A state feedback based closed-loop night target tracking method and device

Abstract

The application belongs to the technical field of night target image processing, and discloses a closed-loop night target tracking method and device based on state feedback. A low-illumination image sequence is sent into a low-light perception enhancement network, and an enhanced image is generated based on Retinex decomposition and structure-preserving feature enhancement. The enhanced image is input into a target tracking network, the target position is output based on the template-search interaction of the Transformer, and the target attention map and the tracking confidence are generated. The target attention map and the confidence are sent into a tracking state feedback network, the target feedback modulation information is obtained by encoding, and is fed back to the target modulation fusion module, the enhancement process is dynamically modulated to strengthen the target region and suppress the background noise, and the closed-loop night target tracking of 'enhancement-tracking-feedback-reenhancement' is realized. The application can improve the tracking stability, noise resistance and robustness in a complex low-light scene, and is suitable for low-illumination visual tasks such as night monitoring, unmanned aerial vehicle night inspection, automatic driving replication perception and the like.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, target tracking, and nighttime image processing, and particularly to a closed-loop nighttime target tracking method and apparatus based on state feedback that drives low-light enhancement through target tracking state feedback and introduces tracking discrimination consistency constraints during the training phase to achieve collaborative optimization of low-light enhancement and tracking. Background Technology

[0002] Low-light scenes are typical scenarios where images are acquired in environments with light intensity less than or equal to 100 lux. The imaging process in these scenes often suffers from insufficient illumination, significant sensor noise, low target contrast, and severe detail degradation. Nighttime target tracking is widely used in scenarios such as drone inspections, nighttime security monitoring, autonomous driving assisted perception, maritime patrols, and border monitoring. However, existing target tracking algorithms generally suffer from tracking drift, target loss, and insufficient robustness in nighttime scenes.

[0003] In low-light imaging, the Retinex model is commonly used to describe the image formation process. Let the first... The input image at night is The reflection component is The light component is The noise term is Then we have: , in, This represents element-wise multiplication. For ease of estimation, it can also be represented in the logarithm field as: , in To prevent logarithmic singularity in the stability constant, This represents noise perturbation in the logarithmic domain. Under nighttime conditions. The values ​​are significantly low and spatially uneven. The increased relative proportion reduces the separability of the target area and the background area in both the pixel domain and the feature domain. This is particularly evident in the fact that edge textures are submerged by noise and dark area structures are missing, making it difficult for the tracking network to stably maintain the target identity.

[0004] like Figure 1As shown, existing serial methods typically include two independent processing stages: a low-light enhancement module and a target tracking module. The nighttime input image first passes through the low-light enhancement module to obtain an enhanced image, which is then input into the target tracker for target localization. In this type of method, the enhancement module cannot obtain state information such as target location, attention distribution, and tracking confidence, thus easily leading to indiscriminate enhancement of background noise and the target region. This serial combination of "low-light enhancement + tracking"—that is, enhancing the input image first and then processing it with a general tracker—has the following specific shortcomings: (1) There is a lack of information interaction between the enhancement module and the tracking module. The enhancement process cannot perceive the target position and tracking status, and it is easy to perform indiscriminate enhancement on the target area and the background area, thereby amplifying the background interference; (2) Augmentation networks often use pixel-level reconstruction error, perceptual quality or natural image statistics as optimization targets, and their output does not necessarily improve the target-background discrimination distance; (3) In complex nighttime scenes, the enhanced model may produce halos, overexposure, and texture artifacts, which may reduce tracking stability. (4) The tracker lacks modeling of the enhancement uncertainty, and is prone to drift and false feedback when the enhancement fails.

[0005] Therefore, there is an urgent need for a new low-light target tracking method that can feed back the target tracking status to the low-light enhancement process and optimize the target with tracking discrimination capability as the core, so as to achieve truly task-oriented low-light enhancement and tracking synergy. Summary of the Invention

[0006] To address the aforementioned shortcomings of existing technologies, this invention proposes a closed-loop nighttime target tracking method and apparatus based on state feedback. This invention feeds back the target tracking state to the low-light enhancement process, enabling the enhancement module to perceive the target's position, attention distribution, and tracking reliability. Furthermore, during the training phase, target discrimination consistency constraints are used to improve the target-background separability of the enhancement results.

