An unmanned aerial vehicle investigation method and system based on visual intelligence

Abstract

The application discloses a kind of unmanned aerial vehicle investigation method and system based on visual intelligence, it is related to unmanned aerial vehicle intelligent investigation technical field, including the following steps: collecting unmanned aerial vehicle original investigation data, complete investigation data space registration;Improved MobileViT model is constructed and is processed alignment visual sequence, generates investigation area chart;Establish the relationship of waypoint, generates investigation waypoint chart;The review waypoint insertion operator is introduced into Dubins algorithm, and the candidate investigation track of curvature restriction is generated;The candidate track waypoint cost is revised, and the update candidate investigation track is generated, the final track is determined, and the spatialization investigation situation result is output.The application realizes the regionalization discrimination of occlusion area, weak target area, review area and repeated investigation area by improved MobileViT model and introduced the Dubins algorithm of review waypoint insertion operator, and the visual discrimination result is reversely acted on unmanned aerial vehicle curvature restricted investigation track generation process.

Description

Technical Field

[0001] This invention relates to the field of intelligent drone reconnaissance technology, and in particular to a drone reconnaissance method and system based on visual intelligence. Background Technology

[0002] With the development of unmanned aerial vehicle (UAV) platforms, airborne visual payloads, and edge computing devices, UAVs have been widely applied in mission scenarios such as area patrol, target search, disaster reconnaissance, border patrol, and reconnaissance of complex terrain. Existing UAV reconnaissance technologies typically acquire reconnaissance images using visible light cameras, infrared cameras, or multispectral sensors. These images are then analyzed using target detection, semantic segmentation, or image classification models to identify targets, roads, buildings, obstructions, and abnormal areas. The reconnaissance results are generated by combining the UAV's flight attitude, heading angle, and mission area boundaries. Some solutions also generate UAV tracks based on starting waypoints, ending waypoints, no-fly zones, and minimum turning radii to meet the requirements of continuous flight and turning constraints.

[0003] In existing technologies, visual recognition and trajectory planning are typically processed as two relatively independent steps. The visual model mainly outputs the target bounding box, category, or confidence level, while the trajectory planning algorithm mainly generates the flight path based on the starting waypoint, ending waypoint, and spatial obstacle constraints. Although this type of method can complete basic target recognition and path generation, it still has shortcomings in complex reconnaissance scenarios: On the one hand, the recognition results of a single frame image are difficult to accurately distinguish between occluded areas, low-confidence targets, weak targets, and repeated reconnaissance areas, resulting in some areas being repeatedly covered, while areas that really need to be scanned or reviewed fail to enter the trajectory adjustment process in time; on the other hand, the visual recognition results lack deep correlation with flight pose, mission area boundaries, and historical reconnaissance records, making it difficult to convert abnormal information in the image into a spatialized reconnaissance area state, and thus unable to directly guide the UAV to perform review, scan, detour, or trajectory update.

[0004] Furthermore, existing lightweight vision models for UAV edge reconnaissance tasks primarily focus on image feature extraction and target classification, failing to adequately utilize the correspondence between image patches and task regions, historical reconnaissance status, and region-level bias information. This makes it difficult to develop region-based discrimination attributes within the model for reconnaissance decision-making. While existing curvature-constrained trajectory generation methods can meet flight constraints such as the minimum turning radius of UAVs, they typically lack a verification waypoint insertion process driven by visual discrimination results, making it difficult to dynamically change candidate trajectories based on occlusion, weak targets, and verification requirements.

[0005] Therefore, how to provide a visual intelligence-based drone reconnaissance method and system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a visual intelligence-based UAV reconnaissance method and system. This invention fully utilizes the improved MobileViT model, Dubins algorithm, UAV flight status perception, and multi-source reconnaissance data fusion technology. It details the complete implementation process from reconnaissance image acquisition, spatiotemporal registration, region discrimination attribute generation, reconnaissance waypoint map construction to UAV updating reconnaissance track output. It can reverse the visual recognition results to the UAV reconnaissance path and has the advantages of strong occlusion area recognition capability, timely weak target verification, reduced repeated reconnaissance, smooth track update, and high spatialization of reconnaissance results.

[0007] A visual intelligence-based drone reconnaissance method according to an embodiment of the present invention includes the following steps: Collect continuous reconnaissance images, flight attitude, heading angle, flight speed, mission area boundary, historical reconnaissance records, minimum turning radius, starting waypoint and ending waypoint to obtain the raw reconnaissance data of the UAV; Based on the flight attitude, heading angle, mission area boundary and historical reconnaissance records, the original reconnaissance data of the UAV is spatiotemporally registered to obtain an aligned reconnaissance visual sequence. A reconstructed detection region layer is added after the image patch unfolding layer of the MobileViT model, and a region bias injection layer is added before the global encoding layer of the image patch to form an improved MobileViT model. The aligned detection visual sequence is input into the improved MobileViT model to generate a detection region map containing region discrimination attributes. A reconnaissance waypoint map is constructed based on the reconnaissance area map, the starting waypoint, and the ending waypoint, and the area discrimination attribute is written into the reconnaissance waypoint; The Dubins algorithm, which incorporates a verification waypoint insertion operator, is used to expand candidate waypoints, insert verification waypoints, and connect curvature-constrained segments on the reconnaissance waypoint map, generating candidate reconnaissance tracks. Adjust the waypoint cost based on the regional discrimination attribute of the waypoints in the candidate reconnaissance tracks, and generate updated candidate reconnaissance tracks; Based on the waypoint costs and segment connectivity in the updated candidate reconnaissance tracks, the updated regional reconnaissance situation results with spatial coordinates and the updated UAV reconnaissance tracks are determined and output.

[0008] Optionally, the obtained raw UAV reconnaissance data includes: The system acquires continuous reconnaissance images using an airborne visual payload, reads flight attitude, heading angle, and flight speed from the flight control system, and binds the corresponding acquisition timestamps to generate visual state data. Read the mission area boundary, starting waypoint, ending waypoint and minimum turning radius from mission configuration data, airframe parameters and flight safety configuration to generate mission constraint data; The region number, number of reconnaissances, most recent reconnaissance time, historical anomaly markers, and historical occlusion markers are read from historical reconnaissance records. Visual state data, mission constraint data, and historical reconnaissance records are packaged according to UAV mission number, region number, and data collection timestamp to obtain the original UAV reconnaissance data.

[0009] Optionally, obtaining the aligned reconnaissance visual sequence includes: Based on the image acquisition timestamp, the flight pose, heading angle and flight speed are matched. When the sampling interval between adjacent states is exceeded, linear interpolation is performed to generate image frame state matching data. Based on the flight pose, heading angle and mission area boundary in the image frame state matching data, the reconnaissance coverage center, reconnaissance coverage direction and mission area attribution are determined, and image frame area matching data is generated. Historical reconnaissance records are retrieved based on the task area affiliation, and the number of reconnaissances, the most recent reconnaissance time, historical anomaly markers, and historical occlusion markers are written into the corresponding image frames. The aligned reconnaissance visual sequence is generated by sorting the images by their acquisition timestamps.

[0010] Optionally, forming the improved MobileViT model includes: Read the convolutional input layer, inverse residual local coding layer, image patch unrolling layer, image patch global coding layer, image patch folding layer and task output layer in the MobileViT model, and retain the order of convolutional input, inverse residual local coding, image patch unrolling, global coding and folding; A reconnaissance region reconstruction layer is added between the image patch unfolding layer and the image patch global coding layer. The reconnaissance region reconstruction layer receives the image patch sequence and generates an ordered image patch sequence based on the flight pose, heading angle, mission area boundary and historical reconnaissance records. A region bias injection layer is added between the reconstructed region layer and the global encoding layer of image patches. The region bias injection layer receives the ordered image patch sequence of the region and reads the local reconnaissance features output by the inverse residual local encoding layer. It extracts the occlusion response and weak target response from the local reconnaissance features to generate the bias image patch sequence. The reconnaissance decoding layer replaces the task output layer. The reconnaissance decoding layer receives global reconnaissance features generated by the image patch foldback layer and outputs region discrimination attributes to form an improved MobileViT model. The region discrimination attributes include region category, occlusion state, weak target state, verification state, and repeated reconnaissance state.

[0011] Optionally, the step of aligning the reconnaissance visual sequence input to the improved MobileViT model to generate a reconnaissance region map containing region discrimination attributes includes: An improved MobileViT model is constructed, which includes a lightweight local coding layer, a reconstructed reconstructed region layer, a region bias injection layer, a global foldback fusion layer, and a reconstructed decoding layer. The aligned reconnaissance visual sequence is input into a lightweight local coding layer, where convolutional extraction, inverse residual coding, and image patch unrolling are performed to generate shallow reconnaissance features, local reconnaissance features, and image patch sequences. The image patch sequence is input into the reconstructed reconnaissance area layer. Based on the flight pose, heading angle, mission area boundary and historical reconnaissance records, an image patch region projection index is established. The image patch sequence is rearranged according to the image patch region projection index to generate an ordered image patch sequence and a region historical state vector. The ordered image patch sequence, the historical state vector of the region, and the local reconnaissance features are input into the region bias injection layer. Occlusion response and weak target response are extracted from the local reconnaissance features. The region bias sequence is generated based on the historical state vector of the region, the occlusion response, and the weak target response. The region bias sequence is added to the ordered image patch sequence of the region to generate the bias image patch sequence. The biased image patch sequence is input into the global foldback fusion layer, and global encoding and image patch foldback are performed to generate global detection features. Shallow detection features, local detection features and global detection features are then fused to generate fused detection features. The fused reconnaissance features are input into the reconnaissance decoding layer to generate region discrimination attributes and region spatial coordinates. The region discrimination attributes are written into the region grid according to the task region boundary to generate a reconnaissance region map containing the region discrimination attributes. An improved MobileViT model was trained using reconnaissance training samples with region discrimination annotations. Region classification error, occlusion recognition error, weak target recognition error, verification discrimination error, and repeated reconnaissance recognition error were used as joint training objectives to update the parameters of the lightweight local coding layer, reconnaissance region reconstruction layer, region bias injection layer, global foldback fusion layer, and reconnaissance decoding layer, thus obtaining the trained improved MobileViT model.

