Asteroid automatic identification method suitable for Antarctic unattended station site
By integrating data augmentation processing and improving the target detection network, the problem of rapid deployment and efficient detection of asteroids in the unattended environment of Antarctica was solved, and high-precision asteroid detection was achieved under low signal-to-noise ratio and complex star field conditions.
Patent Information
- Application Number
- CN202511802922.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-03
- Publication Date
- 2026-02-24
AI Technical Summary
Existing technologies are difficult to deploy quickly and meet the requirements of high timeliness and high reliability for automatic asteroid identification in the unattended environment of Antarctica. In particular, under conditions of low signal-to-noise ratio and complex star field, traditional methods are not sensitive to faint targets and have high online inference costs, making it difficult to meet real-time requirements.
An automatic asteroid identification method suitable for unattended sites in Antarctica is designed, including a data processing module and a target detection module. It adopts integrated data augmentation processing, scale-aware feature pyramid network, dual-path query optimization mechanism and scale balance training strategy to improve the visibility and detection accuracy of faint streaks.
It has achieved rapid deployment and efficient and accurate asteroid detection in the unattended environment of Antarctica, significantly improving the recall rate and overall accuracy of faint targets, and reducing the false detection rate and missed detection rate. It is suitable for automated monitoring of long-term unattended sites in Antarctica.
Smart Images
Figure CN121564322A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the design of an automatic asteroid identification method suitable for unmanned sites in Antarctica, specifically to the design of automatic processing and rapid asteroid detection for sky survey data that can be rapidly deployed and operated unmanned for a long period in the Dome A region of Antarctica. This method enables the automatic discovery of faint space targets such as asteroids without human intervention, thereby supporting the rapid deployment and stable operation of long-term unmanned sites in Antarctica. Background Technology
[0002] Asteroid monitoring and planetary defense are crucial for ensuring human space activities and Earth's safety. The number of near-Earth asteroids worldwide continues to increase, a significant portion of which pose a potential impact risk. Events like the Chelyabinsk incident have demonstrated that even small celestial bodies, around 20 meters in size, can cause regional damage and social impact. Therefore, building a sustainable and scalable automatic asteroid identification and early warning system, with a focus on highly efficient and reliable automated monitoring, has become a fundamental capability development goal within planetary defense systems.
[0003] Dome A in Antarctica boasts environmental advantages such as long polar nights, excellent seeing, extremely low water vapor levels, and virtually no light pollution, providing stable, low-disturbance sky backgrounds and high temporal sampling frequency observation conditions during months of continuous polar night. Compared to mid- and low-latitude sites, these natural advantages significantly improve the detectability and orbital accuracy of faint moving targets in space, providing an ideal platform for the rapid discovery of asteroids and other moving targets. However, the short working window and short research stay in Antarctica often result in Dome A being unmanned. A heavily human-assisted working mode cannot fully utilize the continuous observation period during the polar night and fails to meet the timeliness, reliability, and availability requirements of planetary defense. Therefore, an automatic detection algorithm with near-real-time alarm capabilities is needed to support long-term unmanned detection operations.
[0004] Traditional asteroid detection methods largely rely on motion feature analysis, such as digital tracking, temporal index filtering, and inter-frame differencing. While deep learning methods have made progress on some datasets in recent years, they generally suffer from insensitivity to low signal-to-noise ratio (SNR) and faint streaks, insufficient consideration of multi-scale targets, high online inference costs, and difficulty meeting the real-time requirements of unattended operations in polar nights. Therefore, it is essential to design an automatic asteroid identification method that can be rapidly deployed under low SNR and complex star field conditions and is suitable for unattended operations in Antarctica. Summary of the Invention
[0005] This invention proposes an automatic asteroid detection algorithm design for rapid deployment in the extreme Antarctic environment, which is suitable for long-term unattended operation. The design mainly includes two parts: data processing and target detection, which solves the problem of rapid deployment and long-term unattended operation of observation sites in the extreme Antarctic environment.
