128*256 large array gm-apd active laser imaging signal extraction and point cloud image joint reconstruction method and system based on poisson weight gain

Abstract

A method and system for joint reconstruction of active laser imaging signal extraction and point cloud images based on Poisson weighted gain using a 128×256 large-array GM-APD is presented. This method, within the field of signal processing, addresses the shortcomings of existing reconstruction algorithms in maintaining target integrity while filtering noise. It employs a collaborative processing flow from temporal statistical suppression and spatial geometric filtering to image domain structure restoration. The method includes: calculating weighted gain based on the signal statistical histogram using a Poisson process to suppress abnormally high-probability peaks triggered by noise at the range gate front; converting the processed range image into a 3D point cloud and applying density clustering to separate and remove spatially discrete noise points; back-projecting the denoised point cloud to generate a binary mask, and then optimizing the mask through convolution filtering and morphological operations to compensate for information loss and smooth the target contour; finally, applying the optimized mask to extract the final range image reconstruction result. This method is also applicable to remote sensing mapping, autonomous driving, and national defense security.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, specifically to a method and system for joint reconstruction of active laser imaging signals and point cloud images using a 128×256 large array GM-APD based on Poisson weighted gain. Background Technology

[0002] Single-photon detection technology, with its single-photon-level detection sensitivity, has broken through the performance bottleneck of traditional photoelectric detection technology in weak signal scenarios. This extreme sensitivity enables it to capture the faint light signals reflected by distant targets (such as low-altitude aircraft at kilometer ranges), low-reflectivity targets (such as non-metallic components with rough surfaces), and small targets (such as micro-drones and space debris), thereby achieving accurate detection and early warning for these targets. Geiger-mode avalanche photodiodes (GM-APDs), with their single-photon-level detection capability and nanosecond-level time response characteristics, are key to achieving high-precision 3D imaging. In recent years, GM-APD detectors have been developing towards large-scale arrays, with arrays of 128×256 and even larger sizes now available. Large-scale arrays not only significantly improve the spatial resolution of imaging but also significantly increase the point cloud density, laying a solid hardware foundation for acquiring rich 3D information such as target surface texture and fine structure.

[0003] However, large-scale GM-APD arrays still face severe noise challenges in actual imaging. Under fixed-distance gate control, due to the high trigger probability at the initial stage of gate opening, noise factors such as dark counting, afterpulse, and background stray light easily form significant noise peaks at the front of the gate width in the signal statistical histogram. These noise peaks not only drown out weak signals in neighboring areas but also introduce a large number of spatially isolated spurious points into the 3D point cloud reconstruction results. These abnormal noise points intertwine with the real target signals, severely degrading the signal-to-noise ratio and integrity of the image, and causing serious interference to subsequent target detection, classification, and recognition.

[0004] To address the noise problem of large-scale GM-APD arrays, traditional reconstruction methods such as peaking, matched filtering, and cross-neighborhood methods can achieve some noise filtering effect in medium-to-high signal-to-noise ratio (SNR) scenarios, but their performance limitations become increasingly apparent under extremely low SNR conditions. Specifically, traditional methods struggle to accurately model the distribution characteristics of noise peaks and the statistical regularities of the signal, resulting in insufficient distinction between noise and the true signal. As large-scale GM-APD arrays evolve towards higher resolution and longer detection distances, traditional methods are no longer sufficient to meet the demands of high-quality imaging. Therefore, developing a new method capable of accurately modeling noise characteristics and achieving efficient noise filtering based on this model has become crucial for overcoming the performance bottleneck of large-scale GM-APD arrays. Summary of the Invention

