Dynamic target detection method based on deep learning

Abstract

The present invention proposes a dynamic target detection method based on deep learning, including: extracting dynamic targets through unsupervised background modeling and morphological processing, combining a deep feature network with a lightweight visibility network to output the target structured state; constructing a trajectory velocity modeling network through differential equations based on the bounding box and semantic features to generate the target motion trajectory, and obtaining the predicted spatial position by numerically integrating the differential equations; fusing the semantic feature vector with the predicted spatial position to generate the future appearance features of the target; when a frame image is input, constructing a candidate area based on the predicted spatial position, calculating the matching confidence between the candidate area and the future appearance features of the target to confirm the target position and target features; if the matching confidence is lower than a preset threshold, retaining the previous frame state without updating, so as to improve the robustness in occluded or blurred states.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning, and in particular relates to a dynamic target detection method based on deep learning. Background Art

[0002] With the rapid development of computer vision and artificial intelligence technologies, deep learning-based object detection methods have been widely applied in a variety of dynamic visual scenarios, including intelligent surveillance, autonomous driving, drone vision, and intelligent transportation. In these scenarios, objects are often in continuous motion, accompanied by complex factors such as illumination variations, occlusions, deformation, scale changes, and background interference, posing significant challenges to detection systems. Current mainstream object detection technologies fall into two main categories: static frame detection models, such as YOLO, SSD, and Faster R-CNN, use convolutional neural networks to extract spatial features from images and output bounding boxes and class labels. Video object detection methods, such as Track-RCNN and FGFA (Flow-guided Feature Aggregation), attempt to incorporate temporal information to enhance multi-frame fusion capabilities, thereby improving the detection accuracy and continuity of dynamic objects.

[0003] Although numerous methods have attempted to improve detection performance in dynamic scenes by leveraging temporal context, multi-frame feature fusion, and motion information modeling in recent years, several core technical bottlenecks remain unresolved. First, most current methods rely on concatenating features from previous and subsequent frames within a fixed window or on optical flow estimation to model target motion. However, these strategies are prone to failure in complex scenarios such as severe occlusion, rapid movement, and sudden background changes. This is because they lack the ability to continuously model the target's motion trends and cannot effectively predict the target's potential future location. Second, existing methods generally rely on the target's visual appearance information in the current frame for recognition and classification. When the target is blurred, in low light, occluded, or only partially visible, detectors based on local image features struggle to stably output the target's location and identity, leading to false detections, missed detections, and even target tracking failures. Furthermore, most detector designs fail to model the deep semantic relationship between target motion and appearance features. This results in reduced discriminative ability and insufficient robustness when the target undergoes drastic posture changes or semantic deformations.

[0004] Therefore, existing dynamic object detection technology faces two pressing challenges: first, how to model the continuous motion of the target and predict its possible trajectory to achieve temporal continuity and occlusion recovery during the detection process; second, how to enhance the visual features of the target so that it can still be identified and tracked under non-ideal conditions, thereby improving the stability and fault tolerance of the entire detection system. The combination of these two issues directly restricts the application effectiveness and deployment value of deep object detection systems in real dynamic scenes. Summary of the Invention

[0005] The purpose of this invention is to propose a dynamic target detection method based on deep learning, which realizes continuous, stable and high-precision detection of fast-moving or partially occluded targets by introducing a collaborative framework of motion trend modeling mechanism and visual feature enhancement strategy.

[0006] In order to achieve the above object, the present invention provides a dynamic target detection method based on deep learning, which includes the following steps:

[0007] Acquire a target image, extract dynamic targets through unsupervised background modeling and morphological processing, and combine a deep feature network with a lightweight visibility network to output the target structured state; the target structured state includes: target bounding box, semantic feature vector, and visibility score;

[0008] Constructing a trajectory velocity modeling network using differential equations based on the bounding box and semantic features to generate a target motion trajectory, introducing an acceleration regularization term to smooth trajectory changes, and finally obtaining a predicted spatial position by numerically integrating the differential equations;

[0009] Obtain a semantic feature vector, fuse it with the predicted spatial position, generate the target's future appearance features through a trajectory-guided feature generation module, and introduce a structural alignment loss and a local attribute diversity preservation term to improve feature quality;

[0010] When a frame image is input, a candidate area is constructed based on the predicted spatial position, and the matching confidence between the candidate area and the future appearance features of the target is calculated to confirm the target position and target features; if the matching confidence is lower than the preset threshold, the previous frame state is retained and no update is performed to improve robustness in occluded or blurred states.