[0006] To achieve the above objectives, the system solution adopted by the present invention is as follows: An automatic asteroid identification method applicable to unattended Antarctic sites includes a data processing module and a target detection module, wherein: The data processing module is used to perform integrated enhancement processing on the ground-based optical observation sequence under a unified spatiotemporal reference. It includes an intra-span enhancement unit, a reference system 1 and star registration unit, a temporal enhancement unit, and a clipping and loading unit to improve the visibility of faint streaks under low signal-to-noise ratio conditions. The target detection module is communicatively connected to the data processing module and is used to receive the image sub-image after integrated enhancement and perform real-time asteroid streak detection. The target detection module includes at least a feature fusion and scale adaptation module, a query decoding and target generation module, and a scale balance training strategy. The target detection network returns a list of detected targets for each sub-image, and each target includes a category, confidence level, and pre-selected box.
[0007] Furthermore, the integrated data augmentation processing includes the following steps: Step 1: Perform outlier correction, range normalization, and inverse hyperbolic sine grayscale stretching on the original observation frames; Step 2: Perform automatic world coordinate system calculation and write-back to unify the celestial reference system, and complete star registration and geometric consistency under the constraint of reference frame; Step 3: Within the time window T, take the maximum value pixel by pixel of the registered sequence to obtain the enhanced result of the sliding window maximum value superposition; Step 4: Crop the enhanced results into fixed-size sub-images and input them into the target detection module.
[0008] Furthermore, in step 3, the time window T is adjusted according to the asteroid's motion speed, and is generally selected as any integer between 5 and 20 frames, preferably 5 to 15 frames; the maximum value of the sliding window is superimposed on the pixel position to take the maximum value of the pixels within the window, so as to highlight the continuous bright band of the moving object and suppress random noise.
[0009] Furthermore, the fixed size of the cropped sub-image is 640×640 pixels.
[0010] Furthermore, the target detection module includes the following steps: (a) Feature fusion and scale adaptation module, used to perform cross-layer fusion and scale attention based on feature pyramid to enhance details of dark streaks; (b) The query decoding and target generation module adopts a dual-path decoding strategy of dense and sparse queries to output candidate boxes, categories and confidence scores; (c) Scale-balanced training strategy: introduce scale weights and constraints in the matching, loss weighting and sample sampling stages to improve the recall rate of small targets and low signal-to-noise ratio targets without increasing inference overhead.
[0011] Furthermore, the improved detection network model based on real-time detection transformer (SORT-RTDETR) retains the end-to-end and real-time characteristics of the real-time detection transformer model (RT-DETR). Its detection head still adopts the query-based decoding and cross-union ratio-based quality assessment mechanism of the real-time detection transformer model (RT-DETR). On this basis, the improved detection network model based on real-time detection transformer introduces scale-aware feature pyramid and dual-path query optimization, and adopts a scale-balanced training strategy during the training period to improve the detection accuracy and stability in low signal-to-noise ratio scenarios.
[0012] Furthermore, the improved detection network model based on real-time detection transformer (SORT-RTDETR) extracts multi-scale features in the backbone network and replaces the hybrid encoder with a scale-aware feature pyramid network at the encoding end. The scale-aware feature pyramid network consists of three parts: an adaptive scale attention mechanism, a cross-scale semantic interaction module, and a dynamic weighted fusion module. Among them, the adaptive scale attention mechanism uses multi-receptive field modeling and dynamic weighted noise suppression and signal enhancement within each scale to highlight dark and weak streaks; the cross-scale semantic interaction module uses attention to establish long-range dependencies and semantic alignment between features at each level, bridging the gap between details and semantics; the dynamic weighted fusion module performs gated recalibration of multi-scale channels based on global statistics, so that high-level semantics and low-level details are adaptively matched according to the scene, and finally outputs cross-layer fused features in a unified semantic space.
[0013] Furthermore, at the decoding end, a dual-path query optimization mechanism replaces the single-path query selection. This dual-path query optimization mechanism includes dense path query and sparse path query. Specifically, the dense path query constructs a regularized geometric prior grid based on high-resolution features and injects positional scale embedding to spatially cover weak and narrow streaks. The sparse path query uses a lightweight candidate network to perform top K saliency screening and positional refinement at the deep semantic level, obtaining a more robust semantic prior for complex backgrounds. The two types of queries are fused by a query fusion module to achieve geometric-semantic alignment in a unified representation space using self-attention and cross-attention, serving as the initial query and reference point input for the decoder. This improves matching stability and convergence speed in scenarios with weak signals and noise interference.
[0014] Furthermore, the scale-balanced training strategy incorporates a scale factor into Hungarian matching to increase the priority of positive samples for small targets, and adaptively weights the samples by scale in classification, regression, and focus loss, combined with scale-balanced sampling, to suppress long-tail bias.
