Micro flying target detection method based on trajectory dynamic characteristics and appearance fusion

Abstract

This invention discloses a method for detecting miniature flying targets based on the fusion of trajectory dynamic features and appearance, belonging to the field of computer vision and video target detection technology. The method includes: employing an improved single-stage detection network integrating backbone feature extraction, multi-scale enhancement, attention modulation, and cross-scale fusion to extract the appearance features of the miniature flying target frame by frame, completing localization and preliminary classification, and outputting results including spatial location and appearance confidence scores. Based on these results, a continuous motion trajectory of the target is constructed by combining spatial location and motion consistency constraints. Multi-dimensional temporal features are then extracted and segmented to form a sequence suitable for modeling. Multi-scale temporal modeling, bidirectional dependency capture, and global correlation modeling are used to distinguish between real targets and interference. Two-dimensional confidence scores are statistically analyzed and dynamically weighted and fused. The category is determined by threshold comparison, and the final result is output to achieve accurate identification. This invention can improve the accuracy of miniature target detection in complex backgrounds, suppress interference, and reduce false positives.

Description

Technical Field

[0001] This invention relates to the field of computer vision and video target detection technology, and in particular to a method and apparatus for detecting miniature flying targets based on the fusion of trajectory dynamic features and appearance. Background Technology

[0002] With the rapid popularization of drones in the civilian sector, drone activities in low-altitude airspace have raised the bar for public safety and airspace management. In practical applications, micro-drones often operate against complex backgrounds such as buildings, mountains, and forests, and their pixel count is extremely low at long distances, resulting in poor salience and easy interference from the background. Furthermore, micro-drones are highly similar to birds in scale, appearance, and local motion characteristics, making the stable detection and accurate differentiation of these two types of flying targets in complex scenarios a significant technical challenge.

[0003] To address the aforementioned issues, existing technologies primarily employ visual target detection methods, relying on deep learning models to locate and classify flying targets in images or video sequences. Some solutions enhance the perception of small targets through multi-scale feature extraction, super-resolution reconstruction, or the introduction of short-term motion information; others attempt to utilize temporal information to model target motion, improving detection stability in dynamic scenes. While these methods have improved the detection performance of small flying targets to some extent and have been applied to some low-altitude monitoring and security systems, significant shortcomings still exist.

[0004] On the one hand, most methods rely on static appearance features or short-term time-series information, making it difficult to reliably represent small targets in complex backgrounds and under strong interference conditions. They also lack the ability to distinguish between similar targets such as drones and birds. On the other hand, existing technologies often limit the use of target motion information to local or short-term levels, lacking systematic modeling of the target's long-term trajectory dynamic features. Furthermore, the fusion of appearance and motion information is relatively crude, making it difficult to fully leverage their complementary advantages. Consequently, the detection accuracy and robustness in complex scenarios urgently need improvement. Therefore, there is an urgent need for a micro-flying target recognition scheme that can deeply integrate target appearance and long-term motion features, and possess both high-precision detection and similar target differentiation capabilities. Summary of the Invention

[0005] The main objective of this invention is to provide a method for detecting miniature flying targets based on the fusion of trajectory dynamic features and appearance.

[0006] Another objective of this invention is to propose a miniature flying target detection device based on the fusion of trajectory dynamic features and appearance.

[0007] The third objective of this invention is to provide an electronic device.

[0008] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0009] To achieve the above objectives, a first aspect of the present invention proposes a method for detecting miniature flying targets based on the fusion of trajectory dynamic features and appearance, comprising:

[0010] S1 employs an improved single-stage detection network that integrates multi-scale detection enhancement, attention modulation, and cross-scale fusion. It processes video sequence images frame by frame, extracts the spatial appearance features of micro-flying targets, completes target localization and preliminary classification, and outputs single-frame detection results containing spatial location bounding boxes and appearance classification confidence scores. S2, based on the single-frame detection results, combined with spatial location and motion consistency constraints, calculates the target association similarity between adjacent frames, establishes matching relationships to complete cross-frame instance association, performs target trajectory initialization, continuous update and update termination operations when there is no match, and constructs the target's continuous motion trajectory; S3, based on the continuous motion trajectory of the target, extract multi-dimensional temporal features that characterize the dynamic behavior of the target, segment the continuous motion trajectory to form a temporal feature sequence that adapts to motion pattern discrimination, and output trajectory-level temporal feature data; S4 sequentially performs multi-scale temporal modeling, bidirectional time dependency capture, and global time correlation modeling on trajectory-level temporal feature data. The modeling results are used to determine the target motion pattern, distinguish between real micro-flying targets and pseudo-motion interference, and output motion discrimination results including motion classification confidence. S5 calculates the confidence scores for appearance and motion classification to obtain the overall appearance and overall motion confidence scores. It dynamically adjusts the weights based on the scene and calculates the final classification probability by weighted fusion. The target category is determined by threshold comparison, and the detection results containing location and category information are output to achieve the identification of micro-flying targets.

[0011] Optionally, the cross-scale fusion further includes: Scale alignment is performed on the attention-modulated feature maps at each level using a scale transformation operator. Adjust the spatial resolution of features at different levels to keep the dimensions of features at each level consistent; A layer-by-layer fusion approach is adopted to integrate information from scale-aligned multi-scale features, fusing spatial detail information from low-level features with semantic information from high-level features to form a unified multi-scale feature representation. ,in, Indicates the number after attention enhancement Layer feature mapping, This represents the total number of layers in the feature mapping. It outputs a unified multi-scale feature representation, providing comprehensive feature support for target spatial localization and appearance classification confidence prediction, and improving the robustness of the single-stage detection network for detecting micro-flying targets of different scales.

[0012] Optionally, the initialization, continuous updating, and termination of cross-frame instance association and trajectory also include: The matching threshold is set based on the calculation results of the target association similarity between adjacent frames, and the effective matching pairs that meet the matching threshold are selected with the detection box overlap ratio or the detection box center point distance as the core metric. The single-frame detection results of the target instances in the effective matching pairs are assigned to the same instance identifier, and continuous trajectory segments of the target are accumulated in the time dimension to achieve continuous trajectory updates; For a single-frame detection result that is newly appearing in the current frame and has no historical target instance matching record, initialize a new target instance and its corresponding motion trajectory; For target instances that have no matching objects in several consecutive frames, their trajectory updates are terminated and the trajectory data is archived to ensure the integrity and temporal continuity of the constructed continuous motion trajectory of the target.

