Feature Extraction System and Method for LiDAR Point Clouds

Abstract

This invention relates to the field of lidar ranging technology, and more particularly to a feature extraction system and method for lidar point clouds. The system extracts feature information from the current three-dimensional point cloud voxel tensor and the pulse signal from the previous moment to obtain the current three-dimensional simulated input current. Based on the current three-dimensional simulated input current of the neuron and the previous simulated membrane potential value, an adaptive threshold is used to control the neuron's output of the current discrete pulse signal. Based on the discrete pulse signal, the simulated membrane potential values ​​of adjacent moments are sparsely aggregated to obtain the output features that fuse the spatiotemporal dynamics and local topological depth attributes of the point cloud data stream. This invention uses an adaptive threshold to preserve the local three-dimensional topological and temporal dynamic information in the three-dimensional point cloud obtained by lidar ranging; and uses a total loss function including temporal balance loss and membrane potential utilization loss to dynamically balance the neuronal activity in the spiking neural network, avoiding abnormal decay or hill effect.

Description

Technical Field

[0001] This invention belongs to the field of event data processing technology, and in particular relates to a feature extraction system and method for lidar point clouds. Background Technology

[0002] In scenarios such as autonomous driving and UAV 3D perception and control, LiDAR has become an indispensable core sensor due to its high-precision 3D spatial ranging and geometric topology acquisition capabilities. However, LiDAR faces a severe physical challenge in practical applications: limited by the laser divergence angle and target distance, point cloud data exhibits an extreme "density imbalance" characteristic in spatial distribution. Specifically, the point cloud reflected from nearby targets is extremely dense, while the echo point cloud from distant or low-reflectivity targets is extremely sparse.

[0003] To handle the aforementioned unstructured sparse point clouds, existing 3D convolutional neural networks (3D CNNs) or Transformer-based architectures have significant limitations. On the one hand, the dense high-precision floating-point operations result in enormous energy consumption and computational latency, making it difficult to meet the stringent energy efficiency requirements of drones or automotive embedded edge computing platforms. On the other hand, when dealing with extremely sparse, long-distance point clouds, these networks are highly susceptible to the problem of weak long-distance geometric depth features being submerged or directly lost due to continuous pooling downsampling or fixed feature aggregation mechanisms, leading to serious issues of missed detection of long-distance targets.

[0004] Spiking neural networks (SNNs) offer a novel paradigm for efficient processing of 3D spatial data due to their event-driven nature and extremely low power consumption in addition operations. However, existing SNNs still have significant drawbacks when processing point cloud data streams: when using conventional fixed activation thresholds or traditional learnable embedding mechanisms, the activity of neurons within the network is easily unbalanced due to drastic changes in point cloud density. When facing dense point clouds nearby, strong input stimulation leads to overactivation of neurons (the "hill effect"), resulting in a severe waste of computational power; while when facing sparse point clouds at a distance, neurons often become quiescent due to insufficient input stimulation, resulting in "abnormal decay" in deep feature extraction, directly causing the collapse of the system's long-distance perception capability.

[0005] In summary, there is an urgent need in the existing technology for a point cloud data stream feature extraction scheme that can keenly perceive the depth distribution in three-dimensional space, has adaptive sampling and feature balancing capabilities, and has significant energy efficiency advantages, so as to overcome the contradiction between sensing distant sparse targets and computational power consumption in complex scenarios. Summary of the Invention

[0006] In view of this, the present invention aims to provide a feature extraction system and method for lidar point clouds, applicable to spatial perception and measurement and control of three-dimensional sensors such as lidar. Based on pulse-triggered coding, it employs trainable parameters and adaptive thresholds for depth perception to achieve fine-grained feature extraction at the three-dimensional voxel level, effectively preserving local three-dimensional topology and temporal dynamic information in the three-dimensional point cloud obtained from lidar ranging. A total loss function including time balance loss and membrane potential utilization loss is set to dynamically balance the neuron activity in the spiking neural network, fundamentally overcoming the abnormal attenuation of long-range features and the waste of computing power at close range caused by the extreme imbalance of point cloud spatial density.

