Laser radar multi-target ranging method and system based on adaptive waveform matching

Abstract

The invention discloses a laser radar multi-target ranging method and system based on adaptive waveform matching, and relates to the technical field of laser ranging, and the technical scheme is characterized in that the adaptive waveform matching mechanism proposed by the invention dynamically adjusts convolution kernel sampling points through a deformable convolutional network (DCN), and the adaptive waveform matching is realized; various nonlinear waveform distortions caused by electric field interference and complex target reflection characteristics can be actively adapted, and a customized matching template is generated for each input waveform. Equivalently, the system is endowed with the capability of eliminating counterfeit and keeping true in a strong electromagnetic interference environment, so that the robustness to special working conditions of ultra-high voltage hot-line work is fundamentally improved.

Description

Technical Field

[0001] This invention relates to the field of laser ranging technology, and more specifically, to a multi-target ranging method and system for lidar based on adaptive waveform matching. Background Technology

[0002] LiDAR ranging technology, as a core means of environmental perception, has significant application value in ultra-high voltage (UHV) live-line working robot systems. In this scenario, the LiDAR system needs to achieve precise ranging under complex conditions of strong electric field interference, dense distribution of power equipment, and high safety requirements. However, traditional LiDAR ranging methods face fundamental challenges at the technical principle level under the special operating conditions of UHV.

[0003] For lidar ranging technology, existing technologies include echo matching using predefined pulse-coded sequences, which relies on a template matching mechanism with fixed time intervals and energy ratios. However, when strong electric fields cause nonlinear waveform distortion or target reflection characteristics lead to waveform broadening, a severe mismatch occurs between the fixed template and the actual echo. This rigid matching-based technique cannot adapt to dynamically changing waveform characteristics. Furthermore, existing technologies also describe processing saturated waveforms using hybrid neural networks, which primarily focus on local waveform reconstruction but fail to address signal separation when multiple targets are densely distributed along the time axis. This method has a complex neural network architecture, and its processing does not consider the clustering characteristics of multiple physical target echoes in the time domain.

[0004] Therefore, there is an urgent need for a lidar ranging technology that can fundamentally solve the above-mentioned technical problems in order to meet the comprehensive requirements of anti-interference and multi-target resolution in ultra-high voltage live-line work. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a multi-target ranging method and system for lidar based on adaptive waveform matching. By applying an adaptive waveform matching mechanism and dynamically adjusting the sampling points of the convolution kernel through a deformable convolutional network (DCN), the system can actively adapt to various nonlinear waveform distortions caused by electric field interference and complex target reflection characteristics. It generates a customized matching template for each input waveform, giving the system the ability to distinguish between true and false signals in environments with strong electromagnetic interference, thus fundamentally improving its robustness to special working conditions of ultra-high voltage live-line operations.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: Firstly, a multi-target ranging method for lidar based on adaptive waveform matching is provided, including the following steps: The raw analog echo signals acquired by the lidar are digitized and preprocessed to generate standard digital signals; The standard digital signal is adaptively matched using a deformable convolutional network to generate an enhanced feature map, wherein the deformable convolutional network dynamically adjusts the waveform matching template to resist signal distortion. Peak detection is performed on the enhanced feature map to identify potential target echo locations and extract the corresponding waveform segments; Density clustering algorithm is used to perform cluster analysis on the detected peak points, separating the waveform segment into multiple independent target clusters; Multidimensional features are extracted for each target cluster, and the features are weighted and fused using a deep regression network to generate a preliminary distance estimate. The deep regression network is subjected to lightweight optimization, the final physical distance is calculated based on the clustering results, and structured multi-target ranging results are output.

[0007] Furthermore, the digitization and preprocessing of the raw analog echo signal acquired by the lidar includes: The original analog echo signal is discretized according to a preset sampling interval and converted into a digital signal sequence; The digital signal sequence is filtered and denoised to suppress noise interference; The amplitude of the denoised signal is normalized to bring it within a preset range.

[0008] Furthermore, the adaptive waveform matching processing of the standard digital signal using a deformable convolutional network includes: The standard digital signal is input as a one-dimensional feature map into a deformable convolutional network; The deformable convolutional network learns the local deformation offset of the waveform signal; The sampling position of the convolution kernel is dynamically adjusted according to the local deformation offset, deformable convolution operation is performed, and the enhanced feature map is generated.

[0009] Furthermore, the peak detection of the enhanced feature map includes: Local maxima in the enhanced feature map are identified based on a dynamic thresholding strategy, wherein the threshold in the dynamic thresholding strategy is adaptively determined based on the statistical characteristics of the signal. Centered on the detected peak point, a fixed-length waveform segment is extracted, preserving complete waveform context information.

[0010] Furthermore, the clustering analysis of the detected peak points using the density clustering algorithm specifically employs the DBSCAN algorithm, including: Set the neighborhood radius parameter based on the lidar system parameters and target distribution characteristics; The minimum number of points is set according to the target echo characteristics to identify the core points; Peak points are assigned to different target clusters based on density reachability.

[0011] Furthermore, the multidimensional features include: Signal strength characteristics are obtained by calculating the power ratio of the signal to the noise. Waveform morphology features are obtained by comparing the waveform with a preset template for similarity. The temporal distribution characteristics are obtained by calculating the temporal dispersion between the peak point and the cluster center. Waveform width characteristics are obtained by analyzing the width of the waveform's energy distribution.

[0012] Furthermore, the feature weighted fusion through a deep regression network includes: The multidimensional features of each target cluster are aggregated into a feature vector of fixed dimensions; The feature vector is nonlinearly transformed using a multilayer perceptron network to output a distance estimate. By optimizing network parameters through end-to-end training, the network can adaptively learn feature weight allocation strategies.

[0013] Furthermore, the lightweight optimization process for the deep regression network includes: The network model is pruned to remove redundant network connections and parameters; The network parameters are quantized to convert floating-point parameters into low-precision integer representations; The processing flow is organized using a parallel pipeline architecture.

[0014] Furthermore, the calculation of the final physical distance based on the clustering results includes: Calculate the statistical center value of the peak time index in each of the target clusters; Based on the statistical center value, system sampling interval, and speed of light, the final physical distance is calculated according to the time-of-flight principle.

[0015] Secondly, a lidar multi-target ranging system based on adaptive waveform matching is provided, including: The signal processing module is used to digitize and preprocess the raw analog echo signals acquired by the lidar to generate standard digital signals; The waveform matching module is used to perform adaptive waveform matching processing on the standard digital signal through a deformable convolutional network to generate an enhanced feature map, wherein the deformable convolutional network dynamically adjusts the waveform matching template to resist signal distortion. The peak detection module is used to perform peak detection on the enhanced feature map, identify potential target echo locations, and extract the corresponding waveform segments. The target clustering module is used to perform cluster analysis on the detected peak points using a density clustering algorithm, and to separate the waveform segment into multiple independent target clusters; The feature fusion module is used to extract multi-dimensional features for each target cluster and perform feature weighted fusion through a deep regression network to generate a preliminary distance estimate. The real-time processing module is used to perform lightweight optimization on the deep regression network, calculate the final physical distance based on the clustering results, and output structured multi-target ranging results.

[0016] Thirdly, a computer terminal is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the adaptive waveform matching-based multi-target ranging method for lidar as described in any one of the first aspects.