[0013] Optionally, the extraction of multidimensional temporal features characterizing the dynamic behavior of the target further includes: From the constructed continuous motion trajectory of the target, the center coordinates, width, and height geometric parameters of the target bounding box at each moment are extracted to construct the target position features and characterize the spatial position change pattern of the target. Based on the difference in geometric parameters between adjacent frames and the time interval between adjacent frames Calculate the instantaneous velocity vector of the target at each moment. Construct target velocity characteristics to reflect the target's spatial displacement trend; The scale change rate of the target at each time step is calculated by using the ratio of the difference in the area of ​​the target bounding box between the previous and next frames to the area of ​​the previous frame. To construct target scale features and characterize the relative motion relationship between the target and the imaging device; The target's position features, velocity features, and scale features are combined to form a multidimensional temporal feature vector for each time step. k is the target instance identifier.

[0014] Optionally, the multi-scale temporal modeling further includes: Construct multiple parallel one-dimensional temporal convolutional or equivalent temporal feature extraction structures, and configure different sizes of temporal receptive fields for each structure to adapt to the feature capture requirements of short-term fast motion and relatively stable motion, respectively. By synchronously extracting features from trajectory-level temporal feature sequences through various parallel structures, local motion feature representations corresponding to different time scales are obtained. To mine the motion pattern features of the target in different time dimensions, among which, Let X be the temporal feature extraction operator at the s-th time scale, where X is the trajectory-level temporal feature sequence and S is the total number of parallel structures. The extracted multi-scale local motion features are normalized to eliminate interference caused by differences in feature scales, and the fused multi-scale local temporal features are output, laying a data foundation for subsequent bidirectional temporal dependency modeling.

[0015] Optionally, the statistical appearance classification confidence score and motion classification confidence score further include: Confidence sequence of appearance classification for the same target instance across multiple frames Perform traversal statistics and use the average value calculation method. Eliminate single-frame detection errors to obtain a stable overall appearance confidence score, where m is the number of detection frames for the same target instance; Motion classification confidence sequence of trajectory segments for the same target instance Statistical analysis was conducted, and weights were assigned based on the length of trajectory segments. Weighted statistics or average statistics were used to calculate the overall motion confidence, taking into account the discriminative contribution of motion characteristics of each segment. The overall appearance confidence and overall motion confidence obtained from the statistics are validated, and abnormal confidence values ​​are removed to ensure the reliability of the confidence data and provide an accurate basis for subsequent weighted fusion.

[0016] Optionally, the comparison of the final classification probability with the preset classification threshold and the determination of the target category also include: A preset classification probability threshold is obtained by training on sample data of micro-flying target detection. The threshold is adapted to the classification boundary between micro-flying targets and pseudo-motion interference objects. according to The final classification probability is obtained by performing weighted fusion calculations. ,in, This is the appearance weighting coefficient. For motion weighting coefficients, and + =1; The final classification probability is compared with the preset classification threshold. If the final classification probability is higher than the threshold, it is determined to be a valid micro-flying target. If it is lower than the threshold, it is determined to be a pseudo-motion interference object. Among multiple categories related to micro-flying targets, the category with the highest final classification probability is selected as the final category of the target. Combined with the target's spatial location information, the final detection result is output, thereby achieving accurate identification of micro-flying targets.

[0017] To achieve the above objectives, a second aspect of the present invention provides a miniature flying target detection device based on the fusion of trajectory dynamic features and appearance, comprising: The appearance feature detection module is used to process video sequence images frame by frame using an improved single-stage detection network that integrates multi-scale detection enhancement, attention modulation and cross-scale fusion, extract the spatial appearance features of the micro-flying target, complete the target localization and preliminary classification, and output the single-frame detection results containing the spatial location bounding box and appearance classification confidence. The cross-frame trajectory construction module is used to calculate the similarity of targets in adjacent frames based on the detection results of a single frame, combined with spatial location and motion consistency constraints, and establish a matching relationship to complete the cross-frame instance association. It performs the initialization, continuous update and update termination operations when there is no match for the target trajectory, and constructs the continuous motion trajectory of the target. The temporal feature extraction module is used to extract multi-dimensional temporal features representing the dynamic behavior of the target based on the continuous motion trajectory of the target, segment the continuous motion trajectory to form a temporal feature sequence adapted to motion mode discrimination, and output trajectory-level temporal feature data. The motion pattern discrimination module is used to sequentially perform multi-scale temporal modeling, bidirectional time dependency capture and global time correlation modeling on trajectory-level temporal feature data. It judges the target motion pattern through the modeling results, distinguishes between real micro-flying targets and pseudo-motion interference, and outputs motion discrimination results including motion classification confidence. The fusion decision recognition module is used to calculate the confidence scores for appearance classification and motion classification, obtain the confidence scores for overall appearance and overall motion, dynamically adjust the weights according to the scene and calculate the final classification probability by weighted fusion, determine the target category by threshold comparison, and output the detection results containing location and category information to realize the recognition of micro flying targets.

[0018] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0019] To achieve the above objectives, a third aspect of this application provides an electronic device, including a processor and a memory; wherein the processor runs a program corresponding to the executable program code stored in the memory, in order to implement a micro-flying target detection method based on trajectory dynamic features and appearance fusion as described in the first aspect embodiment.

[0020] To achieve the above objectives, the fourth aspect of this application proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a micro-flying target detection method based on trajectory dynamic features and appearance fusion as described in the first aspect embodiment.