[0007] To achieve the above objectives, the technical solution created by this invention is implemented as follows: A feature extraction system for lidar point clouds is provided for extracting features from spatiotemporal three-dimensional point cloud data streams acquired by lidar ranging. The system includes: an input current generation module, which extracts feature information from the current three-dimensional point cloud voxel tensor and the pulse signal from the previous moment, and performs gating activation on the feature information to obtain the current three-dimensional simulated input current; a neuron module, which updates the current simulated membrane potential value based on the current simulated input current and the previous simulated membrane potential value; if the current simulated membrane potential value is greater than an adaptive threshold related to spatial depth distribution, the neuron outputs the current discrete pulse signal; and an information aggregation module, which sparsely aggregates the current and previous simulated membrane potential values ​​within the three-dimensional spatiotemporal local receptive field based on the discrete pulse signal to obtain output features that fuse the spatiotemporal dynamics and local topological depth attributes of the point cloud data stream.

[0008] Furthermore, in the input current generation module: the three-dimensional point cloud voxel tensor at the current moment and the pulse signal at the previous moment are subjected to three-dimensional convolution processing to obtain their respective gating information and feature information; the two gating information are added together and then activated by sigmoid to obtain normalized gating information; the two feature information are added together and then activated by ReLU to obtain the underlying analog current information; the normalized gating information is used to gating and activate the underlying analog current information to obtain the three-dimensional analog input current at the current moment.

[0009] Furthermore, the adaptive threshold is the product of the learnable decay factor and the base threshold. The value of the learnable decay factor is negatively correlated with the spatial depth distribution of the current 3D point cloud voxels, and the dynamic adjustment mechanism is based on this.

[0010] Furthermore, the negative correlation dynamic adjustment mechanism includes: obtaining the normalized spatial depth value of the current 3D point cloud voxels; the learnable attenuation factor is composed of the product of the basic attenuation parameter and the depth prior weight; the depth prior weight is: Wz=exp(-γ×Z norm ); Where Wz represents the depth prior weight, γ represents the learnable hyperparameter controlling the decay curvature, and Z... norm This represents the normalized spatial depth value.

[0011] Furthermore, in the neuron module, the simulated membrane potential value at the current moment is updated according to the leakage integral mechanism. After the neuron module outputs the discrete pulse signal at the current moment, the simulated membrane potential value is reset to zero.

[0012] Furthermore, in the information aggregation module, discrete pulse signals are used to sample the simulated membrane potential values ​​at the current and previous moments; the sampled simulated membrane potential values ​​are then locally averaged and aggregated to obtain the high-dimensional output features corresponding to the current three-dimensional point cloud voxel positions.

[0013] A feature extraction method for lidar point clouds includes: S1: Obtain the continuous spatiotemporal point cloud dataset acquired by lidar ranging, and perform three-dimensional spatial voxelization preprocessing on the dataset to obtain the training set; S2: Construct a feature extraction system for lidar point clouds as provided in this invention; S3: Use the training set obtained in step S1 to train the system obtained in step S2 to obtain the feature extraction model; S4: Use a lidar to measure the distance to the target, then process the obtained continuous 3D point cloud data according to step S1. Input the resulting 3D point cloud voxel tensor into the feature extraction model obtained in step S3 to obtain the corresponding 3D output feature stream. Furthermore, the preprocessing process in step S1 includes: discretizing the continuous spatiotemporal point cloud data stream using a time window and a three-dimensional spatial grid, extracting local spatial state features, and obtaining a three-dimensional point cloud voxel tensor containing initial simulation feature values; collecting multiple frames of three-dimensional point cloud voxel tensors and forming a training set.

[0014] Furthermore, in the process of discretizing the continuous spatiotemporal point cloud data stream using time windows and three-dimensional spatial grids: the local spatial occupancy density is obtained by counting and summing the lidar point clouds contained in each three-dimensional spatial grid within a given time window, and the local spatial occupancy density is the initial simulated feature value of the three-dimensional spatial voxel position.