[0015] Furthermore, the normalized output of the object detection network for each subgraph includes: a list of detected objects, each object containing a class label, confidence score, and pre-selected bounding box; and the subgraph index, frame number, and batch number associated with that subgraph.
[0016] Furthermore, the method, while balancing speed and accuracy, is particularly suitable for rapid deployment in unattended environments in Antarctica.
[0017] The beneficial effects and advantages of this invention are as follows: 1. The modular design adopted in this invention is conducive to the rapid deployment of unmanned stations in Antarctica while taking into account both the accuracy and efficiency of the algorithm; 2. The present invention employs an integrated data augmentation processing module, which enables the algorithm to significantly improve the visibility and separability of faint streaks in low signal-to-noise ratio and dense star backgrounds, thereby reducing the false negative rate.
[0018] 3. This invention introduces improved strategies such as scale-aware feature pyramid network, dual-path query optimization mechanism and scale-balanced training strategy at the target detection end, so that the detector can achieve multi-scale robust modeling and stable decoding on streaks of different scales and small targets with low signal-to-noise ratio. Without increasing inference overhead, it significantly improves the recall rate and overall accuracy of small targets and targets with low signal-to-noise ratio, and further reduces false detection and false negative detection. Attached Figure Description
[0019] Figure 1 Flowchart for data augmentation processing; Figure 2 This is an overlay of the maximum values of the sliding window. Figure 3 The structure diagram of the improved detection network model based on the real-time detection transformer (SORT-RTDETR); Figure 4 A schematic diagram of a scale-aware feature pyramid network; Figure 5 This is a schematic diagram of the adaptive scale attention mechanism module; Figure 6 This is a schematic diagram of a cross-scale interaction module; Figure 7 This is a schematic diagram of the dual-path query optimization mechanism module; Figure 8 Figure showing experimental results for the improved detection network model based on the real-time detection transformer (SORT-RTDETR); Figure 9Examples of detection results for various typical scenarios are provided. Detailed Implementation
[0020] The invention will now be further described with reference to the accompanying drawings.
[0021] An automatic asteroid identification method applicable to unattended sites in Antarctica is proposed. This method is suitable for rapid deployment in unattended environments and includes two main modules: a data processing module and a target detection module.
[0022] An automatic asteroid identification method applicable to unattended Antarctic sites includes a data processing module and a target detection module, wherein: The data processing module is used to perform integrated enhancement processing on the ground-based optical observation sequence under a unified spatiotemporal reference. It includes an intra-span enhancement unit, a reference system 1 and star registration unit, a temporal enhancement unit, and a clipping and loading unit to improve the visibility of faint streaks under low signal-to-noise ratio conditions.
[0023] The integrated data augmentation process includes the following steps: Step 1: Perform outlier correction, range normalization, and inverse hyperbolic sine grayscale stretching on the original observation frames; Step 2: Perform automatic world coordinate system calculation and write-back to unify the celestial reference system, and complete star registration and geometric consistency under the constraint of reference frame; Step 3: Within the time window T, the maximum value is taken for each pixel of the registered sequence to obtain the enhanced result of the sliding window maximum value superposition; the time window T is adjusted according to the asteroid's motion speed, and is generally selected as any integer between 5 and 20 frames, preferably 5 to 15 frames; the sliding window maximum value superposition takes the maximum value of the pixels within the window at each pixel position to highlight the continuous bright band of the moving object and suppress random noise; Step 4: Crop the enhanced results into fixed-size sub-images and input them into the target detection module.
[0024] The target detection module is communicatively connected to the data processing module and is used to receive the image sub-image after integrated enhancement and perform real-time asteroid streak detection. The target detection module includes at least a feature fusion and scale adaptation module, a query decoding and target generation module, and a scale balance training strategy. The target detection network returns a list of detected targets for each sub-image, and each target includes a category, confidence level, and pre-selected box.
[0025] The target detection module includes the following steps: (a) Feature fusion and scale adaptation module, used to perform cross-layer fusion and scale attention based on feature pyramid to enhance details of dark streaks; (b) The query decoding and target generation module adopts a dual-path decoding strategy of dense and sparse queries to output candidate boxes, categories and confidence scores; (c) Scale-balanced training strategy: introduce scale weights and constraints in the matching, loss weighting and sample sampling stages to improve the recall rate of small targets and low signal-to-noise ratio targets without increasing inference overhead.