[0005] To overcome the shortcomings of existing reconstruction algorithms in maintaining target integrity while filtering noise, this invention proposes a joint reconstruction method based on Poisson weighted gain signal extraction and point cloud collaborative filtering for large-scale GM-APD arrays such as 128×256. This method aims to deeply suppress anomalous noise peaks at the front of the range gate, effectively eliminate discrete noise points inside and around the target, and maximize the preservation of the target's original geometric contours and integrity during the noise filtering process.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: Option 1: This invention proposes a method for joint reconstruction of active laser imaging signals and point cloud images using a 128×256 large-array GM-APD based on Poisson weighted gain. The method includes the following steps: Step 1: Statistically analyze the multi-frame echo data acquired by the GM-APD detector and generate a time-trigger count histogram; Step 2: Based on the characteristics of the Poisson distribution, establish a probability distribution model for noise-triggered events and calculate the weighted gain function; Step 3: Apply the weighted gain function calculated in Step 2 to the time-trigger count histogram described in Step 1 to generate an enhanced histogram after noise suppression, and extract the initial range image of the target. D init ; Step 4: Extract the initial distance image of the target from Step 3. D init Mapping to 3D space to generate an initial point cloud P init ; Step 5: Perform initial point cloud analysis based on the DBSCAN density clustering algorithm. P init Denoising is performed to separate and remove discrete noise points; Step 6: Re-project the denoised point cloud onto the two-dimensional image plane to generate a preliminary binary mask. M init ; Step 7: Apply the initial binary mask M init Convolutional filtering and morphological filtering are performed within the contour area to generate the final mask. M refined ; Step 8: Apply the final mask M refined Image of initial distance D init Element-wise dot product is performed to filter out regions marked as background by the mask, while retaining the distance information of the target region, resulting in a high-quality reconstructed distance image. Dfinal .

[0007] Furthermore, a preferred embodiment is provided, wherein the probability distribution model of the noise-triggered event in step 2 is as follows:

[0008] Where B = N / n, N is the average number of noise events within the entire distance gate, and n is the total number of time bins.

[0009] Furthermore, a preferred embodiment is provided, wherein the method for calculating the weight gain function in step 2 is as follows: Let the first p The average number of primary electrons generated by noise within each time chamber is: (1) For a Poisson process, it occurs within this timebox. m The probability of each event is: (2) The GM-APD detector in the j The probability that a timebox is triggered by noise is: (3) Assuming the noise rate is constant, i.e. M i =B=N / n ,in N Let be the average number of noise events within the entire door width. Then, equation (3) simplifies to: (4) Constructing the weighted gain function W(j) for P n (j) The reciprocal: (5).

[0010] Furthermore, a preferred embodiment is provided, in step 4, the initial distance image of the target extracted in step 3 is... D init Mapping to 3D space to generate an initial point cloud P init The method is as follows: Coordinates of each pixel in the distance image (u, v) and their corresponding distance values z = D init (u, v) Through coordinate system transformation, that is, ,in δ x , δ yFor pixel physical size, c Generate an initial point cloud set P at the speed of light. init ={p i =(x i ,y i ,z i )}.

[0011] Furthermore, a preferred embodiment is provided, wherein in step 5, the initial point cloud is processed based on the DBSCAN density clustering algorithm. P init The method for noise reduction is as follows: The enhanced histogram obtained after calculating the weight gain. H enhanced (t) Estimating the global signal-to-noise ratio SNR global : (6) in, P signal To enhance the average intensity of the peak region in the histogram, P noise The average intensity of the noise-dominant region at the front end of the door; based on SNR global Adaptive determination of neighborhood radius for the DBSCAN algorithm And the minimum number of points threshold MinPts: (7) in, max 、 MinPts min These are preset upper and lower limits. α and β It is a positive adjustment factor; For any point in the point cloud p i Calculate its Points within the neighborhood N (p i );like N (p i If )≥MinPts, then define p i Using any core point as the starting point, recursively visit all points reachable from the core point density to form the target point cloud. Points not belonging to any cluster are marked as noise points and removed, resulting in a denoised point cloud. P denoised .

[0012] Furthermore, a preferred embodiment is provided, wherein in step 7, the preliminary mask is... M init Convolutional filtering and morphological filtering are performed within the contour area to generate the final mask. M refined The method is as follows: Step 7.1: Use a small average or Gaussian convolution kernel K. M init Perform convolution operations: (8) Step 7.2: Set a threshold θ to... M conv Binarization yields the intermediate mask. M bin : (9) Step 7.3, for M bin Perform morphological closing operations to obtain the optimized final mask. M refined : (10) In this context, SE is a structural element.

[0013] Furthermore, a preferred embodiment is provided in which a high-quality reconstructed distance image is obtained in step 8. D final The method is as follows: (11) in, This indicates element-wise multiplication.