[0015] Furthermore, in step S3, the system is trained using the following loss function: L total =αL balance +βL unit ; Among them, L total L represents the loss function of the training system. balanceL represents the time balance loss. unit This represents the loss in membrane potential utilization, with α and β representing the loss weights. The time balance loss is: ; Where T represents the time series length of the three-dimensional point cloud voxel tensor, and R t and R t+1 R represents the pulse emission rate of the neuron module at time t and time t+1, respectively. late The SoftPlus function represents the pulse emission rate in the second half of the time step, δ represents the target attenuation factor, and ρ represents the set minimum pulse rate threshold. SoftPlus(x) = log(1 + e) x ); Where x represents the independent variable of the softPlus function; The loss in membrane potential utilization is: ; Where, r t This represents the ratio between the simulated membrane potential at the current moment and the adaptive threshold at the current moment, [U lower U upper ] represents the ratio r t The preset ideal range in which it is located.

[0016] Compared with the prior art, the present invention can achieve the following beneficial effects: (1) The feature extraction system for lidar point clouds created in this invention breaks through the bottlenecks of extremely high computing power consumption and easy loss of long-distance features faced by traditional 3D deep learning when processing unstructured lidar point clouds. By introducing direct input encoding and depth-sensing adaptive threshold, 3D fine-grained sampling with resistance to point cloud density interference is realized: the excitation threshold is dynamically increased in dense areas near the point to suppress neuron overload and eliminate computing power waste; the threshold is dynamically decreased in sparse areas far away to transform into a high-sensitivity probe and successfully capture weak long-distance geometric features; (2) The feature extraction method of lidar point cloud created in this invention uses spatiotemporal dynamic balance loss and membrane potential utilization loss to optimize in a coordinated manner during training, forcing neurons to fire smoothly within the simulated time step, which completely solves the problem of abnormal decay and large-area neuron death when existing spiking neural networks process spatially sparse data, and greatly improves the stability of the model and the efficiency of deep feature transmission in three-dimensional perception tasks. (2) The feature extraction system and method for lidar point clouds created by the present invention are highly compatible with application scenarios such as autonomous driving and UAV 3D perception that have strict requirements for power consumption and long-distance high-precision spatial feature extraction. Its sparse computing architecture of pure spiking neural network has a much lower energy consumption than traditional 3D convolutional neural network, which greatly extends the endurance of the edge computing platform and has strong scalability. This system can be used as the underlying high-dimensional feature extraction backbone network and can be seamlessly connected to downstream high-level spatial perception networks such as pure spiking 3D target detection, point cloud semantic segmentation, and sparse depth completion and dense reconstruction. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic diagram of the feature extraction system for lidar point clouds described in the embodiments of the present invention; Figure 2 A schematic diagram of the average polymerization described in the embodiments of the present invention; Figure 3 This is a schematic flowchart of the feature extraction method for lidar point clouds described in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0019] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] like Figure 1As shown in the embodiment of the present invention, the feature extraction system for lidar point clouds is used to extract features from event streams captured by an event camera. It includes an input current generation module, a neuron module, and an information aggregation module. The input current generation module extracts feature information from the current three-dimensional point cloud voxel tensor and the pulse signal from the previous moment, and performs gating activation on the feature information to obtain the current three-dimensional simulated input current. The neuron module updates the current simulated membrane potential value based on the current simulated input current and the previous simulated membrane potential value. If the current simulated membrane potential value is greater than an adaptive threshold related to spatial depth distribution, the neuron outputs the current discrete pulse signal. The information aggregation module sparsely aggregates the current and previous simulated membrane potential values ​​within the three-dimensional spatiotemporal local receptive field based on the discrete pulse signal to obtain output features that fuse the spatiotemporal dynamics and local topological depth attributes of the point cloud data stream.

[0021] In this embodiment of the invention, the three-dimensional point cloud voxel tensor at time t is represented as V. t (x,y,z), where (x,y,z) represents the voxel grid coordinates after discretization in three-dimensional space, and the initial simulation feature value stored inside this tensor is the point cloud spatial occupancy density within this local voxel.

[0022] In some embodiments, the feature processing in the input current generation module includes: performing three-dimensional convolution processing on the three-dimensional point cloud voxel tensor at the current time and the pulse signal at the previous time to obtain their respective gating information and feature information; adding the two gating information and then activating it with sigmoid to obtain normalized gating information; adding the two feature information and then activating it with ReLU to obtain the underlying analog current information; and using the normalized gating information to perform gating activation on the underlying analog current information to obtain the three-dimensional analog input current at the current time.