[0021] The embodiments of the present invention have the following beneficial effects: 1. To address the pain points of small scale and weak salience of micro-flying targets in complex backgrounds, this paper enhances the ability to distinguish micro-targets from the background through multi-scale detection enhancement, attention modulation and cross-scale fusion mechanisms, thereby significantly improving the target positioning accuracy; 2. Relying on cross-frame trajectory association and hybrid temporal modeling structure, the long-range motion law of the target is fully explored, which effectively makes up for the deficiency of single-frame appearance detection in distinguishing similar targets such as UAVs and birds, and significantly reduces the misjudgment rate of similar targets; 3. An adaptive weighted fusion strategy is adopted, which can dynamically adjust the weight ratio of appearance and motion features according to the scene, adapt to the detection requirements under different imaging conditions, and improve the robustness and versatility of the method in complex and highly interference scenarios. 4. The overall technical solution combines high-precision detection with efficient differentiation capabilities, and can be widely applied in fields such as low-altitude security and airspace management. It provides reliable technical support for the intelligent monitoring of micro-flying targets and has high practical value and promising prospects for promotion. Attached Figure Description

[0022] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a method for detecting miniature flying targets based on the fusion of trajectory dynamic features and appearance, provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the overall framework of a micro-flying target detection method based on trajectory dynamic features and appearance fusion, provided in an embodiment of the present invention. Figure 3 This is a functional structure diagram of the appearance inspection model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the motion detection model structure based on trajectory dynamic features provided in an embodiment of the present invention; Figure 5 This is a structural diagram of a miniature flying target detection device based on the fusion of trajectory dynamic features and appearance, provided in an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] The following description, with reference to the accompanying drawings, describes a method and apparatus for detecting miniature flying targets based on the fusion of trajectory dynamic features and appearance, according to an embodiment of the present invention.

[0026] Example 1 This invention provides a method for detecting miniature flying targets based on the fusion of trajectory dynamic features and appearance. Figure 1 This is a flowchart illustrating a method for detecting miniature flying targets based on the fusion of trajectory dynamic features and appearance, as provided in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the overall framework of a micro-flying target detection method based on the fusion of trajectory dynamic features and appearance, as provided in an embodiment of the present invention. Figure 1 , Figure 2 As shown, the method includes the following steps: Step S1: An improved single-stage detection network integrating multi-scale detection enhancement, attention modulation, and cross-scale fusion is used to process video sequence images frame by frame, extract the spatial appearance features of the micro-flying target, complete the target localization and preliminary classification, and output the single-frame detection results containing the spatial location bounding box and appearance classification confidence.

[0027] In this embodiment, for the input video frame sequence, single-frame target detection processing based on appearance features is first performed on each frame. This is the fundamental step in the entire detection process, aiming to accurately capture the spatial appearance information of micro-flying targets from a single-frame perspective. Let the input single-frame image be represented as... H and W correspond to the height and width parameters of the image, respectively, and C represents the number of channels in the image. This single-frame image will be directly input into the appearance detection model, which will complete the three core tasks of extracting the spatial appearance features of the micro-flying target, target localization, and preliminary classification, providing basic data support for subsequent cross-frame trajectory construction and multi-dimensional feature fusion.

[0028] The appearance detection model in this embodiment adopts an improved single-stage target detection network structure, such as... Figure 3As shown, compared to traditional single-stage detection networks, this model improves its adaptability to micro-targets through model-level optimization. It consists of four core components: a feature extraction backbone network, a multi-scale feature enhancement model, an attention enhancement model, and a cross-scale feature fusion model. Each model performs its specific function and works in concert to ensure the integrity of feature extraction while enhancing the feature representation capability of micro-targets. Specifically, the input image I is first fed into the feature extraction backbone network, and after layers of convolution and pooling operations, a set of feature maps at different spatial resolutions is generated. , ,… }; among them, low-level features (such as , Focusing on spatial details in images, it can accurately capture subtle features such as the local contours and textures of miniature flying targets, providing precise information for target localization. Meanwhile, higher-level features (such as...) , After multiple feature aggregations, it possesses stronger semantic information, enabling a preliminary judgment of the target category and laying the foundation for subsequent classification tasks.

[0029] To further enhance the detection capability of micro-targets in high-resolution images and address the issues of small scale, low feature ratio, and susceptibility to background information overload in micro-targets, this application introduces a multi-scale detection enhancement mechanism during feature extraction. This mechanism adds a fine-grained feature layer to the high-resolution branch of the network, reducing the receptive field of the detection network and enabling it to focus more precisely on the region containing the micro-target, preventing the target features from being diluted by a large area of ​​background information. Specifically, the shallow feature maps output by the backbone network are first downsampled to adjust the spatial resolution of the features. Then, the number and dimensionality of feature channels are optimized through channel dimension adjustment, ultimately constructing a high-resolution feature representation suitable for micro-target detection. This significantly improves the separability of micro-flying targets in the feature space, providing a high-quality feature foundation for subsequent accurate classification and localization.

[0030] After obtaining the multi-scale feature maps, to further improve the effectiveness of the features and filter background interference, this embodiment of the application further performs attention enhancement processing on the multi-scale feature maps. The attention enhancement model strengthens the network's ability to focus on the target region and its context, thereby enhancing effective features and suppressing interfering features. Specifically, for any feature map F in the set, the attention enhancement model calculates the corresponding attention weights in both the channel and spatial dimensions: the channel dimension weights are used to filter feature channels effective for identifying micro-flying targets and eliminate invalid information from redundant channels; the spatial dimension weights are used to accurately locate the target region and increase the weight ratio of the target region features. Subsequently, the original feature maps are weighted and modulated through element-wise multiplication to obtain the enhanced feature representation. Its mathematical expression is: ,in This is a spatial attention function used to generate a spatial dimension weight matrix. This is a channel attention function used to generate channel-dimensional weight vectors. This represents element-wise multiplication. Through this dual attention mechanism, the feature regions associated with the micro-flying target are significantly enhanced, while background interference features (such as buildings, trees, clouds, etc.) are effectively suppressed, greatly improving the signal-to-noise ratio of the features and providing high-purity features for subsequent cross-scale fusion.

[0031] After attention enhancement processing, a cross-scale feature fusion model is needed to integrate information from feature maps at different levels and scales, achieving complementary fusion of low-level detail features and high-level semantic features. In this embodiment, the core operation of the cross-scale feature fusion model consists of two steps: the first step is scale alignment, achieved through a scale transformation operator. Feature mappings at each level after attention enhancement ( , … The first step involves spatial resolution adjustment to ensure consistent dimensionality in height, width, and number of channels across different feature maps, eliminating fusion barriers caused by scale differences. The second step is layer-by-layer fusion, employing a combination of top-down and bottom-up approaches. This involves information exchange and aggregation of the aligned multi-scale features, ensuring that the fused features retain both the spatial detail advantages of lower-level features and the semantic information advantages of higher-level features. The mathematical expression for this process is: ,in This represents the attention-enhanced feature map of the i-th layer. It encompasses both scale transformation and feature alignment operations. Through this fusion method, the appearance detection model can balance target localization accuracy and category judgment accuracy, significantly improving the robustness of detecting micro-flying targets of different scales and adapting to the needs of micro-target detection under different distances and imaging conditions.

