithm is used to divide the entire detection area into several continuous or discontinuous regional units. The partitioning can be based on a regular geometric grid or a clustering algorithm adapted to the detection channel density, with the goal of forming spatially continuous unit blocks with relatively homogeneous internal features.

[0071] For each target detection channel, its corresponding area cell needs to be located. Then, the average signal strength of all detection channels within that area cell is calculated. Simultaneously, the average signal strength of several adjacent area cells is calculated. The definition of "adjacent" can be set according to actual application requirements, typically including cells sharing edges or corners. The signal strength offset between the target channel's area cell and a specific adjacent cell is the absolute value of the difference between the average signal strengths of these two cells. In practice, the offsets between the target cell and all adjacent cells may be calculated, and the maximum, minimum, or average value may be selected as representative values.

[0072] The next step is to calculate the cumulative signal strength offset of the target detection channel across all timestamps throughout the entire scanning cycle. This cumulative value is not a simple summation, but rather considers the continuous impact over time, possibly using a weighted accumulation method, with more recent timestamps assigned higher weights. The calculation can be expressed as:

[0073]

[0074] in: This represents the cumulative value of the signal strength offset of the c-th detection channel in the p-th scan cycle. This represents the total number of timestamps within the scanning period. This represents the signal strength offset of the channel calculated under the given scan period and timestamp t. It is a weighting coefficient function associated with the timestamp t, used to express the time weight distribution.

[0075] Simultaneously, it is necessary to generate a time interval variation feature based on the signal strength time-series curve of this channel. This feature is used to quantify the non-uniformity of signal variation in the time dimension. Its extraction process first involves identifying key inflection points of the time-series curve, including local maxima, local minima, and points where curvature changes significantly. The identification algorithm can employ mathematical methods based on the first derivative (gradient) and second derivative. After identifying all key inflection points, the time interval between adjacent inflection points is calculated. The statistical characteristics of these time interval sequences (such as standard deviation, coefficient of variation, or entropy) are used to construct the time interval variation feature. The larger the value of this feature, the more irregular the temporal pattern of signal variation and the more non-uniform the intervals.

[0076] The cumulative signal strength offset and time interval variation characteristics calculated above are normalized. Normalization aims to eliminate dimensional differences and map the values ​​to a uniform scale for easier subsequent fusion. Normalization methods can employ Min-Max scaling or Z-score standardization. The two normalized quantities are then combined linearly or nonlinearly to ultimately generate the spatial correlation difference of the detection channel during the scanning period. This entire process, through quantitative analysis of the spatial neighborhood relationships of the detection channel and capture of its time-varying characteristics, effectively extracts the spatial correlation characteristics and degree of difference of the echo signal, providing crucial spatial feature input for subsequent signal quality assessment.

[0077] Example 3: See Figure 4The process of generating signal quality assessment values ​​for each scan cycle hinges on effectively fusing the channel characteristics calculated in the preceding steps to form a comprehensive index representing the signal quality of the entire scan cycle. This process begins with a weighted fusion of the dynamic changes in the signal and spatial correlation differences of each detection channel. For each detection channel c in each scan cycle p, its dynamic changes in signal... Differences in spatial association The signals are assigned appropriate weights and combined accordingly. The weight allocation strategy can be preset based on prior knowledge of the application scenario, or it can be adaptively adjusted using a data-driven approach. The result of the weighted fusion is a new intermediate quantity that comprehensively reflects the signal characteristics of the channel in both time and space dimensions. This fusion calculation can be expressed as:

[0078]

[0079] in: This represents the weighted fusion result of the c-th probe channel in the p-th scan cycle. It is the dynamic change of the signal in this channel. It is the difference in spatial correlation of this channel. and These are the weight coefficients assigned to these two features, and they satisfy... The specific values ​​of these features determine their relative importance in the final evaluation.

[0080] To obtain the weighted fusion result of all detection channels within one scan cycle. Next, these channel-level results need to be aggregated into period-level evaluation values. Signal quality evaluation values. Defined as all C probe channels in the scan period p The accumulated value. Accumulation operations typically use arithmetic summation, expressed as: This value A scalar index is provided to quantify the overall quality level of the echo signal throughout the entire scan cycle. A higher value generally indicates that the signal contains more effective information or has higher reliability within that cycle; a lower value may suggest that the signal is subject to more severe noise interference or has poorer information quality.