[0007] The technical solution of this invention is implemented as follows: A closed-loop nighttime target tracking method based on state feedback includes a closed-loop nighttime target tracking process of "enhancement-tracking-feedback", the specific steps of which are as follows: S1. During the training and testing phases, acquire the input image of frame t in a nighttime scene. H is the image height, W is the image width, and 3 represents the RGB three channels. The input image... Feedback modulation information from the previous moment The image is fed into the Low Light Perception Enhancement Network (LPEN) to obtain an enhanced image. Among them, LPEN includes at least the Dynamic Illumination Decomposition Subnetwork DLDS, the Structure Preservation State Enhancement Subnetwork SP-SSEN, and the Target Modulation Fusion Module TGMF. S2, the enhanced image Input the target tracking network TITN, extract the interaction features between the target template and the search region, and output the target bounding box of the current frame. Target Attention Map and tracking confidence ; S3, The target attention map Tracking confidence Feedback modulation information from the previous moment Input the tracking state feedback network (TSFN) to obtain the current frame feedback modulation information. ; S4. Based on the current frame feedback modulation information Dynamic modulation of TGMF in LPEN enables the enhancement process to adaptively enhance the target region and suppress background interference during iterative enhancement in subsequent frames or the current frame. S5. During the training phase, a target discrimination consistency constraint module (DCCM) is introduced to enhance the image. Constraining the target-background discrimination capability in the feature space, we jointly optimize the low-light perception enhancement network LPEN and the target tracking network TITN.

[0008] Preferably, the overall mapping relationship of the method is expressed as: , , , in: , and These represent LPEN enhancement mapping, TITN tracking mapping, and TSFN feedback mapping, respectively. For the target template; , , These are the learnable network parameters for LPEN, TITN, and TSFN, respectively. The modulation information is fed back from the previous moment; The modulation information is fed back to the current frame for use in the enhancement process of the next frame, or for the iterative re-enhancement process of the current frame; the initial frame's Set to preset initial modulation information.

[0009] Preferably, the Dynamic Illumination Decomposition Subnetwork (DLDS) generates spatially adaptive dynamic illumination maps through multi-scale feature extraction and nonlinear mapping. The initial reflection component is obtained by element-wise division. ,in: , , in, This is a multi-scale feature extraction and fusion network in DLDS. These are the learnable network parameters for DLDS; For Sigmoid or normalization functions; It is a stability constant used to prevent division by zero and maintain numerical stability.

[0010] Preferably, the structure-preserving state enhancement subnetwork SP-SSEN includes a structure-preserving state evolution module SP-SEM, which evaluates the initial reflection components. By performing structural preservation enhancement, an enhanced reflection component is obtained. ,in: , in, To maintain the enhanced mapping for the SP-SSEN structure; These are the learnable network parameters for SP-SSEN.

[0011] Preferably, the target modulation fusion module TGMF utilizes the target attention map. Tracking confidence level and current frame feedback modulation information Enhanced features Perform joint modulation to obtain modulated features And by the modulated features Generate enhanced images .

[0012] Preferably, the target tracking network TITN includes a target-background interactive encoder TBIE and a target state estimation head OSEH. TBIE models the interaction between target template features and search region features through self-attention and cross-attention.

[0013] Preferably, the Tracking State Feedback Network (TSFN) is used to extract target attention distribution and confidence information from the Target Tracking Network (TITN), and output the current frame feedback modulation information after normalization and time smoothing. It is used to modulate the enhancement strategy in the iterative re-enhancement process of the next frame or the current frame.

[0014] Preferably, the target discrimination consistency constraint module DCCM is based on a frozen target tracking feature extraction network. The method constructs a feature space to constrain the distance difference between the target region and the background region before and after enhancement, thereby improving the target-background separability of the enhancement result.

