A highly reliable real-time tracking method and system for light spots in complex scenes
Patent Information
- Application Number
- CN202611425687.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-15
- Publication Date
- 2026-10-20
AI Technical Summary
[0005]综上，现有技术的根本缺陷在于均以单一方法同时承担光斑跟踪与中心精算双重任务
（1）打破精度-速度矛盾
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the field of optoelectronic reconnaissance technology, specifically to a highly reliable real-time tracking method and system for light spots in complex scenarios. Background Technology
[0002] In the field of optoelectronic reconnaissance, spot tracking is a key technology for target detection, localization and tracking. Spots usually originate from reflected light after laser irradiation of a target or from actively emitted light sources. Under ideal conditions (uniform dark background, single spot), spot tracking can be achieved through simple threshold segmentation and centroid calculation. However, in real-world environments, spot tracking faces severe challenges: (1) Ambient light interference, including various stray light sources such as sunlight, lamplight, and reflected light, significantly reduces the contrast between the spot and the background; (2) Multiple spot confusion, when multiple spot targets exist simultaneously in the scene, traditional methods struggle to accurately distinguish and stably track them.
[0003] Existing technical solutions can be mainly divided into the following two categories: The first category is traditional image processing methods, which mainly include thresholding, centroid method, Gaussian fitting, and correlation filtering.
[0004] The second category comprises deep learning methods, primarily object detection algorithms based on convolutional neural networks (CNNs). YOLO, for example, directly predicts the bounding box of the target using a single neural network, achieving significantly higher tracking accuracy than traditional methods. However, to achieve high-precision spot center localization, the deep learning network requires fine-grained calculations of its output, resulting in high computational demands for model inference and hindering its ability to meet the demands of high-frame-rate real-time tracking. Subsequent improved versions, such as YOLOv8 and YOLOv11, have improved inference speed, but speed bottlenecks still exist in high-frame-rate electro-optical reconnaissance scenarios.
[0005] In summary, the fundamental flaw of existing technologies lies in their reliance on a single method to simultaneously handle both spot tracking and central precision calculation. Traditional methods suffer from inaccurate positioning due to their simple algorithms, while deep learning methods are slow in inference due to their pursuit of accuracy. This creates a technical contradiction where accuracy and speed are mutually exclusive. Therefore, there is an urgent need for a spot tracking solution that can overcome this contradiction and achieve both high accuracy and high real-time performance in complex scenarios. Summary of the Invention
[0006] The purpose of this invention is to provide a highly reliable real-time tracking method for light spots in complex scenes, which can be used for real-time and high-precision light spot tracking in complex scenes.
[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention proposes a highly reliable real-time tracking method for light spots in complex scenes, the method comprising the following steps: Step 1: Image acquisition, acquiring image data containing light spots, and preprocessing the image data; Step 2: Coarse tracking. Based on the improved YOLOv11 network, coarse tracking of light spots is performed on the preprocessed image data to obtain the approximate position of each light spot, the light spot bounding box, and the corresponding light spot region data. Step 3: Fine calculation. Based on the weighted Gaussian fitting module, the spot region data and spot boundary boxes obtained from coarse tracking are subjected to weighted Gaussian fitting under the constraints of the coarse positioning region to calculate the high-precision center coordinates of each spot.
[0008] Furthermore, the improved YOLOv11 network described in this invention includes a backbone network. Neck network and detection head The improved YOLOv11 network expression is as follows:
[0009] Wherein, the symbol . indicates function composition, with the output of the preceding module serving as the input of the following module; the detection head Output the bounding box and confidence score of the light spot, and extract the sub-image of the region corresponding to the light spot as the input for subsequent weighted Gaussian fitting.
[0010] Furthermore, the neck network described in this invention A coordinate convolution module is introduced; this module concatenates the input feature map with normalized coordinate channels along their channel dimensions before performing a two-dimensional convolution; for the input feature map... The coordinate convolution module satisfies:
[0011] in, This indicates a splicing operation along the channel dimension. Conv2D For standard two-dimensional convolution, and Normalize to the interval [-1, 1] respectively.