[0081] Signal quality assessment values ​​for each scan cycle obtained from calculation The next step is to classify and label all scan cycles. This process uses a density clustering algorithm. Unlike distance-based partitioning methods, density clustering can discover clusters of arbitrary shapes and identify noise points, making it very suitable for distinguishing between valid signals and noise interference. The algorithm first requires determining initial classification center points. These center points are not randomly initialized but based on signal quality assessment values. The distribution characteristics are analyzed. By analyzing the density distribution of all periodic evaluation values, regions with high density are identified, and the center points of these regions are selected as initial classification center points. The number of possible center points is usually preset to two categories, corresponding to valid echo signals and noise interference, but in some complex scenarios, more than two categories of center points may be identified to correspond to finer-grained classification.

[0082] After determining the initial classification center, the algorithm iterates through each scan cycle in the dataset, based on its signal quality evaluation value. Based on their relative relationship with each center point, they are assigned to the nearest cluster. The distance metric is typically Euclidean distance, but in one-dimensional data... In the case of [missing information], it is simplified to absolute difference. The clustering process is usually an iterative optimization process. The algorithm will continuously adjust the position of the centroid and redistribute the data points until the convergence condition is met, such as the change in the position of the centroid is less than a certain threshold or the cluster assignment no longer changes.

[0083] After the clustering process is complete, each scan cycle is assigned a cluster label. The next step is to interpret the semantics of these clusters, distinguishing which cluster represents a valid echo signal scan cycle and which represents a noisy interference scan cycle. The basis for this distinction is usually the cluster center point. Numerical magnitude. Since effective echo signals typically exhibit higher signal quality and more stable characteristics, their corresponding... The expected value will be significantly higher than the value corresponding to the noise interference period. Therefore, it has a higher... Clusters at the center point are labeled as belonging to the "effective echo signal" category, while those with lower... The cluster at the center point is then labeled as the "noise interference" category. Ultimately, each scan cycle is assigned a distinct binary category label based on the cluster it belongs to (e.g., 1 for valid, 0 for noise).

[0084] The key to this implementation lies in reducing the dimensionality of multi-channel features into a single comprehensive evaluation value through weighted fusion, and automatically discovering the inherent class structure within the data using unsupervised clustering methods. This approach reduces reliance on large amounts of manually labeled data, adaptively classifies the quality of LiDAR scanning cycles, and provides a high-quality labeled data foundation for subsequent training of supervised deep learning models. (Weight coefficients) and The setting of parameters and the selection of density clustering parameters (such as neighborhood size and minimum number of samples) have a significant impact on the final classification results. It is usually necessary to determine the optimal combination of parameters through cross-validation or by debugging on historical datasets.

[0085] Example 4: This example involves training a deep learning network through scan cycles with assigned category labels and includes a feedback optimization mechanism. A specific example is described in detail below. Assume that in a LiDAR system, after processing as described in the previous example, 1000 scan cycles have been classified, with 700 labeled as "valid echo signals" and 300 labeled as "noise interference scan cycles".

[0086] The process of training a deep learning network begins with the partitioning of the dataset. All valid echo signal scan periods are divided into training and validation subsets according to a preset ratio. This preset ratio is typically set to 80% for training and 20% for validation. In this example, 560 of the 700 valid periods are randomly selected and assigned to the training subset, and the remaining 140 are assigned to the validation subset. Similarly, all noisy scan periods are divided into training and validation subsets according to the same preset ratio. 240 of the 300 noisy periods go to the training subset, and 60 go to the validation subset. This proportional partitioning by category ensures balanced representativeness across categories during training and validation, avoiding model bias caused by uneven data distribution.

[0087] The deep learning network used for training can have various architectures, such as one-dimensional convolutional neural networks or long short-term memory networks, depending on the characteristics of the echo signal sequence. The network input is the raw echo signal sequence for a single scan cycle, containing signal strength data from all probe channels at various timestamps, typically organized into a multidimensional tensor. The output is a binary probability representing the likelihood that the cycle belongs to a valid echo or noise interference. The training process uses a subset of data to optimize network parameters through backpropagation to minimize the difference between the predicted and true labels; cross-entropy loss is a commonly used loss function. The validation subset is used to monitor the training process, evaluate the model's generalization ability, and help prevent overfitting. Training is performed in multiple iterations until the model's performance on the validation set stabilizes.