[0015] This invention also discloses a nighttime target tracking device, comprising a Low Light Perception Enhancement Network (LPEN), a Target Tracking Network (TITN), a Tracking State Feedback Network (TSFN), and a Target Discriminant Consistency Constraint Module (DCCM) for the training phase. LPEN generates enhanced images for tracking discrimination based on Retinex imaging decomposition and state space structure preservation enhancement. TITN outputs target position, attention distribution, and tracking confidence based on Transformer template-search interaction modeling. TSFN encodes attention and confidence into target feedback modulation information, dynamically modulating the target modulation fusion module in LPEN. DCCM constrains the relative change in target-background discrimination distance before and after enhancement in the feature space.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The Low Light Perception Enhancement Network (LPEN) proposed in this invention is based on Retinex decomposition, combined with a structure-preserving state enhancement subnetwork, and introduces a structure-preserving gating function to enhance the target edge and texture structure and suppress noise during the state update process. The target feedback closed-loop enhancement mechanism proposed in this invention enables the enhancement process to have target perception capabilities, prioritizes the allocation of enhancement resources to the target area, and reduces the risk of background noise and irrelevant areas being over-enhanced; This invention uses the Target Discriminant Consistency Constraint Module (DCCM) to enable enhanced targets to directly serve tracking and discrimination needs, thereby improving target-background separability. This invention can suppress noise amplification and texture artifacts in complex nighttime scenes, improve tracking stability and robustness, and is suitable for engineering scenarios such as security monitoring, drone nighttime inspection, and low-light vehicle-mounted sensing. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of an existing serial "low-light enhancement-target tracking" method; Figure 2 This is a schematic diagram of the "enhancement-tracking-feedback" method of the present invention; Figure 3 This is a schematic diagram of the overall closed-loop framework of the "enhancement-tracking-feedback" of the present invention; Figure 4 This is a schematic diagram of the module structure of the Low Light Perception Enhancement Network (LPEN) of the present invention; Figure 5 This is a schematic diagram of the target tracking network TITN and the tracking state feedback network TSFN of the present invention. Detailed Implementation

[0018] The present invention will now be clearly and completely described with reference to the accompanying drawings of the embodiments. The described embodiments are used to explain the present invention, and not to limit the scope of protection of the present invention.

[0019] like Figure 2 , 3 As shown, the method of this invention establishes a closed-loop connection between the Low Light Perception Enhancement Network (LPEN), the Target Tracking Network (TITN), and the Tracking Status Feedback Network (TSFN). The target attention map and tracking confidence output by the TITN are fed into the TSFN. The TSFN generates target feedback modulation information and sends this information back to the Target Modulation Fusion Module (TGMF) in the LPEN, thereby enabling the enhancement process to adaptively modulate according to the target position and tracking reliability.