[0012] Furthermore, the backbone network described in this invention Each feature extraction stage is followed by a CBAM hybrid attention module; the CBAM hybrid attention module includes a channel attention submodule and a spatial attention submodule; The channel attention submodule satisfies:
[0013] The spatial attention submodule satisfies:
[0014] in, This represents the activation function. , Represents the weight matrix. and These represent the global average pooling feature and the max pooling feature along the channel dimension, respectively. Represents a 7×7 convolution; and These represent the global average pooling feature and the max pooling feature in the spatial dimension, respectively.
[0015] Furthermore, the backbone network described in this invention It is set Large kernel separable convolutional modules; as described in the training phase The module includes three parallel branches: a 31×31 depthwise separable convolution branch, a 1×1 pointwise convolution branch, and a batch normalization layer; the module output during the training phase satisfies:
[0016] in, 31 31-depth separable convolution, For 1×1 pointwise convolution, The batch normalization layer consists of three branches responsible for large-scale spatial feature extraction, channel transformation, and residual connection, respectively.
[0017] Furthermore, the inference phase of this invention employs structural reparameterization technology, which transforms the training phase... Multiple parallel branches of a module are equivalently merged into a single-volume active core; The module is equivalent to a single 31×31 standard convolution operation, satisfying:
[0018] in, This is the equivalent 31×31 standard convolution kernel obtained by merging after training through algebraic equivalent transformations.
[0019] Furthermore, the weighted Gaussian fitting described in this invention includes constructing a weighted two-dimensional Gaussian model, for the first... k Sub-image of the region of the target light spot The constructed weighted two-dimensional Gaussian model satisfies:
[0020] in, Let be the coordinates of the center of the light spot to be solved. This represents the peak amplitude. As background base, and They are respectively x direction and y The Gaussian width parameter in the direction.
[0021] Furthermore, the weighted Gaussian fitting described in this invention employs a distance-based Gaussian weighting strategy for the region sub-image. Each pixel Assign weights Construct a weighted least squares objective function:
[0022] Weight Differentiated allocation is performed based on the position of pixels in the light spot; Using a distance-based Gaussian weighted strategy satisfies:
[0023] in, For pixels Distance to the center of the light spot prediction This is the weight decay coefficient.
[0024] Furthermore, this invention minimizes the weighted least squares objective function using a nonlinear optimization algorithm to obtain the optimal parameter set:
[0025] theta_k = argmin L(theta_k) in, Extracting high-precision spot center coordinates from the optimal parameter set. And output it.
[0026] Based on the same inventive concept, this invention provides a highly reliable real-time tracking system for light spots in complex scenes, used to implement the highly reliable real-time tracking method for light spots in complex scenes described above. The system includes an image acquisition module, a preprocessing module, an improved YOLOv11 module, and a weighted Gaussian fitting module. The image acquisition module is used to acquire scene image data containing light spots in real time; The preprocessing module is used to receive image data output by the image acquisition module and preprocess the image data to obtain preprocessed image data. The improved YOLOV11 module is used to receive preprocessed image data, complete coarse spot tracking through the built-in backbone network, neck network and detection head, and output the coarse position of each spot, spot bounding box and corresponding spot region data; the improved YOLOV11 module integrates coordinate convolution, CBAM hybrid attention mechanism and 31×31 large kernel separable convolution improved structure with structure reparameterization. The weighted Gaussian fitting module is used to receive the spot boundary box and spot region data output by the improved YOLOV11 module. Using the spot boundary box as a coarse localization region constraint, it calculates and outputs the high-precision center coordinates of each spot by constructing a weighted two-dimensional Gaussian model, a distance Gaussian weighting strategy, and a nonlinear optimization solution method.
[0027] The beneficial effects of this invention are as follows: (1) Breaking the precision-speed contradiction Traditional methods sacrifice accuracy for speed, while deep learning methods sacrifice speed for accuracy. This invention decouples the localization and fine-mapping tasks, allowing deep learning to perform only coarse localization, significantly improving inference speed; weighted Gaussian fitting performs fine-mapping under coarse localization constraints, achieving sub-pixel accuracy at the center.
[0028] (2) High reliability in complex scenarios The coordinate convolution of this invention provides explicit spatial location information, while the 31×31 large-kernel separable convolution expands the perception capability of a wide range of spatial contexts. Spatial channel attention enhances the expression of the spatiotemporal correlation features of light spots. The spatial constraints provided by coarse localization protect the weighted Gaussian fitting from multi-spot confusion and background clutter interference. These three elements work together to suppress interference from complex scenes in different dimensions, improving the robustness and stability of tracking.