[0023] The feature processing procedure in the input current generation module of this invention specifically includes: The three-dimensional point cloud voxel tensor at the current moment is represented as V t Perform 3D convolution on (x, y, z), then split the features obtained from the convolution into channels to obtain the corresponding gating information g. in (x,y,z) and feature information c in (x,y,z); For the discrete pulse signal s of the previous moment t-1 The (x, y, z) region is subjected to 3D convolution, and the resulting features are then split into channels to obtain the corresponding gating information g. rec (x,y,z) and feature information c rec (x,y,z); Gating information g in (x,y,z) and gating information g rec The sum of (x, y, z) is then activated by a sigmoid function to obtain the normalized gated information g. t (x,y,z); feature information c in (x,y,z) and feature information c rec After adding (x, y, z) and activating it with ReLU to ensure its non-negativity, the underlying analog current information I is obtained. t (x,y,z); Normalized gating information g t (x,y,z) and underlying analog current information I t Multiplying (x, y, z) and performing normalized gating information, then gating the current information, we obtain the current three-dimensional analog input current i at the current moment. t (x,y,z).

[0024] In this embodiment of the invention, an asymmetric activation strategy is employed to process gating information and feature information (i.e., sigmoid activation is used for gating information, and ReLU activation is used for feature information). Sigmoid can ensure gating consistency of the original input sampled information. Simultaneously, from the perspective of gradient propagation, ReLU ensures effective signal transmission by maintaining the gradient in the positive interval without decaying, while its gradient truncation in the negative interval promotes feature sparsity. From a biological perspective, this design conforms to the principle of excitation-inhibition pathway separation: the gating pathway simulates inhibitory regulation, while the current pathway simulates excitation driving, which enhances the expressive performance of the system.

[0025] In some embodiments, the adaptive threshold is the product of a learnable decay factor and a base threshold. The value of the learnable decay factor is negatively correlated with the spatial depth distribution (i.e., Z-coordinate) of the current 3D point cloud voxels, and a dynamic adjustment mechanism is used. That is, the deeper the spatial depth (the larger the Z-coordinate value), the smaller the learnable decay factor, thereby adaptively reducing the trigger threshold to enhance the system's trigger sensitivity to far-field sparse point cloud features. The simulated membrane potential value at the current moment is updated according to the leakage integral mechanism. After the neuron module outputs the discrete pulse signal at the current moment, the simulated membrane potential value is reset to zero.

[0026] In some embodiments, the negative correlation dynamic adjustment mechanism includes: obtaining the normalized spatial depth value of the current 3D point cloud voxel, wherein the learnable attenuation factor is composed of the product of a basic attenuation parameter and a depth prior weight, and the depth prior weight is as follows: Wz=exp(-γ×Z norm ); Where Wz represents the depth prior weight, γ represents the learnable hyperparameter controlling the decay curvature, and Z...norm This represents the normalized spatial depth value.

[0027] Using this inverse exponential mapping function, when facing an extremely dense point cloud in the vicinity (i.e., Z...), norm Approaching 0), the system increases the trigger threshold to suppress neuronal overfire; when faced with extremely sparse echo point clouds at long distances (i.e., Z... norm Approaching 1), the system significantly reduces the threshold through a learnable attenuation factor, transforming it into a highly sensitive probe.

[0028] When the simulated membrane potential value m t When (x,y,z) exceeds this adaptive threshold, the controlling neuron emits a pulse signal s. t (x,y,z), and then the simulated membrane potential value is hard reset to zero. This invention improves the traditional leakage integral ignition neuron model by incorporating an adaptive threshold related to spatial depth distribution into the leakage integral ignition neuron model, thereby enabling the leakage integral ignition neuron model to possess depth perception capabilities.

