Target detection method and device, equipment and storage medium
By extracting dynamic visual features from infrared image sequences and fusing feature maps, the problem of accurate identification of small drone targets in complex scenarios was solved, and high-precision detection was achieved in sensitive areas such as airports.
Patent Information
- Application Number
- CN202511595201.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-03
- Publication Date
- 2026-02-27
AI Technical Summary
Existing technologies struggle to accurately identify small drone targets in complex scenarios, especially in sensitive areas such as airports, where they suffer from low detection accuracy and high false negative rates.
By acquiring infrared image sequences, dynamic visual features are extracted to generate motion indicator maps, which are then fused with target appearance feature maps and dynamic feature maps to generate target detection results, enabling accurate identification of small targets.
It effectively filters out complex background interference, significantly improves the accuracy and reliability of small target detection, and reduces the false negative rate.
Smart Images
Figure CN121582607A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to a target detection method, apparatus, device, and storage medium. Background Technology
[0002] Currently, sensitive areas such as airports are frequently harassed by "black flight" drones. These targets are characterized by low flight altitude, variable movement trajectory, small radar cross-section, and occupy only a few pixels when imaging at long distances, with weak signals.
[0003] Current methods for detecting drones mainly include: spatial domain augmentation methods, temporal domain modeling methods, and deep learning fusion methods (which can also be described as spatiotemporal feature fusion methods). Spatial domain augmentation methods enhance the local contrast of the target by manually designing filters; temporal domain modeling methods rely on temporal motion information to extract the target's motion trajectory using frame difference or autoregressive models; and deep learning fusion methods improve detection performance through feature fusion.
[0004] However, the filter parameter adjustment in spatial domain enhancement methods relies on experience and has poor generalization ability in complex backgrounds (such as swaying trees and cloud interference); temporal domain modeling methods can only mark the location of motion and cannot distinguish between real target motion and background disturbance; and the features of small targets (small drones) are significantly attenuated after fusion in deep learning fusion methods. Summary of the Invention
[0005] This application provides a target detection method, apparatus, device, and storage medium to solve the technical problem of accurate identification of small targets in complex scenarios.
[0006] In a first aspect, embodiments of this application provide a target detection method, comprising: acquiring an infrared image sequence of a target region, the infrared image sequence being used to provide spatiotemporal motion information of the target within the target region; performing dynamic visual feature extraction on each frame of the infrared image to generate a corresponding motion indicator map, the motion indicator map being used to mark the potential motion region of the target; and obtaining a target detection result based on the target appearance feature map extracted from the infrared image and the target dynamic feature map extracted from the motion indicator map.
[0007] In one embodiment, dynamic visual feature extraction is performed on each frame of infrared image to generate a corresponding motion indication map. Specifically, this can be achieved by: extracting visual features from each frame of infrared image to obtain a visual feature map; performing background suppression processing on the visual feature map to obtain a background suppression feature map; extracting temporal motion features based on the background suppression feature maps between adjacent frames to obtain a temporal motion feature map, where temporal motion features refer to the motion change features of the target in the time dimension; and performing filtering and thresholding processing on the temporal motion feature map to generate a motion indication map.
[0008] In another embodiment, preliminary visual feature extraction is performed on each frame of infrared image to obtain a visual feature map. Specifically, this can be achieved by: calculating the mean and standard deviation of grayscale values in the neighborhood of each pixel in each frame of infrared image; and performing normalization and mean filtering on the pixel values based on the mean and standard deviation of grayscale values to obtain the visual feature map.
[0009] In another embodiment, temporal motion feature extraction is performed based on the background suppression feature map between adjacent frames to obtain a temporal motion feature map. Specifically, this can be achieved by performing pixel-level difference operations on the background suppression feature map of the current frame infrared image and the background suppression feature map of the previous frame infrared image to obtain a temporal motion feature map.
[0010] In another embodiment, the temporal motion feature map is filtered and thresholded to generate a motion indicator map. Specifically, this can be achieved by: performing Gaussian filtering on the temporal motion feature map to obtain a filtered feature map; and performing binarization filtering on the filtered feature map based on a preset grayscale threshold to generate a motion indicator map. The binarization filtering is used to retain regions with grayscale values greater than the preset grayscale threshold.
[0011] In another embodiment, a target detection result is obtained based on the target appearance feature map extracted from the infrared image and the target dynamic feature map extracted from the motion indicator map. Specifically, this can be achieved by: performing multi-scale feature extraction on the target static visual information in the infrared image to obtain target appearance feature maps at least two scales; performing convolution processing on the target dynamic information in the motion indicator map to obtain target dynamic feature maps at least two scales corresponding to the target appearance feature maps; performing channel stitching and weighted fusion on the target appearance feature map and the target dynamic feature map of the same scale to generate a fused feature map; and performing fused feature detection on the fused feature map to obtain a target detection result, which includes the target location, target category, and confidence level.
[0012] Secondly, embodiments of this application provide a target detection device, including: an acquisition module, a feature extraction module, and a detection module.
[0013] The aforementioned acquisition module is used to acquire infrared image sequences of the target area, which provide spatiotemporal motion information of the target within the target area. The aforementioned feature extraction module is used to: perform dynamic visual feature extraction on each frame of infrared image to generate a corresponding motion indicator map, which is used to mark the potential motion area of the target; The aforementioned detection module is used to: extract target appearance feature maps based on infrared images and target dynamic feature maps based on motion indicator maps, to obtain target detection results.
[0014] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the program to implement the steps of the target detection method of the first or second aspect.
[0015] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the target detection method of the first or second aspect.
[0016] Fifthly, embodiments of this application provide a computer program product, including a computer program, which, when executed by a processor, implements the steps of the target detection method of the first or second aspect.