[0029] (3) Cascaded closed-loop complementary effect The coarse localization outputs the spot bounding box and spot region data, providing directly usable spatial constraints for fine calculations. The fine calculations output high-precision center coordinates for each target spot. This compensates for the insufficient accuracy of coarse positioning. The two are not simply sequential, but interdependent and mutually reinforcing—the quality of coarse positioning determines the efficiency of fine calculation, and the accuracy of fine calculation determines the final performance of the entire system, forming a synergistic effect of 1+1>2.
[0030] This invention provides a highly reliable real-time tracking method for spot targets in complex scenarios such as cluttered backgrounds and multiple spot patterns. It can be applied to scenarios such as optoelectronic reconnaissance, laser guidance, and optical measurement. Attached Figure Description
[0031] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0032] Figure 1 This is a flowchart of a highly reliable real-time tracking method for light spots in complex scenes according to the present invention; Figure 2 This is a schematic diagram of the workflow of the present invention; Figure 3 The improved YOLOV11 network block diagram is presented in this invention.
[0033] Among them, R31 is a 31×31 separable convolutional module with structure reparameterization; C3k and C3k2 are residual feature modules; SPPF is a fast spatial pyramid pooling module; CBAM is a convolutional block attention module; CoordConv is a coordinate convolution module; PSABlock is a position-sensitive attention module; C2PSA is a position-sensitive attention cross-stage feature fusion module; convolutional layers, feature splitting, residual modules, feature concatenation, 2d max pooling, and upsampling are internal network operation units; Detect-P3, Detect-P4, and Detect-P5 are detection outputs at scales P3, P4, and P5, respectively. Detailed Implementation
[0034] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The following examples will help those skilled in the art to further understand the present invention, but do not limit the present invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.
[0035] Implementation Method 1: Combination Figure 1 , Figure 2 and Figure 3 This invention provides a detailed description of a highly reliable real-time spot tracking method for complex scenes. By fully disclosing the method execution flow through this embodiment, those skilled in the art can clearly reproduce the technical solution of this invention. It also clarifies that this invention solves the technical contradiction between accuracy and real-time performance in existing technologies, which use a single algorithm to simultaneously complete spot tracking and spot center calculation. Traditional image processing methods have insufficient positioning accuracy, and deep learning methods have high inference overhead, making it difficult to meet the high frame rate real-time requirements of photoelectric reconnaissance.
[0036] The present invention proposes a highly reliable real-time spot tracking method in complex scenes, comprising the following steps: Step 1: Acquire image data containing light spots and preprocess the image data; Step 2: Perform coarse spot tracking on the preprocessed image data based on the improved YOLOv11 network to obtain the approximate position of each spot, the spot bounding box, and the corresponding spot region data; Step 3: Based on the weighted Gaussian fitting module, the spot region data and spot bounding box obtained from coarse tracking are subjected to weighted Gaussian fitting under the constraint of coarse positioning region, and the high-precision center coordinates of each spot are calculated.
[0037] Furthermore, in one specific embodiment, the workflow of the high-reliability real-time spot tracking method proposed in this invention is as follows: Figure 2 As shown, this invention presents a collaborative tracking architecture combining deep learning coarse tracking and weighted Gaussian fitting for fine calculation. Its core idea is to decouple the dual tasks of localization and fine calculation inherent in traditional single-method approaches—deep learning is only responsible for coarse localization, without needing to pursue precise coordinates, thus significantly improving inference speed; the coarse localization result further provides spatial constraints for weighted Gaussian fitting, enabling high-precision center calculation within the constrained region.
[0038] Furthermore, in one specific implementation, such as Figure 1 As shown, it includes the following steps: Step 1: Image Acquisition. Acquire image data containing light spots using a high frame rate camera; let the acquired image be the... t Frame image is ,in This represents pixel coordinates, and the image resolution is W×H.
[0039] Preprocessing involves processing the acquired image data. Linear filtering preprocessing operation, to the preprocessed image Preprocessing may include denoising operations such as Gaussian filtering, median filtering, or bilateral filtering, and its expression is:
[0040] in, This represents a filtering operator.