[0029] In some embodiments, the feature processing in the information aggregation module includes: sampling the simulated values ​​of the membrane potential at the current time and the previous time using discrete pulse signals; The sampled simulated values ​​of the two membrane potentials were then processed as follows: Figure 2 The local average aggregation shown yields the high-dimensional output features corresponding to the current 3D point cloud voxel position. For example... Figure 2As shown, this invention employs an existing adaptive averaging aggregation strategy based on pulse-triggered intervals (Z. Wang, Z. Wang, H. Li, L. Qin, R. Jiang, D. Ma, and H. Tang, “Eas-snn: End-to-end adaptive sampling and representation for event-based detection with recurrent spiking neural networks,” in European Conference on Computer Vision. Springer, 2024, pp. 310–328.). First, the total number of pulse triggers for each spatial voxel (x, y, z) during the network's forward inference process is counted. Then, using the pulse excitation event as a local time scale, the average membrane potential accumulated from the initial state to the first pulse trigger is assigned to the first feature map; the membrane potential between the first and second pulse triggers is assigned to the second feature map, and so on. In the final aggregated feature sequence, the features of each simulated time step are dynamically sampled from the local pulse firing frequency, thereby balancing the extreme density differences in the three-dimensional spatial distribution of the lidar point cloud at the underlying mechanism.

[0030] In this embodiment of the invention, the information aggregation module will use the current input current i at the current moment. t (x,y,z) and the input current i from the previous moment t-1 (x,y,z) are processed as follows Figure 2 The average aggregation shown yields the event tensor e at the current time. t The output feature a corresponding to (x,y,z) t (x,y,z).

[0031] This invention also provides a feature extraction method for lidar point clouds, combined with Figure 1 and Figure 3 ,include: S1: Obtain the continuous spatiotemporal point cloud dataset acquired by lidar ranging, and perform three-dimensional spatial voxelization preprocessing on the dataset to obtain the training set.

[0032] In this embodiment of the invention, publicly available autonomous driving LiDAR datasets (such as KITTI or nuScenes datasets) are used. The data is collected by non-repeating scanning or mechanically rotating LiDAR mounted on a mobile platform. The acquisition environment covers normal driving, long-distance small target capture, and low-reflectivity material scenes.

[0033] Since the original point cloud is an extremely sparse and unstructured 3D coordinate scattering, this invention employs a 3D spatial voxelization method for preprocessing to transform it into a form that can be processed by neural networks. Specifically, in some embodiments, the 3D spatial voxelization preprocessing includes: discretizing the continuous spatiotemporal point cloud data stream using a time window and a 3D spatial grid, extracting local spatial state features, and obtaining a 3D point cloud voxel tensor containing initial simulated feature values; collecting multiple frames of 3D point cloud voxel tensors and constructing a training set. During the discretization of the continuous spatiotemporal point cloud data stream using a time window and a 3D spatial grid: counting and summing the LiDAR point clouds contained within each 3D spatial grid within a given time window to obtain the local spatial occupancy density, which is the initial simulated feature value of the 3D spatial voxel position.

[0034] In this embodiment of the invention, the specific physical implementation logic of discretizing the continuous spatiotemporal point cloud data stream using a three-dimensional spatial grid in step S1 includes: First, pre-defining the physical size boundaries of the three-dimensional spatial voxels (e.g., setting high length, width, and resolution), and dividing the continuous detection area of ​​the lidar into a regular three-dimensional voxel grid array; Second, for each independent three-dimensional grid cell, performing extremely low-latency spatial occupancy statistics, that is, accurately counting the total number of lidar point clouds whose physical coordinates fall within the range of the three-dimensional grid within a given simulation time window. This total number of lidar point clouds directly characterizes the local spatial occupancy density, objectively reflecting the probability of physical entities existing in the local three-dimensional region and the density of the geometric contours; Finally, using the spatial occupancy density value of each grid as the initial simulation feature value, constructing a regularized three-dimensional point cloud voxel tensor V. t (x,y,z). This geometric counting-based aggregation method cleverly eliminates the radar echo intensity property, which is highly susceptible to interference from target material and ambient light. This not only filters out optical noise from the bottom up, greatly reducing the data preprocessing latency of the edge computing platform, but more importantly, the extracted pure spatial occupancy features naturally align with the spiking neural network's dependence on high-contrast, sparse excitation features, providing an extremely robust three-dimensional geometric data source for subsequent modules to generate deep analog currents.

[0035] S2: Construct a feature extraction system for lidar point clouds as provided in this invention.

[0036] S3: Use the training set obtained in step S1 to train the system obtained in step S2 to obtain the feature extraction model.

[0037] This invention designs a membrane potential loss function, which consists of two parts: time balance loss and membrane potential utilization loss. Specifically, in some embodiments, the system is trained using the following loss function: L total =αLbalance +βL unit ; Among them, L total L represents the loss function of the training system. balance L represents the time balance loss. unit This represents the loss in membrane potential utilization, with α and β representing the loss weights.