[0026] The standardized output of the object detection network for each subgraph includes: a list of detected objects, each object containing a class label, confidence score, preselected bounding box, and the subgraph index, frame number, and batch number associated with that subgraph.
[0027] To improve the detectability of asteroid trajectories under low signal-to-noise ratio conditions, the data processing module of this invention constructs an end-to-end workflow based on real-world data captured by the AST3 Antarctic Survey Telescope. The data processing module is as follows: Figure 1 As shown, the original frame is first subjected to outlier correction, unified normalization, and inverse hyperbolic sine grayscale stretching, and then solved and rewritten using an automatic world coordinate system to achieve a unified celestial reference frame. Based on this, a star registration strategy driven by "unified normalization and inverse hyperbolic sine grayscale stretching" is proposed. Global geometric alignment is achieved using the Astroalign astronomical image automatic registration algorithm library (Astroalign), significantly enhancing the matching stability and accuracy of weak signal star fields, while simultaneously estimating affine transformations. Where 𝐴 describes rotation, scaling, and shearing, and 𝑏 represents translation), the registration result inherits the world coordinate system information of the reference frame; subsequently, multi-frame sliding window maximum superposition is performed on a geometrically consistent sequence to amplify the continuous bright bands of the moving object and suppress random noise, the enhancement form being as follows: (t is the starting frame, w is the window length); finally, the enhanced full-frame image is cropped into a 640×640 sub-image and input into the detection end. An example of image data processing is shown below. Figure 2 As shown, Figure (a) is the preprocessed image, and Figure (b) is the superimposed and enhanced image.
[0028] The target detection module is based on the Transformer-based end-to-end detection framework RT-DETR. Although this network has good global modeling and real-time performance, it still has limitations when directly applied to asteroid detection: it is sensitive to weak signals and scale changes, and is easily affected by stellar background and noise interference; multi-scale feature fusion is insufficient, resulting in a low detection rate for faint targets; and query selection is prone to unstable convergence under complex noise fields. Therefore, this invention proposes an improved detection network model based on a real-time detection transformer (SORT-RTDETR) (e.g., Figure 3While retaining the end-to-end and real-time advantages of the Real-Time Detection Transformer (RT-DETR) model, this paper combines the characteristics of astronomical images, introduces a scale-aware feature pyramid network and a dual-path query optimization mechanism, and adopts a scale-balanced training strategy during training to significantly improve the detection accuracy and stability in low signal-to-noise ratio scenarios. The principle is as follows: First, multi-scale features {P2, P3, P4, P5} are extracted by the backbone network. Addressing the characteristics of astronomical images—low signal-to-noise ratio, dense star distribution, and large scale span—a scale-aware feature pyramid network is used to replace the original hybrid encoder at the encoding end. Figure 4 Its core consists of three parts: an adaptive scale attention mechanism, a cross-scale semantic interaction module, and a dynamic weighted fusion strategy.
[0029] Adaptive scale attention mechanisms (e.g.) Figure 5 Within each scale, multi-receptive field modeling and dynamic weighting are used for noise suppression and signal enhancement to highlight faint streaks. Specifically, (1) Input and branch construction: The input features are denoted as Three scale branches are constructed in parallel at the same spatial resolution to obtain representations of different receptive fields: the small-scale branch... Apply dilated convolution or equivalent stacking of small convolution kernels to enhance local edges and narrow streaks; mesoscale branch pairs Apply standard 3x3 convolutions to balance local details with contextual structure; large-scale branch pairs A larger receptive field is applied to introduce stronger global semantics and long-range dependencies. All three branches contain the smallest units of convolution, normalization, and activation functions, with outputs denoted as follows: To ensure the feasibility of element-wise weighting, the three-branch output and Maintain consistency in the number of channels and spatial size. (2) Scale predictor: for Global average pooling is applied to obtain the channel description vector; this vector is then passed through a lightweight multilayer perceptron or a 1x1 convolution to generate three-dimensional scale weights. It is then normalized using the Softmax function to satisfy... The corresponding formula is: This step is essentially an adaptive gating of the importance of different receptive fields under global statistical constraints, in order to suppress the influence of stellar point noise on the attention distribution. (3) Weighted fusion and residual connection: The three-branch outputs are summed element-wise according to the weights 𝛼, and the residuals are added to the input feature 𝑍 to obtain: . Among them, 𝑍 is the input feature, and the residual structure can effectively avoid information loss in the multi-branch fusion process and ensure that weak signals can still be transmitted. (4) Post-processing steps: In order to improve numerical stability and transferability, post-processing steps can be performed. Then, lightweight normalization and activation operations are added, or channel recalibration is performed to ensure consistency with the feature dimensions of subsequent cross-scale interaction modules. (5) Mechanism of action: The small-scale branch mainly strengthens narrow streaks and local edge responses; the medium-scale branch retains the main structure and texture details; the large-scale branch captures global semantics and long-range dependencies. The scale predictor based on global average pooling will give lower weights to random noise similar to point sources in low signal-to-noise ratio star fields, while giving higher weights to moving streaks with continuous bright bands. This mechanism effectively improves the saliency and separability of faint streaks without changing the inference resolution and computational complexity, while reducing the false detection rate and false alarm rate in dense star backgrounds. The entire module is end-to-end trainable, the increase in parameters is controllable, and only constant-level branches and element-wise fusion are introduced during inference, without affecting the overall real-time detection performance.