[0012] Optionally, the construction of the reconnaissance waypoint map and the writing of region discrimination attributes at the reconnaissance waypoints include: Read the region grid, region spatial coordinates and region discrimination attributes in the reconnaissance area map, convert the region grid center point, boundary turning point, starting waypoint and ending waypoint into reconnaissance waypoints, and generate reconnaissance waypoint coordinates; Establish candidate route connection relationships based on reconnaissance waypoint coordinates and regional grid adjacency relationships, delete candidate routes that cross mission area boundaries and impassable areas, and generate an initial reconnaissance waypoint map; The region's spatial coordinates are used to match the region's discrimination attributes to the reconnaissance waypoints, and the starting waypoint marker, ending waypoint marker, region affiliation marker, and candidate segment connection relationship are written to generate a reconnaissance waypoint map.

[0013] Optionally, generating candidate reconnaissance tracks includes: Read the reconnaissance waypoint map and the original reconnaissance data of the UAV, extract the reconnaissance waypoint status from the area discrimination attributes, filter the reconnaissance waypoints that have the verification status, the obstruction status or the weak target status but do not have the repeated reconnaissance status, and generate a set of waypoints to be inserted for verification. Taking candidate segments as the processing object, the feasibility of insertion is judged by combining the set of waypoints to be inserted for verification, the heading angle and the minimum turning radius. Insertable waypoints for verification are retained, and the verification waypoint insertion operator is constructed by insertable verification waypoints, corresponding candidate segments and segment reconnection order. The verification waypoint insertion operator is introduced into the Dubins algorithm. The Dubins algorithm expands candidate waypoints from the starting waypoint. The verification waypoint insertion operator inserts insertable verification waypoints into the candidate segments and splits the candidate segments to generate a sequence of verification-expanded waypoints. The Dubins algorithm connects the starting waypoint, the extended waypoint sequence, and the ending waypoint based on the heading angle and minimum turning radius to generate an initial curvature segment set. It then deletes curvature-restricted segments that cross repeated reconnaissance status areas or cross area ownership marker boundaries, generating a candidate curvature segment set. Candidate reconnaissance tracks are generated by connecting the starting waypoint, the extended waypoint sequence, and the ending waypoint according to the set of candidate curvature segments.

[0014] Optionally, generating updated candidate reconnaissance tracks includes: Read the reconnaissance waypoints, candidate curvature segments, verification extended waypoint sequences and area discrimination attributes from the candidate reconnaissance tracks, and form waypoint attribute vectors according to verification status, occlusion status, weak target status, repeated reconnaissance status and area spatial coordinates; Based on the waypoint attribute vector, perform itemized corrections on the basic waypoint cost, reduce the passage cost corresponding to the verification state, the occlusion state and the weak target state, increase the passage cost corresponding to the repeated reconnaissance state, and generate the corrected waypoint cost; Write the corrected waypoint cost back to the corresponding reconnaissance waypoint, summarize the corrected waypoint costs of adjacent reconnaissance waypoints based on the candidate curvature segment connectivity, delete candidate reconnaissance tracks that lack candidate curvature segment connectivity, and generate updated candidate reconnaissance tracks.

[0015] Optionally, determining and outputting the updated area reconnaissance situation results with spatial coordinates and the updated UAV reconnaissance track includes: Read and update the reconnaissance waypoints, corrected waypoint costs, candidate curvature segment connections, area discrimination attributes, and area spatial coordinates from the candidate reconnaissance tracks. Summarize the corrected waypoint costs and segment lengths in the connection order from the starting waypoint to the ending waypoint to generate track evaluation values. Sort the tracks from low to high according to their evaluation values, filter out the continuous and matching start and end waypoints of the updated candidate reconnaissance tracks, and determine the target reconnaissance track. Bind the reconnaissance waypoints, curvature-restricted segments, area discrimination attributes, and area spatial coordinates in the target reconnaissance track to generate updated area reconnaissance situation results with spatial coordinates, and output the updated UAV reconnaissance track.

[0016] According to an embodiment of the present invention, a visual intelligence-based unmanned aerial vehicle (UAV) reconnaissance system includes the following modules: The raw reconnaissance data acquisition module is used to receive data from the airborne visual payload, flight control system, mission configuration terminal and historical record database, and generate raw reconnaissance data of UAV; The spatiotemporal registration module is used to perform temporal matching and spatial correlation on the raw UAV reconnaissance data to generate aligned reconnaissance visual sequences; The reconnaissance region discrimination module is used to construct an improved MobileViT model and perform image patch region reconstruction, region bias injection and reconnaissance decoding to generate a reconnaissance region map; The reconnaissance waypoint map construction module is used to extract reconnaissance waypoints and candidate flight segments, write them into the area discrimination attributes, and generate a reconnaissance waypoint map; The verification track generation module is used to introduce the verification waypoint insertion operator into the Dubins algorithm to generate candidate reconnaissance tracks; The waypoint cost correction module is used to correct waypoint costs based on area discrimination attributes and generate updated candidate reconnaissance tracks. The reconnaissance result output module is used to determine the updated reconnaissance trajectory of the UAV based on waypoint costs and segment connectivity, and generate updated area reconnaissance situation results with spatial coordinates.

[0017] The beneficial effects of this invention are: This invention acquires continuous reconnaissance images, flight attitude, heading angle, flight speed, mission area boundaries, and historical reconnaissance records, and performs spatiotemporal registration on the raw UAV reconnaissance data to form an aligned reconnaissance visual sequence. This establishes a correspondence between UAV visual data, flight status data, mission area data, and historical reconnaissance data within the same reconnaissance mission chain. Compared to traditional UAV reconnaissance methods that rely solely on single-frame images for target identification, this invention combines reconnaissance images with spatial location, heading status, and historical coverage, preventing visual recognition results from existing independently of the mission area and flight status. This improves the spatial consistency of reconnaissance data and the reliability of subsequent trajectory adjustments.

[0018] This invention adds a reconnaissance region reorganization layer after the image patch unfolding layer of the MobileViT model and a region bias injection layer before the image patch global encoding layer, forming an improved MobileViT model. This allows the image patch sequence to be reorganized according to the reconnaissance region relationship, and incorporates occlusion response, weak target response, and historical reconnaissance state into the image patch global encoding process. Through these internal model improvements, this invention no longer only outputs ordinary target categories or detection confidence, but generates a reconnaissance region map containing region discrimination attributes. This enables occluded regions, weak target regions, regions to be verified, and repeated reconnaissance regions to be distinguished at the region level, improving the reconnaissance and discrimination capabilities of UAVs in complex backgrounds, partial occlusion, and weak target scenarios.

[0019] This invention converts the reconnaissance area map into a reconnaissance waypoint map and introduces a verification waypoint insertion operator into the Dubins algorithm, enabling region discrimination attributes to participate in candidate waypoint expansion, verification waypoint insertion, and curvature-constrained segment connection processes. Compared to traditional path planning that generates paths solely based on start and end waypoints and obstacle constraints, this invention can proactively insert regions requiring verification into the UAV reconnaissance path based on visual discrimination results. Simultaneously, it adjusts waypoint costs based on repeated reconnaissance status, reducing invalid duplicate coverage and improving the efficiency of rescanning occluded and weak target areas. By outputting updated region reconnaissance situational results with spatial coordinates and updated UAV reconnaissance paths, this invention achieves reverse adjustment of UAV reconnaissance paths by visual intelligent perception results, improving the accuracy, continuity, and autonomous decision-making level of UAV reconnaissance tasks. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a visual intelligence-based drone reconnaissance method proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved MobileViT model, which is a visual intelligence-based UAV reconnaissance method proposed in this invention. Figure 3 This is a schematic diagram of the structure of a visual intelligence-based unmanned aerial vehicle (UAV) reconnaissance system proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1 and Figure 2A visual intelligence-based drone reconnaissance method includes the following steps: Collect continuous reconnaissance images, flight attitude, heading angle, flight speed, mission area boundary, historical reconnaissance records, minimum turning radius, starting waypoint and ending waypoint to obtain the raw reconnaissance data of the UAV; Based on the flight attitude, heading angle, mission area boundary and historical reconnaissance records, the original reconnaissance data of the UAV is spatiotemporally registered to obtain an aligned reconnaissance visual sequence. A reconstructed detection region layer is added after the image patch unfolding layer of the MobileViT model, and a region bias injection layer is added before the global encoding layer of the image patch to form an improved MobileViT model. The aligned detection visual sequence is input into the improved MobileViT model to generate a detection region map containing region discrimination attributes. A reconnaissance waypoint map is constructed based on the reconnaissance area map, the starting waypoint, and the ending waypoint, and the area discrimination attribute is written into the reconnaissance waypoint; The Dubins algorithm, which incorporates a verification waypoint insertion operator, is used to expand candidate waypoints, insert verification waypoints, and connect curvature-constrained segments on the reconnaissance waypoint map, generating candidate reconnaissance tracks. Adjust the waypoint cost based on the regional discrimination attribute of the waypoints in the candidate reconnaissance tracks, and generate updated candidate reconnaissance tracks; Based on the waypoint costs and segment connectivity in the updated candidate reconnaissance tracks, the updated regional reconnaissance situation results with spatial coordinates and the updated UAV reconnaissance tracks are determined and output.