[0017] Fourthly, a computer-readable medium is provided having a computer program stored thereon, the computer program being executed by a processor to implement the adaptive waveform matching-based multi-target ranging method for lidar as described in any one of the first aspects.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. The adaptive waveform matching-based multi-target ranging method for lidar provided by this invention applies an adaptive waveform matching mechanism. By dynamically adjusting the sampling points of the convolution kernel through a deformable convolutional network (DCN), it can actively adapt to various nonlinear waveform distortions caused by electric field interference and complex target reflection characteristics. It generates a customized matching template for each input waveform, giving the system the ability to distinguish between true and false signals in a strong electromagnetic interference environment, and fundamentally improving the robustness to special working conditions of ultra-high voltage live-line work. 2. This invention introduces a temporal density clustering algorithm, which does not require a preset target number and can perform unsupervised clustering based on the natural density distribution of echoes along the time axis. Multiple reflection points originating from the same physical component (such as an insulator) will form a high-density region and be identified as a cluster; there are sparse intervals between different components. This method can clearly distinguish densely distributed electrical equipment, providing a reliable perception basis for achieving accurate obstacle avoidance and operation. 3. This invention employs an end-to-end deep regression network for intelligent fusion, which can learn the complex and nonlinear mapping relationship between multi-dimensional features such as signal-to-noise ratio, waveform similarity, temporal consistency, and echo width and the actual distance. Through network training, an adaptive fusion strategy can be automatically learned. For example, when strong electric field interference causes a general decrease in the signal-to-noise ratio of some echoes, more emphasis is placed on relying on the stability of waveform morphology features to improve estimation accuracy, thereby meeting the stringent requirements for ranging accuracy in ultra-high voltage live-line work. 4. The lightweight technology combination used in this invention removes model redundancy through pruning and converts FP32 calculations to UINT8 calculations through quantization, which can greatly reduce model complexity and computational load. This allows complex algorithms to run efficiently on resource-constrained embedded processors, ensuring that the system can process lidar data in real time and provide timely and continuous environmental distance information to the upper-level control system, ultimately meeting the ultimate requirements of ultra-high voltage live-line work for system response speed and reliability. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart from Embodiment 1 of the present invention; Figure 2 This is a system block diagram in Embodiment 2 of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0021] Example 1: A multi-target ranging method for lidar based on adaptive waveform matching, such as... Figure 1 As shown, it includes the following steps: S1: Digitize and preprocess the raw analog echo signal acquired by the lidar to generate a standard digital signal; S2: Adaptive waveform matching processing is performed on the standard digital signal through a deformable convolutional network to generate an enhanced feature map, wherein the deformable convolutional network dynamically adjusts the waveform matching template to resist signal distortion. S3: Perform peak detection on the enhanced feature map, identify potential target echo locations, and extract the corresponding waveform segments; S4: Density clustering algorithm is used to perform cluster analysis on the detected peak points, separating the waveform segments into multiple independent target clusters; S5: Extract multi-dimensional features for each target cluster, and perform weighted fusion of features through a deep regression network to generate a preliminary distance estimate; S6: Perform lightweight optimization on the deep regression network, calculate the final physical distance based on the clustering results, and output structured multi-target ranging results.

[0022] In step S1, the original analog echo signal is digitized and preprocessed, including: discretizing the original analog echo signal according to a preset sampling interval to convert it into a digital signal sequence; filtering and denoising the digital signal sequence to suppress noise interference; and normalizing the amplitude of the denoised signal to make it within a preset value range.

[0023] Signal input and preprocessing involves converting the raw, noisy analog voltage signal captured by the lidar receiver into a digital signal, providing high-quality input data for subsequent deep learning models. Its core effect is to preserve the true characteristics of the target signal to the greatest extent possible, while suppressing environmental noise and hardware-introduced interference, laying a solid foundation for the high-precision ranging of the entire system.

[0024] Step S11: Digitization Converting continuous analog signals into discrete digital signals.

[0025] ; in, The original echo voltage signal, i.e., the original analog echo signal, is time-dependent. A continuous function, usually measured in volts (V), contains information about the energy of the photons reflected by the target, but is also mixed with various noises.

[0026] The sampling interval is measured in seconds (s). It is the time interval between two adjacent samples, and its reciprocal is called the sampling frequency. According to the Nyquist sampling theorem, the sampling frequency... The signal frequency must be greater than twice that of the highest frequency component to recover the signal without distortion. In lidar, The selection of the smallest resolvable unit directly determines the distance measurement. : ,in It's the speed of light.

[0027] The sampling point index is an integer representing the sampling point index. One sampling point.

[0028] It is the total number of sampling points, and Together, they determine the length of the signal processing time window. : .

[0029] It is the digitized input signal sequence.

[0030] Precise This ensures that distance information can be accurately discretized, and that sufficient... This ensures that the complete echo waveform can be captured. The output of this step... It is the starting point for all subsequent algorithm processing.

[0031] Step S12: Preprocessing To improve signal quality, suppress or eliminate noise and interference introduced during signal acquisition and transmission, and to standardize data scale, thereby enhancing the robustness and convergence speed of subsequent deep learning models.

[0032] Step S121: Noise Reduction This method filters out high-frequency random noise and power frequency interference from the signal, smooths the waveform, and makes the peak values ​​and other characteristics of the true target echo more prominent. In particular, under ultra-high voltage live-line working conditions, the preprocessing algorithm needs to additionally consider and suppress periodic narrow-pulse interference that may be caused by strong electric fields, as this interference characteristic can be confused with the target echo. This invention analyzes the interference spectrum characteristics and adds corresponding band-stop characteristics to the filter core design to enhance signal purity in this special environment.

[0033] Linear or nonlinear filtering techniques such as Gaussian filtering or wavelet denoising are employed. Gaussian filtering is an effective smoothing filter that can preserve the edge (peak) characteristics of a signal relatively well.

[0034] ; in, The amplitude of the original echo signal at the nk-th sampling point; For the filter kernel index; These are the weights for the filter kernel. For Gaussian filtering, It follows a Gaussian distribution (normal distribution), that is... ,in This is the standard deviation that controls the smoothness. The weighting coefficients usually need to be normalized to ensure... Weights are concentrated in Nearby, the farther a point is from the center, the smaller its weight. This is the half-width of the filter kernel. The total length of the filter kernel is... . The larger, Generally, the larger the value, the stronger the smoothing effect, but this can cause waveform details (such as two adjacent peaks) to become blurred. Therefore, The choice needs to strike a balance between noise reduction effect and signal fidelity. This is the denoised signal sequence.

[0035] After filtering, random spikes in the signal are effectively suppressed, the waveform becomes smooth, the signal-to-noise ratio is improved, and subsequent operations such as peak detection are facilitated.

[0036] Standardization of step S122: Eliminating differences in overall signal amplitude caused by factors such as varying target reflectivity and distance, and standardizing the data scale to a fixed range, such as [0,1], can prevent the model from being difficult to train due to inconsistent input data dimensions, accelerate convergence, and improve the model's generalization ability. Mini-maximum standardization is used.

[0037] ; This formula linearly maps the signal amplitude to... The interval. and These are the minimum and maximum values ​​of the denoised signal sequence, respectively.

[0038] Regardless of the original signal strength, the standardized signal has a uniform amplitude range, allowing deep learning models to focus more on learning waveform features such as shape and timing, rather than absolute strength.

[0039] The final output of this step It is a standardized discrete echo signal, which will serve as the direct input to the core innovation—the adaptive waveform matching module.

[0040] The quality of the signal input and preprocessing output directly determines the performance ceiling of the subsequent "adaptive waveform matching" module. Even with a powerful matching network, it's difficult to achieve accurate feature extraction from a noisy, baseline-drifted raw signal. This step, through effective digitization and preprocessing, ensures that subsequent algorithms can achieve maximum efficiency, ultimately working together to achieve high-precision, highly robust multi-target ranging.

[0041] In step S2, adaptive waveform matching processing of the standard digital signal is performed through a deformable convolutional network, including: inputting the standard digital signal as a one-dimensional feature map into the deformable convolutional network; learning the local deformation offset of the waveform signal through the deformable convolutional network; dynamically adjusting the sampling position of the convolutional kernel according to the local deformation offset, performing deformable convolution operation, and generating an enhanced feature map.