[0011] Furthermore, the method of extracting dynamic targets through unsupervised background modeling and morphological processing, combining deep feature networks with lightweight visibility networks, and outputting target structured states specifically includes:

[0012] Model the image through unsupervised background, perform background subtraction on each frame, and generate a foreground mask;

[0013] Performing morphological processing on the foreground mask to enhance the integrity of the target boundary, thereby obtaining an optimized foreground mask;

[0014] Extracting a connected region from the optimized foreground mask and recording the connected region as a target region; a bounding box of the target region is a target bounding box;

[0015] Cutting the target area from the target image and feeding it into a deep feature network to extract semantic features as a semantic feature vector;

[0016] Extracting a structure vector from each target region in the optimized foreground mask; concatenating the structure vector with a deep feature map of the semantic feature vector and inputting it into a lightweight visibility network to output a visibility score;

[0017] Combining the object bounding box, semantic feature vector and visibility score, the output is the state of each object.

[0018] Furthermore, the structure vector includes four dimensions: normalized area, aspect ratio, boundary tightness and isolated object identification.

[0019] Furthermore, the morphological processing includes erosion and dilation operations performed sequentially to optimize the object boundary integrity.

[0020] Furthermore, the unsupervised background modeling adopts Gaussian modeling or sliding average method to generate an initial foreground mask through pixel-level difference threshold segmentation.

[0021] Furthermore, the differential equation is expressed as:

[0022] ;

[0023] in, , To predict the spatial position, it means the target is in time Position status; The trajectory velocity modeling network is parameterized as ; is the observation weight factor, expressed as:

[0024] ;

[0025] The formula represents limiting the visibility score to the lower limit Between 1 and 1 to prevent the track from freezing when it is completely occluded;

[0026] The acceleration regularization term constrains the second-order differences of adjacent elements of the time series, suppresses trajectory jumps and conforms to physical inertia logic.

[0027] Furthermore, the future appearance features of the target are generated by the trajectory guidance feature generation module, which is specifically expressed as follows:

[0028] ;

[0029] in, represents the future appearance features of the generated target i; Represents semantic feature vector; To predict spatial position; Represents the spatial position encoding module; Represents a fusion model consisting of several layers of attention networks and feature transformation layers; represents the learnable parameters of the TGFS module;

[0030] Furthermore, the trajectory-guided feature generation module introduces a local attribute diversity preservation term, which is a structural alignment loss of adjacent dimensions of the future appearance features, and is used to enforce the difference in feature sub-dimensional distribution.

[0031] Furthermore, the spatial position encoding module is MLP or Sinusoidal Embedding.

[0032] Furthermore, the candidate region is generated with the predicted spatial position as the center and constructed through a multi-scale sliding window or a region proposal network;

[0033] The matching confidence is obtained by performing cosine similarity scoring on the semantic feature vector extracted from the candidate region in the current frame image and the future appearance feature of the target.

[0034] The beneficial technical effects of the present invention are at least as follows:

[0035] To address the dual bottlenecks of existing dynamic target detection technologies in motion modeling and visual representation, the present invention proposes a deep learning target detection method for dynamic scenes. By introducing a collaborative framework of motion trend modeling mechanism and visual feature enhancement strategy, it achieves continuous, stable and high-precision detection of fast-moving or partially occluded targets. The core innovations of the present invention are mainly reflected in the following two aspects: First, a modeling mechanism for target states in continuous frames is proposed, which can learn the motion laws of targets in the time dimension. The auxiliary detection module performs forward-looking position inference when the target is lost or occluded, thereby improving the system's continuous perception of dynamic targets; second, the present invention designs a feature completion mechanism for enhancing target representation. When the target is blurred, low-resolution or changes in appearance, it can still generate high-quality target features, thereby ensuring the detector's stable recognition of the target's identity. This design of collaborative optimization from the two dimensions of "motion trend prediction" and "visual completion enhancement" can effectively compensate for the discontinuity and instability of existing technologies in the dynamic detection process, significantly improve the detection accuracy and robustness in complex actual scenes, and has strong engineering practical value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The present invention is further described with reference to the accompanying drawings. However, the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative effort.