[0030] Cross-scale semantic interaction modules (such as) Figure 6 ) Use attention to establish long-range dependencies and semantic alignment between features at different levels, bridging the gap between details and semantics. Specifically: (1) Input and feature preparation: This module receives a multi-scale feature set output from the adaptive scale attention mechanism. The feature maps at each scale are flattened into a two-dimensional sequence for input into the attention structure. To explicitly represent the scale properties of different resolution layers, a learnable scale embedding vector is introduced for each layer's features. This is then added element-wise to the flattened features, thereby achieving explicit encoding of scale information. The corresponding formula is: This step ensures that features from different resolution layers have independent scale labels before entering the attention mechanism, enabling the model to distinguish the source of hierarchical semantics. (2) Multi-head attention input construction: The current layer features The features from the remaining layers are concatenated as the query (Q) and serve as the key (K) and value (V), achieving cross-scale information alignment. The specific form is as follows: This operation establishes a global association between features at different scales, enabling the model to simultaneously perceive detailed information and semantic context in a unified feature space, thereby capturing long-range dependencies. (3) Cross-scale semantic aggregation: using multi-head attention mechanism to... Weighted aggregation is performed to achieve adaptive information fusion between features at different scales. Its calculation form is as follows: The attention weights are adaptively determined by the similarity between the query and the key, thereby dynamically integrating contextual information from other scales at each location. Through this mechanism, high-resolution layer features can obtain semantic supplementation from low-resolution layers, while low-resolution layer features can obtain local detail enhancement from high-resolution layers, achieving bidirectional fusion of semantics and structure. (4) Interlayer feature reorganization and output: fused output After reshaping, the original spatial size is restored so that it can be connected with the subsequent dynamic weighting module. Lightweight normalization and residual connections can be added at this stage to ensure feature stability and training convergence. The final output cross-scale fusion feature set is significantly improved in terms of global semantic consistency and local structural expression. (5) Mechanism advantages and function description: This module breaks through the local receptive field limitation of traditional convolutional fusion and can establish long-range dependencies between hierarchical features in the global scope. The high-resolution layer retains more spatial details and target contour information, while the low-resolution layer aggregates global context and semantic information, thereby achieving semantic alignment and detail compensation. Compared with fixed convolution or linear fusion, the cross-scale semantic interaction module can significantly improve the detection robustness and feature discrimination of dark streaks and small moving targets in complex star fields and low signal-to-noise ratio scenes, providing more consistent and separable semantic input for the subsequent dynamic fusion stage.