[0014] Option 2: A joint system for active laser imaging signal extraction and point cloud image reconstruction using a 128×256 large array GM-APD based on Poisson weighted gain, the system comprising: The data statistics module is used to perform statistical analysis on multi-frame echo data acquired by the GM-APD detector and generate a time-trigger count histogram. The weighted gain function calculation module is used to combine the characteristics of Poisson distribution to establish a probability distribution model of noise-triggered events and calculate the weighted gain function. The initial distance image extraction module applies the weighted gain function calculated by the weighted gain function calculation module to the time-trigger count statistical histogram described by the data statistics module, generating an enhanced histogram after noise suppression, and extracting the initial distance image of the target. D init ; The point cloud processing module is used to process the initial range image of the target extracted by the initial range image extraction module. D init Mapping to 3D space to generate an initial point cloud P init ; The noise reduction module is used to process the initial point cloud based on the DBSCAN density clustering algorithm. P init Denoising is performed to separate and remove discrete noise points; The preliminary binary mask generation module is used to re-project the denoised point cloud onto the two-dimensional image plane to generate a preliminary binary mask. M init ; The final mask generation module is used to generate the initial binary mask. M init Convolutional filtering and morphological filtering are performed within the contour area to generate the final mask. M refined ; Image optimization module, used to optimize the final mask M refined Image of initial distance D init Element-wise dot product is performed to filter out regions marked as background by the mask, while retaining the distance information of the target region, resulting in a high-quality reconstructed distance image. D final .

[0015] Option 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in Option 1.

[0016] Option 4: A computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the method described in Option 1.

[0017] The advantages of this invention are: The method for joint reconstruction of active laser imaging signal extraction and point cloud image based on Poisson weight gain of 128×256 large array GM-APD described in this invention aims to deeply suppress abnormal noise peaks at the front end of the gate, effectively remove discrete noise points inside and around the target, and maximize the preservation of the original geometric contour and integrity of the target during the noise filtering process.

[0018] The method for joint extraction of active laser imaging signals and joint reconstruction of point cloud images based on Poisson weighted gain using a 128×256 large-array GM-APD described in this invention achieves deep suppression of multi-dimensional noise. Specifically, the Poisson weighted model accurately suppresses systematic noise peaks at the gate front based on temporal statistical characteristics; DBSCAN filters out discrete and isolated noise points from the three-dimensional spatial distribution; and morphological filtering smooths residual local irregular noise from the two-dimensional image structure. The synergy of these three methods achieves a multi-level and multi-angle joint attack on noise, resulting in a significant improvement in overall noise suppression rate (including front-end noise peaks and spatially discrete noise) compared to traditional methods.

[0019] The method for joint reconstruction of active laser imaging signal extraction and point cloud image based on Poisson weight gain of 128×256 large array GM-APD described in this invention maximizes the preservation of target integrity, that is, Poisson weight enhancement avoids the signal being masked by noise peaks; DBSCAN is based on density rather than global threshold during noise filtering, which can better preserve sparse but continuous real target point cloud; morphological operation, as a "repair" step, is specifically designed to compensate for and smooth internal holes or edge burrs in the target that may be caused by previous processing.

[0020] The method described in this invention can not only improve the quality and reliability of single-photon three-dimensional imaging, but also fully unleash the imaging potential of large-scale GM-APD arrays, promoting their practical application in fields such as remote sensing mapping, autonomous driving, and national defense security. It has important research significance and engineering application value. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the noise triggering probability distribution Pn(j) and its corresponding weight gain function W(j) as described in Implementation Method 1.

[0022] Among them, (a) is a schematic diagram of the probability distribution of noise-triggered events with negative exponential decay of e, and (b) is a schematic diagram of the weighted gain model based on the Poisson distribution.

[0023] Figure 2 This is a comparison chart showing the effect of applying the weight gain described in Implementation Method 1 to the statistical histogram before and after.

[0024] (a) is the original histogram, and (b) is the enhanced histogram.

[0025] Figure 3 This is a schematic diagram of the initial point cloud during the point cloud collaborative filtering process described in Implementation Method 1.

[0026] Figure 4 This is a schematic diagram of the point cloud collaborative filtering process described in Implementation Method 1.