[0088] The feedback optimization step is performed in real time after model deployment. The system continuously acquires the echo signal sequence of new scan cycles and inputs it into the pre-trained echo signal classification model. The model outputs a prediction result, which is the probability value of whether the new cycle is judged as valid or noise. A probability threshold is set (e.g., 0.5). When the predicted probability is lower than this threshold, the scan cycle is determined to be a noise interference scan cycle.

[0089] When the predicted scan period is a noisy interference period, the subsequent optimization process is initiated. First, the system performs all the aforementioned signal processing steps in parallel for this new scan period (i.e., calculating its signal fluctuation index, signal dynamic change, spatial correlation difference, and finally fusing them to generate its signal quality assessment value). This calculated value is called the actual signal quality assessment value for that scan period. Then, the model's prediction result (a binary category label, e.g., 0 representing noise) is compared with this actual signal quality assessment value. Since the actual signal quality assessment value is a continuous scalar, while the prediction result is a discrete category, it is necessary to map the prediction result to a comparable continuous value scale. One feasible approach is to use the probability value output by the model, or to use a reference value related to the average assessment value of noisy cycles in the training set.

[0090] The deviation between the predicted result and the actual signal quality assessment value is calculated. This deviation is a scalar value reflecting the degree of difference between the model prediction and the actual assessment calculated based on signal features. Based on this deviation, an adaptive compensation factor is generated. This factor can be a linear function value proportional to the deviation, or it can be a more complex nonlinear mapping output. Its purpose is to quantify the magnitude of the current prediction error and provide a basis for subsequent corrections.

[0091] Table 1 below presents a simplified example illustrating the model predictions, actual evaluation values, and calculated bias and compensation factors across multiple consecutive new scan cycles. Here, "Cycle ID" is a unique identifier; "Prediction Result" is the model output category (0: noise, 1: valid); "Prediction Confidence" is the model output probability value; "Actual SigQual Evaluation Value" is a calculated continuous scalar; "Bias" is the calculated difference; and "Compensation Factor" is an adjustment parameter generated based on the bias.

[0092] Table 1: Feedback Optimization Data Table for New Scan Cycle.

[0093] Referring to Table 1, for a new scan period with period ID N001, the model predicts it as noise (0) with a low confidence of 0.15, while the calculated actual signal quality assessment value is 12.5. The calculation of the deviation may be based on the assumption that the ideal actual assessment value for a period predicted as noise should be close to a low reference value (e.g., 10.4). Therefore, the deviation is calculated as |12.5-10.4|=2.1. The compensation factor is set to one-tenth of the deviation, i.e., 0.21. For periods predicted as valid (such as N002 and N005), the deviation calculation and compensation process is not triggered.

[0094] The entire feedback optimization process is a closed-loop system, which enables the echo signal classification model to be not only a static classifier, but also to self-examine and adjust based on real-time data feedback. By continuously monitoring the deviation between the prediction and the actual signal assessment and generating corresponding compensation parameters, the system provides an adaptive mechanism to cope with environmental changes or signal characteristic drift.

[0095] Example 5: This example further expands and deepens the feedback optimization mechanism, aiming to enable the system to continuously self-optimize and adapt to environmental changes by introducing historical trend analysis and dynamic adjustments. Based on Example 4, this example adds long-term monitoring of deviations, generation of comprehensive correction coefficients, and iterative updates to model parameters.

[0096] After processing new scan cycles in real time and generating adaptive compensation factors, the feedback optimization step moves to a more advanced processing stage. The system continuously runs and records the deviation over multiple scan cycles within a continuous time period. The number of "multiple consecutive" cycles is a configurable parameter, which can be set to, for example, the most recent 50 or 100 scan cycles determined to be noise interference. The recorded data includes not only the raw values ​​of the deviation but also corresponding timestamps to analyze their trends over time. The system maintains a fixed-length historical deviation queue; when new data is added, the oldest data is removed, ensuring that the queue always reflects the system performance over a recent period.