[0037] Figure 1 This is a flow chart of the dynamic target detection method based on deep learning of the present invention. DETAILED DESCRIPTION

[0038] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0039] like Figure 1 As shown, the embodiment of the present invention provides a dynamic target detection method based on deep learning, and the method includes:

[0040] S1. Acquire a target image, extract dynamic targets through unsupervised background modeling and morphological processing, combine a deep feature network with a lightweight visibility network, and output the target structured state; the target structured state includes: target bounding box, semantic feature vector and visibility score.

[0041] Specifically, this step proposes a Hybrid Structured Detection Module (HSDM) that combines unsupervised background modeling, morphological foreground optimization, and deep feature extraction to achieve robust detection of dynamic targets in complex scenes. The specific steps of the solution are as follows:

[0042] Construct background model by Gaussian modeling or sliding average , for each frame of image Perform background subtraction to generate a foreground mask :

[0043]

[0044] in: : background model; T: pixel difference threshold; : Initial foreground binary mask.

[0045] Further, Perform erosion and dilation operations to obtain the optimized foreground mask , enhance the target boundary integrity:

[0046]

[0047] Further, in Extracting connected regions , record its bounding box . Cut the area from the original image and feed it into the deep feature network (such as ResNet) extract semantic features:

[0048]

[0049] Further, from Each target area Extract the following structural distribution features to form a structural vector :

[0050]

[0051] in: : The height and width of the image; Dimension 1: Normalized area ( is the number of pixels in the region); the second dimension: aspect ratio; the third dimension: boundary compactness index, reflecting the degree of structural fragmentation; the fourth dimension: whether it is an isolated target (1 means not attached to other targets); each dimension has a clear physical meaning and constitutes a low-dimensional structural distribution descriptor.

[0052] Furthermore, the structural features and deep feature maps Splicing Input Lightweight Visibility Network , output visibility score :

[0053]

[0054] in, : High-level features of the target semantic region output by the backbone network; : splicing operation; : Two-layer fully connected network; : Sigmoid function, output normalized score; : Indicates whether the target is visible, the higher the degree of occlusion / blur, the lower the score. This score will be used as a trajectory speed adjustment factor in step 2 , enhancing the system's dynamic modeling capabilities under occlusion; it also serves as a weighting factor for the temporal consistency regularization term in this step, used to measure cross-frame semantic stability.

[0055] Furthermore, to enhance the stability of cross-frame features, a regularization term is introduced to encourage semantic consistency between the current frame features and the matching targets of the previous frame:

[0056]

[0057] This regularization term serves as an additional loss in the training phase and helps improve tracking accuracy and feature stability.

[0058] Finally, the state of each target i is represented as:

[0059]

[0060] in, : target bounding box; : semantic feature vector; : Visibility score.

[0061] Through the above mechanism, this step realizes the joint modeling of the target's spatial position, semantic appearance and observation quality, so that the detection results are not only suitable for target positioning in this frame, but also have the continuity, traceability and discriminability of backward transmission, supporting subsequent motion prediction and visual completion modules.

[0062] S2. Construct a trajectory velocity modeling network using differential equations based on the bounding box and semantic features to generate the target motion trajectory. An acceleration regularization term is introduced to smooth the trajectory changes. Finally, the predicted spatial position is obtained by numerically integrating the differential equations.

[0063] Specifically, this step focuses on the target structured state generated in step 1 , establish a time-continuous trajectory modeling system to predict the potential position of the target in the next frame This paper proposes a "Gated Differential Track Prediction Model (GODE-Tracker)", the core innovation of which is to use the visibility score output in step 1 The speed weight of the trajectory update is controlled so that the model can automatically slow down or suppress trajectory disturbances when the target is occluded or blurred, thereby enhancing prediction stability.

[0064] First, construct the trajectory change differential equation of target i:

[0065]

[0066] in: : Target in time Position status (upper left corner coordinates and width and height); : The trajectory velocity modeling network customized in this step has parameters ; : Observation weight factor, expressed as:

[0067]

[0068] Indicates that the visibility score in step 1 is Limited to the lower limit (e.g. 0.1) to 1 to prevent the track from freezing when fully occluded.