[0023] In this embodiment, obtaining the original UAV reconnaissance data includes: The system acquires continuous reconnaissance images using an airborne visual payload, reads flight attitude, heading angle, and flight speed from the flight control system, and binds the corresponding acquisition timestamps to generate visual state data. The mission area boundary, starting waypoint, ending waypoint, and minimum turning radius are read from mission configuration data, airframe parameters, and flight safety configuration to generate mission constraint data. Specifically, the generation of mission constraint data includes: Read the boundary point numbers, latitude, longitude, and altitude values ​​of the mission area boundary. Based on the mission coordinate system origin, east axis, north axis, and altitude reference in the mission configuration data, convert the boundary point latitude, longitude, and altitude values ​​into boundary point coordinates in the mission coordinate system, and connect them according to the boundary point numbers to form a closed boundary point sequence. Read the latitude, longitude, altitude values, and heading angle of the starting waypoint and ending waypoint, and convert them into starting waypoint coordinates, ending waypoint coordinates, and heading constraints according to the mission coordinate system. Read the maximum roll angle, cruise speed, and safety margin from the airframe parameters, and combine them with the minimum turning radius limit in the flight safety configuration to determine the minimum turning radius. Encapsulate the closed boundary point sequence, starting waypoint coordinates, ending waypoint coordinates, heading constraints, and minimum turning radius according to the mission number to generate mission constraint data. The raw UAV reconnaissance data is obtained by retrieving the region ID, number of reconnaissance attempts, most recent reconnaissance time, historical anomaly markers, and historical occlusion markers from historical reconnaissance records. Visual state data, task constraint data, and historical reconnaissance records are then packaged according to the UAV mission ID, region ID, and acquisition timestamp. Specifically, the raw UAV reconnaissance data is as follows: A mission data packet is created using the UAV mission number, and mission constraint data is written into the mission constraint field of the mission data packet. A region record is created within the mission data packet using the region number, and the number of reconnaissances, the most recent reconnaissance time, historical anomaly markers, and historical occlusion markers from the historical reconnaissance records are written into the historical status field of the region record. A frame record is created within the region record using the acquisition timestamp, and the continuous reconnaissance images, flight attitude, heading angle, and flight speed from the visual status data are written into the visual status field of the frame record. An index identifier is generated according to the UAV mission number, region number, and acquisition timestamp. The mission data packet, region record, frame record, and index identifier are merged to obtain the original UAV reconnaissance data.

[0024] In this embodiment, obtaining the aligned detection visual sequence includes: Based on the image acquisition timestamp, flight pose, heading angle, and flight velocity are matched. When the sampling interval between adjacent states is exceeded, linear interpolation is performed to generate image frame state matching data. Specifically, the generation of image frame state matching data is as follows: Using the image acquisition timestamp of each consecutive reconnaissance image as the matching benchmark, the first state record with a timestamp no greater than the image acquisition timestamp and the smallest time difference is retrieved in the flight state data, and the second state record with a timestamp no less than the image acquisition timestamp and the smallest time difference is retrieved. When the time difference between the first state record or the second state record and the image acquisition timestamp is within the adjacent state sampling interval, the flight attitude, heading angle and flight speed in the state record with the smaller time difference are selected as the matching state. When the image acquisition timestamp is located between the first state record and the second state record and the time difference exceeds the sampling interval between adjacent states, the proportion of the image acquisition timestamp relative to the first state record timestamp in the time interval between the two sets of state records is calculated. The position coordinates, altitude values, heading angles and flight speeds in the two sets of state records are interpolated item by item according to the proportion to obtain the interpolated flight pose, interpolated heading angle and interpolated flight speed. Bind the image frame number, image acquisition timestamp, matching status or interpolation status to the corresponding continuous reconnaissance image to generate image frame status matching data; Based on the flight pose, heading angle, and mission area boundary in the image frame state matching data, the reconnaissance coverage center, reconnaissance coverage direction, and mission area attribution are determined, and image frame region matching data is generated. Specifically, determining the reconnaissance coverage center, reconnaissance coverage direction, and mission area attribution, and generating image frame region matching data, involves the following steps: Read the flight attitude, heading angle and flight altitude of each image frame in the image frame state matching data, and read the installation pitch angle, installation roll angle and field of view of the airborne visual payload; calculate the direction vector of the imaging center line in the mission coordinate system based on the position coordinates, heading angle, installation pitch angle and installation roll angle in the flight attitude, and take the intersection of the imaging center line and the ground plane of the mission area as the reconnaissance coverage center; The horizontal projection direction corresponding to the heading angle is taken as the reconnaissance coverage direction, and the image frame coverage area is calculated according to the field of view and flight altitude. The image frame coverage area is represented as a clockwise sequence of coverage vertices, and the task region boundary is represented as a clockwise sequence of boundary vertices. Coverage edges in the coverage vertex sequence and boundary edges in the boundary vertex sequence are taken in turn. The directed area sign of the two endpoints of the coverage edge relative to the boundary edge is calculated, and the directed area sign of the two endpoints of the boundary edge relative to the coverage edge is also calculated. When both sets of directed area signs have positive and negative changes and the horizontal projection intervals and vertical projection intervals of the two edges overlap, the edge intersection point is recorded. The overlapping vertices located within the task region boundary, the boundary vertices located within the image frame coverage area, and the edge intersections are merged into an overlapping polygon. When the area of ​​the overlapping polygon is greater than 0, the corresponding region number is read. The region matching index is composed of image frame number, region number and image acquisition timestamp. The detection coverage center, detection coverage direction, image frame coverage range and coverage overlap polygon are written under the region matching index to generate image frame region matching data. Historical reconnaissance records are retrieved based on the task area affiliation. The number of reconnaissances, the most recent reconnaissance time, historical anomaly markers, and historical occlusion markers are written into the corresponding image frames. An aligned reconnaissance visual sequence is generated by sorting the images according to their acquisition timestamps. Specifically, the process of generating the aligned reconnaissance visual sequence by sorting the images according to their acquisition timestamps is as follows: Read the image frames written to the historical reconnaissance record, and extract the image frame number, image acquisition timestamp, continuous reconnaissance images, image frame status matching data, image frame region matching data, reconnaissance count, most recent reconnaissance time, historical anomaly markers, and historical occlusion markers to form an image frame record; convert the image acquisition timestamp into a millisecond-level time value, arrange the image frame records in ascending order of millisecond-level time value, and arrange the image frame records in ascending order of image frame number if the millisecond-level time values ​​are the same; write the sequence number, the previous image frame number, and the next image frame number to each image frame record according to the arrangement result, and store the arranged image frame records continuously according to the sequence number to generate an aligned reconnaissance visual sequence.

[0025] In this embodiment, forming the improved MobileViT model includes: Read the convolutional input layer, inverse residual local coding layer, image patch unrolling layer, image patch global coding layer, image patch folding layer and task output layer in the MobileViT model, and retain the order of convolutional input, inverse residual local coding, image patch unrolling, global coding and folding; A reconnaissance region reconstruction layer is added between the image patch unpacking layer and the image patch global coding layer. This layer receives the image patch sequence and generates an ordered sequence of image patches based on the flight pose, heading angle, mission area boundaries, and historical reconnaissance records. Specifically, the reconnaissance region reconstruction layer between the image patch unpacking layer and the image patch global coding layer is as follows: The original transmission relationship between the image patch unfolding layer and the image patch global encoding layer is broken. The image patch sequence generated by the image patch unfolding layer is input into the reconnaissance region reconstruction layer. The reconnaissance region reconstruction layer reads the flight attitude, heading angle, mission area boundary, and historical reconnaissance records. It calculates the corresponding line-of-sight direction according to the row and column position of each image patch in the image. Combining the flight attitude and heading angle, the center of the image patch is projected onto the mission area coordinate system to determine the mission area number corresponding to the image patch. Historical reconnaissance records are retrieved according to the mission area number, and the number of reconnaissances, historical anomaly markers, and historical occlusion markers are written for the image patch. The image patch sequence is rearranged according to the mission area number, historical occlusion markers, historical anomaly markers, number of reconnaissances, and the original order of the image patches to generate a region-ordered image patch sequence. The region-ordered image patch sequence is input into the image patch global encoding layer to complete the addition of the reconnaissance region reconstruction layer. A region bias injection layer is added between the reconstructed region layer and the global image patch coding layer. This layer receives an ordered sequence of image patches and reads the local reconnaissance features output by the inverse residual local coding layer. It then extracts occlusion and weak target responses from these local features to generate a biased image patch sequence. Specifically, the region bias injection layer between the reconstructed region layer and the global image patch coding layer is as follows: The original transmission relationship between the output of the reconstructed region layer and the global encoding layer of the image patch is broken. The ordered image patch sequence generated by the reconstructed region layer is input into the region bias injection layer. The local reconnaissance features output by the inverse residual local encoding layer are simultaneously input into the region bias injection layer, retaining the spatial row and column indices and channel indices of the local reconnaissance features. The region bias injection layer performs channel mean compression and response mapping on the local reconnaissance features to generate occlusion response maps and weak target response maps. According to the coverage of the image patch in the local reconnaissance features, the average response values ​​of the occlusion response map and the weak target response map are calculated within each image patch range to obtain the image patch occlusion response value and the image patch weak target response value. The image patch occlusion response value and the image patch weak target response value are converted into region bias coefficients. According to the image patch number, the region bias coefficients are added to the feature values ​​of the corresponding image patch in the ordered image patch sequence to generate a biased image patch sequence. The biased image patch sequence is input into the global encoding layer of the image patch. The reconnaissance decoding layer replaces the task output layer. This layer receives global reconnaissance features generated by the image patch foldback layer and outputs region discrimination attributes, forming an improved MobileViT model. These region discrimination attributes include region category, occlusion state, weak target state, re-detection state, and repeated reconnaissance state. Specifically, replacing the task output layer with the reconnaissance decoding layer involves: The classification mapping parameters, pooling parameters, and category output fields corresponding to the task output layer in the MobileViT model are removed, while the global detection features output by the image patch foldback layer are retained. A detection decoding layer is connected after the image patch foldback layer. The detection decoding layer receives the global detection features and performs channel compression, region grid mapping, and multi-branch discriminant mapping on the global detection features. Channel compression uses pointwise convolution to compress the global detection features into the region discriminant channel. Region grid mapping maps the region discriminant channel to the region grid position according to the task region boundary. Multi-branch discriminant mapping generates region category response, occlusion response, weak target response, solid response, and repeated detection response, respectively. The region category response, occlusion response, weak target response, solid response, and repeated detection response are merged into a region discriminant attribute, which is then output by the detection decoding layer.