[0097] Generating a moving average of historical deviations is one of the core operations at this stage. The calculation of the moving average aims to smooth out short-term fluctuations and reveal the overall trend of deviations over a medium-term range. Instead of a simple arithmetic average, the calculation considers the time decay effect, giving higher weight to recent data so that the moving average better reflects the current state of the system. A common approach is to use an exponentially weighted moving average, which uses a decay coefficient to control the influence of historical data, making its calculation more sensitive to recent changes. This historical deviation moving average serves as an important trend indicator, quantifying the stability and directionality of the system's prediction deviations.

[0098] The system needs to combine the adaptive compensation factor generated for the latest cycle in the preceding steps with the calculated historical moving average of the deviation to generate a dynamic correction coefficient. This combination process is not a simple addition or averaging, but a comprehensive decision-making process. The logic behind generating the dynamic correction coefficient is to consider both the immediate deviation of the current single cycle (reflected by the adaptive compensation factor) and the average deviation level over a recent period (reflected by the historical moving average). For example, if the current deviation is large, but the historical average deviation is small, it may be judged as an occasional anomaly, and the generated dynamic correction coefficient will not be too aggressive; conversely, if both the current deviation and the historical average deviation remain large, it indicates that the system may have persistent performance degradation or environmental adaptation problems. In this case, a stronger dynamic correction coefficient is needed to drive the system to make more significant adjustments. This coefficient is a comprehensive adjustment parameter, and its magnitude and sign together determine the direction and magnitude of subsequent adjustments.

[0099] After generating the dynamic correction coefficient, its primary application is to adjust the calculation weights of the signal quality assessment value. The signal quality assessment value is obtained by fusing the dynamic changes in the signal with spatial correlation differences, and its fusion weights were originally preset fixed values. The dynamic correction coefficient will be introduced into this weight allocation logic. For example, the system can proportionally fine-tune the weight α of the dynamic changes in the signal and the weight β of the spatial correlation differences based on the magnitude of the dynamic correction coefficient. If the dynamic correction coefficient consistently indicates a large deviation between the assessment of the time dimension feature (the dynamic changes in the signal) and the model prediction, its weight α can be appropriately reduced, while the weight β of the spatial dimension feature (spatial correlation differences) can be correspondingly increased, and vice versa. This adjustment makes the calculation process of the signal quality assessment value no longer static, but can dynamically optimize its feature fusion strategy based on actual feedback, thereby more accurately reflecting the actual situation of the signal.

[0100] More importantly, the dynamic correction coefficients need to be fed back into the fusion process of the dynamic changes in the signal and the differences in spatial correlation. This means that the fusion algorithm itself receives an external control signal. This control signal is not only used to adjust the weights, but in some implementations, it may even affect the calculation method of fusion itself, such as switching from linear fusion to a nonlinear fusion mode to better handle the special cases indicated by the dynamic correction coefficients.

[0101] Ultimately, all the above adjustments aim to optimize the echo signal classification model. Based on the new signal quality assessment values ​​calculated after the adjustments, the system has a more accurate standard for measuring the quality of the scanning cycle. These new assessment values, along with the deviation records from previous model predictions, together constitute the basis for updating the model parameters. Model updates can employ various strategies. For models with online learning capabilities, new data (including new cycle features and their adjusted assessment values ​​or labels) can be used directly for incremental learning, fine-tuning the model parameters in real time. For models with cycle-based batch updates, the system accumulates a certain number of new samples and feedback data before initiating a retraining process. Using the original training set plus new feedback data samples, the model is retrained, thus solidifying the knowledge gained from feedback optimization into the next generation of models. Through this mechanism, the echo signal classification model can gradually adapt to new environmental conditions, changes in signal characteristics, and deviations in its previous judgments, achieving continuous performance improvement and enhanced robustness. The implementation method in this embodiment constructs a complete "monitoring-evaluation-adjustment-learning" closed loop, enabling the lidar signal processing system to possess advanced adaptive capabilities for long-term stable and reliable operation in real, changing environments.

[0102] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0103] 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 method of processing a laser radar return signal, characterized by, The method comprises the following steps: obtaining a plurality of original echo signal sequences of scanning cycles, each original echo signal sequence containing signal intensity data of all detection channels at each time stamp; extracting a signal fluctuation index of each detection channel in each scanning cycle by comparing the signal intensity change mode of each detection channel in each scanning cycle with that of all detection channels in all scanning cycles; analyzing the signal intensity distribution interval of each detection channel in each scanning cycle and the signal fluctuation index to generate a signal dynamic change amount of each detection channel in each scanning cycle; extracting a spatial correlation difference of each detection channel in each scanning cycle by comparing the signal spatial distribution characteristics of each detection channel with those of all other detection channels, and the signal intensity gradient difference between each detection channel and its adjacent detection channel; fusing the signal dynamic change amount and the spatial correlation difference of all detection channels in each scanning cycle to generate a signal quality evaluation value of each scanning cycle; assigning a category label to all scanning cycles based on the signal quality evaluation value, and training a deep learning network through the scanning cycles after the category label is assigned to obtain an echo signal classification model.