[0020] The closed-loop overall framework of this invention includes a Low-Light Perception Enhancement Network (LPEN), a Target Tracking Network (TITN), a Tracking State Feedback Network (TSFN), and a Target Discriminant Consistency Constraint Module (DCCM) used during the training phase. The nighttime input image first enters the LPEN, which generates an enhanced image through a Dynamic Illumination Decomposition Subnetwork (DLDS), a Structure Preservation State Enhancement Subnetwork (SP-SSEN), and a Target Modulation Fusion Module (TGMF). The enhanced image is then input into the TITN, which outputs a target bounding box, a target attention map, and a tracking confidence score based on the interaction between the target template and the search region. The target attention map and tracking confidence score are further input into the TSFN, which outputs target feedback modulation information and passes this information to the TGMF in the LPEN to guide iterative enhancement in subsequent frames or the current frame. The DCCM is used only during the training phase to constrain the discriminant distance between the target region and the background region in the feature space before and after enhancement. The specific method steps of this invention are as follows: S1. During the training and testing phases, acquire the input image of frame t in a nighttime scene. H is the image height, W is the image width, and 3 represents the RGB three channels. The input image... Feedback modulation information from the previous moment The image is fed into the Low Light Perception Enhancement Network (LPEN) to obtain an enhanced image. Among them, LPEN includes at least the Dynamic Illumination Decomposition Subnetwork DLDS, the Structure Preservation State Enhancement Subnetwork SP-SSEN, and the Target Modulation Fusion Module TGMF. S2, the enhanced image Input the target tracking network TITN, extract the interaction features between the target template and the search region, and output the target bounding box of the current frame. Target Attention Map and tracking confidence ; S3, The target attention map Tracking confidence level Feedback modulation information from the previous moment Input the tracking state feedback network (TSFN) to obtain the current frame feedback modulation information. ; S4. Based on the current frame feedback modulation information Dynamic modulation of TGMF in LPEN enables the enhancement process to adaptively enhance the target region and suppress background interference during iterative enhancement in subsequent frames or the current frame. S5. During the training phase, a target discrimination consistency constraint module (DCCM) is introduced to enhance the image. Constraining the target-background discrimination capability in the feature space, we jointly optimize the low-light perception enhancement network LPEN and the target tracking network TITN.

[0021] like Figure 4 As shown, the Low Light Sensing Enhancement Network (LPEN) includes the Dynamic Illumination Decomposition Subnetwork (DLDS), the Structure Preservation State Enhancement Subnetwork (SP-SSEN), the Target Modulation Fusion Module (TGMF), and the Damage Repairer. Figure 4 The "nighttime low-light frame" in the image corresponds to the t-th frame of the input image. The "illuminance estimator" corresponds to the Dynamic Illumination Decomposition Subnetwork (DLDS); the "illuminance feature map" corresponds to the brightness-related intermediate features extracted by DLDS. "Light map" corresponds to spatial adaptive dynamic lighting map Its numerical range is normalized to (0, 1); "Structure Preservation State Enhancement Module" corresponds to SP-SSEN; "SP-SEM" corresponds to Structure Preservation State Evolution Module; "Target Modulation Fusion Node" corresponds to the fusion node in TGMF that receives target state feedback; "Damage Repairer" corresponds to the decoding mapping that repairs and reconstructs based on the modulated features; the dashed box on the right represents the internal hierarchical structure of SP-SSEN, that is, the input features pass through SP-SSEN, SP-SEM, repair and reconstruction mapping and output layer in sequence.

[0022] exist Figure 4 In the embodiment shown, DLDS first inputs an image from a low-light nighttime frame. Estimated illumination map and illumination feature map SP-SSEN according to and Perform structure preservation enhancement and output structure enhancement features. The target modulation fusion node receives the target attention map from the TSFN. Tracking confidence level and current frame feedback modulation information The enhancement features are then used to strengthen the target region; the damage repairer fuses the modulated enhancement features, illumination feature map, and structural enhancement features to output the final enhanced image. Let the intermediate enhancement feature of the SP-SSEN output be... TITN output target attention map Normalized to: , in, This is the normalization function; This is the normalized target attention map. Then, Aligned to via bilinear interpolation Spatial dimensions. The modulation weighting map is defined as follows: , in, For the modulation weighting diagram; and The modulation coefficient is used to control the intensity of the influence of tracking confidence and feedback modulation information on the enhanced features. This maps the feedback modulation information to a projection function with the same spatial or channel dimensions as the enhanced features. Feature modulation is represented as: , and Figure 4 The formula in the text corresponds to the damage repairer input features when using only attention and confidence for simplified modulation. The modulation form is represented as: , in, Enhanced features for damage repair devices, This represents the enhanced features after target modulation. The formula illustrates that when attention is focused on the target region and tracking confidence is high, the target region features are preferentially enhanced; when tracking confidence is low, the modulation intensity decreases, thus avoiding enhancement shift caused by false feedback.

[0023] In one implementation, TGMF further incorporates a channel gating structure: , , in Indicates global average pooling; and These are learnable weight parameters in channel gating. It is a non-linear activation function; This is the channel gating vector; This is an enhanced feature following channel gating.