[0017] The target detection method, apparatus, device, and storage medium provided in this application can completely preserve the target's motion trajectory in the time dimension and positional changes in the spatial dimension by acquiring infrared image sequences of the target area, avoiding the loss of motion features due to isolated information in a single frame image; and combined with the motion indication map generated by dynamic visual feature extraction (precisely marking potential motion areas), it can effectively lock small targets from complex backgrounds, significantly reducing the problem of missed detection caused by low target pixel ratio and weak features.
[0018] Finally, by fusing the target appearance feature map generated from the infrared image (depicting static attributes such as target grayscale, contour, and texture) with the target dynamic feature map generated from the motion indicator image (focusing on dynamic attributes such as target movement trend and displacement), a dual constraint of static appearance localization and dynamic motion verification is achieved. Compared to single feature detection, this fusion method can more comprehensively depict the features of small targets, reduce false detections caused by background interference (such as slight movement of leaves or changes in lighting) or similar target shapes, and significantly improve the accuracy and reliability of small target detection. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 A schematic flowchart of the target detection method provided in the embodiments of this application; Figure 2 A schematic diagram illustrating the principle of the target detection method provided in the embodiments of this application; Figure 3 This is a schematic diagram of the target detection device provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0022] In the description of this application, it should be understood that the terms "upper," "lower," "left," "right," "front," "rear," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or relative positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and for simplification, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Unless otherwise specified, the above-mentioned orientational descriptions can be flexibly set in practical applications, provided that the relative positional relationships shown in the accompanying drawings are satisfied.
[0023] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.
[0024] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "communication" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection. They can refer to a direct connection or an indirect connection through an intermediate medium, or a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.
[0025] In embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, article, or apparatus that includes that element.
[0026] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.
[0027] In the embodiments of this application, at least one can also be described as one or more, and multiple can be two, three, four or more, and this application does not impose any restrictions.
[0028] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.
[0029] To facilitate understanding, the terms used in the embodiments of this application will be explained first.
[0030] You Only Look Once (YOLO) is a real-time object detection technology based on deep learning. Its core is to simultaneously predict the target location (bounding box regression) and determine the target category (classification) through a single forward propagation of the neural network. It breaks the traditional two-stage detection mode (first generate candidate regions, then classify) and greatly improves the computing speed while ensuring detection accuracy. It is suitable for real-time detection scenarios such as video surveillance, autonomous driving, and drone recognition.
[0031] Tiny-Intersection over Union Loss (Tiny-IoULoss) is an improved bounding box regression loss function designed to address the localization error and bounding box offset issues in the detection of tiny targets (target size is typically less than 0.12% of the image).
[0032] Generalized Intersection over Union (GIoU) is an extension of the traditional Intersection over Union (IoU). It introduces the area difference between the predicted bounding box and the ground truth bounding box to solve the gradient vanishing problem of non-overlapping boxes in the traditional IoU, and provides a basic localization error measurement capability for bounding box regression.
[0033] Commonly used object detection methods in the industry mainly include spatial domain augmentation methods, temporal domain modeling methods, and deep learning fusion methods (which can also be described as spatiotemporal feature fusion methods).
[0034] Among them, spatial domain enhancement methods and temporal domain modeling methods mostly rely on local grayscale, texture or morphological features of images to enhance point-like or weakly bright targets through methods such as difference and filtering. The core is based on classical image processing (such as Top-Hat transform and wavelet transform) or traditional machine learning (such as sparse coding and support vector machine).
[0035] Subsequently, deep learning frameworks, represented by the YOLO series and Transformer, broke through traditional limitations, achieving end-to-end unification of feature learning and target classification, eliminating the need for manual feature design and significantly promoting the development of target detection technology. Specifically in the field of infrared small target detection, the research focus has further shifted to automatic detection frameworks based on feature learning and spatiotemporal fusion: some schemes introduce inter-frame difference, optical flow estimation, or multi-frame convolution to capture dynamic features.
[0036] However, existing methods still have core limitations in complex dynamic scenes (such as urban, forest, mountain, and water environments), significantly limiting their detection performance. These limitations can be summarized in four points: Weak features are easily attenuated: At long distances, targets such as micro-sized drones occupy only a few pixels, and weak features are easily attenuated and lost after multiple convolutions. Although existing methods compensate through multi-scale detection heads in the later stage, there is a lack of targeted weak signal enhancement mechanisms in the low-level feature extraction stage, which cannot preserve weak target information from the source.
[0037] Difficulty in distinguishing targets and high false alarm rates: Complex terrain textures and fog interference reduce the signal-to-noise ratio between targets and the background. Combined with atmospheric turbulence and momentary occlusion, targets appear discontinuously in the sequence, making it difficult to distinguish between drones with weak appearance features and small targets such as birds, ultimately leading to high false alarm rates, high false alarm rates, and high misjudgments.
[0038] Dynamic background interference is difficult to distinguish: Most methods rely on simple temporal modeling such as inter-frame difference and short-time convolution, which can only capture explicit motion features and cannot effectively distinguish between real target motion and background disturbances such as cloud drift, water ripples, and leaf swaying, resulting in frequent false alarms in non-rigid dynamic scenes.
[0039] Insufficient feature fusion capability: Traditional networks often use simple feature splicing or linear weighting to fuse spatiotemporal features, lacking the ability to adaptively adjust features of different dimensions. This not only fails to fully highlight the saliency of the target, but also easily introduces feature redundancy, which restricts the overall performance improvement of the model.