[0041] Step 2: Coarse Tracking – Based on the improved YOLOv11 network, the preprocessed image data is... The data is transmitted to an improved YOLOv11 network for coarse spot tracking, obtaining the approximate location, bounding box, and corresponding spot region data for each spot. The improved YOLOv11 network in this step incorporates three improvements over the original YOLOv11: (a) the introduction of CoordConv to enhance the network's spatial awareness; (b) the introduction of a Hybrid Attention Mechanism (CBAM) module in the backbone network; and (c) the introduction of 31×31 large-kernel separable convolutions in the backbone network, combined with structural reparameterization techniques to maintain inference efficiency.
[0042] The overall mathematical expression of the improved YOLOv11 network is as follows:
[0043] The dot (.) indicates function composition (i.e., the output of the previous module becomes the input of the next module). Backbone network, through the introduction of That is, a 31×31 large kernel separable convolution module is used to enhance the network's CSP feature extraction capability, and a CBAM hybrid attention module is connected in series after each stage to enhance the network's spatial awareness capability; The Neck network introduces a coordinate convolution module to inject explicit coordinate information into the shallow feature map, and combines the FPN+PAN structure to enhance the network's multi-scale feature fusion capability. Detection Head: Outputs the spot bounding box and confidence score.
[0044] Furthermore, three improvements to the original YOLOV11 are explained; (1) In the neck network By concatenating additional coordinate channels into the input feature map, the convolutional kernel can explicitly utilize the absolute spatial location information of pixels for inference.
[0045] For feature maps The coordinate convolution module can be represented as:
[0046] in, This indicates a splicing operation along the channel dimension. Conv2D For standard two-dimensional convolution, and Normalized to the interval [-1, 1], it can be specifically expressed as:
[0047]
[0048] in, The spatial index of the pixel.
[0049] (2) After each feature extraction stage of the backbone network, a hybrid attention mechanism (CBAM) module is introduced. The CBAM module is composed of a channel attention submodule and a spatial attention submodule combined sequentially.
[0050] Among them, the channel attention submodule Represented as:
[0051] in, This represents the activation function. , Represents the weight matrix. and These represent the global average pooling feature and the max pooling feature of the channel dimension, respectively.
[0052] Spatial Attention Submodule Represented as:
[0053] in, This represents the activation function. Represents a 7×7 convolution. and These represent the global average pooling feature and the max pooling feature in the spatial dimension, respectively.
[0054] (3) To enhance the network's ability to perceive large-scale spatial context, a 31×31 large-kernel separable convolutional module is introduced into the backbone network. Large kernel convolutions can capture long-distance spatial dependencies between light spots and the background, but directly using standard 31×31 convolutions will lead to a sharp increase in the number of parameters and computational cost. Therefore, this invention uses depthwise separable convolutions instead of standard convolutions, and combines this with structural re-parameterization technology to convert the high expressive power of the training phase into efficient computational power for the inference phase.
[0055] Training phase, The module consists of three parallel branches: a 31×31 depthwise separable convolution branch, a 1×1 pointwise convolution branch, and a batch normalization layer; represented as:
[0056] in, 31 31. Depthwise separable convolution (channel-wise convolution). For 1×1 pointwise convolution, This is a batch normalization layer. Three branches are responsible for large-scale spatial feature extraction, channel transformation, and residual connection, respectively.
[0057] During the inference phase, the three branches are equivalently merged into a single convolutional operation through structural reparameterization:
[0058] in, This is an equivalent 31×31 standard convolution kernel obtained after training through algebraic equivalent transformations. Only one standard convolution operation is required during the inference phase, with the number of parameters and computational cost comparable to a single branch during training, ensuring inference efficiency.
[0059] (4) Improve the YOLOV11 module for preprocessing images Perform inference and output the bounding box of the target light spot. and its confidence level This step prioritizes detection speed and does not require precise center coordinates. Simultaneously, it extracts the sub-image of the region corresponding to the target spot. This serves as the input for step 3.
[0060] Furthermore, Figure 3 The diagram shown is of the improved YOLOv11 network. The left side shows the improved backbone network, which is stacked. The large kernel separable convolutional module and C3k2 module are connected to the CBAM hybrid attention module after the output of each stage; SPPF module and C2PSA module are set at the end of the backbone; the middle part is the neck network, which adopts the FPN+PAN architecture and introduces the CoordConv coordinate convolution module to realize explicit coordinate information injection; the right side is the detection head Detect-P3, Detect-P4, Detect-P5; the figure also shows the internal sub-module structure of C3k2, SPPF, and C2PSA.