[0024] When the tracking confidence is high and attention is focused on the target region, the features of the target region are preferentially enhanced; when the tracking confidence is low, the modulation intensity decreases, thereby avoiding enhancement shift caused by false feedback.

[0025] The Dynamic Illumination Decomposition Subnetwork (DLDS) is based on the Retinex model, and its goal is to decompose nighttime inputs. Estimated light component DLDS employs a multi-scale convolutional encoder to extract brightness-related features and generates a spatially adaptive illumination map through nonlinear mapping. , in, This represents a multi-scale feature extraction and fusion network; These are the learnable network parameters for DLDS; Ensure the output is within the preset value range.

[0026] In one implementation, Convert to luminance channel In conjunction with color channels, including RGB channels or Lab color space channels, to improve the stability of illumination estimation. Then, calculate the initial reflection: , The core of the Structure Preservation State Enhancement Subnetwork (SP-SSEN) is the Structure Preservation State Evolution Module (SP-SEM). This is due to the nighttime reflection component. Often, noise and texture degradation exist; SP-SSEN uses state-space iteration to enhance structural information. The input features at step k are obtained through shallow encoding. Let the state vector be... The state at step k represents the structural enhancement state, and its state update is expressed as: , , in, Input features for the k-th step; This represents the structural enhancement state at step k. This is the structural enhancement state of the previous iteration step; Output the status; , and The learnable transformation parameters in the state space update are implemented using linear transformations or 1×1 convolutional mappings.

[0027] Next, to achieve structure preservation and noise immunity, this invention introduces a structure preservation gating function. : , in, Maintain gating weights for the structure; This represents a gradient operator, which is implemented using the Sobel operator or a learnable convolution operator; Indicates the convolution operation; This indicates channel dimension splicing. Gated update is represented as: , in, This represents the enhanced structural state after gating. This indicates element-wise multiplication. In areas with prominent edges and textures, Larger values ​​result in more thorough enhancement and updates; in noisy or flat regions, Smaller values ​​preserve more historical states, thus suppressing noise amplification. The output features are decoded to obtain the enhanced reflection component: , in The number of iterations. This represents the structural enhancement state after the Kth gate step; For decoder and reconstruction mapping; These are the learnable network parameters for the decoder. The final enhanced image is represented as: , in, To enhance the image; This means cropping the pixel values ​​to the effective pixel range.

[0028] like Figure 5 As shown, the target tracking network TITN includes a patch embedding module, a segmentation and patch embedding module, a target-background interactive encoder TBIE, a target state estimation head OSEH, a target-background interactive decoding module, and a feedback network TSFN connected to it. Figure 5 In this context, "target(128,128)" represents the target template region obtained by cropping from the initial target bounding box or the prediction bounding box of the previous frame. The size is set to 128×128; "Search (320,320)" represents the search area obtained by cropping the current frame image according to the prediction box of the previous frame. The size is set to 320×320; "H×W×3" means that the input image or feature retains the RGB three-channel format; "Patch Embedding" and "Segmentation and Patch Embedding" mean that the target template and search area are divided into image blocks and mapped to tokens; the light red token in the legend represents the target token, and the blue-green token represents the background token.

[0029] Figure 5 In this context, "concatenating tokens" refers to concatenating the target token with the background token to form a joint sequence that inputs into the subsequent attention module; "Stage 1," "Stage 2," and "Stage 3" represent different scale-based hierarchical feature processing stages, with feature sizes of [sizes to be filled in]. , and Where C represents the number of channels; "Self-attention" means performing self-attention modeling on the concatenated tokens; "Target Background Interaction Encoding" corresponds to TBIE, which is used to model the interaction between the target template and the search region; "Target State Estimation" corresponds to OSEH, which is used to output the target bounding box, target attention map and tracking confidence; "Feedback Network" corresponds to TSFN, which is used to generate feedback modulation information based on the target attention map and tracking confidence and send it back to LPEN.