[0042] The purpose of adaptive waveform matching is to replace the fixed matched filter used in traditional lidar ranging systems. In complex real-world applications, the laser echo waveform undergoes various unpredictable distortions, broadening, and fracturing due to target material, incident angle, atmospheric conditions (such as rain, fog, and dust), and the system's own characteristics. In ultra-high voltage live-line working scenarios, waveform distortion may also originate from the instantaneous influence of strong electric fields on photoelectric detection devices, as well as minute changes in refractive index caused by ionized air around the target (such as high-voltage conductors). Traditional matched filters use a fixed "ideal echo" template, and their performance degrades sharply when the actual echo mismatches with the template. The deformable convolutional network (DCN) introduced in this step acts like an intelligent template with "shape adaptation capabilities." It can dynamically adjust its "perception posture" to actively match the actually received, potentially distorted waveform, thereby extracting the essential features of the target signal more accurately from noise. The effect of this step is to significantly improve the system's tolerance to waveform distortion in complex environments.

[0043] Step S21: Network Input and Feature Map Initialization The preprocessed, standardized, and cleaned one-dimensional time-series signal Prepare a format suitable for deformable convolutional networks and initialize the network's forward propagation process.

[0044] Step S211, Input Preparation: Input the length of... Discrete signal sequences Consider it as a height of 1 and a width of Feature map with 1 channel.

[0045] Step S212, Network forward propagation starts: The pre-trained lightweight DCN begins to work. Unlike convolutional neural networks (CNNs), the DCN of this invention learns the required geometric deformation offset at each position of the input signal through a parallel auxiliary network before convolution calculation. The auxiliary network can be a fully connected layer or a convolutional layer.

[0046] This step marks the beginning of activating DCN (Deformable Convolutional Networks). Instead of passively scanning the signal with fixed convolutional kernels, the network begins analyzing the local characteristics of the input signal, preparing for subsequent adaptive convolution operations. Furthermore, this step completes the interface conversion of data from the traditional signal processing domain to the deep learning domain and initiates the deformation-aware process of the DCN, preparing the data and parameters for the core deformable convolution operation.

[0047] Step S22: Deformable Convolution Operation By performing the core operation of deformable convolution, the offset learned by the network is used to enable the sampling points of the convolution kernel to adaptively shift according to the local features of the input signal, thereby achieving dynamic and intelligent waveform matching.

[0048] For each position m on the output feature map, its value is calculated using the following deformable convolution formula: ; in, To output the feature map at the th The response value at each location (time point). The magnitude of this value reflects the degree of matching between the intelligent template learned by the network and the actual signal in that region of the input signal. The larger the value, the higher the matching degree, which is a strong indication of the presence of the target.

[0049] This refers to the size (width) of the convolution kernel. It defines the number of input signal points that participate in the computation during a single convolution operation, i.e., the length of the template. For example, This indicates the use of a 5-point wide convolution kernel.

[0050] For the convolution kernel in the th The weight parameters at each position. These weights are learned through the backpropagation algorithm during network training and essentially define the ideal waveform or feature pattern that the network seeks. They can be viewed as an optimizable template.

[0051] A predefined, fixed offset within the convolution kernel. For one-dimensional convolution, typically... For example, when hour, Possibly This ensures that the convolution kernel is in its current position. Sampling is performed centered on the sample.

[0052] Learned by the network, and related to the current input The relevant adaptive offset. It is a real number, which can be decimal, so the sampling points of the convolution kernel can fall on non-integer (sub-pixel) positions, and the signal value at that position is obtained through bilinear interpolation. This makes the convolution kernel no longer a regular grid, but able to dynamically adjust its sampling position according to the shape deformation of the input waveform, thereby accurately aligning and matching those distorted echo waveforms.

[0053] Step S22, the essence of the operation is a variable cross-correlation calculation. This is achieved by introducing... The flexibility of the convolution kernel is enhanced. For example, when the echo waveform is broadened, the DCN can learn appropriate... This expands the convolution kernel accordingly to maintain a high correlation response. This enhances robustness to waveform distortion and produces a sharper, more accurate peak response compared to a fixed convolution kernel.

[0054] Step S2, the deformable convolution operation, is the core of this innovative algorithm. It introduces a learnable offset. This gives the model the ability to dynamically adjust its receptive field, achieving a leap from fixed template matching to adaptive template matching, and significantly improving the ability to extract features from distorted waveforms.

[0055] Step S23: Feature Map Generation and Output Complete the entire input signal The scanning and convolution calculations generate the final high-dimensional feature map. This will be the output of this step.

[0056] Step 231, Full image scan: Apply the deformable convolution operation from step S22 to all valid locations of the input signal. (Boundary padding usually needs to be considered), calculate the output feature map. every point The value of .

[0057] Step S232, Feature map characteristics: generated feature map It is a one-dimensional sequence of the same length as the input signal (or slightly shorter, depending on padding and step size). In this feature map: High response region: This corresponds to the region in the input signal that highly matches the target features learned by the network, i.e., the location of the potential target echo.

[0058] Low response region: Corresponds to noise or background region.

[0059] Step S233, Nonlinear activation: In actual networks, convolution output A non-linear activation function (such as ReLU) is typically used to introduce a non-linear transformation, thereby enhancing the expressive power of the model.

[0060] Feature map after DCN processing Its signal-to-noise ratio compared to the original input Further improvements were achieved. The peaks of the target signal became more prominent and concentrated, while noise and irrelevant responses were effectively suppressed. This provided a high-quality input for the next step (peak detection), directly reducing the probability of missed and false detections.

[0061] Step S23 generates an output feature map containing rich and enhanced target feature information. This feature map can be viewed as the result of the original signal being cleaned and enhanced, with the target location information significantly highlighted.

[0062] Step S2, adaptive waveform matching, is the first advanced intelligent processing module in the entire signal processing chain, and its output feature map... The quality of this step directly determines the performance ceiling of all subsequent modules. It provides a clear and separable peak sequence for "Step S4: Multi-target Signal Separation and Clustering" and a high-quality feature source for "Step S5: Intelligent Feature Weighted Fusion". The successful implementation of this step ensures that the entire system can maintain stable and reliable high-precision ranging capabilities when facing complex and ever-changing real-world environments, and is the key to achieving the two beneficial effects of "stronger environmental adaptability" and "higher ranging accuracy" of the solution.

[0063] In step S3, peak detection is performed on the enhanced feature map, including: identifying local maxima points in the enhanced feature map based on a dynamic threshold strategy, wherein the threshold in the dynamic threshold strategy is adaptively determined according to the signal statistical characteristics; and taking the detected peak points as the center, truncating waveform segments of fixed length, while retaining complete waveform context information.

[0064] Peak detection and target signal generation primarily transform the enhanced feature map generated by the adaptive waveform matching step in the previous step into specific target candidate points. The purpose of this step is to accurately locate each peak that might represent a physical target from the continuous signal waveform and to completely extract these peaks and their surrounding waveforms, providing source data for subsequent multi-target separation and accurate ranging. The effect of this step is to generate a set of high-quality candidate target segments containing complete waveform information. The completeness (no missed detections) and accuracy (no false detections) of its detection directly determine the number and quality of targets that the entire system can ultimately identify and range. If adaptive waveform matching is about highlighting target signals, then this step is about selecting these highlighted signals.

[0065] For targets with small dimensions and large differences in reflective cross-section, such as insulator discs, that need to be detected during ultra-high voltage and extra-high voltage live-line work, this invention employs an adaptive dynamic threshold strategy in this step. The threshold calculation not only considers the global noise floor but also incorporates the signal statistical characteristics within a local window, thereby specifically improving the system's detection sensitivity for small targets with weak reflections, while avoiding excessive false alarms due to excessively low thresholds.