[0040] Based on this, this application proposes a target detection method, comprising: acquiring an infrared image sequence of a target region, wherein the infrared image sequence is used to provide spatiotemporal motion information of the target within the target region; performing dynamic visual feature extraction on each frame of infrared image to generate a corresponding motion indication map, wherein the motion indication map is used to mark the potential motion region of the target; and obtaining a target detection result based on the target appearance feature map extracted from the infrared image and the target dynamic feature map extracted from the motion indication map.
[0041] The target detection method provided in this application is particularly suitable for the accurate identification of moving targets in complex scenes. By acquiring infrared image sequences, it can fully explore the spatiotemporal motion information of the target, making up for the information gap of single-frame static detection. It extracts dynamic visual features to generate motion indication maps, which can accurately mark potential motion areas and effectively filter background disturbances such as tree swaying and cloud interference. The fused appearance and dynamic feature maps achieve dual-dimensional feature complementarity, which can highlight the weak features of small targets at a distance and avoid feature decay. Based on dual feature map detection, it can still stably identify targets in complex dynamic backgrounds.
[0042] The target detection method is performed by a device and / or platform with data processing and model running capabilities, but is not limited thereto, and the embodiments of this application do not impose specific limitations on this. For example, it can be performed by an embedded infrared monitoring device.
[0043] Figure 1 This is a schematic flowchart illustrating the target detection method provided in an embodiment of this application. (Refer to...) Figure 1 This application provides a target detection method, which may include: Step S101: Obtain the infrared image sequence of the target area.
[0044] The target area refers to the specific spatial range in which target detection needs to be performed.
[0045] The size of the target area can be limited according to actual needs. This application embodiment does not impose specific limitations on this. For example, it can be determined comprehensively based on the spatial scale of the detection scene, the detection range of the infrared imaging device, and the activity radius of the target (such as a drone) to ensure complete coverage of the target's potential movement trajectory.
[0046] Optionally, the target area can be an area prone to low-altitude small targets, such as an airport airspace, a border control and monitoring zone, or the airspace surrounding sensitive facilities, but it is not limited to these areas. This application embodiment does not impose specific restrictions on this.
[0047] An infrared image sequence refers to a collection of multiple frames of infrared images continuously acquired by infrared imaging equipment for a specific target area (such as airport airspace or airspace surrounding sensitive facilities) in target detection tasks (such as UAV detection).
[0048] In this embodiment of the application, an infrared image sequence can be obtained by continuously acquiring frames of the target area through a device equipped with an infrared imaging sensor (such as an infrared camera, an UAV-borne infrared detection module, a fixed infrared monitoring device, etc.).
[0049] It is important to note that during data acquisition, a reasonable frame rate and imaging resolution should be set based on the target area, the estimated size of the target, and its movement speed to ensure that the infrared image sequence can completely record the target's movement process in the time dimension, providing continuous and effective image data support for subsequent extraction of spatiotemporal motion information.
[0050] Step S102: Extract dynamic visual features from each frame of infrared image to generate a corresponding motion indicator map.
[0051] Among them, dynamic visual features refer to the feature information related to the target's motion extracted from each frame of the infrared image sequence, simulating the neural information processing mechanism of biological retinal Y cells (photoreceptor layer - horizontal cell layer - bipolar cell layer - amacrine cell layer - ganglion cell layer), which is used to reflect the target's dynamic changes in the time dimension and its motion attributes in the spatial dimension.
[0052] For example, by performing layered processing on each frame of an infrared image sequence, feature information that reflects the motion attributes of a target (such as a micro-drone) can be extracted from low signal-to-noise ratio, complex dynamic backgrounds (such as swaying leaves or cloud movement).
[0053] A motion indicator map is a single-channel image with the same size as the input infrared image, used to visually mark the potential motion area of a target within the target region.
[0054] Specifically, visual features are extracted from each frame of infrared image to obtain a visual feature map; background suppression processing is performed on the visual feature map to obtain a background suppression feature map; temporal motion features are extracted based on the background suppression feature maps between adjacent frames to obtain a temporal motion feature map; and filtering and thresholding processing are performed on the temporal motion feature map to generate a motion indication map.
[0055] Visual feature extraction refers to the low-level features that reflect the basic visual attributes of an image and are directly extracted from an infrared image. These features mainly include basic information at the pixel level and in local areas (such as the mean gray value, standard deviation, local gray value distribution, and simple texture outlines of the pixel neighborhood).
[0056] A visual feature map is a structured feature carrier that represents the underlying visual information of an image after preliminary visual feature extraction of each frame of an infrared image. Its core is the visualization and structured presentation of the basic visual attributes of the pixel level and local areas of the infrared image.
[0057] Visual feature extraction can enhance the saliency of features of small targets (such as micro drones) under low signal-to-noise ratio conditions, laying the foundation for subsequent processing.
[0058] In this embodiment of the application, the detailed steps for extracting visual features from each frame of infrared image to obtain a visual feature map include: calculating the mean and standard deviation of grayscale values in the neighborhood of each pixel in each frame of infrared image; and performing normalization and mean filtering on the pixel values based on the mean and standard deviation of grayscale values to obtain a visual feature map.
[0059] For example, for each pixel in the infrared image, a 3×3 local window is taken with that pixel as the center, and the mean and standard deviation of the gray values of all pixels within the local window are calculated.
[0060] The grayscale mean is the average of the grayscale values of all pixels within the local window.
[0061] Based on the standard deviation formula, the standard deviation is obtained by taking the square root of the sum of the squared differences between the gray values of each pixel and the mean gray value within a local window, and then taking the average of the sums.
[0062] By subtracting the mean gray value of its neighborhood from the original gray value of the target pixel, and then dividing by the sum of the standard deviation and the zero constant, a normalized infrared image is obtained, making the gray value difference between the target pixel and the background more significant.