[0027] Among them, (a) is a schematic diagram of the point cloud after DBSCAN clustering and denoising, and (b) is a schematic diagram of the two-dimensional distance image generated by projection.

[0028] Figure 5 This is a schematic diagram of mask optimization during the point cloud collaborative filtering process described in Implementation Method 1.

[0029] Among them, (a) is a schematic diagram of the two-dimensional distance image generated by projection, (b) is a schematic diagram of the initial mask generated by projection, (c) is a schematic diagram of the mask after convolution, and (d) is a schematic diagram of the optimized mask after morphological processing.

[0030] Figure 6 This is a comparison chart of the reconstruction results of the method of the present invention and other traditional methods.

[0031] Among them, (a) is the peak method, (b) is the cross-neighborhood method, (c) is the matched filtering method, (d) is the cross-neighborhood combined with matched filtering method, and (e) is the algorithm described in this invention.

[0032] Figure 7 The graph shows the quantitative evaluation curves of image restoration accuracy and image signal-to-noise ratio for different methods. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0034] Implementation Method 1: The method for joint reconstruction of active laser imaging signal extraction and point cloud image based on Poisson weighted gain using a 128×256 large array GM-APD specifically includes the following steps: Step 1: Signal extraction based on the Poisson distribution weighted gain model: S1, regarding GM Statistical analysis was performed on multiple frames of echo data collected by the APD detector, and the generation time was... Histogram of trigger counts.

[0035] Within the repetition cycle of a laser pulse, a fixed distance gate is set, and the time within the gate is uniformly discretized into n time-resolved intervals (timeboxes). The detector responds to the first trigger event within the gate and records its position within the timebox. After accumulating multiple frames of data, the number of triggers in each timebox is counted to form a statistical histogram. H(t) ,in t =1,2,…,n.

[0036] S2. Based on the Poisson distribution process, establish a probability distribution model for noise-triggered events and derive its weight gain function.

[0037] Under stable detection conditions, the occurrence of noise events such as background noise photons and dark counting within the gate can be modeled as a Poisson process. Let the average generation rate of noise events per unit time be... λ n Then in the first j The probability of generating at least one noise-triggered event within a time bin P n (j) It depends not only on the expectations of the box itself. M j It is also subject to the constraint that none of the preceding timeboxes have triggered this condition. The calculation method is as follows: Let the first p The average number of primary electrons generated by noise within each time chamber is: (1) For a Poisson process, it occurs within this timebox. m The probability of each event is: (2) Considering the characteristic of GM-APD that it only responds to the first trigger, the detector in the second... j The probability that a timebox is triggered by noise is: (3) Assuming the noise rate is constant, i.e. M i =B=N / n ,in N Let be the average number of noise events within the entire door width. Then equation (3) simplifies to: (4) visible, P n (j) Follow j The gain increases exponentially, resulting in a high-probability noise peak at the beginning of the gate width. To suppress this noise peak and enhance subsequent potential signals, a weighted gain function is constructed. W(j) for P n (j) The reciprocal: (5) S3. Apply the weighted gain function to the original statistical histogram to generate an enhanced histogram after noise suppression, and extract the initial range image of the target. D init .

[0038] WeightW(j) With histogram H(t) Multiplying point by point yields the enhanced histogram. H enhanced (t) = H(t)·W(t) .right H enhanced (t) Peak search and other processing are performed to obtain the target distance value corresponding to each pixel, forming an initial two-dimensional distance profile. D init .

[0039] Step 2: Collaborative Noise Filtering Based on Point Cloud Clustering and Morphological Filtering S4, Initial distance image D init Mapping to 3D space to generate an initial point cloud P init .

[0040] Based on the coordinates of each pixel in the distance image (u, v) and their corresponding distance values z = D init (u, v) Through coordinate system transformation (e.g.: , among which δ x , δ y For pixel physical size, c (at the speed of light), generating a 3D point cloud set P init ={p i =(x i ,y i ,z i )}.

[0041] S5. Based on the DBSCAN density clustering algorithm, the initial point cloud is denoised to separate and remove discrete noise points.

[0042] First, the enhanced histogram is obtained based on the weighted gain. H enhanced (t) Estimating the global signal-to-noise ratio SNR global : (6) in, P signal To enhance the average intensity of the peak region in the histogram, P noise This represents the average intensity of the noise-dominant region at the front end of the door.