[0030] Let the target template image be The search area is Search area It is obtained by cropping the current frame image according to the prediction box of the previous frame. and Segmented into image patches And map it to a token: , in, For the i-th image patch; For image patch embedding functions; Embed for the i-th token; For positional encoding. Single-head attention is defined as: , Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively; The dimension of the key vector; This is a normalized exponential function. Bullish attention is defined as: , , in, Let h be the i-th attention head; h is the number of attention heads. , , For the query, key, and value mapping parameters corresponding to the i-th attention head; Output mapping parameters for multi-head attention; This represents the concatenation operation. TBIE combines the template token and the search token, then applies several layers of self-attention and cross-attention to obtain the fused representation. : , in, A collection of template tokens; A set of tokens for the search area; These are the learnable network parameters for TBIE. OSEH outputs the bounding box, target attention map, and confidence score based on the fused features: , in, , and These are the bounding box regression head, the classification evaluation head, and the attention aggregation map, respectively. , and These are the corresponding learnable parameters.

[0031] The target attention map received by the Tracking State Feedback Network (TSFN) from the TITN output and tracking confidence The attention map is normalized and thresholded, and the tracking confidence is temporally smoothed to obtain stable target feedback modulation information. The confidence smoothing is expressed as: , in, The tracking confidence score after smoothing the current frame; The tracking confidence score after smoothing from the previous frame; This is the time smoothing coefficient, with a value ranging from 0 to 1; Track the confidence level for the current frame. Attention map normalization and threshold suppression are represented as follows: , in, This is the processed stable attention map; The attention threshold is used. The final output feedback modulation information is represented as: , in, For the feedback coding network in TSFN; This indicates that the stable attention map, smoothed confidence, and feedback modulation information from the previous time step are concatenated or jointly encoded. The feedback strength coefficient is represented as: , in, For adaptive feedback intensity coefficient; and These are learnable scalar parameters.

[0032] The Target Discriminant Consistency Constraint Module (DCCM) directly constrains the improvement of the target-background feature distance before and after enhancement during the training phase, because image quality loss alone cannot guarantee improved tracking discriminant performance. Let the frozen target tracking feature extraction network be φ, which includes a TITN front-end encoder or an independent backbone network, and keep its parameters frozen during training; let the target region mask be... The background area mask is For image enhancement The mean features of the target region and the background region are expressed as: , , Similarly, for the original low-light input The mean features of the target region and the background region are expressed as: , , in, Indicates regional average pooling. This represents the mean feature value of the corresponding region. The target-background feature distance is defined as: , , in, The characteristic distance metric function is either Euclidean distance or cosine distance. To enhance the target-background feature distance in an image; The target-background feature distance in the original low-light image. The expected relative increase in distance after enhancement is at least the discrimination interval. The discriminant consistency loss is defined as follows: , minimize prompt Greater than This means enhancing the target-background separability of the results, which is more conducive to target tracking at night.

Claims

1. A closed-loop nighttime target tracking method based on state feedback, characterized in that, The closed-loop nighttime target tracking process, including "enhancement-tracking-feedback," consists of the following steps: S1. During the training and testing phases, acquire the input image of frame t in a nighttime scene. H is the image height, W is the image width, and 3 represents the RGB three channels. The input image... Feedback modulation information from the previous moment The image is fed into the Low Light Perception Enhancement Network (LPEN) to obtain an enhanced image. Among them, LPEN includes at least the Dynamic Illumination Decomposition Subnetwork DLDS, the Structure Preservation State Enhancement Subnetwork SP-SSEN, and the Target Modulation Fusion Module TGMF. S2, the enhanced image Input the target tracking network TITN, extract the interaction features between the target template and the search region, and output the target bounding box of the current frame. Target Attention Map and tracking confidence ; S3, The target attention map Tracking confidence Feedback modulation information from the previous moment Input the tracking state feedback network (TSFN) to obtain the current frame feedback modulation information. ; S4. Based on the current frame feedback modulation information Dynamic modulation of TGMF in LPEN enables the enhancement process to adaptively enhance the target region and suppress background interference during iterative enhancement in subsequent frames or the current frame. S5. During the training phase, a target discrimination consistency constraint module (DCCM) is introduced to enhance the image. Constraining the target-background discrimination capability in the feature space, we jointly optimize the low-light perception enhancement network LPEN and the target tracking network TITN.