[0063] Specifically, the normalized infrared image satisfies the following formula: in, This represents the pixel values of the normalized infrared image. This represents the pixel values of the input infrared image. This represents the average grayscale value of a local window region centered at pixel (x, y). ϵ represents the standard deviation of the local window region centered at pixel (x, y), and ϵ represents the constant excluding zero.
[0064] Furthermore, the normalized infrared image is subjected to mean filtering to simulate the center-periphery antagonistic receptive field characteristics of biological retinal Y cells. The specific process is as follows: For the normalized infrared image, two two-dimensional Gaussian kernels G with different standard deviations are used. σ1 With G σ2 (Where σ1 and σ2 are the standard deviations of the two Gaussian kernels, and σ1 < σ2), first obtain G through convolution operation. σ1 and The convolution result, G σ2 and The convolution result of G; then G σ2 and The convolution result is multiplied by the suppression coefficient k (k is a preset constant, ranging from 0.5 to 0.9, used to adjust the peripheral suppression intensity), and then multiplied by G. σ1 and The convolution results are subtracted; then the result is added to the normalized infrared image to obtain the visual feature map. The corresponding calculation formula is: in, Represents visual feature maps. This represents a Gaussian kernel with a standard deviation of σ1. This represents a Gaussian kernel with a standard deviation of σ². represents the normalized infrared image, and * represents the two-dimensional convolution operation.
[0065] The target detection method provided in this application enhances the target signal in the central region and suppresses background noise in the peripheral region through normalization and mean filtering, thus completing the initial separation of the target and the background and laying the foundation for subsequent dynamic visual feature extraction.
[0066] Furthermore, the detailed steps for performing background suppression processing on the visual feature map to obtain the background suppression feature map include: performing a difference operation between the visual feature map and the infrared image after Gaussian filtering to obtain the background suppression feature map.
[0067] In the background suppression feature map, the highlighted areas represent the regions where the target is located.
[0068] Specifically, through a large-scale Gaussian kernel G σl To achieve lateral suppression and simulate background suppression, the formula for calculating the background suppression feature map is as follows: in, Represents the background suppression feature map. This represents the visual feature map, and α represents the suppression intensity coefficient (used to adjust the degree of background suppression). represents a Gaussian kernel with a standard deviation of σ1, and ∗ represents a two-dimensional convolution operation.
[0069] The target detection method provided in this application can significantly reduce the interference of large-area flat backgrounds (such as sky and ground textures) through background suppression processing, making the potential moving target area more prominent.
[0070] Furthermore, the detailed steps for extracting temporal motion features based on the background suppression feature maps between adjacent frames to obtain the temporal motion feature map include: performing pixel-level difference operations on the background suppression feature map of the current frame infrared image and the background suppression feature map of the previous frame infrared image to obtain the temporal motion feature map.
[0071] Temporal motion feature map refers to the motion change characteristics of a target in the time dimension.
[0072] Motion change characteristics refer to the dynamic differences in visual signals generated by the movement of a target; they are descriptions of the changes that occur during the movement of a target.
[0073] Optionally, motion change features include pixel grayscale changes in the dynamic region where the target is located between adjacent frames, subtle deformations of the target outline caused by motion, and changes in the contrast difference between the dynamic region and the static background, but are not limited thereto, and the embodiments of this application do not impose specific limitations on this.
[0074] Specifically, the temporal motion feature map satisfies the following formula: in, Represents temporal motion feature maps. This represents the image after performing Gaussian kernel difference operations on the background suppression feature map at two different scales. This represents a Gaussian kernel with a standard deviation of σ1. This represents a Gaussian kernel with a standard deviation of σ², where σ₁ is less than σ², and Δt represents the time step, typically set to 1, i.e., the time between adjacent image frames. This represents the background suppression feature map.
[0075] The target detection method provided in this application amplifies pixel changes caused by target motion (such as drone movement) by subtracting the current frame from the previous frame and taking the absolute value, while filtering out interference from static backgrounds (such as fixed buildings) and slowly dynamic backgrounds (such as slow-moving clouds), and finally extracts the real motion area in the scene.
[0076] Furthermore, the detailed steps for filtering and thresholding the temporal motion feature map to generate a motion indicator map include: performing Gaussian filtering on the temporal motion feature map to obtain the filtered feature map; and performing binarization filtering on the filtered feature map based on a preset grayscale threshold to generate the motion indicator map.
[0077] Among them, binarization filtering is used to retain regions with gray values greater than a preset gray value threshold.
[0078] Specifically, the motion indicator graph satisfies the following formula: in, Indicates motion indication diagram, Representing the rectified linear unit, it is a non-linear activation function that sets values less than 0 to 0, simulating the firing threshold characteristics of biological neurons (a response is triggered only when the signal strength exceeds the threshold). Represents temporal motion feature maps. This represents a Gaussian kernel with a standard deviation of σ1. This indicates the activation threshold.
[0079] The target detection method provided in this application integrates the instantaneous motion features extracted from the temporal motion feature map by smoothing and thresholding, filtering redundant noise, and finally outputting a motion indication map that can intuitively mark the potential motion area of the target.
[0080] Step S103: Obtain the target detection result by extracting the target appearance feature map based on the infrared image and the target dynamic feature map based on the motion indicator map.
[0081] Among them, the target appearance feature map refers to the feature map that reflects the spatial appearance attributes of the target (such as the shape and texture of the target) extracted from the input infrared image (original infrared image) by the convolutional neural network feature extraction module (using a lightweight structure such as depthwise separable convolution).
[0082] Optionally, the target appearance feature map includes spatial dimension information of the target in a single frame image, such as grayscale distribution, contour shape, local texture, etc., but is not limited thereto, and the embodiments of this application do not impose specific limitations on this.