[0066] Step S31: Peak Detection Feature map generated in step S2 The system automatically identifies all local maxima, which are signal points that are higher than their left and right neighbors in a specific neighborhood, and determines these points as markers of the arrival time of potential target echoes.

[0067] The peak detection process is a rule-based dual-criteria screening mechanism that ensures that only significant and prominent peaks are selected.

[0068] Local maxima criterion: First, a candidate point must be the highest point in its immediate neighborhood. This is a fundamental condition to ensure that what is detected is the "peak" of the waveform rather than a "valley" or "slope".

[0069] Dynamic threshold criterion: Secondly, the height of the peak value must exceed a dynamically calculated threshold. This threshold is not a fixed value, but is adaptively calculated based on the global or local statistical characteristics of the signal (such as the level of background noise). For example, it can be set to the noise mean plus 3 to 5 times its standard deviation. This can effectively filter out small fluctuations that are actually noise, even if they are local extrema.

[0070] Candidate peak point set Defined by the following formula: ; in, For the first Each peak point in the feature map The time index (position) is an integer. This index directly corresponds to half of the light pulse's flight time and is used for subsequent calculations of the target distance. The core raw data. Among them... It is the sampling time interval. It's the speed of light. For the first The amplitude value of each peak point, i.e. This amplitude value characterizes the strength of the signal response at that location, reflecting the energy of the echo, and is typically related to factors such as target reflectivity and distance. In subsequent processing, the amplitude can serve as an important feature for measuring the reliability of this peak point, for example, in signal-to-noise ratio estimation or weighting. Local maximum condition. Indicates position The immediate preceding ( ) and immediately following ( The amplitude of the echo is considered. This condition requires that the candidate point be higher than its immediate left and right neighbors to ensure that a sharp peak is detected. For wider echoes, a larger neighborhood range (such as two points on each side) is sometimes considered.

[0071] This is the dynamic detection threshold. This is a crucial parameter, directly impacting detection sensitivity and false alarm rate. Setting it too high may cause weak targets (such as distant or low-reflectivity targets) to be missed; setting it too low may misclassify noise as a target, increasing false alarms. Its dynamic nature is reflected in its ability to be calculated in real-time based on the noise floor of the current signal frame or local signal, for example... ,in and These are the estimated mean and standard deviation of the background noise. It is an adjustable scaling factor.

[0072] The peak detection step uses both local salience and absolute salience criteria to accurately filter out all points of interest that may represent the target from a continuous waveform. Its output... It is an initial target list containing target location and intensity information, which completes the conversion from analog signal features to discrete target points.

[0073] Step S32: Signal interception For each peak point detected in step S31 Centered on the feature map A fixed-length waveform segment is extracted. The purpose is not only to preserve the positional information of the peak point, but more importantly, to preserve the complete waveform context information surrounding the peak point.

[0074] Importance of contextual information: The echo of a light spot is not an ideal impulse pulse, but rather has a certain width and shape. Knowing the peak position... Knowing only partial information, and extracting waveform fragments This contains complete features. Features such as the shape, width, and symmetry of waveform segments are crucial for subsequent steps (such as waveform similarity calculation and distinguishing closely adjacent targets).

[0075] Extraction operation: by each Take as the center, and take forward and backward respectively. points, forming a length of . subsequence.

[0076] The obtained waveform segment The definition is as follows: ; in, For the first A vector of waveform segments centered at each peak point. It is a ( A 1 / 2-dimensional vector that contains complete waveform information of the peak point and its left and right neighbors.

[0077] This is used to truncate half the length of the window. This is a preset hyperparameter. The size needs to be carefully selected: if If the size is too small, the captured waveform segment will be too short and may not contain the complete echo waveform features, resulting in the loss of useful shape information.

[0078] if If the size is too large, the fragment will contain too much irrelevant background signal, which may introduce interference from other targets or additional noise, and will also increase the amount of computation.

[0079] In this invention The settings should be based on the pulse width of the lidar system and the expected echo broadening, ensuring coverage of the effective energy range of most target echoes. For example, when the laser hits a slope or rough object, the echo will broaden. If... If the setting is too rigid (exactly equal to the transmitted pulse width), then when encountering complex targets, the edges of the echo will be truncated, causing the DCN to be unable to extract complete "body features." Therefore, The value of must be such that The value should be no less than 1.5 times the laser pulse half-width and no more than 3 times the laser pulse half-width.

[0080] The signal truncation step reconstructs discrete peak points into continuous waveform segments. These waveform segments... It serves as the foundational data unit for subsequent, more refined analyses (such as waveform clustering and feature fusion).

[0081] Step S3 processes the enhanced feature map output from step S2, producing a series of clear target proposals (waveform segments). The quality of these proposals (i.e., the accuracy of peak detection and the completeness of the waveform segments) is directly input into the "multi-target signal separation and clustering" module in step four. If this step misses a peak (e.g., misclassifying two adjacent target peaks as one), step four will be unable to separate them correctly; if this step generates a large number of false alarm peaks caused by noise, it will increase the clustering burden and misjudgment risk of step four. Finally, this set of candidate target waveform segments... It is then sent to the next stage for deeper intelligent analysis.

[0082] In step S4, a density clustering algorithm is used to perform cluster analysis on the detected peak points. Specifically, the DBSCAN algorithm is used, which includes: setting the neighborhood radius parameter according to the lidar system parameters and target distribution characteristics; setting the minimum number of points parameter according to the target echo characteristics to identify core points; and dividing the peak points into different target clusters based on density reachability.

[0083] In complex scenarios, a single laser pulse may simultaneously illuminate multiple targets at different distances with varying reflection characteristics, or reflect back from multiple parts of an extended target, forming multiple echo peaks that are closely adjacent or partially overlapping in time. In the candidate peak set generated in step three, these echoes originating from the same physical target naturally cluster together in the time dimension. In ultra-high voltage live-line work, a typical scenario involves a single laser beam simultaneously illuminating different points on a high-voltage conductor, multiple ceramic plates on an insulator string, or the conductor and adjacent grading rings. These targets are very close in distance, resulting in a high density of echo peaks in time.

[0084] The purpose of this step is to utilize the distribution characteristics of these echo peaks along the time axis and, through advanced clustering algorithms, automatically and accurately classify them into their respective physical target categories. This step achieves precise separation of multi-target signals, grouping a series of discrete peak points into clusters, laying a solid foundation for subsequent independent distance estimation of each target, and is crucial for the entire system to process multiple targets in parallel.

[0085] Step S41: Data Preparation and Measurement Definition , .

[0086] Prepare input data for the clustering algorithm and define metrics for similarity or distance between data points. This is fundamental to cluster analysis, ensuring the algorithm can correctly understand the relationships between data points.

[0087] Step S411 Input Data: The direct input for this step is the set of candidate peak points output from step three. For clustering tasks, the primary focus is on the time index of each peak point. This is because multiple reflected echoes originating from the same physical target will have very similar arrival times (i.e., time indices), which is the fundamental basis for classifying them into one category.

[0088] Step S412 Distance Measurement: Clustering algorithms need to quantify the distance or similarity between points. In this invention, two peak points... and The similarity between them is determined by the proximity of their time indices. Therefore, Euclidean distance is used as the distance metric. ; This distance The unit is the number of sampling points, which directly reflects the time difference between two echoes. The smaller the value, the closer the two echo peaks are in time, and the greater the likelihood that they originate from the same target.

[0089] This step clarifies that the input to the clustering algorithm is the peak time index sequence. It also defines the absolute value of the time difference as the distance metric. This prepares the data for the application of the DBSCAN algorithm, ensuring that the algorithm can perform clustering based on the temporal proximity of the echoes.