[0031] The dynamic weighted fusion strategy then performs gated recalibration on the multi-scale channels based on global statistics, so that the high-level semantics and low-level details are adaptively matched according to the scene, and finally outputs cross-layer fusion features under a unified semantic space, which enhances the separability of short streaks and low-contrast small targets without significantly increasing computation. Specifically: (1) Module input and design purpose: After the adaptive scale attention mechanism and cross-scale semantic interaction module, a multi-level cross-scale fusion feature set is obtained. Since the contribution of features at different scales to the detection task varies, if a fixed weighting method is used, small targets and weak signals will be masked by high-level semantics, or low-level detail noise will be excessively amplified. Therefore, this module establishes a scale weight learning mechanism based on global statistical information to realize the dynamic recalibration of multi-scale features at the semantic level. (2) Global statistical extraction: Global average pooling is performed on the output feature map of each scale to obtain its global statistical description. Global average pooling compresses the spatial dimension of each feature channel into a single statistic, forming a vector representing the overall response intensity of the scale. The global average pooling results of all scales are concatenated according to the channel dimension to obtain the global feature description. The result summarizes the average activation information of features at different scales in the global space, providing input basis for subsequent weight generation. (3) Scale weight generation: The concatenated global statistical vector is mapped to the same dimension as the number of scales through a 1×1 convolutional layer, and the normalized scale gating coefficient vector is generated through the Sigmoid activation function. The corresponding formula is: This step achieves adaptive modeling of the importance of features at each scale. The Sigmoid function constrains the weight range to [0,1] to facilitate smooth adjustment during the fusion stage. (4) Scale-wise fusion and output generation: using the generated weight coefficients For corresponding scale features Perform element-wise weighted fusion to obtain the output features: ,in This indicates an element-wise multiplication operation. Through this operation, the contribution ratio of features at different scales can be adaptively adjusted according to the scene, thereby dynamically balancing high-level semantics and low-level structural details during the detection process. (5) Mechanism advantages and effect description: The dynamic weighted fusion module learns the optimal combination of multi-scale features in end-to-end training through a global statistical-driven gating mechanism. This module can effectively reduce the interference of stellar background noise on the detector in low signal-to-noise ratio star fields and enhance the response of weak streaks and small targets. High-level semantic features provide robust contextual information, while low-level detail features supplement texture and structural clues. The two are optimally fused after gating weight adjustment. Compared with the traditional fixed weighted or simple summation feature pyramid structure, the dynamic weighted fusion module has stronger scale adaptability and feature discrimination ability. While maintaining real-time inference performance, it can significantly improve the detection recall and localization accuracy of weak streaks and small targets, and further enhance the robustness and generalization of the model in complex star fields.
[0032] Then, at the decoding end, a dual-path query optimization mechanism is proposed to replace the single-path query selection (e.g., Figure 7 ): Dense path query generator, dense path is based on high-resolution features to construct regular geometric prior grid and inject position scale embedding to ensure spatial coverage of weak and narrow stripes; sparse path query generator, sparse path query uses a lightweight candidate network to perform top K saliency screening and position refinement in deep semantics to obtain more robust semantic priors for complex backgrounds; the two types of queries are dynamically weighted and fused in a unified representation space to complete geometric-semantic alignment with self-attention and cross-attention, which serves as the initial query and reference point input for the decoder, thereby significantly improving matching stability and convergence speed in weak signal and noise interference scenarios. Specifically: (1) Dense path generation: Dense path takes high-resolution shallow features as input and constructs regular geometric grid in the full field of view to form a set of candidate reference points. Each grid point is injected with position embedding and scale embedding information, and combined with the results of local convolution feature extraction, a dense query representation with geometric constraints is generated. The query generated by this path is spatially continuous and has fine-grained geometric priors, which can ensure full coverage and response enhancement of weak and narrow stripes. (2) Sparse path generation: The sparse path takes deep semantic features as input, predicts saliency scores and spatial location parameters through a lightweight candidate network, and performs top K saliency screening based on the scores to select the most representative semantic candidates as sparse queries. Then, a location refinement network is used to perform sub-pixel-level position correction on the candidate points to generate the final sparse query set. This path can highlight the main moving targets and potential trail features, suppress low-confidence noise areas, and provide robust semantic priors for the model. (3) Query fusion and geometric-semantic alignment: The queries generated by the two paths are respectively denoted as dense queries. With sparse queries Its reference point set is In the fusion phase, the two types of queries are first mapped to a unified feature dimension, and a self-attention mechanism is used to achieve information interaction and complementarity between geometric and semantic features. Then, using the fused query as input and the reference point position embedding as the key, a cross-attention mechanism is executed to achieve explicit alignment of geometric structure and semantic features. This fusion process can be represented as: ,in This represents the set of reference points after fusion. (4) Output and performance advantages: fused query and its reference points Both serve as the initial input to the decoder. Since geometric-semantic pre-alignment is completed before entering the decoding stage, this mechanism can significantly improve matching stability and decoding convergence speed in observation scenarios with low signal-to-noise ratios, dense star formations, or complex noise, reducing the number of iterations and suppressing false matches. Furthermore, the dual-path query module has low computational overhead, enhancing the detectability of weak-signal targets and overall detection accuracy without sacrificing inference speed, further improving the model's reliability and practicality in unattended Antarctic observation missions.