[0061] The input image is fed into the backbone network, which consists of multiple groups. The modules are constructed by alternating stacking of C3k2 modules; The backbone network employs a 31×31 large-kernel separable convolutional module, combined with structural reparameterization technology, to enhance the network's long-range spatial context capture capability while maintaining inference speed. After each feature extraction stage, the output feature map of the backbone network is followed by a CBAM hybrid attention module to strengthen the target features of the light spot and suppress background clutter interference. The backbone network's backend is configured with SPPF and C2PSA modules to perform high-level feature aggregation; the C2PSA module output is also enhanced by the CBAM module. Multi-scale features from the backbone are fed into the neck network, which is based on an FPN+PAN architecture. Upsampling and concatenation are used to achieve bidirectional fusion of multi-scale features. Furthermore, a CoordConv coordinate convolutional module is introduced into the neck network to concatenate normalized pixel coordinate channels to the feature map, enabling the convolution to utilize absolute pixel spatial location information and improve the light spot location perception capability. The multi-scale features after fusion by the neck network are fed into the Detect-P3, Detect-P4, and Detect-P5 detection heads respectively. The detection heads output the bounding box and confidence score of the spot to complete the coarse spot tracking task. They output the approximate position of the spot, the bounding box of the spot, and the sub-image of the region corresponding to the spot, which are then passed to the subsequent weighted Gaussian fitting module to solve for the spot center with sub-pixel accuracy.
[0062] Step 3: Fine calculation – Based on the weighted Gaussian fitting module, the spot region data obtained from coarse tracking is processed. and the light spot boundary frame The data is transmitted to the spot fitting module, where a weighted Gaussian fit is performed on each spot under the constraint of the coarse positioning region to calculate the high-precision center coordinates of each spot. Specifically, this includes: (1) Construction of weighted Gaussian model: For the target light spot k Region sub-image Construct a weighted two-dimensional Gaussian model:
[0063] in, Let be the coordinates of the center of the light spot to be solved. This represents the peak amplitude. As background base, and They are respectively x direction and y The Gaussian width parameter in the direction.
[0064] (2) Weighting strategy: for regional sub-images Each pixel in Assign weights Construct a weighted least squares objective function:
[0065] Weight Differentiated allocation is performed based on the position of the pixel in the light spot.
[0066] Using a distance-based Gaussian weighted strategy satisfies:
[0067] in, For pixels Distance to the center of the light spot prediction is the weight attenuation coefficient. This strategy assigns higher weights to pixels in the center of the spot and lower weights to pixels in the edge areas, thereby effectively suppressing the impact of edge noise and background clutter on fitting accuracy in complex scenes.
[0068] (3) Parameter solution: Minimize the objective function L using nonlinear optimization methods (such as the Levenberg-Marquardt algorithm) to solve for the optimal parameter set:
[0069] theta_k = argmin L(theta_k) in .
[0070] Output the high-precision center coordinates of each target spot. ,in k =1,2,...,K.
[0071] Through the above design, the highly reliable real-time tracking method for light spots in complex scenes proposed in this invention has the following technical innovations: (1) Balancing real-time tracking with sub-pixel-level positioning accuracy, breaking the technical contradiction of mutual constraint between accuracy and speed in traditional solutions. This invention decouples the spot tracking task into two stages: coarse positioning and high-precision center solving. The improved YOLOV11 network is used to undertake only the coarse spot tracking task. The network only outputs the rough position of the spot, the bounding box, and the spot area data, without outputting the high-precision spot center, which greatly reduces the inference burden of the deep learning network and ensures high frame rate real-time processing capability. Then, the effective calculation area is defined by the spot bounding box output by the improved YOLOV11. Under the coarse positioning constraint, a weighted Gaussian fitting based on the distance weighting strategy is performed. Nonlinear optimization is only performed on the local area of the spot within the bounding box, avoiding the computational power consumption caused by full-image calculation. Finally, the sub-pixel-level spot center coordinates are obtained, realizing the simultaneous improvement of algorithm inference speed and positioning accuracy.