[0043] based on SNR globalAdaptive determination of neighborhood radius for the DBSCAN algorithm And the minimum number of points threshold MinPts: (7) in, max 、 MinPts min These are preset upper and lower limits. α and β This is a positive adjustment coefficient. This mapping strategy ensures that clustering conditions are relaxed at low signal-to-noise ratios to avoid target fragmentation, while conditions are tightened at high signal-to-noise ratios to improve noise filtering accuracy.

[0044] For any point in the point cloud p i Calculate its Points within the neighborhood N (p i ).like N (p i If )≥MinPts, then define p i Using any core point as the central point, recursively visit all points reachable from the core point density, forming a cluster (i.e., the target point cloud). Points not belonging to any cluster are marked as noise points and removed, resulting in a denoised point cloud. P denoised .

[0045] S6, Denoising Point Clouds P denoised Back-projecting onto the two-dimensional image plane generates a preliminary binary mask. M init .

[0046] Denoising point cloud P denoised Points in the image are mapped back to image coordinates based on their original projection relationships. If a pixel location has a corresponding point in the point cloud, the mask value is set to 1 (target); otherwise, it is set to 0 (background / missing), thus obtaining a preliminary binary mask. M init ∈{0,1} H×W .

[0047] S7. Perform convolution filtering and morphological operations on the initial mask to compensate for missing regions and smooth the target contour, generating an optimized mask. M refined .

[0048] First, use a small average or Gaussian convolution kernel K pair Minit Perform convolution operations to fill in tiny holes or contour depressions inside the target that may be caused by clustering: (8) Secondly, by setting a threshold θ M conv Binarization yields the intermediate mask. M bin : (9) Finally, for M bin Perform morphological closing operations (dilation followed by erosion) to smooth target boundaries, connect adjacent regions, and remove small protrusions or burrs, resulting in the optimized final mask. M refined : (10) In this context, SE is a structural element.

[0049] S8. Apply an optimized mask to extract the final reconstruction result from the initial distance image.

[0050] Final mask M refined Image of initial distance D init Element-wise dot product is performed to filter out regions marked as background by the mask, while retaining the distance information of the target region, resulting in a high-quality reconstructed distance image. D final : (11) in, This indicates element-wise multiplication. This step achieves effective noise suppression while preserving the geometric structure and contour integrity of the target to the greatest extent possible.

[0051] Step 3: Difficulties and Challenges in the Collaborative Integration Process In the process of integrating Poisson distribution weighted gain signal extraction, point cloud clustering algorithm (DBSCAN), and morphological filtering algorithm into a unified processing flow, the following key technical difficulties and challenges are encountered: S9.1: Challenges of Cross-Domain Feature Transfer and Information Fidelity The Poisson weighted gain algorithm operates on the trigger histogram of each pixel in the time domain, and its core output is a two-dimensional distance image corrected by a probability model. The subsequent DBSCAN point cloud clustering algorithm, on the other hand, operates on the point set in the three-dimensional geometric domain. The mapping process from the two-dimensional image domain to the three-dimensional point cloud domain (based on the camera model and distance values) introduces inherent information transformations.

[0052] The main challenge lies in the fact that each pixel output by the Poisson processing carries a "confidence level" or "signal enhancement level" assigned by the weighting model. However, after converting it to 3D points, this continuous confidence information is difficult to integrate seamlessly into DBSCAN's clustering criteria based on pure geometric distance (such as Euclidean distance). This can lead to real signal points that are correctly enhanced in the time domain but still appear sparse being misjudged as discrete noise and filtered out by DBSCAN in the spatial domain due to insufficient "density," resulting in the loss of detailed target information.

[0053] S9.2: Challenges of Multi-Algorithm Parameter Coupling and Adaptive Tuning This joint method involves three core processing modules, each with its own key parameters. These parameters are intricately coupled and difficult to optimize independently. Poisson weighted model parameters: mainly involve the average rate of noise. N or equivalent parameters B The estimate. B The value directly affects the weighting function. W(j) The rate of attenuation, i.e., the strength of suppression of noise at the front end of the door. If B If the estimate is too low, the suppression will be insufficient; if the estimate is too high, the weak signal at the front end may be over-suppressed.