[0033] A scale-balanced training strategy is introduced on the training side: by introducing scale-adaptive mechanisms at the matching, weighting, and sampling levels, the bias of the model towards large-scale targets is alleviated, while the learning of small-scale and weak-signal targets is strengthened. First, a scale weight factor (1.5 times for small targets, 1 times for medium targets, and 0.8 times for large targets) is embedded in the Hungarian matching stage to increase the priority of positive samples from small targets. Second, scale-adaptive weighting is introduced into the classification, regression, and focus loss calculation stages, defining weight coefficients. , , And an adaptive scaling focus loss is adopted: ,in To balance gradient contributions, the scale varies with the target scale. Furthermore, a scale-balancing sampler is introduced during the sample sampling phase to ensure that samples of different scales appear evenly in each batch, preventing small targets from being weakened due to sparsity. The overall loss function is defined as: This strategy is only enabled during the training phase and does not increase inference overhead. Its detection head still uses query-based decoding of the Real-Time Detection Transformer (RT-DETR) model and a quality assessment mechanism based on intersection-union ratio, which significantly improves the recall rate and overall detection accuracy of small-scale and low-contrast targets while maintaining real-time performance.
[0034] In summary, this invention provides an automatic asteroid identification algorithm suitable for unattended sites in Antarctica. Addressing the challenges of polar nights, low temperatures, long-term continuous observations, and complex star fields with low signal-to-noise ratios, it proposes an integrated process of multi-frame registration enhancement and end-to-end detection: First, star point registration and sliding window maximum value superposition are completed based on feature matching to amplify the micro-displacement signal of the moving object relative to the star. On this basis, an improved detection network model based on a real-time detection transformer (SORT-RTDETR) is constructed. Its core components include: a scale-aware feature pyramid network to enhance the representation of weak streaks and multi-scale targets; a dual-path query optimization mechanism to jointly use geometric and semantic priors to stabilize query initialization and convergence in low signal-to-noise ratio scenarios; and a scale-balanced training strategy to introduce scale-adaptive weights at matching, loss, and sampling points to improve the recall of small targets without increasing inference overhead.
[0035] Experimental results of the improved detection network model based on real-time detection transformer (SORT-RTDETR) are as follows: Figure 8 As shown; examples of detection results for typical scenarios are as follows. Figure 9 As shown. Based on real-world data captured by the AST3 Antarctic Survey Telescope at Kunlun Station in Antarctica, the algorithm significantly outperforms the baseline Real-Time Detection Transformer (RT-DETR) model on small targets and low signal-to-noise ratio subsets, achieving an accuracy of 98.7%, a recall of 90.2%, and an average precision improvement while maintaining a real-time performance of 162.3 frames per second. This makes it suitable for automated deployment in long-term unattended Antarctic survey missions.
[0036] The above embodiments are merely typical implementations of the present invention and are not intended to limit the present invention. All equivalent substitutions or improvements made within the scope of the claims of the present invention are within the protection scope of the present invention.
Claims
1. A method for automatic asteroid identification at unattended sites in Antarctica, characterized in that, Includes a data processing module and a target detection module: The data processing module is used to perform integrated enhancement processing on the ground-based optical observation sequence under a unified spatiotemporal reference. It includes an intra-span enhancement unit, a reference system 1 and star registration unit, a temporal enhancement unit, and a clipping and loading unit to improve the visibility of faint streaks under low signal-to-noise ratio conditions. The target detection module is communicatively connected to the data processing module and is used to receive the image sub-image after integrated enhancement and perform real-time asteroid streak detection. The target detection module includes at least a feature fusion and scale adaptation module, a query decoding and target generation module, and a scale balance training strategy. The target detection network returns a list of detected targets for each sub-image, and each target includes a category, confidence level, and pre-selected box.
2. The method for automatic asteroid identification at unattended Antarctic sites according to claim 1, characterized in that, The integrated enhancement process includes the following steps: Step 1: Perform outlier correction, range normalization, and inverse hyperbolic sine grayscale stretching on the original observation frames; Step 2: Perform automatic world coordinate system calculation and write-back to unify the celestial reference system, and complete star registration and geometric consistency under the constraint of reference frame; Step 3: Within the time window T, take the maximum value pixel by pixel of the registered sequence to obtain the enhanced result of the sliding window maximum value superposition; Step 4: Crop the enhanced results into fixed-size sub-images and input them into the target detection module.