[0072] (2) Multi-dimensional collaborative suppression of interference from complex scenes such as ambient stray light and multiple spot confusion improves the robustness and reliability of spot tracking. On the one hand, coordinate convolution provides explicit spatial location information, 31×31 large kernel separable convolution expands the perception of a wide range of spatial context, and spatial channel attention enhances the expression of the spatiotemporal correlation features of the spot; the spatial constraints provided by coarse localization protect the weighted Gaussian fitting from interference from multiple spot confusion and background clutter. The three work together to suppress interference from complex scenes in different dimensions, improving the robustness and stability of tracking. On the other hand, the spot bounding box output by coarse tracking provides an independent spatial constraint region for subsequent weighted Gaussian fitting. Different spots correspond to their own independent fitting calculation intervals, avoiding mutual interference between multiple spots; at the same time, the weighted Gaussian fitting adopts a distance-based Gaussian weighting strategy, which increases the weight of the center pixel of the spot and reduces the weight of the noise pixel at the edge of the spot, further suppressing the interference of background clutter and edge noise on the fitting results. The combination of multiple technical means significantly improves the stability of spot tracking in complex scenes of photoelectric reconnaissance.
[0073] (3) Coarse localization and fine calculation form a cascaded closed-loop synergistic gain, achieving a combined technical effect of 1+1>2. The improved YOLOV11 coarse tracking output spot bounding boxes and spot region sub-images are directly used as inputs for weighted Gaussian fitting, providing spatial constraints and local image data for the fitting operation, limiting the effective range of the fitting solution, and reducing the participation of invalid pixels in the operation. The detection effect of coarse localization directly determines the operational efficiency and convergence stability of subsequent fitting. The weighted Gaussian fitting solution obtains the spot center coordinates with sub-pixel precision, making up for the inherent defect of YOLO network detection head that can only output bounding boxes and cannot obtain high-precision centers, and outputting the final high-precision localization result of the system. The two are not simply sequential executions. Coarse localization provides the computational basis for fine calculation, and fine calculation makes up for the accuracy shortcomings of deep learning detection. The two techniques complement each other and enhance each other. The overall system performance is better than using deep learning detection alone or using Gaussian fitting algorithm alone, achieving a synergistic amplification effect.
[0074] Implementation Method 2: This implementation method provides a highly reliable real-time spot tracking system for complex scenes, used to implement the highly reliable real-time spot tracking method for complex scenes proposed above. The system adopts a cascaded architecture of "deep learning coarse localization + Gaussian fitting fine calculation", which effectively solves the technical contradiction of insufficient accuracy of traditional algorithms and poor real-time performance of pure deep learning algorithms. It can achieve high-precision, high-frame-rate stable spot tracking in complex optoelectronic reconnaissance scenes with stray light interference and multiple spots coexisting.
[0075] Specifically, the system includes an image acquisition module, a preprocessing module, an improved YOLOv11 module, and a weighted Gaussian fitting module, which are connected in sequence.
[0076] The image acquisition module is used to acquire scene image data containing the target light spot in real time, providing raw image input for subsequent light spot tracking.
[0077] The preprocessing module is connected to the image acquisition module and is used to receive raw image data and perform image preprocessing operations, including denoising operations such as Gaussian filtering, median filtering or bilateral filtering, to improve the stability and accuracy of subsequent network feature extraction.
[0078] The improved YOLOV11 module, serving as the core unit for coarse localization, receives preprocessed image data and collaborates with the optimized backbone network, neck network, and detection head to complete the coarse spot tracking task. Specifically, this module integrates multiple optimized structures based on the original YOLOV11: the backbone network introduces a 31×31 large-kernel separable convolution with structure reparameterization to enhance the network's ability to perceive large-scale spatial context; and after each feature extraction stage of the backbone network, a hybrid attention mechanism (CBAM) module is introduced to enhance the network's spatial perception ability. The neck network introduces a coordinate convolution module to inject explicit coordinate information into the shallow feature map, and combines it with an FPN+PAN structure to enhance the network's multi-scale feature fusion capability; the detection head is only responsible for quickly outputting the coarse position of each spot, the spot bounding box, and the corresponding spot region sub-image, without performing high-precision center solving, greatly reducing the network inference overhead and ensuring the system's real-time performance.