[0054] DBSCAN clustering parameter: neighborhood radius The minimum number of points (MinPts) and the minimum point threshold (MinPts) are the two parameters that together determine the granularity of the algorithm's cluster identification. They need to be adaptively adjusted based on the spatial distribution density of the point cloud (affected by target distance, angle, and the aforementioned Poisson processing effect).

[0055] Morphological filtering parameters: type (e.g., square, circle) and size of the structuring element (SE), and the convolution kernel. K Size and binarization threshold θ These parameters determine the degree of mask restoration and contour smoothing.

[0056] The aforementioned parameter set forms an interconnected chain. For example, if excessive smoothing during Poisson processing leads to sparse target point clouds, then the DBSCAN parameter needs to be increased. Alternatively, reducing MinPts can prevent target breakage; however, this may lead to more noise being mistakenly included. Morphological operations, as the final repair step, need to have an intensity (such as SE size) that matches the mask quality (hole size, edge burr degree) output by DBSCAN. A lack of coordinated parameter optimization strategies easily leads to a dilemma of "filtering causing target damage" or "under-filtering causing residual noise."

[0057] S9.3: Fidelity Challenges in Handling Repeated Dimensional Transformations The method described in this invention involves multiple dimensional transformations: "2D range image - 3D point cloud - 2D mask - 2D final range image". Each transformation may introduce errors or information loss. 2D-3D projection distortion: Converting a distance image into a point cloud relies on accurate system calibration parameters (intrinsic and extrinsic parameters). Any calibration error or model simplification will lead to deviations in the position of 3D points, thus affecting the accuracy of DBSCAN clustering.

[0058] Aliasing and blurring in 3D-2D backprojection: When backprojecting a 3D point cloud onto a 2D image plane to generate a mask, multiple 3D points may be projected onto the same pixel, or a single pixel may be missing due to the lack of a corresponding point. This can lead to jagged edges or unexpected holes in the mask.

[0059] S9.4: The Dilemma of Co-processing Noise-Signal Transition Region Under extremely low signal-to-noise ratio conditions, points in weakly reflective regions at the target edge or within the target may exhibit statistical characteristics (in the Poisson histogram) and spatial distribution characteristics (in the point cloud) very similar to those of points with strong noise. Collaborative processes face decision-making difficulties in such areas. The Poisson weighting model may not provide sufficient gain to highlight these weak signal points in the histogram. Even if highlighted, these points may still appear sparse and discontinuous in 3D space, easily identified as noise by DBSCAN. While subsequent morphological operations can repair some of the missing data, if too many underlying signal points are missing, morphological repair may introduce false structures or cause contour distortion.

[0060] This invention achieves a better synergistic enhancement effect by effectively coordinating three algorithms, designing a parameter linkage mechanism, introducing feedback adjustment based on intermediate results in key stages, and optimizing the dimension transformation algorithm.

[0061] Through the close coordination of the above steps, the method of the present invention overcomes the limitations of using each algorithm independently and achieves the following significant performance improvements: Deep suppression of multi-dimensional noise: The Poisson weighted model accurately suppresses systematic noise peaks at the gate front based on temporal statistical characteristics; DBSCAN filters out discrete and isolated noise points from the three-dimensional spatial distribution; and morphological filtering smooths residual local irregular noise from the two-dimensional image structure. The three technologies work together to achieve a multi-level and multi-angle joint attack on noise, and the overall noise suppression rate (including front-end noise peaks and spatial discrete noise) is significantly improved compared with traditional methods.

[0062] Maximizing the preservation of target integrity: Poisson weight enhancement avoids the signal being masked by noise peaks; DBSCAN, based on density rather than global threshold, can better preserve sparse but continuous real target point clouds during noise filtering; morphological operations, as a "repair" step, are specifically designed to compensate for and smooth internal holes or edge burrs that may be caused by previous processing.

[0063] Example 1: This example uses a 128×256 pixel GM-APD lidar system to collect data. The distance gate width is 1250 ns (corresponding to a distance of 187.5 meters), and the time within the gate is discretized into 1250 time boxes (each 1 ns).

[0064] Step 1 execution process: For each pixel, calculate the statistical histogram of its triggering events in each time bin.