[0083] The target dynamic feature map refers to a feature map that reflects the target's temporal motion attributes (such as the target's trajectory and instantaneous changes) and is extracted based on the motion indication map through a feature extraction path that runs parallel to the target appearance feature map.
[0084] The target dynamic feature map inherits the potential motion region information of the target in the motion indication map, and further enhances the expression of motion features through feature extraction.
[0085] Specifically, multi-scale feature extraction is performed on the static visual information of the target in the infrared image to obtain target appearance feature maps at least two scales.
[0086] Static visual information refers to the visual attributes in an image that do not change over time, mainly including the inherent visual features of the target such as shape, texture, grayscale distribution, and contour structure.
[0087] In this embodiment of the application, taking UAV infrared images as an example, static visual information can be represented as follows: 1. Shape characteristics: the fuselage outline of the UAV (such as the rotor layout shape of a multi-rotor, the wingspan shape of a fixed-wing aircraft), the structural proportions of the wings and fuselage, etc.
[0088] 2. Texture features: Infrared texture of the drone surface material (such as uniform grayscale texture of metal fuselage, relatively rough texture of plastic parts), local texture differences at the rotor shaft, etc.
[0089] 3. Gray-scale distribution characteristics: the difference in infrared gray-scale between the UAV as a whole and the background (such as the sky and ground buildings), the gray-scale gradient distribution of different parts of the UAV body (such as the hot area of the engine and the cold area of the fuselage), etc.
[0090] In this embodiment, the target appearance feature map uses the current frame in the infrared image sequence as input. To ensure the integrity of target feature extraction while controlling computational overhead, a staged partially connected network structure is adopted. The specific implementation method and principle are as follows: Based on the different feature extraction requirements of different network layers, the structure configures the convolution operation method differently: for shallow layers (i.e., the initial stage of feature extraction, corresponding to the layer with larger feature map size and focusing on the extraction of local details of the target), grouped convolution or depthwise separable convolution is used; for deep layers (i.e., the later stage of feature extraction, corresponding to the layer with smaller feature map size and focusing on the integration of global semantic information of the target), standard convolution is used.
[0091] Among them, grouped convolution divides the channel dimension of the input feature map into several independent groups, so that each channel only performs convolution operation with the convolution kernel of the corresponding group, reducing redundant calculations between channels; depthwise separable convolution splits the standard convolution into two steps: depthwise convolution (single-channel convolution kernel processes the corresponding input channel separately) and pointwise convolution (1×1 convolution kernel integrates the features of each channel), which greatly reduces the number of parameters and the amount of computation.
[0092] The application of these two types of convolution in the shallow layers of the network can effectively preserve high-frequency detail features such as the edges of small targets and local grayscale changes in infrared images while suppressing computational costs, thus avoiding the loss of details caused by excessive compression.
[0093] Specifically, feature extraction can be performed using the following formula: in, A partial convolutional block is a network module that includes operations such as grouped convolution, depthwise separable convolution, or standard convolution. It is used to perform convolution operations on input features to extract features at different scales or levels. This represents the output target appearance feature map of the a-th feature branch in stage i. This represents the target appearance feature map input to the a-th feature branch in stage i-1, meaning that the input of the current stage is the output of the previous stage.
[0094] Furthermore, the target dynamic information in the motion indication map is convolved to obtain target dynamic feature maps at least two scales corresponding to the target appearance feature map.
[0095] Target dynamic information refers to the visual changes of a target over time as it moves.
[0096] In this embodiment of the application, taking UAV infrared images as an example, the target dynamic information can be represented as follows: Motion trajectory / displacement: such as the path the drone flies (straight line / turn), how many pixels the position moves in each frame of the image; Changes in shape with motion: for example, the blurred areas created when the fuselage tilts or the rotors spin during flight; Gray-scale changes during motion: for example, the shift of the engine hot zone with motion, and the alternation of light and dark gray caused by rotor rotation.
[0097] Specifically, by using lightweight convolutional layer feature mapping and channel attention mechanisms, motion features are adaptively extracted and enhanced, ultimately generating a dynamic feature map of the target.
[0098] Lightweight convolutional layer feature mapping is performed using the following formula: in, This represents the initial dynamic feature map of the target. The weights of lightweight convolutional layers are represented by depthwise separable convolutions or grouped convolutions to reduce computational cost. This represents a motion indicator diagram.
[0099] Lightweight convolutional layer feature mapping maps motion feature maps to initial dynamic feature maps, preserving the basic features of the motion region.
[0100] Subsequently, a channel attention mechanism is introduced to adaptively calibrate the feature channels, using attention weights s to adjust the initial dynamic feature map. Channel-by-channel weighting is performed to enhance the features of channels strongly correlated with the moving target and suppress irrelevant noise channels, ultimately generating a dynamic feature map of the target.
[0101] For feature maps Its attention weights s satisfy the following formula: Where s represents the attention weight, and GAP represents global average pooling. and These are the weights of the fully connected layer. It is the ReLU activation function. It is the Sigmoid function.
[0102] The weighted feature map satisfies the following formula: in, Representation of feature map The feature map after channel attention weighting is also known as the target dynamic feature map.
[0103] Furthermore, channel stitching and weighted fusion are performed on the target appearance feature map and the target dynamic feature map at the same scale to generate a fused feature map. The fused feature map is then used for fused feature detection to obtain the target detection result.
[0104] The detection result refers to the set of information about the target's location, classification, confidence level, etc., output after fusing and analyzing the target's appearance feature map and target dynamic feature map.
[0105] Optionally, the detection results may include the target's location, target category, and confidence level, but are not limited thereto, and the embodiments of this application do not impose specific limitations on this.