[0079] The weighted Gaussian fitting module, serving as the system's high-precision calculation unit, is data-connected to the improved YOLOv11 module and receives the spot boundary boxes and spot region data output from the coarse tracking. This module uses the spot boundary boxes as coarse localization region constraints, limiting the fitting operation to only within the effective spot region, thus avoiding interference from multiple spot crossovers and large-area background noise. By constructing a weighted two-dimensional Gaussian model and employing a pixel-distance-based Gaussian weighting strategy to allocate pixel weights, it suppresses fitting interference from spot edge noise and background clutter. Finally, a nonlinear optimization algorithm iteratively solves for the optimal Gaussian model parameters, accurately outputting the sub-pixel-level high-precision center coordinates of each spot, achieving highly reliable real-time tracking of spots in complex scenes.
[0080] This system uses a dual-module cascaded complementary architecture to split the spot tracking task into two independent stages: high-speed coarse positioning and high-precision fine calculation. It relies on the improved YOLOV11 module to ensure the tracking frame rate and target anti-interference capability, and relies on the weighted Gaussian fitting module to make up for the lack of positioning accuracy of deep learning bounding boxes. It achieves the best performance in both real-time and positioning accuracy, and is suitable for the engineering application requirements of complex optoelectronic reconnaissance scenarios.
[0081] Implementation Method 3: This implementation method provides a highly reliable real-time spot tracking method in complex optoelectronic reconnaissance scenarios, specifically including the following steps: Step 1: Image acquisition.
[0082] A high frame rate industrial camera is used to acquire image sequences of the target scene. Let the acquired t-th frame be... The image size is W=1920, H=1200, and the pixel depth is 8-bit grayscale.
[0083] Step 2: Preprocessing.
[0084] For the acquired images Gaussian filtering was used for noise reduction, with a kernel size of 3×3 and a standard deviation of [missing value]. The preprocessed image is as follows:
[0085] Among them, symbols This represents the convolution operation. This is a two-dimensional Gaussian kernel function. This preprocessing step effectively suppresses camera sensor noise while preserving the grayscale distribution characteristics of the light spot. Alternatively, a bilateral filter can be used to denoise while maintaining the edge of the light spot.
[0086] Step 3: Coarse positioning based on improved YOLOV11.
[0087] The overall computational process of the improved YOLOv11 network can be represented as follows:
[0088] Step 4: Fine calculation based on weighted Gaussian fitting.
[0089] As mentioned above, the sub-pixel precision center coordinates of each spot are obtained through initial parameter estimation, weighted matrix construction, and LM nonlinear optimization. .
[0090] Step 5: Output high-precision center coordinates.
[0091] Output the center coordinates of all K light spots to the downstream task.
[0092] The above description of the technical solution provided by the present invention through several specific embodiments is intended to highlight the advantages and benefits of the technical solution provided by the present invention. However, the above-described specific embodiments are not intended to limit the present invention. Any reasonable modifications and improvements to the present invention, reasonable combinations of implementation methods and equivalent substitutions based on the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A highly reliable real-time tracking method for light spots in complex scenes, characterized in that, Includes the following steps: Step 1: Image acquisition, acquiring image data containing light spots, and preprocessing the image data; Step 2: Coarse tracking. Based on the improved YOLOv11 network, coarse tracking of light spots is performed on the preprocessed image data to obtain the approximate position of each light spot, the light spot bounding box, and the corresponding light spot region data. Step 3: Fine calculation. Based on the weighted Gaussian fitting module, the spot region data and spot boundary boxes obtained from coarse tracking are subjected to weighted Gaussian fitting under the constraints of the coarse positioning region to calculate the high-precision center coordinates of each spot.
2. The highly reliable real-time tracking method for light spots in complex scenes according to claim 1, characterized in that, The improved YOLOv11 network includes a backbone network. Neck network and detection head The improved YOLOv11 network expression is as follows: Wherein, the symbol . indicates function composition, with the output of the preceding module serving as the input of the following module; the detection head Output the bounding box and confidence score of the light spot, and extract the sub-image of the region corresponding to the light spot as the input for subsequent weighted Gaussian fitting.