[0065] Estimate the total noise rate based on the system's calibrated average dark count rate and background light intensity. λ n , and we get B.

[0066] Calculate the weight function according to formula (5) W(j) .

[0067] Will W(j) The histogram applied to each pixel is used to obtain an enhanced histogram.

[0068] A local peak search is performed on the enhanced histogram, and the time bin index corresponding to the maximum peak is converted into a distance value to generate an initial distance profile of 128×256. D init .

[0069] Step Two Execution Process: Set camera intrinsics and pixel size, D init Convert to 3D point cloud P init .

[0070] Set DBSCAN parameters: And the minimum number of points threshold, MinPts. P init After clustering and removing noise points, the following results are obtained: P denoised .

[0071] Will P denoised Projecting back onto the image plane generates a preliminary mask. M init .

[0072] Use the 3×3 mean to check. Minit Perform convolution and take a threshold. θ Perform binarization.

[0073] Morphological closing operations are performed on the binarized result using a circular structuring element with a radius of 2 pixels, resulting in... M refined .

[0074] According to formula (11), M refined and D init Dot product yields the final reconstructed distance image. D final .

[0075] Figure 6 The reconstruction result (e) of this embodiment is shown in comparison with the results of the peak method (a), the cross-neighborhood method (b), the matched filtering method (c), and the cross-neighborhood combined with matched filtering method (d). It can be seen that the method of the present invention effectively eliminates discrete noise points inside the target and in the background while best maintaining the continuity and integrity of the target contour.

[0076] Figure 7 Quantitative evaluation results are presented, where the black curve represents image restoration degree (measuring the degree of target structure preservation), and the red curve represents image signal-to-noise ratio. The proposed method achieves optimal values ​​for both metrics.

[0077] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or technical solutions of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. This is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0078] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended technical solutions are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the present invention. Clearly, those skilled in the art can make various modifications and variations to the present invention without departing from its spirit and scope. Thus, if these modifications and variations of the present invention fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A method for joint reconstruction of active laser imaging signal extraction and point cloud image based on Poisson weighted gain of a 128×256 large array GM-APD, characterized in that, The method includes the following steps: Step 1: Statistically analyze the multi-frame echo data acquired by the GM-APD detector and generate a time-trigger count histogram; Step 2: Based on the characteristics of the Poisson distribution, establish a probability distribution model for noise-triggered events and calculate the weighted gain function; Step 3: Apply the weighted gain function calculated in Step 2 to the time-trigger count histogram described in Step 1 to generate an enhanced histogram after noise suppression, and extract the initial range profile of the target. D init ; Step 4: Extract the initial distance image of the target from Step 3. D init Mapping to 3D space to generate an initial point cloud P init ; Step 5: Perform initial point cloud analysis based on the DBSCAN density clustering algorithm. P init Denoising is performed to separate and remove discrete noise points; Step 6: Re-project the denoised point cloud onto the 2D image plane to generate a preliminary binary mask. M init ; Step 7: Apply the initial binary mask M init Convolutional filtering and morphological filtering are performed within the contour area to generate the final mask. M refined ; Step 8: Apply the final mask M refined Image of initial distance D init Element-wise dot product is performed to filter out regions marked as background by the mask, while retaining the distance information of the target region, resulting in a high-quality reconstructed distance image. D final .

2. The method for joint reconstruction of active laser imaging signal extraction and point cloud image based on Poisson weighted gain for a 128×256 large array GM-APD according to claim 1, characterized in that, The probability distribution model for the noise-triggered event mentioned in step 2 is as follows: Where B = N / n, N is the average number of noise events within the entire distance gate, and n is the total number of time bins.