[0026] In this embodiment, the step of inputting the aligned reconnaissance visual sequence into the improved MobileViT model to generate a reconnaissance region map containing region discrimination attributes includes: An improved MobileViT model is constructed, comprising a lightweight local coding layer, a reconnaissance region reconstruction layer, a region bias injection layer, a global foldback fusion layer, and a reconnaissance decoding layer. Specifically, the improved MobileViT model is constructed as follows: The convolutional input parameters, inverse residual coding parameters, image patch unrolling parameters, image patch global coding parameters, and image patch folding parameters from the MobileViT base model are used as the initial parameters for the lightweight local coding layer and the global folding fusion layer. Convolutional input processing, inverse residual coding processing, and image patch unrolling processing are combined into a lightweight local coding layer, and image patch global coding processing, image patch folding processing, and multi-scale feature fusion processing are combined into a global folding fusion layer. A detection region reconstruction layer is connected after the lightweight local coding layer, and a region partial... An injection layer is placed, followed by a global foldback fusion layer, and then a reconnaissance decoding layer. A reconnaissance region reconstruction layer is configured to read image patch sequences, flight pose, heading angle, mission region boundaries, and historical reconnaissance records. A region bias injection layer is configured to read ordered image patch sequences and local reconnaissance features. A reconnaissance decoding layer is configured to output region discrimination attributes and region spatial coordinates, forming an improved MobileViT model comprising a lightweight local coding layer, a reconnaissance region reconstruction layer, a region bias injection layer, a global foldback fusion layer, and a reconnaissance decoding layer. The aligned reconnaissance visual sequence is input into a lightweight local coding layer, where convolutional extraction, inverse residual coding, and image patch unfolding are performed to generate shallow reconnaissance features, local reconnaissance features, and image patch sequences. Specifically, the generation of shallow reconnaissance features, local reconnaissance features, and image patch sequences is as follows: Read consecutive reconnaissance images from the aligned reconnaissance visual sequence. Divide the pixel values ​​by 255 for each of the red, green, and blue channels, and subtract the mean of the corresponding channel from the training samples to obtain a standardized image tensor. Input the standardized image tensor into the convolutional processing part of the lightweight local coding layer. Use a 3×3 convolution kernel with a stride of 2 to extract edge, texture, and feature contour responses. Unify the number of channels through pointwise convolution to generate shallow reconnaissance features. Input the shallow reconnaissance features into the inverse residual coding processing part. Perform channel expansion using 1×1 convolution, extract local spatial responses using 3×3 depthwise convolution, and perform channel compression using 1×1 pointwise convolution. Add the channel compression result to the corresponding input features using residuals of the same size to generate local reconnaissance features. Divide the local reconnaissance features into non-overlapping image blocks according to the image block size. Record the row and column numbers, spatial range, and channel features of each image block. Expand the channel features of each image block into image block vectors in row-major order. Arrange all image block vectors by row and column numbers to generate an image block sequence. The image patch sequence is input into the reconnaissance region reconstruction layer. An image patch region projection index is established based on flight pose, heading angle, mission area boundary, and historical reconnaissance records. The image patch sequence is then rearranged according to the image patch region projection index to generate an ordered image patch sequence and a region historical state vector. The reconnaissance region reconstruction layer includes: Image patch coordinate resolution unit: Reads the row and column number, spatial range and channel features of each image patch in the image patch sequence, and determines the pixel position of the image patch center in the continuous reconnaissance images; Region projection index unit: Based on the flight attitude, heading angle and mission area boundary, the pixel position of the center of the image patch is mapped to the mission area coordinate system to generate the image patch region projection index; Historical state encoding unit: Retrieves historical reconnaissance records based on the image patch region projection index, and encodes the number of reconnaissances, the most recent reconnaissance time, historical anomaly markers and historical occlusion markers into a region historical state vector; Region order rearrangement unit: The image block sequence is rearranged according to the image block region projection index, region history state vector and the original row and column order of the image blocks to generate a region-ordered image block sequence; In the reconstructed reconstructed area layer, the image patch coordinate parsing unit reads the image patch number, row and column number, spatial range, and channel features from the image patch sequence, and takes the center pixel of the spatial range as the image patch center; the region projection indexing unit reads the position coordinates and altitude values ​​in the flight pose, reads the heading angle and mission area boundary, converts the image patch center into a plane offset relative to the imaging center according to the image row and column position, rotates the plane offset to the mission area coordinate system according to the heading angle, and adds it to the position coordinates in the flight pose to obtain the mission coordinates of the image patch center; the region projection indexing unit determines the inclusion relationship between the mission coordinates of the image patch center and the mission area boundary, reads the region number containing the image patch center, and establishes the image patch region projection between the image patch number and the region number. The index and historical status encoding unit retrieve historical reconnaissance records according to the region number, read the number of reconnaissances, the most recent reconnaissance time, historical anomaly markers, and historical occlusion markers, normalize the number of reconnaissances into a number encoding, convert the most recent reconnaissance time into a time interval encoding, convert historical anomaly markers and historical occlusion markers into status encodings, and concatenate the number of reconnaissances, time interval encodings, and status encodings to generate a region historical status vector; the region order rearrangement unit rearranges the image block sequence according to the arrangement rules of region number from small to large, historical occlusion marker priority, historical anomaly marker priority, reconnaissance count from few to many, and the original row and column order of image blocks, generates a region ordered image block sequence, and writes the region historical status vector into the image block position with the same region number in the region ordered image block sequence; The ordered image patch sequence, the historical state vector of the region, and local reconnaissance features are input into the region bias injection layer. Occlusion response and weak target response are extracted from the local reconnaissance features. A region bias sequence is generated based on the historical state vector, occlusion response, and weak target response. The region bias sequence is then added to the ordered image patch sequence to generate a bias image patch sequence. The region bias injection layer includes: Response extraction unit: Performs channel compression and response mapping on local detection features to extract occlusion response maps and weak target response maps; Regional State Fusion Unit: Performs numerical normalization, state embedding, and linear mapping on the regional historical state vector to generate regional state codes; Bias coefficient generation unit: Aggregates occlusion response, weak target response and region state coding according to the image patch coverage area to generate region bias coefficient; Image block offset writing unit: Writes the region offset coefficient into the ordered image block sequence according to the image block number to generate the offset image block sequence; In the region bias injection layer, the response extraction unit reads local detection features, averages the channel features according to their spatial row and column positions, generates a spatial response base map, and performs 3×3 convolution mapping and normalization on the spatial response base map to generate occlusion response maps and weak target response maps; the region state fusion unit reads the detection count code, time interval code, historical abnormal state code, and historical occlusion state code from the region historical state vector, performs maximum value normalization on the detection count code and time interval code, performs discrete state embedding on the historical abnormal state code and historical occlusion state code, and performs linear mapping between the normalization result and the embedding result to generate the region state code; the bias coefficient generation unit... Based on the spatial range of each image patch in the ordered image patch sequence, response values ​​falling within the spatial range are extracted from the occlusion response map and the weak target response map. The average occlusion response value and the average weak target response value are calculated and concatenated with the region state code corresponding to the same region number. After linear mapping and normalization, region bias coefficients are generated. The image patch bias writing unit reads the image patch features in the ordered image patch sequence according to the image patch number, expands the region bias coefficients to the same channel dimension as the image patch features, and adds them to the image patch features channel by channel to generate bias image patch features. The bias image patch features are summarized according to the arrangement order of the ordered image patch sequence to generate the bias image patch sequence. The biased image patch sequence is input into the global foldback fusion layer, where global encoding and foldback of image patches are performed to generate global detection features. These features are then fused with shallow detection features, local detection features, and global detection features to generate fused detection features. Specifically, the generation of fused detection features involves: The biased image patch sequence is input into the global foldback fusion layer. Linear mapping is performed on the features of each biased image patch to generate an image patch query vector, an image patch key vector, and an image patch value vector. The correlation score between the image patch query vector and the image patch key vector is calculated, and the correlation score is normalized to obtain the image patch association weight. The image patch value vector is then weighted and summed according to the image patch association weight to generate a globally encoded image patch sequence. Based on the row and column numbers and region order of the image patches recorded in the reconstructed region layer, the globally encoded image patch sequence is written back to its two-dimensional spatial location one by one. Generate global detection features; map shallow detection features, local detection features, and global detection features to the same number of channels through 1×1 convolutions, and perform average pooling on shallow detection features according to the spatial size of global detection features, and perform bilinear interpolation on local detection features based on the distance weights of the four adjacent pixel features to obtain shallow alignment features, local alignment features, and global alignment features with consistent size; concatenate the shallow alignment features, local alignment features, and global alignment features according to the channel direction, and fuse them through 3×3 convolutions, batch normalization, and nonlinear activation to generate fused detection features; The fused reconnaissance features are input into the reconnaissance decoding layer to generate region discrimination attributes and region spatial coordinates. The region discrimination attributes are then written into the region grid according to the task region boundaries, generating a reconnaissance region map containing the region discrimination attributes. Specifically, generating the reconnaissance region map containing the region discrimination attributes involves: The fused reconnaissance features are input into the reconnaissance decoding layer. The layer performs pointwise convolution and channel compression on these features to generate region discrimination features. Region category mapping, occlusion state mapping, weak target state mapping, verification state mapping, and repeated reconnaissance state mapping are then performed on these features to generate region category response, occlusion response, weak target response, verification response, and repeated reconnaissance response, respectively. A region grid is established according to the task region boundaries. The boundary and center coordinates of each region grid are read, and the response values ​​falling within the same region grid range are averaged. All values ​​are aggregated to obtain the region category value, occlusion status value, weak target status value, verification status value, and repeated reconnaissance status value corresponding to each region grid. The region category value, occlusion status value, weak target status value, verification status value, and repeated reconnaissance status value are combined into a region discrimination attribute, and the boundary coordinates and center coordinates of the region grid are combined into a region spatial coordinate. A region record is established by region number, and the region spatial coordinates and region discrimination attributes are written into the region record. The region connection relationship is written according to the adjacency relationship between the region grids to generate a reconnaissance region map containing the region discrimination attribute. An improved MobileViT model was trained using reconnaissance training samples with region-specific annotations. Region classification error, occlusion recognition error, weak target recognition error, verification error, and repeated reconnaissance recognition error were used as joint training objectives to update the parameters of the lightweight local coding layer, reconnaissance region reconstruction layer, region bias injection layer, global foldback fusion layer, and reconnaissance decoding layer, resulting in the trained improved MobileViT model. Specifically, the training of the improved MobileViT model using reconnaissance training samples with region-specific annotations was conducted as follows: Read reconnaissance training samples with region discrimination annotations. The training samples include reconnaissance images, flight pose, heading angle, mission area boundary, historical reconnaissance records and region discrimination annotations. The region discrimination annotations include region category annotations, occlusion status annotations, weak target status annotations, verification status annotations and repeated reconnaissance status annotations. The training samples are divided into training samples and validation samples according to the collection time sequence. The training samples are then input into the improved MobileViT model and pass through the lightweight local coding layer, the reconstructed detection region layer, the region bias injection layer, the global foldback fusion layer, and the detection decoding layer in sequence to obtain the region category response, occlusion response, weak target response, verification response, and repeated detection response. Align the region category response with the region category label region by region, calculate the difference between the region category response value and the corresponding region category label value for each region, take the absolute value of the difference for all regions and average it to obtain the region classification error; Align the occlusion response with the occlusion status label region by region, calculate the difference between the occlusion response value and the corresponding occlusion status label value for each region, take the absolute value of the difference for all regions and average it to obtain the occlusion recognition error; Align the weak target response with the weak target status label region by region, calculate the difference between the weak target response value and the corresponding weak target status label value in each region, take the absolute value of the difference for all regions and average it to obtain the weak target identification error. Align the verification response with the verification status label region by region, calculate the difference between the verification response value and the corresponding verification status label value in each region, take the absolute value of the difference for all regions and average it to obtain the verification discrimination error. Align the repeated reconnaissance response with the repeated reconnaissance status label region by region, calculate the difference between the repeated reconnaissance response value and the corresponding repeated reconnaissance status label value in each region, take the absolute value of the difference for all regions and average it to obtain the repeated reconnaissance identification error. Training weights are assigned to the region classification error, occlusion recognition error, weak target recognition error, verification discrimination error, and repeated detection recognition error, respectively. Each error is multiplied by its corresponding training weight to obtain the weighted errors for region classification, occlusion recognition, weak target recognition, verification discrimination, and repeated detection. The five weighted errors are summed to form a joint training objective. Gradients are calculated for the convolutional parameters, linear mapping parameters, response mapping parameters, and discriminant mapping parameters in the lightweight local coding layer, detection region reconstruction layer, region bias injection layer, global foldback fusion layer, and detection decoding layer based on the joint training objective. The corresponding parameters are updated according to the product of the learning rate and the gradient. Calculate the region discrimination accuracy and the verification discrimination accuracy on the validation samples. Stop parameter updates when the number of training rounds reaches 80 rounds or the accuracy improvement of the validation samples is less than 0.5% for 5 consecutive rounds, and obtain the improved MobileViT model after training.