[0090] Step 4.2: DBSCAN Clustering Execution The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm is executed to divide peak points into different clusters based on their density distribution characteristics on the time axis, and to identify sparsely distributed noise points.

[0091] DBSCAN is a density-based clustering algorithm that can group density-connected points into clusters, discover clusters of arbitrary shapes, and effectively handle noisy points. Its execution relies on two core parameters and concepts: Neighborhood: Given an object (peak point) ,That The neighborhood is defined as... Centered on, with radius The set of all peak points within the time interval, i.e. In response to the dense target characteristics in ultra-high voltage live-line work, this invention sets the neighborhood radius parameter ϵ of DBSCAN to a relatively small value to ensure that echoes that are very close in time but belong to different physical components (such as adjacent insulator pieces) can be effectively separated, avoiding undersegmentation.

[0092] Core point determination: If a peak point of The number of points contained in the neighborhood (including itself) reaches at least a certain minimum value. If a point is located within a cluster (dense region), then that point is called the core point.

[0093] ; Density directly connected to density: if a point Located at the core point of Within the neighborhood, it is called from Density directly reaches the target. If a point chain exists... ,in , And each from If the density reaches directly, it is called... and Density-connected points. All density-connected points form a cluster.

[0094] Noise points: Points that do not belong to any core point Points in the neighborhood that are not themselves core points are marked as noise points. These points are usually isolated false alarm peaks.

[0095] in, is the neighborhood radius. This is the distance threshold used to determine whether two points are adjacent. The choice is crucial: if If it is too small, it may not be able to group multiple echoes that belong to the same target but are slightly dispersed in time into one category, causing a target to be split into multiple small clusters.

[0096] if If the size is too large, it may mistakenly merge echoes from different but closely spaced targets into a single cluster.

[0097] In this invention The settings need to be selected empirically based on the angular resolution of the lidar, the detection range, and the possible spatial distribution range of the target. Specifically, The value of is generally negatively correlated with angular resolution, detection distance, and spatial distribution range (specifically, the radius of cluster search). Finally, an adaptive selection is made based on application requirements, without specific restrictions.

[0098] The minimum number of neighborhood points required for the core point to form a cluster. Meanwhile, considering that some components (such as a single conductor) may only generate the main echo, while an insulator string may generate multiple echoes, The number can be set to a low value (such as 2), and allows for target clusters with only one echo point to accommodate scene diversity.

[0099] if If the value is too small, noise points may be misidentified as core points, forming meaningless small clusters.

[0100] if If the signal is too large, some small targets that only produce a few echoes may not be able to form clusters and may be misjudged as noise.

[0101] In lidar ranging, a target typically produces more than one reflection (especially extended targets), therefore It can usually be set to a small integer greater than 1 (such as 2 or 3).

[0102] DBSCAN clustering is based on density principles, automatically dividing densely distributed peak points on the time axis into different clusters, each cluster corresponding to a physical target. It does not require pre-specifying the number of clusters (M), can handle clusters of arbitrary shapes, and effectively identify and filter noise points, exhibiting strong robustness.

[0103] Step S43: Cluster Output and Noise Processing Organize and output the clustering results of the DBSCAN algorithm, and properly handle the points marked as noise.

[0104] Cluster output: After the DBSCAN algorithm completes execution, the following results are obtained: Target clusters Each cluster Contains several peak points These points are believed to originate from the same physical target. This is the time index of the first peak point within cluster j. This is the time index of the second peak point within cluster j. The echo intensity of the first peak point within cluster j. The echo intensity is the second peak point within cluster j.

[0105] Noise Point Handling: Isolated peaks marked as noise by DBSCAN are typically removed directly. This is because these points are likely false alarms caused by random noise, shot noise, or other non-target factors. Removing them purifies the candidate target list, reducing the computational burden and risk of misjudgment in subsequent steps. In highly demanding systems, secondary evaluation of these noise points or filtering using more complex rules can also be performed.

[0106] Step S43 ultimately outputs a clear and well-separated cluster of multiple targets, filtering out unreliable noise points. This marks the completion of the multi-target signal separation task.

[0107] Step S4 receives the original candidate target proposals generated in step S3 and outputs clearly separated target signal clusters that correspond one-to-one with the physical targets through intelligent clustering. The quality of this separation directly determines the input quality of step S5, "intelligent feature weighted fusion". If the clustering is accurate, step S5 can independently and accurately estimate the distance for each target; if the clustering is incorrect (e.g., different targets are mixed up, or the same target is fragmented), the fusion result of step S5 will inevitably be biased. Therefore, this step ensures that the system can clearly distinguish multiple closely adjacent objects in the scene, significantly improving the perception capability of the LiDAR in complex environments.

[0108] In step S5, the multidimensional features include: signal strength features, obtained by calculating the power ratio of signal to noise; waveform morphology features, obtained by comparing the similarity of the waveform with a preset template; time distribution features, obtained by calculating the time dispersion of peak points and cluster centers; and waveform width features, obtained by analyzing the energy distribution width of the waveform.

[0109] Feature weighting and fusion through deep regression networks includes: aggregating the multidimensional features of each target cluster into a fixed-dimensional feature vector; performing nonlinear transformation on the feature vector through a multilayer perceptron network to output a distance estimate; and optimizing the network parameters through end-to-end training to enable the network to adaptively learn the feature weight allocation strategy.

[0110] Intelligent feature weighted fusion is the core algorithmic step in this invention to achieve the beneficial effect of "higher ranging accuracy". After the preceding steps, multiple separated target clusters have been obtained, each containing several waveform segments and their peak information generated by the same physical target. Traditional methods typically use simple weighted averaging (such as signal-to-noise ratio weighting) or directly select the strongest peak to estimate the distance. This method fails to fully utilize the rich information contained in the waveform segments, and the fixed weighting strategy cannot adapt to complex and ever-changing real-world scenarios. The purpose of this step is to construct a lightweight deep regression network to learn, end-to-end, the complex nonlinear mapping relationship from the multi-dimensional feature set of each target cluster to the optimal distance estimate. Its core effect is to implement an "intelligent" feature fusion mechanism that can automatically evaluate the reliability of each waveform segment within the cluster (e.g., based on signal-to-noise ratio, waveform quality, time consistency, etc.) and perform optimal weighted fusion based on this, thereby significantly improving the accuracy and robustness of single-target distance estimation.

[0111] Step S51: Multidimensional Feature Extraction For each target cluster Each waveform segment included A set of feature metrics that comprehensively characterize the signal quality, waveform characteristics, and temporal reliability are calculated. These features will serve as inputs to the subsequent intelligent fusion model and will be the basis for the model's decision-making.

[0112] This step quantizes and represents waveform segments from three key dimensions, constructing feature vectors. .

[0113] Step S511 Signal-to-noise ratio characteristics characterize signal strength reliability This quantifies the relative strength of the effective signal power to the background noise power in the waveform segment. A higher signal-to-noise ratio generally indicates a higher likelihood that the segment is generated by a reflection from a real target, and therefore, it should have a greater weight in distance estimation.

[0114] ; in, For the first The first target cluster A vector of waveform segments.

[0115] The sum of squares of all components of the waveform segment vector represents the total energy or power of the segment. Here, the square of the L2 norm is used to approximate the signal power.

[0116] This is an estimate of the background noise power. This value can usually be obtained statistically from targetless regions of the signal (such as the baseline portion far from the peak), and is a global or inter-frame estimate.

[0117] The signal-to-noise ratio is calculated and expressed in decibels (dB). A higher value indicates better quality for the waveform segment.