[0032] Based on the unified multi-scale feature representation after fusion The appearance detection model predicts miniature flying targets in a single frame image, ultimately outputting two core results: first, the target's spatial location bounding box, which accurately locates the target's position in the image using the bounding box's coordinate parameters (such as the coordinates of the top-left and bottom-right corners); and second, the corresponding appearance classification confidence score, which quantifies the probability estimate of the target belonging to different flight categories (such as the probability of belonging to a drone, bird, or other interfering target). These output results will serve as the core foundational input for subsequent multi-frame fusion, cross-frame trajectory association, and motion feature modeling, providing reliable data support for the subsequent processes of the entire detection method.

[0033] Step S2: Based on the single-frame detection results, combined with spatial location and motion consistency constraints, calculate the target association similarity between adjacent frames, establish a matching relationship to complete cross-frame instance association, perform target trajectory initialization, continuous update and update termination operations when there is no match, and construct the target's continuous motion trajectory.

[0034] In this embodiment, after completing the appearance detection of all single-frame images in the video sequence, the detection results of a single frame can only reflect the spatial position and appearance features of the target at a certain moment, and cannot reflect the temporal motion pattern of the target. Therefore, it is necessary to further perform cross-frame association operations on the detection results in different video frames. The core purpose of this operation is to associate and match the detection results of the same target in different frames to form continuous trajectory information of the target in the temporal dimension. This provides core data support for the temporal feature extraction and motion pattern modeling of the subsequent motion detection model, filling the gap in temporal information capture in single-frame detection.

[0035] Let the set of detection results output by the appearance detection model in frame t be represented as: ,in Let be the total number of targets detected in frame t. This represents the i-th detected target in frame t, where each... Each frame contains two core pieces of information: first, the target's spatial location bounding box information, which accurately locates the target's position in the current frame using the bounding box's coordinate parameters (such as the coordinates of the top-left corner, bottom-right corner, or center point coordinates plus width and height); and second, the corresponding appearance classification confidence score, used to quantify the probability that the target belongs to a certain flight category. These two types of information together constitute the core basis for cross-frame association, ensuring that the association process can determine the continuity of the target based on its spatial location, while also using appearance features to assist in verifying the accuracy of the matching.

[0036] To achieve accurate instance-level association of targets between adjacent video frames (frame t and frame t+1) and avoid mismatches between different targets, this application achieves efficient filtering by constructing matching relationships and calculating association similarity. Specifically, for the detected targets in frame t... Detected target in frame t+1 This application combines spatial location constraints and motion consistency constraints to calculate their correlation similarity. This similarity value is used to quantitatively measure whether two targets belong to the same instance; a higher value indicates a greater probability of matching. The correlation similarity can include various measurement factors. Considering the motion characteristics of micro-flying targets, it must at least include measures based on spatial overlap or positional distance to ensure that the correlation results closely match the actual motion trajectory of the targets.

[0037] In one feasible implementation of this application, the spatial consistency of the target bounding box is used as the core metric, and the degree of overlap between detection boxes in adjacent frames is calculated. The distance to the center point or the center point can be used to construct a spatial matching cost function, which inversely reflects the correlation similarity. For example, a detection box in frame t can be defined. With the detection box in frame t+1 Spatial matching cost between Its mathematical expression is: ,in This represents the overlap ratio between bounding boxes, calculated as the ratio of the intersection area to the union area of ​​two bounding boxes. Represents the detection box in frame t. The center point coordinates. When two detection boxes belong to the same target. The value approaches 1, corresponding to the matching cost. Approaching 0; when belonging to different objectives, The value approaches 0, and the matching cost is... Approaching 1. This spatial constraint allows for the rapid elimination of candidate pairs whose matching costs exceed a preset range, significantly reducing invalid matching operations and improving the efficiency and accuracy of cross-frame association.

[0038] After obtaining the matching costs of all detected targets between adjacent frames, cross-frame instance correspondences need to be established through association allocation operations to form continuous trajectories. In this embodiment, the association allocation process follows the principle of "effective matching retention, new target initialization, and disconnected target termination" to ensure the integrity and effectiveness of the trajectory. First, a reasonable matching threshold is set based on the calculated association similarity, and effective matching pairs with matching costs lower than the threshold (or similarities higher than the threshold) are selected. The detected targets corresponding to these matching pairs are assigned to the same target instance identifier, and the detection results of the instance in different frames are accumulated in the time dimension to form continuous trajectory segments.

[0039] Secondly, for newly added detection results in the current frame (frame t+1) that cannot establish a valid match with any target in historical frames (frame t and previous frames), they are determined to be newly appearing target instances. A new instance identifier and corresponding trajectory are immediately initialized, and this detection result is used as the starting node of the new trajectory. Finally, for a given target instance, if it fails to establish a valid match with any detected target in subsequent frames (e.g., 3 or 5 frames, which can be adjusted according to the actual application scenario), it is determined that the target has left the monitoring range or disappeared. Its trajectory update is terminated and archived to avoid invalid trajectories consuming computational resources and to prevent interference with subsequent trajectory segmentation and motion feature modeling.

[0040] Through the complete cross-frame association process described above, this application organizes all detection results belonging to the same target instance in a video sequence in chronological order, ultimately constructing the continuous motion trajectory of the target. The target trajectory can be represented as... ,in This represents the trajectory of the k-th target instance. This indicates the consecutive occurrences of the target in the video sequence, each... The target should be in the The detection information in the frame. This trajectory fully records the target's position changes and appearance confidence changes over time, serving as the core input for subsequent trajectory segmentation, multi-dimensional temporal feature extraction, and motion pattern modeling, providing data support for fully exploring the target's motion patterns.

[0041] Step S3: Based on the continuous motion trajectory of the target, extract multi-dimensional temporal features that characterize the dynamic behavior of the target, segment the continuous motion trajectory to form a temporal feature sequence that adapts to motion pattern discrimination, and output trajectory-level temporal feature data.