[0027] In this embodiment, the construction of the reconnaissance waypoint map and the writing of region discrimination attributes into the reconnaissance waypoints include: Read the region grid, region spatial coordinates, and region discrimination attributes from the reconnaissance area map. Convert the region grid center point, boundary inflection point, starting waypoint, and ending waypoint into reconnaissance waypoints, generating reconnaissance waypoint coordinates. Specifically, the conversion of the region grid center point, boundary inflection point, starting waypoint, and ending waypoint into reconnaissance waypoints and the generation of reconnaissance waypoint coordinates is as follows: Read the region number, boundary coordinates, and regional spatial coordinates of each region grid in the reconnaissance area map. Take the average x-coordinate and average y-coordinate of the boundary coordinates of the same region grid as the plane coordinates of the region grid center point, and read the altitude reference in the regional spatial coordinates as the altitude value of the region grid center point to generate the coordinates of the region center waypoint. Read the adjacent edges on the boundary of the region grid in sequence, calculate the direction vector of the adjacent edges. When the direction vectors of the adjacent edges are not collinear, take the intersection of the adjacent edges as the boundary turning point, and combine the plane coordinates of the boundary turning point with the altitude reference in the regional spatial coordinates to generate the coordinates of the boundary turning waypoint. Read the starting waypoint coordinates and ending waypoint coordinates in the task constraint data, write the starting waypoint coordinates into the starting waypoint record, and write the ending waypoint coordinates into the ending waypoint record. Write waypoint type tags according to the types of region center waypoint, boundary turning waypoint, starting waypoint, and ending waypoint. Merge waypoints with a plane coordinate difference of less than 1m and an altitude difference of less than 0.5m, retain the region number, waypoint type tag, and waypoint coordinates to generate the coordinates of the reconnaissance waypoint. Based on the reconnaissance waypoint coordinates and the regional grid adjacency relationships, candidate flight segment connections are established. Candidate flight segments that cross mission area boundaries and impassable areas are deleted, and an initial reconnaissance waypoint map is generated. The generation of the initial reconnaissance waypoint map specifically involves: Read the coordinates of reconnaissance waypoints, the adjacency relationship of the regional grid, the boundary of the mission area, and the markers of impassable areas. Treat each reconnaissance waypoint as a graph node and write the waypoint number, waypoint coordinates, waypoint type marker, and region number to the graph node. Select reconnaissance waypoints in adjacent regions according to the adjacency relationship of the regional grid. Calculate the planar distance and height difference between adjacent reconnaissance waypoints. When the planar distance is less than the upper limit of the length of a single reconnaissance segment and the height difference is less than the safe height difference threshold, generate a candidate segment. Intersect the line connecting the two ends of the candidate segment with the boundary edge of the mission area segment by segment. If the line connecting the endpoints exceeds the boundary of the mission area, delete the corresponding candidate segment. Intersect the line connecting the two ends of the candidate segment with the boundary edge of the impassable area segment by segment. If the line connecting the endpoints passes through the impassable area, delete the corresponding candidate segment. Write the retained candidate segments into the connection field between the corresponding graph nodes and record the segment length, segment direction, and adjacent region number to generate an initial reconnaissance waypoint map containing the connection relationship between reconnaissance waypoints and candidate segments. The region's spatial coordinates are used to match the region's discrimination attributes to reconnaissance waypoints, and the starting waypoint marker, ending waypoint marker, region affiliation marker, and candidate segment connectivity are written to generate a reconnaissance waypoint map. Specifically, generating the reconnaissance waypoint map involves: The system reads graph nodes, waypoint coordinates, waypoint type markers, and candidate segment connectivity from the initial reconnaissance waypoint map; it also reads the regional spatial coordinates and regional discrimination attributes from the reconnaissance area map. It performs an inclusion relationship check between each waypoint coordinate and the regional boundary coordinates in the regional spatial coordinates. If the waypoint coordinates are within the regional boundary range, it reads the corresponding region number and regional discrimination attribute. If the waypoint coordinates are within the overlapping range of multiple region boundaries, it calculates the distance between the waypoint coordinates and the center coordinates of each region, and selects the region number and regional discrimination attribute with the smallest distance. It writes the read region number into the region attribution marker of the graph node and the region discrimination attribute into the attribute field of the graph node. It performs distance matching between the starting and ending waypoint coordinates and the waypoint coordinates of the graph node, writing the starting and ending waypoint markers for graph nodes with distances less than the waypoint merging threshold. It writes the candidate segment connectivity into the edge field between the corresponding graph nodes, and writes the segment length, segment direction, and adjacent region number into the edge field, generating a reconnaissance waypoint map containing graph nodes, edge fields, region attribution markers, regional discrimination attributes, starting waypoint markers, and ending waypoint markers.