[0069] Furthermore, in order to suppress trajectory jumping and comply with physical inertia logic, an acceleration regularization term is added:

[0070]

[0071] in: , : The position of the target in the previous two frames, derived from the historical state; : Acceleration smoothing factor, used to adjust the model's penalty for drastic trajectory changes; this item can be added to the loss function during training to encourage the predicted trajectory to remain smooth in time.

[0072] Finally, by numerically integrating the above differential system (such as fourth-order Runge-Kutta), the predicted position of each target in frame t+1 can be obtained: : Predict spatial position as the feature generation module for the next step The spatial guidance input.

[0073] S3. Obtain a semantic feature vector, fuse the semantic feature vector with the predicted spatial position, generate the future appearance features of the target through the trajectory-guided feature generation module, and introduce structural alignment loss and local attribute diversity preservation items to improve feature quality.

[0074] Specifically, this step aims to solve the problem of missing target appearance information in dynamic video scenes due to factors such as target occlusion, blur, and rapid changes in viewing angle. In order to improve the system's ability to identify and re-detect targets in invisible or blurred states, the present invention designs a trajectory-guided feature generator (TGFS) that integrates historical features and trajectory prediction information to generate the expected features of the target in the next frame. This feature serves as the “identity template” for the subsequent recognition module and is particularly suitable for situations where the target is temporarily unobservable visually.

[0075] The input of TGFS depends on the output of the first two steps, forming a clear variable transfer path:

[0076] The semantic feature vector output from step 1 ;

[0077] The predicted spatial position output from step 2 .

[0078] The goal of the TGFS module is to build a function ,Will and Fusion, output prediction features at time t+1 , this feature must simultaneously meet the following requirements: consistent with historical semantics, relevant to the predicted spatial location, and robustly distinguishable.

[0079] Furthermore, the present invention defines the feature generation process as follows:

[0080]

[0081] in: : The generated future appearance features of target i; : Historical appearance features extracted in step 1; : The target position predicted in step 2; : Spatial position encoding module (such as MLP or SinusoidalEmbedding); : Fusion model, consisting of several layers of attention networks and feature transformation layers; : Represents the learnable parameters of the TGFS module.

[0082] It can be understood that the present invention introduces spatial position coding It is then fused with semantic features to make the generated features "spatially consistent", that is, no erroneous features that deviate from the true location of the target are generated.

[0083] Furthermore, in order to enhance the model's ability to control the quality of generated feature semantics during the training phase, the present invention introduces a structural alignment loss and innovatively adds a local attribute diversity preservation term. , used to encourage the differences between feature sub-dimensions and improve the discriminative ability in the re-identification process:

[0084]

[0085] in, represents the j-th dimension of future appearance features; is the weight coefficient for diversity loss (recommended to be 0.05); the inner summation encourages distribution differences between different dimensions to prevent all dimensions from converging to similar values.

[0086] In the case of occlusion or blur, the lack of visual information may cause the feature dimension to "collapse", and the model will output homogeneous features with low discriminability. This regularization term forces TGFS to retain individuality in the sub-dimensions, which helps the subsequent matching module to perform effective recognition.

[0087] S4. When a frame image is input, a candidate area is constructed based on the predicted spatial position, and the matching confidence between the candidate area and the future appearance features of the target is calculated to confirm the target position and target features; if the matching confidence is lower than the preset threshold, the previous frame state is retained and no update is performed to improve robustness in occluded or blurred states.

[0088] Specifically, this step is the last execution link of the patent method, and its goal is to: In step 2, the target position is predicted The target future appearance features generated in step 3 , confirm the spatial position and identity of the target, and complete an "enhanced guidance" dynamic target detection output.

[0089] Furthermore, to predict the spatial position Centered in the image Construct a set of local candidate regions , extract corresponding features from each region , using the following method with the target's future appearance features Do similarity matching:

[0090]

[0091] in, is the cosine similarity score, which represents the matching confidence between the jth candidate region and the target template; The future appearance features of the target generated in step 3; The features extracted from the candidate region in the current frame image. Find the best matching region , output the detection result of the target in frame t+1: Final target position: ; Final target features: .

[0092] If the highest match Below threshold (such as 0.6), the system will mark the target as "uncertain state", retain the state of the previous frame, and not update it to improve the robustness of the system in occlusion or blur state.