[0042] In this embodiment, after obtaining the continuous motion trajectory of the target, the trajectory only records the basic detection information of the target at different times and has not yet formed feature data that can be directly used for motion pattern modeling. Therefore, it is necessary to perform professional temporal feature extraction on the trajectory data. The core objective of temporal feature extraction in this embodiment is to mine multi-dimensional time-series features that can characterize the dynamic behavior of the target from the continuous trajectory data, quantify the motion state of the target, and provide high-quality feature input for the subsequent target detection and category differentiation work carried out by the motion detection model based on the dynamic features of the trajectory. This extraction process is based entirely on the spatial position change of the target in continuous video frames and the various motion attributes derived from the position change, which can comprehensively depict the real motion state of the target in the time dimension from multiple dimensions.

[0043] Specifically, for each moment in the trajectory This application embodiment extracts and constructs the multi-dimensional motion features of the target sequentially according to the logical hierarchy of feature extraction, progressively achieving a refined representation of the target's motion state. First, the core geometric parameters of the target bounding box are extracted, specifically including the center coordinates of the bounding box. , and the width of the bounding box. and height These geometric parameters directly reflect the target's spatial position. Constructing the target's positional features based on these parameters allows for accurate capture of the target's positional changes in two-dimensional image space, reflecting the target's spatial trajectory at different times. Building upon these positional features, the target's velocity characteristics are further calculated. These characteristics are derived from the ratio of the change in geometric parameters between adjacent time points to the time interval, quantitatively characterizing the target's speed and direction of spatial displacement. For example, the instantaneous velocity vector of the target between adjacent frames can be specifically represented as... ,in, The fixed time interval between two adjacent frames is used to clearly reflect the real-time motion trend of the target, effectively distinguish different motion states such as uniform speed, acceleration, and deceleration, and explore the dynamic change characteristics of the target motion.

[0044] Furthermore, to further enhance the ability to describe target motion patterns in different application scenarios and compensate for the shortcomings of position and velocity features in representing the relative motion relationship of targets, this application embodiment also specifically introduces temporal features related to target scale changes, using the bounding box size change rate as the core characterization index, to accurately reflect the relative motion relationship between the target and the imaging device, such as the scale changes when the target approaches or moves away from the imaging device. Accordingly, scale change features can be constructed: By combining position features, velocity features, and scale change features in a multi-dimensional manner, the motion behavior of a target within a short time range can be characterized comprehensively and meticulously from three dimensions: spatial position, motion rate, and relative motion. This provides rich and effective feature data for subsequent motion detection models to perform motion pattern discrimination.

[0045] After extracting the multidimensional motion features at each moment in the trajectory, this embodiment integrates and constructs the features at each moment to form a temporal feature sequence that reflects the continuous motion law of the target. Specifically, each moment... The extracted multidimensional motion features, such as position, velocity, and scale changes, are combined in an ordered manner to generate a multidimensional motion feature vector corresponding to that moment, thereby constructing a complete temporal feature sequence corresponding to the target trajectory. ,in, Indicates the first Each target instance at time... The multidimensional motion feature vector contains at least three core features: the target's position, velocity, and scale change. The constructed temporal feature sequence can fully describe the dynamic change pattern of the target in the time dimension, transforming the target's continuous motion process into feature data that can be recognized and processed by the model.

[0046] Meanwhile, considering the inherent differences in the length of the motion trajectories of different targets in actual monitoring scenarios, and the local differences in the motion patterns of the same target in the time dimension, a single complete temporal feature sequence is difficult to adapt to the model's requirements for local motion pattern extraction and long-term time dependency modeling. Therefore, in the temporal feature extraction stage, this embodiment of the application simultaneously segments the complete trajectory of the target. According to the actual detection requirements, the complete trajectory is divided into several trajectory segments of fixed or adaptively adjusted length. Each trajectory segment corresponds to a continuous temporal feature subsequence, and each subsequence independently serves as the modeling basis for the subsequent motion detection model, respectively supporting the local motion pattern extraction and long-term time dependency modeling. By combining the above-mentioned temporal feature extraction with trajectory segmentation, the model's ability to express complex motion patterns can be significantly improved while strictly ensuring the continuity of target motion time. This effectively adapts to the modeling requirements of targets in different motion states. The trajectory-level temporal feature sequence obtained through the above series of processes will be directly used as input data for the motion detection model, providing an accurate and reliable data foundation for target classification and discrimination based on trajectory dynamic features.

[0047] Step S4 involves sequentially performing multi-scale temporal modeling, bidirectional time dependency capture, and global time correlation modeling on the trajectory-level temporal feature data. The modeling results are used to determine the target motion pattern, distinguish between real micro-flying targets and pseudo-motion interference, and output motion discrimination results including motion classification confidence.

[0048] In this embodiment of the application, after extracting the temporal features of the target trajectory, a motion detection model based on the dynamic features of the trajectory is introduced, such as... Figure 4 As shown, the model takes the temporal feature sequence corresponding to the target trajectory as input, and achieves effective representation and classification of target motion behavior through multi-scale temporal modeling and contextual information fusion, thereby distinguishing real moving targets from background interference or pseudo-moving targets and making up for the shortcomings of single-frame appearance detection.

[0049] To fully explore the motion pattern features of the target at different time scales, this application first performs multi-scale local temporal modeling on the temporal feature sequence. In one embodiment of this application, multiple parallel one-dimensional temporal convolutional or equivalent temporal feature extraction structures are used, configured with different sizes of temporal receptive fields to adapt to the feature capture requirements of short-term rapid motion and relatively stable motion. The temporal feature sequence is synchronously extracted through each parallel structure to obtain local motion feature representations corresponding to different time scales. This process can be represented as follows: ,in, This represents the temporal feature extraction operator at the s-th time scale. To capture the local temporal features at the corresponding scale, the parallel extraction of multi-scale features can enhance the model's adaptability to targets with different motion rates and comprehensively capture the local motion changes of the target.