[0028] In this embodiment, generating candidate reconnaissance tracks includes: Read the reconnaissance waypoint map and the original UAV reconnaissance data, extract the reconnaissance waypoint status from the area discrimination attributes, filter out reconnaissance waypoints that have verification status, obstruction status, or weak target status but do not have the repeated reconnaissance status, and generate a set of waypoints to be inserted for verification. Specifically, the generation of the set of waypoints to be inserted for verification is as follows: Read the graph nodes, waypoint coordinates, region discrimination attributes, region attribution markers, and candidate flight segment connections from the reconnaissance waypoint map. Also read the starting waypoint, ending waypoint, heading angle, and minimum turning radius from the original UAV reconnaissance data. Extract the verification status value, occlusion status value, weak target status value, and repeated reconnaissance status value from the region discrimination attributes. Mark graph nodes with at least one verification status value, occlusion status value, or weak target status value that is not less than 0.5 and a repeated reconnaissance status value that is less than 0.5 as candidate verification graph nodes. Delete connections without candidate flight segments. Candidate verification map nodes are selected based on their relationships. Nodes less than 1m from the starting or ending waypoint are deleted. Verification waypoint records are generated based on the waypoint coordinates, region attribution markers, region discrimination attributes, and adjacent candidate segments of the retained candidate verification map nodes. Each verification waypoint record contains the verification waypoint number, verification waypoint coordinates, region attribution marker, verification status value, occlusion status value, weak target status value, and adjacent candidate segment number. The verification waypoint records are then arranged according to the region number and verification waypoint number to generate a set of verification waypoints to be inserted. Taking candidate flight segments as the processing object, the feasibility of insertion is determined by combining the set of waypoints to be inserted for verification, the heading angle, and the minimum turning radius. Insertable waypoints are retained, and a verification waypoint insertion operator is constructed from the insertable verification waypoints, the corresponding candidate flight segments, and the flight segment reconnection order. Specifically, taking candidate flight segments as the processing object and combining the set of waypoints to be inserted for verification, the heading angle, and the minimum turning radius to determine insertion feasibility, and retaining insertable verification waypoints, the process is as follows: Read the coordinates of the starting waypoint, the ending waypoint, and the direction of each candidate segment. Read the coordinates and area affiliation markers of the verification waypoints in the set of verification waypoints to be inserted. Calculate the perpendicular position from the verification waypoint coordinates to the candidate segment. When the perpendicular position is between the starting and ending waypoints of the candidate segment, calculate the forward and backward distances from the verification waypoint to the starting and ending waypoints, respectively. Determine the forward entry direction based on the heading angle, the starting waypoint coordinates, and the verification waypoint coordinates. Determine the backward departure direction based on the verification and ending waypoint coordinates. When both the forward and backward distances are greater than twice the minimum turning radius, and the turning changes formed by the forward entry direction, the backward departure direction, and the heading angle satisfy the minimum turning radius constraint, mark the corresponding verification waypoint as an insertable verification waypoint. Delete verification waypoints that do not meet the distance, turning, or area affiliation marker consistency conditions, resulting in the retained insertable verification waypoints. The verification waypoint insertion operator is constructed from the insertable verification waypoints, the corresponding candidate segments, and the segment reconnection order, specifically as follows: An insertion record is created centered on each insertable verification waypoint. The insertion record contains the verification waypoint number, verification waypoint coordinates, corresponding candidate segment number, origin waypoint number, and destination waypoint number. The corresponding candidate segment is split into a first segment from the origin waypoint to the verification waypoint and a second segment from the verification waypoint to the destination waypoint. The first segment is placed before the second segment to form a segment reconnection order. The insertion records are arranged according to the verification waypoint number to generate an insertion record table. The insertion record table, the first segment, the second segment, and the segment reconnection order are encapsulated into a verification waypoint insertion operator. The verification waypoint insertion operator is introduced into the Dubins algorithm. The Dubins algorithm expands candidate waypoints from the starting waypoint. The verification waypoint insertion operator inserts insertable verification waypoints into the candidate segments and splits the candidate segments to generate a sequence of verified expanded waypoints. Specifically, the generation of the verified expanded waypoint sequence is as follows: Read the insertion record table, the first segment, the second segment, and the segment reconnection order from the verification waypoint insertion operator; read the candidate waypoint sequence obtained by the Dubins algorithm expansion; find the corresponding candidate segment number according to the adjacent waypoint number in the candidate waypoint sequence; when the candidate segment number matches the corresponding candidate segment number in the insertion record table, read the verification waypoint number and verification waypoint coordinates from the insertion record table; arrange the starting waypoint, verification waypoint, and ending waypoint among the adjacent waypoints according to the segment reconnection order, and write the first segment from the starting waypoint to the verification waypoint and the second segment from the verification waypoint to the ending waypoint into the connection position between adjacent waypoints; when the same candidate segment corresponds to multiple verification waypoints, arrange the multiple verification waypoints from the verification waypoint to the starting waypoint in order of distance from the verification waypoint to the starting waypoint; arrange the waypoints after the verification waypoint insertion and candidate segment splitting are arranged continuously in the expansion order from the starting waypoint to the ending waypoint to generate the verification extended waypoint sequence; The Dubins algorithm connects the starting waypoint, the extended waypoint sequence, and the ending waypoint based on the heading angle and minimum turning radius to generate an initial curvature segment set. It then removes curvature-restricted segments that traverse areas with repeated reconnaissance states or cross area ownership boundary markers, generating a candidate curvature segment set. Specifically, the initial curvature segment set is generated as follows: Read the starting waypoint, verify the extended waypoint sequence, end waypoint, heading angle, and minimum turning radius, and form adjacent waypoint pairs according to the order from the starting waypoint to the end waypoint; for each adjacent waypoint pair, read the coordinates of the starting waypoint and the ending waypoint, take the heading angle of the starting waypoint as the entry heading, and take the direction from the starting waypoint to the ending waypoint as the exit heading; generate left-turn centers and right-turn centers on both sides of the starting waypoint and the ending waypoint according to the entry heading and the exit heading, respectively, and take the minimum turning radius as the distance from the center to the corresponding waypoint; calculate the external and internal tangent segments between the left-turn center and the right-turn center, select the tangent points that are continuous with both the entry and exit headings, and generate the entry arc segment, the intermediate straight segment, and the exit arc segment; combine the entry arc segment, the intermediate straight segment, and the exit arc segment according to the direction from the starting waypoint to the ending waypoint, and write them into the segment record of the corresponding adjacent waypoint pair, summarize all segment records, and generate the initial curvature segment set; Generate a set of candidate curvature segments, specifically as follows: Read the segment records, repeated reconnaissance status area boundaries, area attribution marker boundaries, and mission area boundaries from the initial curvature segment set; discretize each entrance and exit circular arc segment into circular arc sampling points at fixed angular intervals, and discretize each intermediate straight segment into straight line sampling points at fixed distance intervals; merge the circular arc sampling points and straight line sampling points to generate a segment sampling point sequence; determine the area number to which each sampling point in the segment sampling point sequence belongs; delete the corresponding segment record when the sampling point falls within the boundary of the repeated reconnaissance status area; delete the corresponding segment record when the area number to which the sampling point belongs is inconsistent with the area attribution marker of the adjacent waypoint; delete the corresponding segment record when the sampling point falls outside the mission area boundary; retain the segment records that have not been deleted, and summarize them according to the order of adjacent waypoint pairs to generate a candidate curvature segment set; By connecting the starting waypoint, the verified extended waypoint sequence, and the ending waypoint according to the candidate curvature segment set, candidate reconnaissance tracks are generated. Specifically, the generation of candidate reconnaissance tracks involves: Read the candidate curvature segment set, starting waypoint, verified extended waypoint sequence, and ending waypoint, and generate an ordered waypoint chain according to the order of the starting waypoint, verified extended waypoint sequence, and ending waypoint. Read adjacent waypoint pairs in the ordered waypoint chain sequentially, and retrieve the corresponding curvature-constrained segment records from the candidate curvature segment set. If an adjacent waypoint pair contains a curvature-constrained segment record, write the curvature-constrained segment record into the segment field of the track record; if no curvature-constrained segment record exists in the adjacent waypoint pair, delete the corresponding ordered waypoint chain. Multiple... When recording curvature-constrained segments, the segments are arranged from shortest to longest, and branch track records are generated for each segment. The segments are written sequentially into the branch track records according to the ordered waypoint chain. The waypoint numbers, waypoint coordinates, segments, segments, segments, directions, and area affiliation markers in the branch track records are sequentially numbered. The coordinates of the endpoints and origins of adjacent segments are checked to see if they are consistent. Branch track records with discontinuous coordinates are deleted, and the retained branch track records are written into the candidate track list to generate candidate reconnaissance tracks.

[0029] In this embodiment, generating updated candidate reconnaissance tracks includes: Read the reconnaissance waypoints, candidate curvature segments, verification extended waypoint sequences, and area discrimination attributes from the candidate reconnaissance tracks. Form waypoint attribute vectors based on verification status, occlusion status, weak target status, repeated reconnaissance status, and area spatial coordinates. Specifically, the formation of waypoint attribute vectors based on verification status, occlusion status, weak target status, repeated reconnaissance status, and area spatial coordinates is as follows: Read the waypoint number, waypoint coordinates, and region number of each reconnaissance waypoint in the candidate reconnaissance track. Based on the region number, read the verification status value, occlusion status value, weak target status value, and repeated reconnaissance status value from the region discrimination attribute. Trim the verification status value, occlusion status value, weak target status value, and repeated reconnaissance status value to the range of 0 to 1 to form a status code. Read the region center coordinates and region boundary coordinates from the region spatial coordinates. Use the minimum x-coordinate, minimum y-coordinate, and altitude reference of the mission region boundary as the coordinate starting value, and use the lateral span, longitudinal span, and altitude span of the mission region boundary as the scaling value to perform coordinate normalization on the waypoint coordinates and generate coordinate codes. Concatenate the verification status code, occlusion status code, weak target status code, repeated reconnaissance status code, and coordinate code in the order of the coordinate codes, and write them into the waypoint number and region number to form a waypoint attribute vector. Based on the waypoint attribute vector, the basic waypoint cost is modified in stages, reducing the passage cost for verification, occlusion, and weak target states, and increasing the passage cost for repeated reconnaissance states, thus generating modified waypoint costs. Specifically, the generation of modified waypoint costs involves: The basic waypoint cost for each reconnaissance waypoint in the candidate reconnaissance track is read. The basic waypoint cost is formed by adding the passage costs corresponding to the adjacent candidate curvature segment lengths, waypoint dwell times, and area affiliation markers. The verification status code, occlusion status code, weak target status code, and repeated reconnaissance status code are read from the waypoint attribute vector. The verification reduction coefficient, occlusion reduction coefficient, weak target reduction coefficient, and repeated reconnaissance enhancement coefficient are read from the mission configuration data. The verification status code is multiplied by the verification reduction coefficient to obtain the verification reduction amount; the occlusion status code is multiplied by the occlusion reduction coefficient to obtain the occlusion reduction amount; the weak target status code is multiplied by the weak target reduction coefficient to obtain the weak target reduction amount; and the repeated reconnaissance status code is multiplied by the repeated reconnaissance enhancement coefficient to obtain the repeated reconnaissance enhancement amount. The verification reduction amount, occlusion reduction amount, and weak target reduction amount are subtracted from the basic waypoint cost, and the repeated reconnaissance enhancement amount is added to obtain the initial correction cost. When the initial correction cost is less than 0, it is set to 0 as the corrected waypoint cost; when the initial correction cost is not less than 0, the initial correction cost is retained as the corrected waypoint cost. Write the corrected waypoint cost back to the corresponding reconnaissance waypoint. Summarize the corrected waypoint costs of adjacent reconnaissance waypoints based on the candidate curvature segment connectivity. Delete candidate reconnaissance tracks that lack candidate curvature segment connectivity and generate updated candidate reconnaissance tracks. Specifically, generating updated candidate reconnaissance tracks involves: Read the candidate track number, reconnaissance waypoint sequence, candidate curvature segment connectivity, and corrected waypoint cost from the candidate reconnaissance track; locate the corresponding reconnaissance waypoints one by one according to the reconnaissance waypoint sequence and write the corrected waypoint cost into the waypoint cost field; retrieve the candidate curvature segment connectivity based on adjacent reconnaissance waypoint numbers, and if a connectivity exists, read the corresponding candidate curvature segment length, segment direction, and segment number to generate an adjacent waypoint cost record; if a connectivity is missing, add the corresponding candidate reconnaissance track to the deletion list; for candidate reconnaissance tracks not written to the deletion list, accumulate the corrected waypoint cost and candidate curvature segment length from the adjacent waypoint cost records in the order from the starting waypoint to the ending waypoint to generate the total track cost and track connectivity sequence; write the total track cost, track connectivity sequence, reconnaissance waypoint sequence, and candidate curvature segment connectivity into the candidate reconnaissance track to obtain the updated candidate reconnaissance track.