3. The highly reliable real-time tracking method for light spots in complex scenes according to claim 2, characterized in that, The neck network A coordinate convolution module is introduced; this module concatenates the input feature map with normalized coordinate channels along their channel dimensions before performing a two-dimensional convolution; for the input feature map... The coordinate convolution module satisfies: in, This indicates a splicing operation along the channel dimension. Conv2D For standard two-dimensional convolution, and Normalize to the interval [-1, 1] respectively.
4. The highly reliable real-time tracking method for light spots in complex scenes according to claim 2, characterized in that, The backbone network Each feature extraction stage is followed by a CBAM hybrid attention module; the CBAM hybrid attention module includes a channel attention submodule and a spatial attention submodule; The channel attention submodule satisfies: The spatial attention submodule satisfies: in, This represents the activation function. , Represents the weight matrix. and These represent the global average pooling feature and the max pooling feature along the channel dimension, respectively. Represents a 7×7 convolution; and These represent the global average pooling feature and the max pooling feature in the spatial dimension, respectively.
5. A highly reliable real-time tracking method for light spots in complex scenes according to claim 2, characterized in that, The backbone network It is set Large kernel separable convolutional modules; as described in the training phase The module includes three parallel branches: a 31×31 depthwise separable convolution branch, a 1×1 pointwise convolution branch, and a batch normalization layer; the module output during the training phase satisfies: in, 31 31-depth separable convolution, For 1×1 pointwise convolution, The batch normalization layer consists of three branches responsible for large-scale spatial feature extraction, channel transformation, and residual connection, respectively.
6. The highly reliable real-time tracking method for light spots in complex scenes according to claim 5, characterized in that, The inference phase employs structure reparameterization, which is used in the training phase. Multiple parallel branches of a module are equivalently merged into a single-volume active core; The module is equivalent to a single 31×31 standard convolution operation, satisfying: in, This is the equivalent 31×31 standard convolution kernel obtained by merging after training through algebraic equivalent transformations.
7. The highly reliable real-time tracking method for light spots in complex scenes according to claim 1, characterized in that, The weighted Gaussian fitting includes constructing a weighted two-dimensional Gaussian model, for the first... k Sub-image of the region of the target light spot The constructed weighted two-dimensional Gaussian model satisfies: in, Let be the coordinates of the center of the light spot to be solved. Peak amplitude, As background base, and They are respectively x direction and y The Gaussian width parameter in the direction.
8. The highly reliable real-time tracking method for light spots in complex scenes according to claim 7, characterized in that, The weighted Gaussian fitting employs a distance-based Gaussian weighting strategy for the region sub-image. Each pixel Assign weights Construct a weighted least squares objective function: Weight Differentiated allocation is performed based on the position of pixels in the light spot; Using a distance-based Gaussian weighted strategy satisfies: in, For pixels Distance to the center of the light spot prediction This is the weight decay coefficient.
9. A highly reliable real-time tracking method for light spots in complex scenes according to claim 8, characterized in that, The optimal parameter set is obtained by minimizing the weighted least squares objective function using a nonlinear optimization algorithm. theta_k = argmin L(theta_k) in, Extracting high-precision spot center coordinates from the optimal parameter set. And output it.
10. A highly reliable real-time tracking system for light spots in complex scenes, characterized in that, The system is the same as the highly reliable real-time tracking method for light spots in complex scenes according to any one of claims 1-9. The system includes an image acquisition module, a preprocessing module, an improved YOLOv11 module, and a weighted Gaussian fitting module. The image acquisition module is used to acquire scene image data containing light spots in real time; The preprocessing module is used to receive image data output by the image acquisition module and preprocess the image data to obtain preprocessed image data. The improved YOLOV11 module is used to receive preprocessed image data, complete coarse spot tracking through the built-in backbone network, neck network and detection head, and output the coarse position of each spot, spot bounding box and corresponding spot region data; the improved YOLOV11 module integrates coordinate convolution, CBAM hybrid attention mechanism and 31×31 large kernel separable convolution improved structure with structure reparameterization. The weighted Gaussian fitting module is used to receive the spot boundary box and spot region data output by the improved YOLOV11 module. Using the spot boundary box as a coarse localization region constraint, it calculates and outputs the high-precision center coordinates of each spot by constructing a weighted two-dimensional Gaussian model, a distance Gaussian weighting strategy, and a nonlinear optimization solution method.