3. The method for joint reconstruction of active laser imaging signal extraction and point cloud image based on Poisson weighted gain for a 128×256 large array GM-APD according to claim 1, characterized in that, The method for calculating the weighted gain function in step 2 is as follows: Let the first p The average number of primary electrons generated by noise within each time chamber is: (1) For the Poisson process, in the... p Events within a timebox m The probability of each event is: (2) The GM-APD detector in the j The probability that a timebox is triggered by noise is: (3) Assuming the noise rate is constant, i.e. M i =B=N / n ,in N Let be the average number of noise events within the entire door width. Then, equation (3) simplifies to: (4) Constructing the weighted gain function W(j) for P n (j) The reciprocal: (5)。 4. The method for joint reconstruction of active laser imaging signal extraction and point cloud image based on Poisson weighted gain for a 128×256 large array GM-APD according to claim 1, characterized in that, In step 4, the initial range image of the target extracted in step 3 is... D init Mapping to 3D space to generate an initial point cloud P init The method is as follows: Coordinates of each pixel in the distance image (u,v) and their corresponding distance values z=D init (u,v) Through coordinate system transformation, that is, ,in δ x , δ y For pixel physical size, c Generate an initial point cloud set P at the speed of light. init ={p i =(x i ,y i ,z i )}.

5. The method for joint reconstruction of active laser imaging signal extraction and point cloud image based on Poisson weighted gain for a 128×256 large array GM-APD according to claim 1, characterized in that, In step 5, the initial point cloud is processed using the DBSCAN density clustering algorithm. P init The method for noise reduction is as follows: The enhanced histogram obtained after calculating the weight gain. H enhanced (t) Estimating the global signal-to-noise ratio SNR global : (6) in, P signal To enhance the average intensity of the peak region in the histogram, P noise The average intensity of the noise-dominant region at the front end of the door; based on SNR global Adaptive determination of neighborhood radius for the DBSCAN algorithm And the minimum number of points threshold MinPts: (7) in, max 、 MinPts min These are preset upper and lower limits. α and β It is a positive adjustment factor; For any point in the point cloud p i Calculate its Points within the neighborhood N (p i );like N (p i If )≥MinPts, then define p i Using any core point as the starting point, recursively visit all points reachable from the core point density to form the target point cloud. Points not belonging to any cluster are marked as noise points and removed, resulting in a denoised point cloud. P denoised .

6. The method for joint reconstruction of active laser imaging signal extraction and point cloud image based on Poisson weighted gain for a 128×256 large array GM-APD according to claim 1, characterized in that, Step 7 involves the initial masking. M init Convolutional filtering and morphological filtering are performed within the contour area to generate the final mask. M refined The method is as follows: Step 7.1: Use a small average or Gaussian convolution kernel K. M init Perform convolution operations: (8) Step 7.2: Set a threshold θ to... M conv Binarization yields the intermediate mask. M bin : (9) Step 7.3, for M bin Perform morphological closing operations to obtain the optimized final mask. M refined : (10) In this context, SE is a structural element.

7. The method for joint reconstruction of active laser imaging signal extraction and point cloud image based on Poisson weighted gain for a 128×256 large array GM-APD according to claim 1, characterized in that, Step 8 yields a high-quality reconstructed distance image. D final The method is as follows: (11) in, This indicates element-wise multiplication.

8. A system for joint reconstruction of active laser imaging signal extraction and point cloud image based on a 128×256 large array GM-APD using Poisson weighted gain, characterized in that... The system includes: The data statistics module is used to perform statistical analysis on multi-frame echo data acquired by the GM-APD detector and generate a time-trigger count histogram. The weighted gain function calculation module is used to combine the characteristics of Poisson distribution to establish a probability distribution model of noise-triggered events and calculate the weighted gain function. The initial distance image extraction module applies the weighted gain function calculated by the weighted gain function calculation module to the time-trigger count statistical histogram described by the data statistics module, generating an enhanced histogram after noise suppression, and extracting the initial distance image of the target. D init ; The point cloud processing module is used to process the initial range image of the target extracted by the initial range image extraction module. D init Mapping to 3D space to generate an initial point cloud P init ; The noise reduction module is used to process the initial point cloud based on the DBSCAN density clustering algorithm. P init Denoising is performed to separate and remove discrete noise points; The preliminary binary mask generation module is used to re-project the denoised point cloud onto the two-dimensional image plane to generate a preliminary binary mask. M init ; The final mask generation module is used to generate the initial binary mask. M init Convolutional filtering and morphological filtering are performed within the contour area to generate the final mask. M refined ; Image optimization module, used to optimize the final mask M refined Image of initial distance D init Element-wise dot product is performed to filter out regions marked as background by the mask, while retaining the distance information of the target region, resulting in a high-quality reconstructed distance image. D final .

9. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-7.

10. A computer device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method of any one of claims 1-7.

Images