[0118] Step S512 Waveform similarity features characterize the reliability of waveform morphology. Measure the current waveform segment With a preset "ideal" or "typical" echo template Similarity in shape. In ultra-high voltage live-line working applications, templates for various typical targets can be pre-stored, such as "conductor templates," "insulator templates," and "metal fitting templates." Waveform similarity calculation can be performed in parallel with multiple templates. The highest similarity value and the corresponding template type can serve as important features, not only for assessing reliability but also for providing auxiliary information for subsequent target identification. The higher the waveform similarity, the more closely the segment matches the system's expectation of the target echo, the less affected by distortion, and the higher the reliability.

[0119] ; in, is the vector of the waveform segment to be evaluated. This is a predefined template waveform vector. It typically represents the ideal echo response of the system measured under standard conditions. The inner product of two vectors reflects their linear dependence. , These are the L2 norms (modulus) of the two vectors. The cosine similarity is calculated with a value ranging from -1 to 1. The closer the value is to 1, the more similar the two waveforms are.

[0120] Step S513: Timing consistency characteristics characterize the reliability of time positions. Evaluate the peak timestamp of this waveform segment Relative to its cluster The degree of dispersion of the central time location. The closer the timestamp of a peak point is to the center of the cluster, the greater the likelihood that it is the main echo of that target, and the higher the reliability of its time information.

[0121] ; in, This is the time index of the peak point of the waveform segment. For clusters The mean of the time indices of all peak points within the period. This represents the approximate center time position of the target echo. For clusters The standard deviation of the time index of all peak points within the range, This characterizes the degree of dispersion in the time distribution of the target echo. The calculated consistency index takes a value range of (0, 1). At a time when the center is reached, the value reaches its maximum of 1; the further away from the center, the smaller the value becomes. This exponential function form maps the time difference to a smoothly decaying weighting coefficient.

[0122] The feature extraction step is for each waveform segment A three-dimensional feature vector was constructed. This vector comprehensively describes the reliability and value of this segment for distance estimation from three orthogonal dimensions: signal strength, waveform quality, and temporal concentration. For the entire cluster... This yielded a set of feature vectors. .

[0123] Step S514 (Adding a new feature dimension): Echo width feature, characterizing target size properties. In ultra-high voltage live-line work, the echo widths of conductors (slender) and insulator discs (smaller) may differ. This invention adds an echo width feature, calculates the full width at half maximum (FWHM) of a waveform segment or the echo energy concentration, as an auxiliary feature describing the target size characteristics. Incorporating this feature into the fusion network helps improve the ability to distinguish densely packed components.

[0124] ; in, Echo width characteristics, This represents the full width at half maximum (FWHM) function used to calculate the waveform. This feature, along with signal-to-noise ratio (SNR), waveform similarity, and other parameters, constitutes an enhanced feature vector. .

[0125] Step S52: Deep Regression Network Fusion Using a lightweight deep neural network (such as a multilayer perceptron MLP), the set of feature vectors extracted in the previous step... As input, the optimal distance estimate corresponding to the target cluster is directly regressed. This step aims to learn a more intelligent, non-fixed-formula feature-weighted fusion strategy using a data-driven approach.

[0126] Step S521, Network Input Processing: A Cluster Include There are waveform segments, therefore there are There are several feature vectors. These feature vectors need to be integrated into a fixed-dimensional representation before being input into the network. A simple approach is to first concatenate all feature vectors into a long vector, but a more common and effective method is to first perform pooling operations (such as calculating the mean and maximum) on each feature dimension within the cluster, forming a fixed-dimensional comprehensive feature vector that summarizes the statistical characteristics of the entire cluster, and then input it into the network.

[0127] Step S522, Network Mapping: ; in, It is a cluster Feature set, The fixed-dimensional feature representation obtained after some kind of aggregation. It is the set of learnable parameters for deep regression networks (such as MLPs), including weights and biases. Indicates Nonlinear function mapping implemented by a multilayer perceptron with parameters. The scalar value output by the network, i.e., for the th... Distance estimates for each target.

[0128] Step S523 End-to-end training: Network parameters It is learned end-to-end through a large amount of labeled data (i.e., the input echo signal and the actual target distance). The training objective is usually to minimize the network's output distance. Distance from reality The loss between features (such as mean squared error, MSE). Through this process, the network not only learns how to interpret the meaning of each feature, but more importantly, it learns how to dynamically and non-linearly determine the "contribution weight" of each waveform segment to the final distance estimate based on the combination of these features. For example, the network may learn that when a segment has a high signal-to-noise ratio and waveform similarity, it should be given a higher weight even if its temporal consistency is slightly poor.

[0129] The deep regression network fusion step leverages the powerful representation learning capabilities of neural networks to replace traditional hand-designed fixed-weight formulas. It can automatically learn the complex, environment-dependent relationships between features and distances, achieving an adaptive and intelligent fusion that promises superior ranging accuracy compared to any fixed-weight strategy.

[0130] Step S5 relies on the output quality of the preceding steps: the good clustering results provided in step S4 are a prerequisite for this step to accurately fuse data independently for each target. This step partially transforms the ranging problem from the traditional signal processing domain into a feature-based machine learning regression problem. This paradigm shift allows the system to utilize data-driven methods to optimize its performance ceiling. Finally, this step outputs a preliminary distance estimate for each target. This estimate has already undergone intelligent optimization. Subsequently, these results will be sent to the final step for lightweight post-processing and output, completing the entire intelligent sensing process from laser echo to multi-target distance information.

[0131] In step S6, the deep regression network is subjected to lightweight optimization, including: pruning the network model to remove redundant network connections and parameters; quantizing the network parameters to convert floating-point parameters into low-precision integer representations; and organizing the processing flow using a parallel pipeline architecture.

[0132] The final physical distance is calculated based on the clustering results, including: calculating the statistical center value of the peak point time index in each target cluster; and calculating the final physical distance based on the statistical center value, the system sampling interval, and the speed of light, according to the time-of-flight principle.

[0133] Step S6 employs a series of lightweight processing techniques to ensure the entire system meets the stringent real-time requirements of LiDAR, typically needing to process a single frame of data within tens of milliseconds. Secondly, it generates and outputs structured, highly available multi-target ranging results, providing direct and reliable environmental distance information to the perception and decision-making modules. The effect of this step is to achieve efficient deployment and operation of the algorithm on resource-constrained embedded platforms while maintaining extremely high ranging accuracy, ultimately achieving the core innovative goal of maintaining real-time processing capabilities.

[0134] For ultra-high voltage live-line working robots or personnel assistance systems, processing speed is directly related to operational safety; any delay can pose a risk. Therefore, the lightweight design and real-time performance of this step are crucial for the applicability of the solution in this scenario.

[0135] Step S61: Model lightweighting The deep learning models used in this invention (mainly the deformable convolutional network DCN in step two and the deep regression network MLP in step five) are compressed and accelerated to significantly reduce their computational complexity and memory usage while maintaining model accuracy as much as possible, enabling them to run efficiently on automotive embedded devices.

[0136] This step employs three mainstream model lightweighting techniques, creating a combined advantage: Step S611, Network Pruning By removing redundant weights, connections, and even entire neuron channels from a neural network, while retaining the most critical parameters, a sparser and more compact network model can be created. This is similar to pruning the branches and leaves of a tree to make it more aesthetically pleasing and healthier.

[0137] After model training is complete, the operation involves evaluating the importance of the weights (e.g., based on their absolute value or their contribution to the final output loss) and resetting or removing weights whose importance is below a preset threshold. Subsequently, the pruned model typically needs fine-tuning to recover any small amount of accuracy lost due to pruning.

[0138] This significantly reduces the number of model parameters and computational cost. A sparse model can leverage sparse computing libraries to accelerate core operations such as matrix multiplication and significantly reduce the memory space required for model storage.

[0139] Step S612, Quantization compression Reducing the precision of model weights and activation values ​​by converting them from the 32-bit floating-point numbers commonly used during training to a lower-precision format (such as 8-bit integers) is one of the most effective methods for model compression, directly reducing memory usage and accelerating computation.