[0093] Finally, this step converts the spatial position of each target at time t+1 into and semantic features As the structured detection output of the patented system, it is used to output result records or further feed back to the trajectory modeling module to form a closed loop.

[0094] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0095] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a division of logical functions. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0096] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0097] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A dynamic target detection method based on deep learning, characterized in that: The method comprises: Acquire a target image, extract dynamic targets through unsupervised background modeling and morphological processing, and combine a deep feature network with a lightweight visibility network to output the target structured state; the target structured state includes: target bounding box, semantic feature vector, and visibility score; Constructing a trajectory velocity modeling network using differential equations based on the bounding box and semantic features to generate a target motion trajectory, introducing an acceleration regularization term to smooth trajectory changes, and finally obtaining a predicted spatial position by numerically integrating the differential equations; Obtain a semantic feature vector, fuse it with the predicted spatial position, generate the target's future appearance features through a trajectory-guided feature generation module, and introduce a structural alignment loss and a local attribute diversity preservation term to improve feature quality; When a frame image is input, a candidate region is constructed based on the predicted spatial position, and the matching confidence between the candidate region and the future appearance features of the target is calculated to confirm the target position and target features. If the matching confidence is lower than a preset threshold, the previous frame state is retained without updating to improve robustness in occlusion or blur. The method of extracting dynamic targets through unsupervised background modeling and morphological processing, combining deep feature networks with lightweight visibility networks, and outputting target structured states specifically includes: Model the image through unsupervised background, perform background subtraction on each frame, and generate a foreground mask; Performing morphological processing on the foreground mask to enhance the integrity of the target boundary and obtain an optimized foreground mask; Extracting a connected region from the optimized foreground mask and recording the connected region as a target region; a bounding box of the target region is a target bounding box; Cutting the target area from the target image and feeding it into a deep feature network to extract semantic features as a semantic feature vector; Extracting a structure vector from each target region in the optimized foreground mask; concatenating the structure vector with a deep feature map of the semantic feature vector and inputting it into a lightweight visibility network to output a visibility score; Combining the object bounding box, semantic feature vector and visibility score, the output is the state of each object.

2. The dynamic target detection method based on deep learning according to claim 1, characterized in that The structure vector includes four dimensions: normalized area, aspect ratio, boundary tightness and isolated object identification.

3. The dynamic target detection method based on deep learning according to claim 1, characterized in that: The morphological processing includes erosion and dilation operations performed sequentially to optimize the object boundary integrity.

4. The dynamic target detection method based on deep learning according to claim 1, characterized in that The unsupervised background modeling adopts Gaussian modeling or sliding average method to generate an initial foreground mask through pixel-level difference threshold segmentation.

5. The dynamic target detection method based on deep learning according to claim 1, characterized in that: The differential equation is expressed as: ； in, , To predict the spatial position, it means the target is in time Position status; The trajectory velocity modeling network is parameterized as ; is the observation weight factor, expressed as: ； in, Indicates that the visibility score is limited to the lower limit Between 1 and 1 to prevent the track from freezing when it is completely occluded; Score visibility. ; The acceleration regularization term constrains the second-order differences of adjacent elements of the time series, suppresses trajectory jumps and conforms to physical inertia logic.

6. The dynamic target detection method based on deep learning according to claim 1, characterized in that: The future appearance features of the target are generated by the trajectory guidance feature generation module, which is specifically expressed as follows: ； in, Indicates the generated target Future appearance features; Represents semantic feature vector; To predict spatial position; Represents the spatial position encoding module; Represents a fusion model consisting of several layers of attention networks and feature transformation layers; represents the learnable parameters of the TGFS module.

7. The dynamic target detection method based on deep learning according to claim 6, characterized in that: The trajectory-guided feature generation module introduces a local attribute diversity preservation term, which is a structural alignment loss of adjacent dimensions of the future appearance feature, and is used to enforce feature sub-dimensional distribution differences.

8. The dynamic target detection method based on deep learning according to claim 6, characterized in that: The spatial position encoding module is MLP or Sinusoidal Embedding.

9. The dynamic target detection method based on deep learning according to claim 1, characterized in that: The candidate region is generated with the predicted spatial position as the center and constructed through a multi-scale sliding window or a region proposal network; The matching confidence is obtained by performing cosine similarity scoring on the semantic feature vector extracted from the candidate region in the current frame image and the future appearance feature of the target.

Images