[0050] After obtaining multi-scale local temporal features, this application further fuses these features to eliminate information redundancy between features of different scales. Simultaneously, a bidirectional temporal dependency modeling mechanism is introduced to characterize the temporal correlation of target motion. In this embodiment, a bidirectional recurrent neural network structure is used to model the fused temporal features, thereby utilizing motion information from both past and future moments. This process can be represented as follows: Where F represents the fused local temporal feature sequence and H is the hidden state sequence containing bidirectional time dependency information. This method can effectively alleviate the problem of insufficient understanding of complex motion trajectories by single-direction modeling and improve the ability to predict the motion trend of the target.

[0051] Furthermore, to capture the global temporal relationships of the target motion over a longer time span, this application introduces a temporal feature modeling model based on a self-attention mechanism, building upon bidirectional temporal modeling, to model key moments in the trajectory and their interrelationships. Specifically, by calculating the correlation weights between features at different times, important motion changes are enhanced, thereby obtaining feature representations containing global temporal dependencies. This process can be abstractly represented as follows: Where Z represents the trajectory-level feature representation after fusing global time-dependent information, the self-attention modeling method can effectively improve the model's ability to discriminate complex, non-uniform motion patterns and accurately capture the overall motion pattern of the target.

[0052] After completing multi-scale local modeling, bidirectional temporal dependency capture, and global correlation modeling, the obtained trajectory-level features are input into the classification and discrimination unit to complete the final determination of the target's motion attributes. In this embodiment, the classification and discrimination unit includes at least one fully connected layer and normalization processing. First, the trajectory-level features Z, which incorporate global temporal dependencies, are input into the fully connected layer. Linear transformation is used to compress the dimensions and reconstruct the information of the features, extracting the core discrimination features. Then, normalization processing is performed on the output features of the fully connected layer to avoid model training instability caused by excessively large feature values, thereby improving feature discrimination capability. Finally, a nonlinear activation function is applied... Calculate the confidence score of whether the target belongs to a real moving target or background interference. The classification output can be represented as follows: Where z is the trajectory-level feature vector, and W and b are learnable parameters. It is a non-linear activation function. Based on the confidence score, it is compared with a preset threshold to complete the discrimination of the target motion attributes and output the motion discrimination result, providing a reliable basis for subsequent fusion decisions.

[0053] Step S5: Calculate the confidence scores for appearance classification and motion classification to obtain the confidence scores for overall appearance and overall motion. Dynamically adjust the weights according to the scene and calculate the final classification probability by weighted fusion. Determine the target category by threshold comparison and output the detection results containing location and category information to achieve the identification of micro flying targets.

[0054] In this embodiment, the appearance detection model extracts and preliminarily distinguishes the target's appearance features in a single frame, while the motion detection model models the dynamic features of the target's trajectory and determines its motion attributes in the time dimension. Both models provide discrimination criteria for micro-flying target detection from different dimensions. After the independent discrimination by the two models is completed, this application specifically introduces a fusion decision model as the core decision-making step of the entire detection method. The core design logic of this model is to fully integrate the discrimination information of the target in both appearance and motion dimensions, systematically processing the single-frame appearance detection results and trajectory motion detection results. This effectively compensates for the shortcomings of a single model in discriminating against complex backgrounds, and significantly improves the ability to distinguish between targets with highly similar appearances and local motion characteristics, such as UAVs and flying birds. Ultimately, it outputs micro-flying target detection results that combine accuracy and reliability.

[0055] Specifically, the workflow of the fusion decision model consists of three core steps: confidence statistics, weighted fusion, and threshold comparison and result output. Each step is progressive, achieving a precise transformation from raw discrimination information to the final detection result. First, confidence statistics are performed on the appearance and motion detection results. The purpose of this step is to integrate and optimize the multiple sets of discrete discrimination confidence scores output by the two models to obtain an overall confidence score that can characterize the overall discrimination result of the model, providing a stable and reliable numerical basis for subsequent fusion calculations. Let the classification confidence score sequence output by the appearance detection model for the k-th target instance be... ,in For this goal in the The classification confidence score output by the appearance detection model in each frame reflects the fluctuation of the discrimination results of the target's appearance features in different frames. To eliminate discrimination errors caused by background interference, low target pixel ratio, and other factors in single-frame detection and obtain stable overall appearance classification results, this application performs a full traversal statistical analysis of the confidence score sequence and uses the arithmetic mean to obtain the overall appearance confidence score of the target. The calculation formula is as follows: This method can effectively smooth the fluctuations in single-frame discrimination results, and greatly improve the stability and reference value of appearance confidence.

[0056] Similarly, let the classification confidence sequence of each trajectory segment output by the motion detection model for the target be... ,in The classification confidence score of the j-th trajectory segment of the target, as determined by the motion detection model, is calculated using an average or weighted statistical method. Considering the varying lengths and richness of motion feature information contained in different trajectory segments, this application performs averaging or weighted statistical analysis on the sequence. During weighted statistical analysis, weights are assigned based on the actual length of the trajectory segments; longer segments with richer motion features receive higher weights. This approach balances the actual contribution of each trajectory segment to motion feature discrimination, ultimately yielding the overall motion confidence score of the target. The calculation formula is as follows: Furthermore, after obtaining the overall appearance confidence level and the overall motion confidence level, this application will also perform validity verification on the two types of confidence level data. By using preset outlier judgment rules, abnormal confidence level values ​​caused by factors such as model detection errors and abnormal trajectory segments will be eliminated to ensure that the confidence level data participating in the final fusion calculation is true and reliable, and to provide accurate numerical basis for subsequent weighted fusion operations.

[0057] After obtaining the optimized overall confidence scores for both appearance and motion dimensions, the fusion decision model employs an adaptive weighted fusion strategy to calculate the fusion of the two confidence scores, yielding the final classification probability of the target and achieving complementary advantages between the two dimensions of discriminative information. The fusion calculation formula is as follows: ,in The weighting coefficient for appearance confidence. Let be the weighting coefficients for the motion confidence, and let both satisfy . The normalization constraint ensures the rationality of the fusion calculation. In this application, the two sets of weight coefficients are not fixed values, but can be dynamically adjusted according to the characteristics of the actual detection scene to achieve adaptive discrimination under different scene conditions. The specific adjustment rule is as follows: in detection scenes where the target image is clear, the appearance information is complete, and the target morphological features are obvious, the discrimination advantage of appearance features is significant; in this case, the weight coefficient should be appropriately increased. The value of allows appearance confidence to have a higher weight in the fusion result, fully leveraging the discriminative role of appearance features in target category identification. In detection scenarios with dense targets, complex backgrounds, indistinct target appearance features, or stable and highly recognizable target motion features, the discriminative reliability of motion features is higher. In these cases, appropriately increasing ... The value of the weight is determined by the dynamic features of the motion trajectory to achieve accurate differentiation of similar targets. Through this dynamic weight adjustment method, the fusion decision model can adapt to detection scenarios of varying complexity, significantly enhancing the robustness and adaptability of the fusion judgment.