[0030] In this embodiment, determining and outputting the updated area reconnaissance situation results with spatial coordinates and the updated UAV reconnaissance track includes: The system reads and updates the reconnaissance waypoints, corrected waypoint costs, candidate curvature segment connectivity, area discrimination attributes, and area spatial coordinates from the candidate reconnaissance tracks. It then summarizes the corrected waypoint costs and segment lengths according to the connection order from the starting waypoint to the ending waypoint, generating a track evaluation value. Specifically, the process of summarizing the corrected waypoint costs and segment lengths according to the connection order from the starting waypoint to the ending waypoint to generate the track evaluation value is as follows: Read the starting waypoint marker, ending waypoint marker, reconnaissance waypoint sequence, corrected waypoint cost, and candidate curvature segment connection relationship from each updated candidate reconnaissance track; using the reconnaissance waypoint corresponding to the starting waypoint marker as the traversal starting point, read adjacent reconnaissance waypoints and corresponding curvature-restricted segments segment by segment according to the candidate curvature segment connection relationship until the reconnaissance waypoint corresponding to the ending waypoint marker is read, forming a track connection order; accumulate the corrected waypoint cost of each reconnaissance waypoint according to the track connection order to obtain the total waypoint cost, and accumulate the segment length of each curvature-restricted segment to obtain the total segment length; read the waypoint cost weight and segment length weight from the mission configuration data, multiply the total waypoint cost by the waypoint cost weight to obtain the waypoint evaluation component, multiply the total segment length by the segment length weight to obtain the segment evaluation component, and add the waypoint evaluation component and the segment evaluation component to generate the track evaluation value; Sort the tracks from low to high based on their evaluation values, and filter out continuous, matching start and end waypoints to determine the target reconnaissance track. Specifically, the process of filtering continuous, matching start and end waypoints for the target reconnaissance track is as follows: Read the updated candidate reconnaissance tracks sorted from low to high based on track evaluation values. Extract the reconnaissance waypoint sequence and candidate curvature segment connectivity from each candidate track. Compare the coordinate distance between the first reconnaissance waypoint in the reconnaissance waypoint sequence and the starting waypoint in the mission constraint data. If the coordinate distance is less than the waypoint matching threshold, write a starting waypoint matching marker. Compare the coordinate distance between the last reconnaissance waypoint in the reconnaissance waypoint sequence and the ending waypoint in the mission constraint data. If the coordinate distance is less than the waypoint matching threshold, write an ending waypoint matching marker. Read adjacent reconnaissance waypoints segment by segment according to the order of the reconnaissance waypoint sequence. Search for candidate curvature segment connectivity between adjacent reconnaissance waypoints. If candidate curvature segment connectivity exists, write a continuous connection marker. Retain updated candidate reconnaissance tracks that simultaneously have a starting waypoint matching marker, an ending waypoint matching marker, and all continuous connection markers. Read the updated candidate reconnaissance track with the lowest track evaluation value from the retained results and determine it as the target reconnaissance track. By binding reconnaissance waypoints, curvature-restricted segments, area discrimination attributes, and area spatial coordinates in the target reconnaissance trajectory, an updated area reconnaissance situation result with spatial coordinates is generated, and the updated UAV reconnaissance trajectory is output. Specifically, generating the updated area reconnaissance situation result with spatial coordinates and outputting the updated UAV reconnaissance trajectory involves: The system reads the reconnaissance waypoint sequence, curvature-restricted segment records, area discrimination attributes, and area spatial coordinates from the target reconnaissance track. It then reads the waypoint number, waypoint coordinates, and area number one by one according to the reconnaissance waypoint sequence, and reads the area boundary coordinates and area center coordinates from the area spatial coordinates according to the area number. The system writes the waypoint number, waypoint coordinates, area number, area boundary coordinates, area center coordinates, and area discrimination attributes into the area situation record, forming a set of area situation records with spatial coordinates. It reads the adjacent waypoint numbers, entrance arc segments, intermediate straight segments, exit arc segments, segment lengths, and segment directions according to the curvature-restricted segment records, and writes them into the track segment record. The system arranges the area situation record set and track segment record according to the connection order from the starting waypoint to the ending waypoint, generating an updated area reconnaissance situation result with spatial coordinates. Finally, it organizes the reconnaissance waypoint sequence, track segment record, waypoint coordinates, segment direction, and segment length into an updated UAV reconnaissance track and outputs it to the UAV track execution terminal.

[0031] refer to Figure 3 A visual intelligence-based drone reconnaissance system includes the following modules: The raw reconnaissance data acquisition module is used to receive data from the airborne visual payload, flight control system, mission configuration terminal and historical record database, and generate raw reconnaissance data of UAV; The spatiotemporal registration module is used to perform temporal matching and spatial correlation on the raw UAV reconnaissance data to generate aligned reconnaissance visual sequences; The reconnaissance region discrimination module is used to construct an improved MobileViT model and perform image patch region reconstruction, region bias injection and reconnaissance decoding to generate a reconnaissance region map; The reconnaissance waypoint map construction module is used to extract reconnaissance waypoints and candidate flight segments, write them into the area discrimination attributes, and generate a reconnaissance waypoint map; The verification track generation module is used to introduce the verification waypoint insertion operator into the Dubins algorithm to generate candidate reconnaissance tracks; The waypoint cost correction module is used to correct waypoint costs based on area discrimination attributes and generate updated candidate reconnaissance tracks. The reconnaissance result output module is used to determine the updated reconnaissance trajectory of the UAV based on waypoint costs and segment connectivity, and generate updated area reconnaissance situation results with spatial coordinates.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to a post-disaster emergency reconnaissance scenario in a mountainous area. The reconnaissance area was 4.8 km², including landslide zones, temporary roads, abandoned buildings, woodlands, and low-lying waterlogged areas. A UAV equipped with a visible light camera and an edge computing unit performed continuous reconnaissance missions. The flight altitude was set to 80m to 120m, the flight speed to 8m / s to 12m / s, the minimum turning radius to 35m, the starting waypoint to the southwest of the area, and the ending waypoint to the northeast of the area.

[0033] During a complete reconnaissance cycle, the system acquires 1860 consecutive reconnaissance images with a resolution of 1920×1080 pixels. Simultaneously, it reads flight attitude, heading angle, and flight speed from the flight control system. The mission area boundary consists of 12 boundary vertices. Historical reconnaissance records contain 96 area numbers, of which 31 areas have historical occlusion markers, 18 areas have historical anomaly markers, and 24 areas have been reconnaissanceed more than 3 times.

[0034] The system assigns image acquisition timestamps to continuous reconnaissance images and status acquisition timestamps to flight attitude, heading angle, and flight speed. The image acquisition frequency is 10 frames / s, and the flight status acquisition frequency is 20 times / s. The system matches flight status data based on image acquisition timestamps. When the time difference between an image frame and the status data exceeds 0.05s, linear interpolation is performed on two adjacent sets of flight status data to obtain the image frame status matching data corresponding to each image frame.

[0035] After completing time matching, the system determines the reconnaissance coverage center and direction for each image frame based on the flight attitude, heading angle, and mission area boundaries. For example, the UAV position corresponding to image frame 425 is (1260.4m, 840.7m, 96.5m) in the mission coordinate system, with a heading angle of 42.3°. The system assigns this image frame to regions 18, 19, and 23. Subsequently, the system retrieves historical reconnaissance records, writing the number of reconnaissances, the most recent reconnaissance time, historical anomaly markers, and historical occlusion markers into the corresponding image frames, generating an aligned reconnaissance visual sequence.

[0036] The system trains the improved MobileViT model using reconnaissance training samples with region discrimination annotations. The training samples consist of 6000 frames, and the validation samples consist of 1200 frames. Region discrimination annotations include normal regions, occluded regions, weak target regions, verification regions, and repeated reconnaissance regions. During training, region classification error, occlusion recognition error, weak target recognition error, and verification discrimination error are used as joint training objectives to update the parameters of the lightweight local coding layer, reconnaissance region reconstruction layer, region bias injection layer, global foldback fusion layer, and reconnaissance decoding layer. After training, the validation set region discrimination accuracy reaches 91.6%, and the verification region recall rate reaches 88.4%.

[0037] In the actual inference process, the system improves the MobileViT model by aligning the reconnaissance visual sequence input. A lightweight local coding layer performs convolutional extraction, inverse residual coding, and image patch unrolling on the image, dividing each frame into 120×68 image patches, forming 8160 image patch sequence positions, while simultaneously generating shallow reconnaissance features and local reconnaissance features. The reconnaissance region reconstruction layer reads the flight pose, heading angle, mission area boundary, and historical reconnaissance records to establish an image patch region projection index. For example, the image patch position (53, 27) in frame 425 is projected to region 19, and the system binds it to the historical occlusion marker and reconnaissance count of region 19. After completing the mapping of all image patches, the system rearranges the image patch sequence according to region number and spatial adjacency, generating an ordered image patch sequence and forming a region historical state vector.

[0038] The region bias injection layer extracts occlusion response and weak target response from local reconnaissance features. In forest areas, some targets are occluded by tree canopies, with an average occlusion response of 0.76. In distant road edge areas, the average weak target response is 0.64, higher than the 0.22 in ordinary background areas. The system combines the region's historical state vector, occlusion response, and weak target response to generate a region bias sequence, and then adds the region bias sequence to the ordered image patch sequence to form a biased image patch sequence.