[0038] It should be noted that, considering the lack of physical time flow in single-frame 3D point cloud data, this invention employs a direct input encoding mechanism. The time series length T and time t in the following formulas do not refer to real-world timestamps, but rather to the algorithm simulation time steps in the forward inference process of the spiking neural network. A single frame of spatially static 3D point cloud voxel tensor is converted into a constant underlying analog current, which continuously excites the neuron modules within a set T consecutive analog time steps.

[0039] The time-balanced loss is used to constrain the pulse emission rate of the system over the entire time series length T, preventing it from falling into absolute quiescence or continuous overload due to extreme unevenness in the local density of the point cloud. Specifically: ; Among them, R t and R t+1 Let δ represent the pulse emission rate of the neuron module at time t and time t+1, respectively; let δ represent the target attenuation factor; and let ρ represent the set minimum pulse rate threshold. late R represents the pulse emission rate of the second half of the time step (if the total time steps are 4, R...). late If the time step is {3, 4}, and the total time step is 6, then R late (Time step = {4, 5, 6}), the SoftPlus function is: SoftPlus(x) = log(1 + e) x ); Here, x represents the independent variable of the softPlus function.

[0040] The time-balanced loss forced the system to maintain continuous integration and successful extraction of sparse features at long distances in the second half of the time step, completely solving the problem of "abnormal decay" of deep features.

[0041] Membrane potential utilization loss is used to optimize the membrane potential state of non-spiking neurons, specifically: ; Where, r t This represents the ratio between the simulated membrane potential at the current moment and the adaptive threshold at the current moment, [U lower U upper ] represents the ratio r tThe preset ideal range is used to avoid neurons falling into a silent or overactive state during spatial depth feature extraction.

[0042] Furthermore, in this embodiment of the invention, the SGD optimizer is used for system training, and the learning rate is 0.00008.

[0043] S4: Use lidar to measure the distance to the target, and then process the obtained continuous three-dimensional point cloud data according to step S1. The resulting three-dimensional point cloud voxel tensor is input into the feature extraction model obtained in step S3 to obtain the corresponding three-dimensional output feature flow.

[0044] The feature extraction system and method for LiDAR point clouds provided by this invention can serve as a underlying backbone network for various downstream tasks based on 3D spatial perception. In practical application, the output features of this system can be directly connected to downstream pure-pulse 3D object detection, point cloud semantic segmentation, or sparse depth completion and dense reconstruction networks. At this point, the entire system constitutes a fully pulsed end-to-end solution from raw scattered points to 3D perception ground truth. Compared to the high multiply-accumulate overhead of traditional 3D convolutional neural networks in 3D convolution calculations, the fully pulsed neural network architecture of this invention only requires pure addition operations to complete deep geometric feature extraction. During the training of the entire joint system, the bounding box regression / depth residual loss function of the downstream task and the internal constraint loss function provided by this invention are jointly optimized to guide the feature extraction system to learn a spatially adaptive and efficient sampling strategy.

[0045] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0046] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A feature extraction system for lidar point clouds, used for feature extraction of spatiotemporal three-dimensional point cloud data streams acquired by lidar ranging, characterized in that, include: The input current generation module extracts feature information from the three-dimensional point cloud voxel tensor at the current moment and the pulse signal at the previous moment, and performs gating activation on the feature information to obtain the three-dimensional simulated input current at the current moment. The neuron module updates the simulated membrane potential value at the current moment based on the three-dimensional simulated input current of the neuron at the current moment and the simulated membrane potential value at the previous moment; if the simulated membrane potential value at the current moment is greater than the adaptive threshold related to the spatial depth distribution, the neuron is controlled to output the discrete pulse signal at the current moment. The information aggregation module performs sparse aggregation of the simulated membrane potential values ​​at the current and previous moments within the three-dimensional spatiotemporal local receptive field based on discrete pulse signals, thereby obtaining the output features of the spatiotemporal dynamics and local topological depth attributes of the fused point cloud data stream.