3. The method for automatic asteroid identification at unattended Antarctic sites according to claim 2, characterized in that, The time window T is adjusted according to the asteroid's motion speed; the maximum value of the sliding window is superimposed, and by taking the maximum value of the pixels within the window at each pixel position, the continuous bright band of the moving object is highlighted and random noise is suppressed.
4. The method for automatic asteroid identification at unattended Antarctic sites according to claim 1, characterized in that, The target detection module employs an improved detection network model based on a real-time detection transformer, referred to as the SORT-RTDETR model, which includes: The feature fusion and scale adaptation module is used to perform cross-layer fusion and scale attention based on feature pyramids to enhance details of dark streaks; The query decoding and target generation module adopts a dual-path decoding strategy of dense and sparse queries to output candidate boxes, categories, and confidence scores. The scale-balanced training strategy introduces scale weights and constraints in the matching, loss weighting, and sample sampling stages to improve the recall rate of small targets and low signal-to-noise ratio targets without increasing inference overhead.
5. The method for automatic asteroid identification at unattended Antarctic sites according to claim 4, characterized in that, The SORT-RTDETR model retains the end-to-end and real-time characteristics of the real-time detection transformer model RT-DETR. Its detection head still adopts the query-based decoding and cross-union ratio-based quality assessment mechanism of the RT-DETR model. On this basis, the SORT-RTDETR model introduces a scale-aware feature pyramid network and a dual-path query optimization mechanism, and adopts a scale-balanced training strategy during the training period to improve the detection accuracy and stability in low signal-to-noise ratio scenarios.
6. The method for automatic asteroid identification at unattended Antarctic sites according to claim 4, characterized in that, The SORT-RTDETR model extracts multi-scale features in the backbone network and replaces the hybrid encoder with a scale-aware feature pyramid at the encoding end. The scale-aware feature pyramid consists of three parts: an adaptive scale attention mechanism, a cross-scale semantic interaction module, and a dynamic weighted fusion module. Specifically, the adaptive scale attention mechanism uses multi-receptive field modeling and dynamic weighted noise suppression and signal enhancement within each scale to highlight dark and weak streaks; the cross-scale interaction module uses attention to establish long-range dependencies and semantic alignment between features at each level, bridging the gap between details and semantics; and the dynamic weighted fusion module performs gated recalibration of multi-scale channels based on global statistics, so that high-level semantics and low-level details are adaptively matched according to the scene, and finally outputs cross-layer fused features in a unified semantic space.
7. The method for automatic asteroid identification at unattended Antarctic sites according to claim 4, characterized in that, At the decoding end, a dual-path query optimization mechanism replaces the single-path query selection. This dual-path query optimization mechanism includes dense path query and sparse path query. Dense path query constructs a regularized geometric prior grid based on high-resolution features and injects positional scale embedding to spatially cover weak and narrow streaks. Sparse path query uses a lightweight candidate network to perform top K saliency screening and positional refinement at the deep semantic level, obtaining a more robust semantic prior for complex backgrounds. The two types of queries are fused by a query fusion module to achieve geometric-semantic alignment in a unified representation space using self-attention and cross-attention, serving as the initial query and reference point input for the decoder. This improves matching stability and convergence speed in scenarios with weak signals and noise interference.
8. The method for automatic asteroid identification at unattended sites in Antarctica according to claim 4, characterized in that, The scale-balanced training strategy involves adding a scale factor to Hungarian matching to increase the priority of positive samples for small targets, adaptively weighting by scale in classification, regression, and focus loss, and combining it with scale-balanced sampling to suppress long-tail bias.
9. The method for automatic asteroid identification at unattended Antarctic sites according to claim 1, characterized in that, The target detection network standardizes the output for each subgraph, and the output includes: a list of detected targets, each target having a category label, confidence level and pre-selection box, as well as the subgraph index, frame number and batch number associated with that subgraph.
10. The method for automatic asteroid identification at unattended sites in Antarctica according to claim 1, characterized in that, The method, which balances speed and accuracy, is particularly suitable for rapid deployment in unattended environments in Antarctica.