2. The closed-loop nighttime target tracking method based on state feedback according to claim 1, characterized in that, The overall mapping relationship of the method is expressed as follows: ， ， ， in: , and These represent LPEN enhancement mapping, TITN tracking mapping, and TSFN feedback mapping, respectively. For the target template; , , These are the learnable network parameters for LPEN, TITN, and TSFN, respectively. The modulation information is fed back from the previous moment; The modulation information is fed back to the current frame for use in the enhancement process of the next frame, or for the iterative re-enhancement process of the current frame; the initial frame's Set to preset initial modulation information.

3. The closed-loop nighttime target tracking method based on state feedback according to claim 1, characterized in that, The Dynamic Illumination Decomposition Subnetwork (DLDS) generates spatially adaptive dynamic illumination maps through multi-scale feature extraction and nonlinear mapping. The initial reflection component is obtained by element-wise division. ,in: , ， in, This is a multi-scale feature extraction and fusion network in DLDS. These are the learnable network parameters for DLDS; For Sigmoid or normalization functions; It is a stability constant used to prevent division by zero and maintain numerical stability.

4. The closed-loop nighttime target tracking method based on state feedback according to claim 3, characterized in that, The structure-preserving state enhancement subnetwork SP-SSEN includes a structure-preserving state evolution module SP-SEM, which evaluates the initial reflection components. By performing structural preservation enhancement, an enhanced reflection component is obtained. ,in: ， in, To maintain the enhanced mapping for the SP-SSEN structure; These are the learnable network parameters for SP-SSEN.

5. The closed-loop nighttime target tracking method based on state feedback according to claim 1, characterized in that, The target modulation and fusion module TGMF utilizes target attention maps. Tracking confidence and current frame feedback modulation information Enhanced features Perform joint modulation to obtain modulated features And by the modulated features Generate enhanced images .

6. The closed-loop nighttime target tracking method based on state feedback according to claim 1, characterized in that, The target tracking network TITN includes a target-background interactive encoder TBIE and a target state estimation head OSEH. TBIE models the interaction between target template features and search region features through self-attention and cross-attention.

7. The closed-loop nighttime target tracking method based on state feedback according to claim 5, characterized in that, The Tracking State Feedback Network (TSFN) is used to extract target attention distribution and confidence information from the Target Tracking Network (TITN), and output the current frame feedback modulation information after normalization and temporal smoothing. It is used to modulate the enhancement strategy in the iterative re-enhancement process of the next frame or the current frame.

8. The closed-loop nighttime target tracking method based on state feedback according to claim 1, characterized in that, The target discrimination consistency constraint module DCCM is based on a frozen target tracking feature extraction network. The method constructs a feature space to constrain the distance difference between the target region and the background region before and after enhancement, thereby improving the target-background separability of the enhancement result.

9. A nighttime target tracking device for performing the method according to any one of claims 1 to 8, characterized in that, It includes a Low-Light Perception Enhancement Network (LPEN), a Target Tracking Network (TITN), a Tracking State Feedback Network (TSFN), and a Target Discriminant Consistency Constraint Module (DCCM) for the training phase. Among them, LPEN generates enhanced images for tracking discrimination based on Retinex imaging decomposition and state space structure preservation enhancement; TITN outputs target position, attention distribution, and tracking confidence based on Transformer template-search interaction modeling; TSFN encodes attention and confidence into target feedback modulation information and dynamically modulates the target modulation fusion module in LPEN; DCCM constrains the relative change of target-background discriminant distance before and after enhancement in the feature space.

Images