[0058] Finally, the fusion decision model performs threshold comparison and outputs the final detection results, achieving accurate identification of micro-flying targets. The preset classification probability threshold in this application is obtained after model training based on a large amount of sample data of micro-flying targets and background interference objects. This threshold accurately matches the classification boundary between micro-flying targets and interference objects, effectively distinguishing between valid targets and invalid interference. The fusion decision model calculates the final target classification probability. The target is compared with a preset threshold. If the classification probability is higher than the threshold, the target is determined to be a valid micro-flying target; if the classification probability is lower than the threshold, the target is determined to be background interference and is directly eliminated, thus achieving preliminary screening of valid targets and interference. After determining the valid target, for detection scenarios with multiple target categories, the fusion decision model selects the category with the highest classification confidence from all valid target discrimination results as the final category determination result for the target. At the same time, the final category of the target is associated with the corresponding spatial location information to form a complete detection result. Through the deep integration of the above appearance and motion dimension information, this application effectively combines the spatial appearance features and long-term trajectory dynamic features of micro-flying targets. It not only leverages the role of appearance features in the initial target discrimination but also utilizes the advantages of motion features in distinguishing similar targets. Under the conditions of complex backgrounds and micro-target detection, it can significantly improve the detection accuracy and recognition reliability of similar flying targets such as UAVs and birds, while effectively suppressing misjudgment problems caused by factors such as appearance similarity and motion interference, ultimately achieving accurate identification of micro-flying targets. The detection method proposed in this application has excellent robustness and practical application value. It can be widely deployed in various scenarios that require monitoring of micro-flying targets, such as low-altitude security, airspace management, park monitoring, and border control, providing reliable technical support for intelligent monitoring in related fields.

[0059] Example 2 This invention provides a miniature flying target detection device based on the fusion of trajectory dynamic features and appearance. Figure 5 This is a schematic flowchart illustrating a miniature flying target detection device based on the fusion of trajectory dynamic features and appearance, provided in an embodiment of the present invention. Figure 5 As shown, the device includes: The appearance feature detection module 100 is used to process video sequence images frame by frame using an improved single-stage detection network that integrates multi-scale detection enhancement, attention modulation and cross-scale fusion, extract the spatial appearance features of the micro-flying target, complete the target localization and preliminary classification, and output the single-frame detection result containing the spatial location bounding box and appearance classification confidence. The cross-frame trajectory construction module 200 is used to calculate the similarity of target association between adjacent frames based on the single-frame detection results and combined with spatial position and motion consistency constraints, thereby establishing a matching relationship to complete the cross-frame instance association, performing target trajectory initialization, continuous update and update termination operations when there is no match, and constructing the continuous motion trajectory of the target. The temporal feature extraction module 300 is used to extract multi-dimensional temporal features representing the dynamic behavior of the target based on the continuous motion trajectory of the target, segment the continuous motion trajectory to form a temporal feature sequence adapted to motion mode discrimination, and output trajectory-level temporal feature data. The motion pattern discrimination module 400 is used to sequentially perform multi-scale temporal modeling, bidirectional time dependency capture and global time correlation modeling on trajectory-level temporal feature data. It judges the target motion pattern through the modeling results, distinguishes between real micro-flying targets and pseudo-motion interference, and outputs motion discrimination results including motion classification confidence. The fusion decision recognition module 500 is used to calculate the confidence scores for appearance classification and motion classification, obtain the confidence scores for overall appearance and overall motion, dynamically adjust the weights according to the scene and calculate the final classification probability by weighted fusion, determine the target category by threshold comparison, and output the detection results containing location and category information to realize the recognition of micro flying targets.

[0060] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0061] Example 3 To implement the methods of the above embodiments, the present invention also provides an electronic device, which includes a memory and a processor; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the various steps of the methods described above.

[0062] Example 4 To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0064] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0065] Furthermore, 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. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for detecting miniature flying targets based on the fusion of trajectory dynamic features and appearance, characterized in that, include: S1 employs an improved single-stage detection network that integrates multi-scale detection enhancement, attention modulation, and cross-scale fusion. It processes video sequence images frame by frame, extracts the spatial appearance features of micro-flying targets, completes target localization and preliminary classification, and outputs single-frame detection results containing spatial location bounding boxes and appearance classification confidence scores. S2, based on the single-frame detection results, combined with spatial location and motion consistency constraints, calculates the target association similarity between adjacent frames, establishes matching relationships to complete cross-frame instance association, performs target trajectory initialization, continuous update and update termination operations when there is no match, and constructs the target's continuous motion trajectory; S3, based on the continuous motion trajectory of the target, extract multi-dimensional temporal features that characterize the dynamic behavior of the target, segment the continuous motion trajectory to form a temporal feature sequence that adapts to motion pattern discrimination, and output trajectory-level temporal feature data; S4 sequentially performs multi-scale temporal modeling, bidirectional time dependency capture, and global time correlation modeling on trajectory-level temporal feature data. The modeling results are used to determine the target motion pattern, distinguish between real micro-flying targets and pseudo-motion interference, and output motion discrimination results including motion classification confidence. S5 calculates the confidence scores for appearance and motion classification to obtain the overall appearance and overall motion confidence scores. It dynamically adjusts the weights based on the scene and calculates the final classification probability by weighted fusion. The target category is determined by threshold comparison, and the detection results containing location and category information are output to achieve the identification of micro-flying targets.