[0106] Target location: Select the coordinate area where the target is located in the infrared image (such as the coordinates of the upper left and lower right corners of the rectangle).
[0107] Target Category: Identify the type of target (e.g., micro drones, birds, etc.).
[0108] Confidence level: The reliability of the model for the detection result (e.g., 0.95 means that there is a 95% probability that it is the target).
[0109] For target dynamic feature maps and target appearance feature maps of the same scale, an adaptive gating mechanism is used for fusion.
[0110] First, calculate the fusion weight graph: in, The fusion weight map at scale l has values between 0 and 1, used to characterize the fusion ratio of appearance features and dynamic features. The sigmoid function represents the activation function, mapping the weights to the 0-1 interval to simulate the gating switching characteristics. [,] represents channel splicing. , This represents the weights and biases of the gated convolutional layer, which are adaptively adjusted through learning. Represents the target appearance feature map at scale l. With target dynamic feature map The channel splicing operation integrates dual-modal features.
[0111] The characteristics after fusion are: in, The fused feature map at scale l is represented by the weight map. Pixel-wise weighting of appearance and dynamic features enables adaptive fusion that emphasizes motion features in dynamic regions and appearance features in static backgrounds.
[0112] Generated at each scale Furthermore, cross-scale feature fusion is performed through a feature pyramid network, and then the detection head (such as the detection layer of the YOLO series) is connected to perform target localization and classification. Finally, the detection result containing the target location (bounding box coordinates), target category, and confidence score is output.
[0113] The target detection method provided in this application can accurately capture the static visual attributes of a target, such as shape, texture, and grayscale distribution, through appearance feature maps. Even in low-contrast and noisy environments common in infrared scenes, it can still effectively distinguish the inherent differences between the target and the background. Through dynamic feature maps, it can focus on the temporal motion attributes of the target, such as its motion trajectory and dynamic region distribution, quickly eliminating interference from static backgrounds (such as fixed buildings and vegetation) and accurately locating potential moving targets. The two methods complement each other, improving the accuracy of target detection in infrared scenes.
[0114] For example, Figure 2 This is a schematic diagram illustrating the principle of the target detection method provided in the embodiments of this application. Figure 2 A bidirectional feature fusion network architecture for infrared small target detection is demonstrated, consisting of four core modules: spatial flow, temporal flow, bidirectional feature fusion, and decoupling head. Details are as follows: 1. Input and dual-branch initialization: The input is an infrared image sequence.
[0115] Spatial Flow: Si represents a single frame of infrared image in an infrared image sequence. Si captures the spatial appearance features (such as grayscale, contour, texture, etc.) of the target through a multi-scale feature extraction module (generating appearance feature maps at different scales P1-P5).
[0116] Temporal Flow: First, the motion extraction module extracts temporal motion feature maps Ti from the image sequence. Then, the multi-scale feature extraction module generates P1-P5 temporal motion feature maps to capture the temporal motion features of the target (such as motion regions, trajectories, etc.).
[0117] Spatial-Temporal Fusion (STF): STF fuses features from spatial and temporal streams across modalities, enhancing the spatiotemporal correlation of features at each scale and providing richer semantic information for subsequent fusion.
[0118] 2. Two-way feature fusion.
[0119] This module uses upsampling and downsampling operations to generate feature maps at different scales (such as P). i+1 P i P i−1 Cross-scale fusion between ).
[0120] For example, large-scale feature maps (such as P) i+1) through upsampling and intermediate scale feature maps (such as P) i ) fusion, small-scale feature maps (such as P) i−1 By fusing downsampling with intermediate-scale feature maps, a multi-scale fused feature map is finally generated, taking into account both details and semantic information.
[0121] 3. Decoupling head.
[0122] This module divides the fused feature map into two parallel branches: Classification Header: Predicts the target category and confidence level through convolutional operations (3×3+1×1 convolution).
[0123] Bounding box regression branch: Predicts the bounding box coordinates of the target through convolution operations (3×3+1×1 convolution), thereby decoupling and optimizing classification and localization and improving detection accuracy.
[0124] The target detection method provided in this application, through the collaborative design of bidirectional feature fusion and decoupling head, can achieve high-precision and high-robustness detection of small targets in complex optoelectronic environments (such as background clutter, small target size, and weak motion).
[0125] It should be noted that for small targets (usually smaller than 0.12% of the image), due to the low pixel ratio and limited feature information, problems such as inaccurate bounding box positioning (e.g., box offset, size imbalance) are prone to occur.
[0126] Tiny-IoU Loss is an improved loss function designed to address localization errors and bounding box offset issues in small object detection. It adds three constraint terms to GIoU and achieves accurate bounding box regression through weighted fusion, specifically including: (1) Contextual constraints.
[0127] Function: Enhances the modeling of the target's surrounding environment and avoids positioning errors caused by the isolation of small target features.
[0128] Principle: Instead of focusing solely on the overlap between the predicted bounding box and the ground truth bounding box, it also encourages a higher IoU between the two in their surrounding context (such as background pixels within a certain range around the target). By associating the spatial relationship between the target and the environment, it assists in the accurate localization of the bounding box.
[0129] Specifically, the formula for the context constraint term loss is as follows: in, This represents the loss value of the context constraint term. This represents the intersection-union ratio (IoU) between the predicted bounding box and the surrounding context area of the ground truth bounding box.
[0130] (2) Distance constraint terms.
[0131] Function: Forces the center point of the predicted bounding box to align with the center point of the ground truth bounding box, avoiding the problem of overall box position offset.
[0132] Specifically, the loss formula for the distance constraint term is as follows: in, This represents the loss value of the distance constraint term. This indicates the coordinates of the center point of the prediction box. This represents the coordinates of the center point of the true bounding box. The symbol represents the Euclidean distance calculation, used to measure the straight-line distance between two center points, and diag represents the length of the image diagonal.