2. The feature extraction system for lidar point clouds according to claim 1, characterized in that, In the input current generation module: The three-dimensional point cloud voxel tensor at the current time and the pulse signal at the previous time are subjected to three-dimensional convolution processing to obtain their respective gating information and feature information; The two gating information are added together and then activated by sigmoid to obtain normalized gating information; After adding the two types of feature information, ReLU activation is performed to obtain the underlying analog current information; By using normalized gating information to gating and activating the underlying analog current information, the three-dimensional analog input current at the current moment can be obtained.

3. The feature extraction system for lidar point clouds according to claim 1, characterized in that, The adaptive threshold is the product of the learnable decay factor and the base threshold. The value of the learnable decay factor is negatively correlated with the spatial depth distribution of the current 3D point cloud voxels, and the dynamic adjustment mechanism is based on this.

4. The feature extraction system for lidar point clouds according to claim 3, characterized in that, Negative correlation dynamic adjustment mechanisms include: Obtain the normalized spatial depth value of the current 3D point cloud voxel; the learnable attenuation factor is composed of the product of the basic attenuation parameter and the depth prior weight. The depth prior weights are: Wz=exp(-γ×Z norm ); Where Wz represents the depth prior weight, γ represents the learnable hyperparameter controlling the decay curvature, and Z... norm This represents the normalized spatial depth value.

5. The feature extraction system for lidar point clouds according to claim 1, characterized in that, In the neuron module, the simulated membrane potential value at the current moment is updated according to the leakage integral mechanism. After the neuron module outputs the discrete pulse signal at the current moment, the simulated membrane potential value is reset to zero.

6. The feature extraction system for lidar point clouds according to claim 1, characterized in that, In the information aggregation module: The simulated values ​​of the membrane potential at the current moment and the previous moment are sampled using discrete pulse signals; The sampled simulated values ​​of the two membrane potentials are locally averaged and aggregated to obtain the high-dimensional output features corresponding to the current three-dimensional point cloud voxel positions.

7. A method for feature extraction from lidar point clouds, characterized in that, include: S1: Obtain the continuous spatiotemporal point cloud dataset acquired by lidar ranging, and perform three-dimensional spatial voxelization preprocessing on the dataset to obtain the training set; S2: Construct a feature extraction system for lidar point clouds as described in any one of claims 1 to 6; S3: Use the training set obtained in step S1 to train the system obtained in step S2 to obtain the feature extraction model; S4: Use lidar to measure the distance to the target, and then process the obtained continuous three-dimensional point cloud data according to step S1. The resulting three-dimensional point cloud voxel tensor is input into the feature extraction model obtained in step S3 to obtain the corresponding three-dimensional output feature flow.

8. The feature extraction method for lidar point clouds according to claim 7, characterized in that, The preprocessing process in step S1 includes: By using time windows and three-dimensional spatial grids to discretize continuous spatiotemporal point cloud data streams, local spatial state features are extracted to obtain three-dimensional point cloud voxel tensors containing initial simulation feature values. Collect multiple frames of 3D point cloud voxel tensors and construct a training set.

9. The feature extraction method for lidar point clouds according to claim 8, characterized in that, In the process of discretizing continuous spatiotemporal point cloud data streams using time windows and three-dimensional spatial grids: The local spatial occupancy density is obtained by counting and summing the lidar point clouds contained in each three-dimensional spatial grid within a given time window. The local spatial occupancy density is the initial simulated feature value of the three-dimensional spatial voxel position.

10. The feature extraction method for lidar point clouds according to claim 7, characterized in that, In step S3, the system is trained using the following loss function: L total =αL balance +βL unit ; Among them, L total L represents the loss function of the training system. balance L represents the time balance loss. unit This represents the loss in membrane potential utilization, with α and β representing the loss weights. The time balance loss is: ； Where T represents the time series length of the three-dimensional point cloud voxel tensor, and R t and R t+1 R represents the pulse emission rate of the neuron module at time t and time t+1, respectively. late The SoftPlus function represents the pulse emission rate in the second half of the time step, δ represents the target attenuation factor, and ρ represents the set minimum pulse rate threshold. SoftPlus(x)=log(1+e x ); Where x represents the independent variable of the softPlus function; The loss in membrane potential utilization is: ； Where, r t This represents the ratio between the simulated membrane potential at the current moment and the adaptive threshold at the current moment, [U lower U upper ] represents the ratio r t The preset ideal range in which it is located.

Images