2. The method according to claim 1, characterized in that, The cross-scale fusion also includes: Scale alignment is performed on the attention-modulated feature maps at each level using a scale transformation operator. Adjust the spatial resolution of features at different levels to keep the dimensions of features at each level consistent; A layer-by-layer fusion approach is adopted to integrate information from scale-aligned multi-scale features, fusing spatial detail information from low-level features with semantic information from high-level features to form a unified multi-scale feature representation. ,in, Indicates the number after attention enhancement Layer feature mapping, This represents the total number of layers in the feature mapping. It outputs a unified multi-scale feature representation, providing comprehensive feature support for target spatial localization and appearance classification confidence prediction, and improving the robustness of the single-stage detection network for detecting micro-flying targets of different scales.

3. The method according to claim 2, characterized in that, The initialization, continuous updating, and termination of cross-frame instance association and trajectory also include: The matching threshold is set based on the calculation results of the target association similarity between adjacent frames, and the effective matching pairs that meet the matching threshold are selected with the detection box overlap ratio or the detection box center point distance as the core metric. The single-frame detection results of the target instances in the effective matching pairs are assigned to the same instance identifier, and continuous trajectory segments of the target are accumulated in the time dimension to achieve continuous trajectory updates; For a single-frame detection result that is newly appearing in the current frame and has no historical target instance matching record, initialize a new target instance and its corresponding motion trajectory; For target instances that have no matching objects in several consecutive frames, their trajectory updates are terminated and the trajectory data is archived to ensure the integrity and temporal continuity of the constructed continuous motion trajectory of the target.

4. The method according to claim 3, characterized in that, The extraction of multidimensional temporal features characterizing the dynamic behavior of the target also includes: From the constructed continuous motion trajectory of the target, the center coordinates, width, and height geometric parameters of the target bounding box at each moment are extracted to construct the target position features and characterize the spatial position change pattern of the target. Based on the difference in geometric parameters between adjacent frames and the time interval between adjacent frames Calculate the instantaneous velocity vector of the target at each moment. Construct target velocity characteristics to reflect the target's spatial displacement trend; The scale change rate of the target at each time step is calculated by using the ratio of the difference in the area of ​​the target bounding box between the previous and next frames to the area of ​​the previous frame. To construct target scale features and characterize the relative motion relationship between the target and the imaging device; The target's position features, velocity features, and scale features are combined to form a multidimensional temporal feature vector for each time step. k is the target instance identifier.

5. The method according to claim 4, characterized in that, The multi-scale temporal modeling also includes: Construct multiple parallel one-dimensional temporal convolutional or equivalent temporal feature extraction structures, and configure different sizes of temporal receptive fields for each structure to adapt to the feature capture requirements of short-term fast motion and relatively stable motion, respectively. By synchronously extracting features from trajectory-level temporal feature sequences through various parallel structures, local motion feature representations corresponding to different time scales are obtained. To mine the motion pattern features of the target in different time dimensions, among which, Let X be the temporal feature extraction operator at the s-th time scale, where X is the trajectory-level temporal feature sequence and S is the total number of parallel structures. The extracted multi-scale local motion features are normalized to eliminate interference caused by differences in feature scales, and the fused multi-scale local temporal features are output, laying a data foundation for subsequent bidirectional temporal dependency modeling.

6. The method according to claim 5, characterized in that, The statistical appearance classification confidence score and motion classification confidence score also include: Confidence sequence of appearance classification for the same target instance across multiple frames Perform traversal statistics and use the average value calculation method. Eliminate single-frame detection errors to obtain a stable overall appearance confidence score, where m is the number of detection frames for the same target instance; Motion classification confidence sequence of trajectory segments for the same target instance Statistical analysis was conducted, and weights were assigned based on the length of trajectory segments. Weighted statistics or average statistics were used to calculate the overall motion confidence, taking into account the discriminative contribution of motion characteristics of each segment. The overall appearance confidence and overall motion confidence obtained from the statistics are validated, and abnormal confidence values ​​are removed to ensure the reliability of the confidence data and provide an accurate basis for subsequent weighted fusion.

7. The method according to claim 6, characterized in that, The final classification probability is compared with the preset classification threshold, and the target category is determined. This also includes: A preset classification probability threshold is obtained by training on sample data of micro-flying target detection. The threshold is adapted to the classification boundary between micro-flying targets and pseudo-motion interference objects. according to The final classification probability is obtained by performing weighted fusion calculations. ,in, This is the appearance weighting coefficient. For motion weighting coefficients, and + =1; The final classification probability is compared with the preset classification threshold. If the final classification probability is higher than the threshold, it is determined to be a valid micro-flying target. If it is lower than the threshold, it is determined to be a pseudo-motion interference object. Among multiple categories related to micro-flying targets, the category with the highest final classification probability is selected as the final category of the target. Combined with the target's spatial location information, the final detection result is output, thereby achieving accurate identification of micro-flying targets.

8. A miniature flying target detection device based on the fusion of trajectory dynamic features and appearance, characterized in that, include: The appearance feature detection module is used to process video sequence images frame by frame using an improved single-stage detection network that integrates multi-scale detection enhancement, attention modulation and cross-scale fusion, extract the spatial appearance features of the micro-flying target, complete the target localization and preliminary classification, and output the single-frame detection results containing the spatial location bounding box and appearance classification confidence. The cross-frame trajectory construction module is used to calculate the similarity of targets in adjacent frames based on the detection results of a single frame, combined with spatial location and motion consistency constraints, and establish a matching relationship to complete the cross-frame instance association. It performs the initialization, continuous update and update termination operations when there is no match for the target trajectory, and constructs the continuous motion trajectory of the target. The temporal feature extraction module is used to extract multi-dimensional temporal features representing the dynamic behavior of the target based on the continuous motion trajectory of the target, segment the continuous motion trajectory to form a temporal feature sequence adapted to motion mode discrimination, and output trajectory-level temporal feature data. The motion pattern discrimination module is used to sequentially perform multi-scale temporal modeling, bidirectional time dependency capture and global time correlation modeling on trajectory-level temporal feature data. It judges the target motion pattern through the modeling results, distinguishes between real micro-flying targets and pseudo-motion interference, and outputs motion discrimination results including motion classification confidence. The fusion decision recognition module is used to calculate the confidence scores for appearance classification and motion classification, obtain the confidence scores for overall appearance and overall motion, dynamically adjust the weights according to the scene and calculate the final classification probability by weighted fusion, determine the target category by threshold comparison, and output the detection results containing location and category information to realize the recognition of micro flying targets.

9. An electronic device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images