[0140] The operation employs post-training quantization or quantization-aware training techniques. For the weight matrix... Each weight value in Remove it from the FP32 range Linearly mapped to the integer range [0, 255] of UINT8 ; in, The original full-precision weight values ​​are typically of type FP32 (single-precision floating-point). While FP32 offers high dynamic range and precision, it occupies 4 bytes of storage and is computationally resource-intensive; therefore, this invention employs mapping processing. , These are the minimum and maximum values ​​in the weight tensor, respectively. They are used to determine the quantization scale and zero point. To normalize the weight values ​​to the [0,1] interval. ×255: Maps the normalized values ​​to the integer range [0,255]. This is a rounding function that rounds the result to the nearest integer. The quantized weight value is an 8-bit unsigned integer that occupies only 1 byte of storage space.

[0141] The model size is reduced by about 75%, memory bandwidth requirements are reduced, and integer operations are much faster than floating-point operations on most hardware, thus significantly improving inference speed.

[0142] Step S613, Parallel Pipeline Design At the system architecture level, the entire signal processing flow (steps S1 to S6) is organized into a multi-stage pipeline, which allows the processing stages of different frame data to overlap, thereby maximizing the utilization of hardware resources and the overall throughput of the system.

[0143] The operation method designs the main modules, such as adaptive waveform matching, peak detection, clustering, and feature fusion, as independent processing units. When the... When performing "Step S5: Intelligent Feature Weighted Fusion" on the frame data, the first... The frame data can be processed simultaneously in "Step S4: Multi-target signal separation and clustering", while the second step... Frame data can then undergo "Step S2: Adaptive Waveform Matching", and so on.

[0144] Although the end-to-end processing latency of a single frame of data remains unchanged, the number of data frames that the system can process per unit time (throughput) is significantly improved, thereby meeting the real-time streaming processing requirements of high-frequency lidar (such as above 10Hz).

[0145] The lightweight model processing steps utilize three key technologies—pruning, quantization, and pipeline design—to deeply optimize the algorithm model and system architecture. This successfully "ports" computationally intensive deep learning algorithms to resource-constrained in-vehicle computing platforms, removing obstacles to real-time processing.

[0146] Step 62: Final Distance Calculation and Result Formatting The preliminary distance estimate output by the deep learning network in step S5 (Or directly calculate based on the cluster center timestamp), convert it into accurate physical distance, and organize the ranging results of all targets into a standardized data structure that is easy for upper-layer applications to parse and use.

[0147] Step S621, final distance calculation: Although the neural network in step S5 has already output a distance estimate, to ensure the accuracy of the physical meaning, a direct calculation method based on the time-of-flight principle is usually adopted. This utilizes the most reliable time information obtained after clustering—the cluster center time index. .

[0148] ; in, No. The center time index of the target cluster. Typically, the time indices of all peak points within the cluster are taken. The average or median represents the most likely time location of the target echo. The system's sampling time interval, measured in seconds, is a fundamental parameter of the ADC sampling system. It determines the time index. Compared with actual time Conversion relationship ( ). The speed of light in a vacuum or air is usually taken as... Because the distance measured by lidar is the two-way distance of the light pulse from emission to the target and back to the receiver, it needs to be divided by 2 to obtain the one-way target distance. The calculated first The final distance estimate for each target, in meters.

[0149] Step S622, format the output of the result: The multi-target ranging results are encapsulated into structured messages, which can be easily sent to upper-level modules such as perception and planning via vehicle networks (such as CAN FD and Ethernet).

[0150] Generate a data packet in a structure or JSON format, containing the following key fields: target_count(target quantity): Integer value , which represents the total number of targets detected in this frame of data.

[0151] targets (target list): An array where each element represents information about a target, typically including: id (target ID): A unique identifier within this frame.

[0152] distance: , in meters.

[0153] type_hint: An enumerated value or confidence vector that, based on the waveform similarity comparison results with multiple templates in step S5, indicates the possible component type of the target (e.g., wire, insulator, hardware, etc.), providing richer perception information for the upper-level system.

[0154] Confidence: A value between 0 and 1, representing the reliability of the ranging result. It can be calculated based on information such as the average signal-to-noise ratio of echoes within the cluster, waveform similarity, or the probability of neural network output.

[0155] timestamp: The timestamp at which data collection or processing is completed.

[0156] Example 2: A multi-target ranging system for lidar based on adaptive waveform matching. This system is used to implement the multi-target ranging method for lidar based on adaptive waveform matching as described in Example 1. Figure 2 As shown, it includes a signal processing module, a waveform matching module, a peak detection module, a target clustering module, a feature fusion module, and a real-time processing module.

[0157] The system comprises the following modules: a signal processing module for digitizing and preprocessing the raw analog echo signals acquired by the lidar to generate standard digital signals; a waveform matching module for adaptive waveform matching of the standard digital signals using a deformable convolutional network to generate enhanced feature maps, where the deformable convolutional network dynamically adjusts the waveform matching template to resist signal distortion; a peak detection module for peak detection of the enhanced feature maps to identify potential target echo locations and extract corresponding waveform segments; a target clustering module for clustering the detected peak points using a density clustering algorithm to separate waveform segments into multiple independent target clusters; a feature fusion module for extracting multi-dimensional features for each target cluster and performing weighted feature fusion using a deep regression network to generate preliminary distance estimates; and a real-time processing module for lightweight optimization of the deep regression network, calculating the final physical distance based on the clustering results, and outputting structured multi-target ranging results.

[0158] The present invention also describes a computer terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the multi-target ranging method of lidar based on adaptive waveform matching as described in Embodiment 1.

[0159] The present invention also describes a computer-readable medium having a computer program stored thereon, which, when executed by a processor, can implement the adaptive waveform matching-based multi-target ranging method for lidar as described in Example 1.

[0160] Working principle: (1) In ultra-high voltage live-line working environments, lidar echo signals face two major challenges: First, strong electric fields may introduce high-frequency pulse noise, causing instantaneous interference to the detector and resulting in atypical waveform distortion; second, the surface material and shape of the working targets (such as insulator strings, wires, and fittings) are complex, and there may be dirt or ice, leading to broadening, distortion, or multiple reflection peaks in the echo waveform. Traditional matched filters based on fixed laboratory templates are severely mismatched under such complex and non-standard waveform distortions, resulting in a sharp drop in the feature extraction signal-to-noise ratio, making it difficult to reliably detect the real target and posing a huge hidden danger to work safety.

[0161] The adaptive waveform matching mechanism proposed in this invention dynamically adjusts the sampling points of the convolution kernel through a deformable convolutional network (DCN), enabling it to proactively adapt to various nonlinear waveform distortions caused by electric field interference and complex target reflection characteristics, generating a customized matching template for each input waveform. This is equivalent to giving the system the ability to distinguish true waveforms from false ones in environments with strong electromagnetic interference, fundamentally improving its robustness to the special working conditions of ultra-high voltage live-line operations.

[0162] (2) In ultra-high voltage live-line working scenarios, laser beams often simultaneously irradiate multiple targets that are extremely densely packed in space, such as individual ceramic pieces in an insulator string, conductors, and adjacent equalizing rings or vibration dampers. The echoes of these targets are highly overlapping or adjacent on the time axis. Traditional peak detection methods based on fixed thresholds are prone to erroneously merging these echoes from different physical components into a single target (undersegmentation) or misjudging multiple reflection points of an extended target as multiple independent targets (oversegmentation). This error can directly lead to the robot's misjudgment of the environmental structure, potentially causing serious safety accidents such as mechanical collisions.

[0163] This invention introduces a temporal density clustering algorithm that does not require a preset target number and can perform unsupervised clustering based on the natural density distribution of echoes along the time axis. Multiple reflection points originating from the same physical component (such as an insulator) will form a high-density region and be identified as a cluster; while there are sparse intervals between different components. This method can clearly distinguish densely distributed electrical equipment, providing a reliable perception basis for achieving precise obstacle avoidance and operation.