[0133] (3) Shape constraints.
[0134] Function: To suppress the size mismatch error between the predicted bounding box and the true bounding box, and to avoid target truncation or redundant background problems caused by box stretching / compression.
[0135] Principle: By measuring the difference between the predicted bounding box aspect ratio and the actual bounding box aspect ratio (such as calculating the logarithmic deviation or squared difference of the two ratios), the degree of imbalance in the proportions is converted into a loss value. The closer the proportions are, the lower the loss value, ensuring that the bounding box shape fits the actual target.
[0136] Specifically, the formula for the shape constraint term loss is as follows: in, This represents the loss value of the shape constraint term. These refer to the width (w) and height (h) dimensions of the bounding box, respectively. This indicates the size of the predicted bounding box in the corresponding dimension (width or height). This represents the dimension (width or height) of the actual bounding box. , Take the minimum and maximum values of the corresponding dimensions of the predicted bounding box and the ground truth bounding box, respectively.
[0137] The total loss is the weighted sum of the above items: in, This represents the total loss value. This represents the basic generalized intersection and union ratio loss. Represents the loss due to the distance constraint term. Represents the loss due to shape constraints. This represents the loss due to context constraints. , , This represents a hyperparameter used to balance the contributions of distance, shape, and context constraints to the total loss. It needs to be determined through experimental tuning based on the specific dataset (such as an infrared small target dataset).
[0138] The target detection method provided in this application can comprehensively balance the multi-dimensional losses of the model, such as the localization error, classification error and feature fusion bias of small targets, through the total loss value. By accurately updating the network parameters through the backpropagation algorithm, the model can focus on the subtle features (such as local grayscale differences and edge contours) of small photoelectric targets (such as weak targets at long distances and with low signal-to-noise ratios), reduce missed detections and false detections, and significantly improve the detection accuracy and localization precision.
[0139] In summary, acquiring infrared image sequences of the target area can completely preserve the target's motion trajectory in the time dimension and positional changes in the spatial dimension, avoiding the loss of motion features due to isolated information in a single frame image. Furthermore, by combining the motion indicator map generated by dynamic visual feature extraction (which accurately marks potential motion areas), small targets can be effectively locked from complex backgrounds, significantly reducing the problem of missed detection caused by low target pixel ratio and weak features.
[0140] By fusing a target appearance feature map generated from infrared images (depicting static attributes such as target grayscale, contour, and texture) with a target dynamic feature map generated from motion indicator images (focusing on dynamic attributes such as target movement trends and displacement), a dual constraint of static appearance localization and dynamic motion verification is achieved. Compared to single feature detection, this fusion method can more comprehensively depict the features of small targets, reduce false detections caused by background interference (such as slight movement of leaves or changes in lighting) or similar target shapes, and significantly improve the accuracy and reliability of small target detection.
[0141] The foregoing mainly describes the solution provided in this application. Accordingly, this application also provides a target detection device for implementing the above-described method embodiments.
[0142] The target detection device provided in the embodiments of this application is described below. The target detection device described below and the target detection method described above can be referred to in correspondence.
[0143] Figure 3 This is a schematic diagram of the target detection device provided in an embodiment of this application. Figure 3 As shown, this application provides a target detection device, which may include: The acquisition module 301 is used to: acquire an infrared image sequence of the target area, wherein the infrared image sequence is used to provide spatiotemporal motion information of the target within the target area; The feature extraction module 302 is used to: perform dynamic visual feature extraction on each frame of infrared image and generate a corresponding motion indicator map, which is used to mark the potential motion area of the target; The detection module 303 is used to obtain target detection results by: extracting target appearance feature maps based on infrared images and extracting target dynamic feature maps based on motion indicator maps.
[0144] In one embodiment, the feature extraction module 302 is specifically used to extract visual features from each frame of infrared image to obtain a visual feature map; perform background suppression processing on the visual feature map to obtain a background suppression feature map; extract temporal motion features based on the background suppression feature maps between adjacent frames to obtain a temporal motion feature map, where temporal motion features refer to the motion change features of the target in the time dimension; and perform filtering and threshold screening processing on the temporal motion feature map to generate a motion indication map.
[0145] In another embodiment, the feature extraction module 302 is specifically used to calculate the mean and standard deviation of grayscale values in the neighborhood of each pixel in each frame of infrared image; and to perform normalization and mean filtering on the pixel values based on the mean and standard deviation of grayscale values to obtain a visual feature map.
[0146] In another embodiment, the feature extraction module 302 is specifically used to perform pixel-level difference operations on the background suppression features of the current frame infrared image and the background suppression features of the previous frame infrared image to obtain a temporal motion feature map.
[0147] In another embodiment, the feature extraction module 302 is specifically used to perform Gaussian filtering on the temporal motion feature map to obtain a filtered feature map; and to perform binarization filtering on the filtered feature map based on a preset grayscale threshold to generate a motion indicator map. The binarization filtering is used to retain regions with grayscale values greater than the preset grayscale threshold.
[0148] In another embodiment, the detection module 303 is specifically used to perform multi-scale feature extraction on the static visual information of the target in the infrared image to obtain target appearance feature maps at least two scales; to perform convolution processing on the target dynamic information in the motion indicator map to obtain target dynamic feature maps at least two scales corresponding to the appearance feature maps; to perform channel splicing and weighted fusion on the target appearance feature map and the target dynamic feature map of the same scale to generate a fused feature map; and to perform fused feature detection on the fused feature map to obtain the target detection result, which includes the target position, target category and confidence level of the target.