2. The lidar return signal processing method of claim 1, wherein, The process of extracting the signal fluctuation index of each detection channel in each scanning cycle comprises the following steps: arranging all signal intensity data of each detection channel in each scanning cycle in chronological order to generate a signal intensity time sequence curve; calculating a global change trend feature quantity of each signal intensity time sequence curve; obtaining a mean value of the global change trend feature quantity of all detection channels in all scanning cycles; the signal fluctuation index is obtained by comparing the global change trend feature quantity of each detection channel in each scanning cycle with the mean value.

3. The lidar return signal processing method of claim 2, wherein, The process of generating the signal dynamic change amount of each detection channel in each scanning cycle comprises the following steps: calculating a signal intensity range value of each detection channel in each scanning cycle; the signal dynamic change amount is in a positive proportional relationship with the signal intensity range value and the signal fluctuation index.

4. The lidar return signal processing method of claim 3, wherein, The process of extracting the spatial correlation difference of each detection channel in each scanning cycle comprises the following steps: constructing a scanning point array distribution model to divide the detection channels arranged in spatial coordinates in each scanning cycle into a plurality of regional units; calculating a signal intensity mean value of all detection channels in each regional unit, and extracting a signal intensity offset amount between a target regional unit and an adjacent regional unit in which a target detection channel is located; statistically calculating a signal intensity offset amount cumulative value of the target detection channel at all time stamps in each scanning cycle; generating a time interval change feature quantity based on a key turning point distribution feature of the signal intensity time sequence curve; normalizing the signal intensity offset amount cumulative value and the time interval change amount to obtain the spatial correlation difference.

5. The lidar return signal processing method of claim 4, wherein, The process of generating the signal quality evaluation value of each scanning cycle comprises the following steps: calculating a weighted fusion result of the signal dynamic change amount and the spatial correlation difference of each detection channel in each scanning cycle; the signal quality evaluation value is an accumulated value of the weighted fusion result of all detection channels in each scanning cycle.

6. The lidar return signal processing method of claim 5, wherein, The process of assigning a category label to all scanning cycles based on the signal quality evaluation value comprises the following steps: determining an initial classification center point by using a density clustering algorithm according to the signal quality evaluation value of each scanning cycle; The initial classification center point is used to cluster all scan cycles, to distinguish valid echo signal scan cycles from noise interference scan cycles, and to assign different class labels.

7. The lidar return signal processing method of claim 6, wherein, The process of training the deep learning network using the scan cycles after assigning class labels includes: Divide all valid echo signal scan cycles into a training subset and a validation subset according to a preset ratio; Divide all noise interference scan cycles into the training subset and the validation subset according to the same preset ratio; Iteratively train and validate the deep learning network using the training subset and the validation subset.

8. The lidar return signal processing method of claim 7, wherein, It also includes a feedback optimization step: Collect the echo signal sequence of a new scan cycle in real time, input the echo signal classification model, and output the prediction result; When the prediction result is a noise interference scan cycle, extract the actual signal quality evaluation value of the scan cycle; Calculate the deviation between the prediction result and the actual signal quality evaluation value; Generate an adaptive compensation factor based on the deviation.

9. The lidar return signal processing method of claim 8, wherein, The feedback optimization step also includes: Record the deviation of a plurality of consecutive scan cycles to generate a historical deviation moving average; Combine the adaptive compensation factor and the historical deviation moving average to generate a dynamic correction coefficient; Adjust the calculation weight of the signal quality evaluation value by using the dynamic correction coefficient.

10. The lidar return signal processing method of claim 9, wherein, It also includes: Feed the dynamic correction coefficient back to the fusion process of the signal dynamic change amount and the spatial correlation difference; Update the parameters of the echo signal classification model based on the adjusted signal quality evaluation value.

Images