[0164] (3) Ultra-high voltage live-line work requires extremely high absolute accuracy and stability in distance measurement (often requiring centimeter-level or even millimeter-level accuracy), as this is the direct basis for the robotic arm to perform delicate actions such as grasping and installation. Even if multiple echoes of a single target are successfully separated, traditional fusion methods using fixed weights (such as signal-to-noise ratio weighting) are difficult to cope with the optimal estimation requirements under complex waveform quality variations. For example, an echo with a slightly lower signal-to-noise ratio but extremely standard waveform shape may reflect the true distance better than an echo with a higher signal-to-noise ratio but slightly distorted. This complex, nonlinear trade-off relationship cannot be solved by fixed strategies.

[0165] This invention employs an end-to-end deep regression network for intelligent fusion, which can learn the complex and nonlinear mapping relationship between multi-dimensional features such as signal-to-noise ratio, waveform similarity, temporal consistency, and echo width and the actual distance. Through network training, an adaptive fusion strategy can be automatically learned. For example, when strong electric field interference causes a general decrease in the signal-to-noise ratio of some echoes, more emphasis is placed on relying on the stability of waveform morphology features to improve estimation accuracy, thereby meeting the stringent requirements for ranging accuracy in ultra-high voltage live-line operations.

[0166] (4) Ultra-high voltage live-line working robot systems have extremely strict limitations on the processing latency of sensing data. Any processing delay may lead to control command delay and cause safety accidents. Deploying advanced deep learning models on vehicle-mounted or airborne embedded platforms faces triple constraints of computing power, power consumption, and real-time performance. If complex models are not optimized, they will not be able to meet the real-time processing requirements of high frequency (usually >10Hz) and low latency (required to be far below 100ms), making it impossible for advanced algorithms to be transformed into actual safety assurance capabilities.

[0167] The lightweight technology combination used in this invention removes model redundancy through pruning and converts FP32 computation to UINT8 computation through quantization, which can greatly reduce model complexity and computational load. This allows complex algorithms to run efficiently on resource-constrained embedded processors, ensuring that the system can process lidar data in real time and provide timely and continuous environmental distance information to the upper-level control system, ultimately meeting the ultimate requirements of ultra-high voltage live-line work for system response speed and reliability.

[0168] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0169] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0170] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0171] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0172] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-target ranging method for lidar based on adaptive waveform matching, characterized in that, Includes the following steps: The raw analog echo signals acquired by the lidar are digitized and preprocessed to generate standard digital signals; The standard digital signal is adaptively matched using a deformable convolutional network to generate an enhanced feature map, wherein the deformable convolutional network dynamically adjusts the waveform matching template to resist signal distortion. Peak detection is performed on the enhanced feature map to identify potential target echo locations and extract the corresponding waveform segments; Density clustering algorithm is used to perform cluster analysis on the detected peak points, separating the waveform segment into multiple independent target clusters; Multidimensional features are extracted for each target cluster, and the features are weighted and fused using a deep regression network to generate a preliminary distance estimate. The deep regression network is subjected to lightweight optimization, the final physical distance is calculated based on the clustering results, and structured multi-target ranging results are output.

2. The multi-target ranging method for lidar based on adaptive waveform matching according to claim 1, characterized in that, The digitization and preprocessing of the raw analog echo signal acquired by the lidar includes: The original analog echo signal is discretized according to a preset sampling interval and converted into a digital signal sequence; The digital signal sequence is filtered and denoised to suppress noise interference; The amplitude of the denoised signal is normalized to bring it within a preset range.

3. The multi-target ranging method for lidar based on adaptive waveform matching according to claim 1, characterized in that, The adaptive waveform matching processing of the standard digital signal using a deformable convolutional network includes: The standard digital signal is input as a one-dimensional feature map into a deformable convolutional network; The deformable convolutional network learns the local deformation offset of the waveform signal; The sampling position of the convolution kernel is dynamically adjusted according to the local deformation offset, deformable convolution operation is performed, and the enhanced feature map is generated.

4. The multi-target ranging method for lidar based on adaptive waveform matching according to claim 1, characterized in that, The peak detection of the enhanced feature map includes: Local maxima in the enhanced feature map are identified based on a dynamic thresholding strategy, wherein the threshold in the dynamic thresholding strategy is adaptively determined based on the statistical characteristics of the signal. Centered on the detected peak point, a fixed-length waveform segment is extracted, preserving complete waveform context information.

5. The multi-target ranging method for lidar based on adaptive waveform matching according to claim 1, characterized in that, The density clustering algorithm used to perform clustering analysis on the detected peak points specifically employs the DBSCAN algorithm, including: Set the neighborhood radius parameter based on the lidar system parameters and target distribution characteristics; The minimum number of points is set according to the target echo characteristics to identify the core points; Peak points are assigned to different target clusters based on density reachability.

6. The multi-target ranging method for lidar based on adaptive waveform matching according to claim 1, characterized in that, The multidimensional features include: Signal strength characteristics are obtained by calculating the power ratio of the signal to the noise. Waveform morphology features are obtained by comparing the waveform with a preset template for similarity. The temporal distribution characteristics are obtained by calculating the temporal dispersion between the peak point and the cluster center. Waveform width characteristics are obtained by analyzing the width of the waveform's energy distribution.

7. The multi-target ranging method for lidar based on adaptive waveform matching according to claim 1, characterized in that, The feature weighting fusion through a deep regression network includes: The multidimensional features of each target cluster are aggregated into a feature vector of fixed dimensions; The feature vector is nonlinearly transformed using a multilayer perceptron network to output a distance estimate. By optimizing network parameters through end-to-end training, the network can adaptively learn feature weight allocation strategies.

8. The multi-target ranging method for lidar based on adaptive waveform matching according to claim 1, characterized in that, The lightweight optimization process for the deep regression network includes: The network model is pruned to remove redundant network connections and parameters; The network parameters are quantized to convert floating-point parameters into low-precision integer representations; The processing flow is organized using a parallel pipeline architecture.

9. The multi-target ranging method for lidar based on adaptive waveform matching according to claim 1, characterized in that, The calculation of the final physical distance based on the clustering results includes: Calculate the statistical center value of the peak time index in each of the target clusters; Based on the statistical center value, system sampling interval, and speed of light, the final physical distance is calculated according to the time-of-flight principle.

10. A multi-target ranging system for lidar based on adaptive waveform matching, characterized in that, include: The signal processing module is used to digitize and preprocess the raw analog echo signals acquired by the lidar to generate standard digital signals; The waveform matching module is used to perform adaptive waveform matching processing on the standard digital signal through a deformable convolutional network to generate an enhanced feature map, wherein the deformable convolutional network dynamically adjusts the waveform matching template to resist signal distortion. The peak detection module is used to perform peak detection on the enhanced feature map, identify potential target echo locations, and extract the corresponding waveform segments. The target clustering module is used to perform cluster analysis on the detected peak points using a density clustering algorithm, and to separate the waveform segment into multiple independent target clusters; The feature fusion module is used to extract multi-dimensional features for each target cluster and perform feature weighted fusion through a deep regression network to generate a preliminary distance estimate. The real-time processing module is used to perform lightweight optimization on the deep regression network, calculate the final physical distance based on the clustering results, and output structured multi-target ranging results.

11. A computer terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the lidar multi-target ranging method based on adaptive waveform matching as described in any one of claims 1-9.

12. A computer-readable medium having a computer program stored thereon, characterized in that, The computer program, when executed by a processor, can implement the adaptive waveform matching-based multi-target ranging method for lidar as described in any one of claims 1-9.

Images