[0149] In some embodiments, the target detection device includes hardware structures and / or software modules corresponding to the execution of each function in order to achieve the above-described functions. Those skilled in the art will readily recognize that, based on the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0150] This application embodiment can divide the target detection device into functional modules according to the above method embodiment. For example, each function can be divided into a separate functional module, or two or more functions can be integrated into a feature extraction module. The integrated module can be implemented in hardware or as a software functional module. It should be noted that the module division in this application embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.
[0151] Figure 4 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application, such as... Figure 4 As shown, the electronic device may include a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute the steps of a target detection method, which includes: acquiring an infrared image sequence of a target area, the infrared image sequence being used to provide spatiotemporal motion information of the target within the target area; performing dynamic visual feature extraction on each frame of the infrared image to generate a corresponding motion indication map, the motion indication map being used to mark the potential motion area of the target; and obtaining a target detection result based on the target appearance feature map extracted from the infrared image and the target dynamic feature map extracted from the motion indication map.
[0152] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0153] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the steps of the target detection method provided in the above embodiments, such as: acquiring an infrared image sequence of a target area, the infrared image sequence being used to provide spatiotemporal motion information of the target within the target area; performing dynamic visual feature extraction on each frame of infrared image to generate a corresponding motion indication map, the motion indication map being used to mark the potential motion area of the target; and obtaining a target detection result based on the target appearance feature map extracted from the infrared image and the target dynamic feature map extracted from the motion indication map.
[0154] On the other hand, embodiments of this application also provide a processor-readable storage medium storing a computer program. The computer program is used to cause the processor to execute the steps of the methods provided in the above embodiments, such as: acquiring an infrared image sequence of a target area, the infrared image sequence being used to provide spatiotemporal motion information of a target within the target area; performing dynamic visual feature extraction on each frame of infrared image to generate a corresponding motion indication map, the motion indication map being used to mark the potential motion area of the target; and obtaining a target detection result based on the target appearance feature map extracted from the infrared image and the target dynamic feature map extracted from the motion indication map.
[0155] Processor-readable storage media can be any available medium or data storage device that the processor can access, including but not limited to magnetic storage (such as floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (such as CDs, DVDs, BDs, HVDs, etc.), and semiconductor storage (such as ROMs, EPROMs, EEPROMs, non-volatile memory (NAND flash), solid-state drives (SSDs)).
[0156] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0157] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.
[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A target detection method, characterized in that, include: Acquire an infrared image sequence of the target area, the infrared image sequence being used to provide spatiotemporal motion information of the target within the target area; Dynamic visual features are extracted from each frame of infrared image to generate a corresponding motion indication map, which is used to mark the potential motion area of the target; The target detection result is obtained based on the target appearance feature map extracted from the infrared image and the target dynamic feature map extracted from the motion indicator map.
2. The target detection method according to claim 1, characterized in that, The step of extracting dynamic visual features from each frame of infrared image to generate a corresponding motion indicator map includes: Visual features are extracted from each frame of the infrared image to obtain a visual feature map; The visual feature map is subjected to background suppression processing to obtain a background suppressed feature map; Temporal motion features are extracted based on the background suppression feature maps between adjacent frames to obtain a temporal motion feature map, wherein the temporal motion features refer to the motion change features of the target in the time dimension. The temporal motion feature map is filtered and thresholded to generate the motion indication map.
3. The target detection method according to claim 2, characterized in that, The preliminary visual feature extraction of each frame of infrared image to obtain a visual feature map includes: Calculate the mean and standard deviation of grayscale values in the neighborhood of each pixel in each frame of the infrared image; The pixel values are normalized and mean filtered based on the mean gray value and the standard deviation to obtain the visual feature map.
4. The target detection method according to claim 2, characterized in that, The temporal motion feature extraction based on the background suppression feature map between adjacent frames, to obtain the temporal motion feature map, includes: The temporal motion feature map is obtained by performing pixel-level difference operations on the background suppression feature map of the current frame infrared image and the background suppression feature map of the previous frame infrared image.
5. The target detection method according to claim 2, characterized in that, The step of filtering and thresholding the temporal motion feature map to generate the motion indication map includes: The time-series motion feature map is subjected to Gaussian filtering to obtain the filtered feature map; The filtered feature map is binarized based on a preset grayscale threshold to generate the motion indicator map. The binarization filtering is used to retain regions with grayscale values greater than the preset grayscale threshold.
6. The target detection method according to any one of claims 1 to 5, characterized in that, The target appearance feature map extracted based on the infrared image and the target dynamic feature map extracted based on the motion indicator map are used to obtain the target detection result, including: Multi-scale feature extraction is performed on the static visual information of the target in the infrared image to obtain target appearance feature maps at least two scales; The target dynamic information in the motion indication map is convolved to obtain the target dynamic feature map at least two scales corresponding to the target appearance feature map; The target appearance feature map and the target dynamic feature map of the same scale are subjected to channel stitching and weighted fusion to generate a fused feature map; The fused feature map is subjected to fused feature detection to obtain the target detection result, which includes the target location, target category and confidence level of the target.
7. A target detection device, characterized in that, include: The acquisition module is used to: acquire an infrared image sequence of a target area, wherein the infrared image sequence is used to provide spatiotemporal motion information of the target within the target area; The feature extraction module is used to: perform dynamic visual feature extraction on each frame of infrared image to generate a corresponding motion indicator map, wherein the motion indicator map is used to mark the potential motion region of the target; The detection module is used to obtain target detection results based on the target appearance feature map extracted from the infrared image and the target dynamic feature map extracted from the motion indicator map.
8. An electronic device comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the target detection method according to any one of claims 1 to 6.
9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the target detection method as described in any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the target detection method according to any one of claims 1 to 6.