[0039] After the biased image patch sequence enters the global foldback fusion layer, the system performs global encoding and image patch foldback to generate global reconnaissance features. Subsequently, shallow reconnaissance features, local reconnaissance features, and global reconnaissance features are fused to obtain the fused reconnaissance features. Compared to the MobileViT model without region bias, the feature response of occluded regions is improved by an average of approximately 24%, and the feature response of weak target regions is improved by an average of approximately 19%.

[0040] The reconnaissance decoding layer generates region discrimination attributes and region spatial coordinates based on the fused reconnaissance features, and writes the region discrimination attributes into the region grid according to the task region boundaries. The system ultimately generates 84 reconnaissance region grids, including 47 normal regions, 13 occluded regions, 9 weak target regions, 11 verification regions, and 4 repeated reconnaissance regions. Each region in the reconnaissance region map is bound to region spatial coordinates, region discrimination attributes, and historical reconnaissance status.

[0041] The system constructs a reconnaissance waypoint map based on the reconnaissance area map, starting waypoint, and ending waypoint. The center point of the area grid, the turning points of the area boundary, the starting waypoint, and the ending waypoint are converted into reconnaissance waypoints, generating a total of 126 reconnaissance waypoints. The system establishes candidate flight segment connections based on the reconnaissance waypoint coordinates and the adjacency relationships of the area grid, and deletes candidate segments that cross the mission area boundary or are impassable, leaving 214 candidate segments. The system then matches the area discrimination attributes to the corresponding reconnaissance waypoints, forming the reconnaissance waypoint map.

[0042] During the candidate reconnaissance track generation phase, the system extracts the reconnaissance waypoint status from the area discrimination attributes, filters out reconnaissance waypoints with verification status, obstruction status, or weak target status but without repeated reconnaissance status, resulting in 27 waypoints to be inserted for verification. The system combines candidate segments, heading angle, and minimum turning radius to determine insertion feasibility, retains 18 insertable waypoints, and constructs a verification waypoint insertion operator from the insertable waypoints, corresponding candidate segments, and segment reconnection order.

[0043] The system incorporates a verification waypoint insertion operator into the Dubins algorithm. The Dubins algorithm expands candidate waypoints starting from the initial waypoint. The verification waypoint insertion operator inserts insertable verification waypoints into the candidate segments and splits the original candidate segments into a first segment and a second segment, generating a sequence of expanded verification waypoints. Subsequently, the system connects the initial waypoint, the sequence of expanded verification waypoints, and the final waypoint based on the heading angle and minimum turning radius, generating an initial set of curvature segments. Curvature-restricted segments that traverse areas with repeated reconnaissance states or cross area ownership boundary markers are removed, generating a set of candidate curvature segments.

[0044] During the waypoint cost correction phase, the system reads reconnaissance waypoints, candidate curvature segments, verification extended waypoint sequences, and area discrimination attributes from candidate reconnaissance tracks, and forms waypoint attribute vectors according to verification status, obstruction status, weak target status, repeated reconnaissance status, and area spatial coordinates. The passage cost of waypoints with verification status, obstruction status, and weak target status is reduced, while the passage cost of waypoints with repeated reconnaissance status is increased. After correction, the number of candidate reconnaissance tracks is reduced from 16 to 7, and the average length of invalid duplicate coverage is reduced from 620m to 210m.

[0045] The system summarizes and corrects waypoint costs and segment lengths according to the connection order from the starting waypoint to the ending waypoint, generating a track evaluation value. After sorting and filtering, the system determines the target reconnaissance track, with a total track length of 6.7km, including 14 verification waypoints and 3 curvature-restricted segments that bypass repeated reconnaissance areas. The system binds the reconnaissance waypoints, curvature-restricted segments, area discrimination attributes, and area spatial coordinates in the target reconnaissance track, generating an updated area reconnaissance situation result with spatial coordinates, and outputs the updated UAV reconnaissance track.

[0046] In the comparative experiment, the system compared the method of this invention with the traditional MobileViT recognition combined with the ordinary Dubins track generation method, keeping the test area, flight altitude, number of images, and mission start and end points consistent. On the same batch of test data, the accuracy of occluded area recognition increased from 76.8% to 89.5%, the recall rate of weak target areas increased from 61.4% to 82.7%, the coverage rate of areas awaiting verification increased from 68.2% to 91.3%, the proportion of repeated reconnaissance segments decreased from 18.6% to 7.4%, the average length of updated reconnaissance tracks decreased from 7.4km to 6.7km, the curvature constraint satisfaction rate increased from 92.1% to 98.6%, and the accuracy of regional reconnaissance situational spatial matching increased from 79.3% to 90.8%.

[0047] As can be seen from the embodiments, the present invention can form a unified processing chain of continuous reconnaissance images, flight poses, mission area boundaries and historical reconnaissance records in actual operation. It generates region discrimination attributes by improving the MobileViT model, and then generates and updates the reconnaissance track by introducing the Dubins algorithm with the verification waypoint insertion operator. This allows occluded areas, weak target areas and areas to be verified to enter the track adjustment process in a timely manner, verifying the engineering feasibility of the present invention in complex UAV reconnaissance scenarios.

[0048] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A visual intelligence-based unmanned aerial vehicle (UAV) reconnaissance method, characterized in that, Includes the following steps: Collect continuous reconnaissance images, flight attitude, heading angle, flight speed, mission area boundary, historical reconnaissance records, minimum turning radius, starting waypoint and ending waypoint to obtain raw UAV reconnaissance data; Based on the flight attitude, heading angle, mission area boundary and historical reconnaissance records, the original reconnaissance data of the UAV is spatiotemporally registered to obtain an aligned reconnaissance visual sequence. A reconstructed detection region layer is added after the image patch unfolding layer of the MobileViT model, and a region bias injection layer is added before the global encoding layer of the image patch to form an improved MobileViT model. The aligned detection visual sequence is input into the improved MobileViT model to generate a detection region map containing region discrimination attributes. A reconnaissance waypoint map is constructed based on the reconnaissance area map, the starting waypoint, and the ending waypoint, and the area discrimination attribute is written into the reconnaissance waypoint; The Dubins algorithm, which incorporates a verification waypoint insertion operator, is used to expand candidate waypoints, insert verification waypoints, and connect curvature-constrained segments on the reconnaissance waypoint map, generating candidate reconnaissance tracks. Adjust the waypoint cost based on the regional discrimination attribute of the waypoints in the candidate reconnaissance tracks, and generate updated candidate reconnaissance tracks; Based on the waypoint costs and segment connectivity in the updated candidate reconnaissance tracks, the updated regional reconnaissance situation results with spatial coordinates and the updated UAV reconnaissance tracks are determined and output.

2. The UAV reconnaissance method based on visual intelligence according to claim 1, characterized in that, The obtained raw drone reconnaissance data includes: The system acquires continuous reconnaissance images using an airborne visual payload, reads flight attitude, heading angle, and flight speed from the flight control system, and binds the corresponding acquisition timestamps to generate visual state data. Read the mission area boundary, starting waypoint, ending waypoint and minimum turning radius from mission configuration data, airframe parameters and flight safety configuration to generate mission constraint data; The region number, number of reconnaissances, most recent reconnaissance time, historical anomaly markers, and historical occlusion markers are read from historical reconnaissance records. Visual state data, mission constraint data, and historical reconnaissance records are packaged according to UAV mission number, region number, and data collection timestamp to obtain the original UAV reconnaissance data.

3. The UAV reconnaissance method based on visual intelligence according to claim 1, characterized in that, The obtained aligned detection visual sequence includes: Based on the image acquisition timestamp, the flight pose, heading angle and flight speed are matched. When the sampling interval between adjacent states is exceeded, linear interpolation is performed to generate image frame state matching data. Based on the flight pose, heading angle and mission area boundary in the image frame state matching data, the reconnaissance coverage center, reconnaissance coverage direction and mission area attribution are determined, and image frame area matching data is generated. Historical reconnaissance records are retrieved based on the task area affiliation, and the number of reconnaissances, the most recent reconnaissance time, historical anomaly markers, and historical occlusion markers are written into the corresponding image frames. The aligned reconnaissance visual sequence is generated by sorting the images by their acquisition timestamps.

4. The UAV reconnaissance method based on visual intelligence according to claim 1, characterized in that, The formation of the improved MobileViT model includes: Read the convolutional input layer, inverse residual local coding layer, image patch unrolling layer, image patch global coding layer, image patch folding layer and task output layer in the MobileViT model, and retain the order of convolutional input, inverse residual local coding, image patch unrolling, global coding and folding; A reconnaissance region reconstruction layer is added between the image patch unfolding layer and the image patch global coding layer. The reconnaissance region reconstruction layer receives the image patch sequence and generates an ordered image patch sequence based on the flight pose, heading angle, mission area boundary and historical reconnaissance records. A region bias injection layer is added between the reconstructed region layer and the global encoding layer of image patches. The region bias injection layer receives the ordered image patch sequence of the region and reads the local reconnaissance features output by the inverse residual local encoding layer. It extracts the occlusion response and weak target response from the local reconnaissance features to generate the bias image patch sequence. The reconnaissance decoding layer replaces the task output layer. The reconnaissance decoding layer receives global reconnaissance features generated by the image patch foldback layer and outputs region discrimination attributes to form an improved MobileViT model. The region discrimination attributes include region category, occlusion state, weak target state, verification state, and repeated reconnaissance state.

5. The UAV reconnaissance method based on visual intelligence according to claim 1, characterized in that, The step of inputting aligned reconnaissance visual sequences into the improved MobileViT model to generate a reconnaissance region map containing region discrimination attributes includes: An improved MobileViT model is constructed, which includes a lightweight local coding layer, a reconstructed reconstructed region layer, a region bias injection layer, a global foldback fusion layer, and a reconstructed decoding layer. The aligned reconnaissance visual sequence is input into a lightweight local coding layer, where convolutional extraction, inverse residual coding, and image patch unrolling are performed to generate shallow reconnaissance features, local reconnaissance features, and image patch sequences. The image patch sequence is input into the reconstructed reconnaissance area layer. Based on the flight pose, heading angle, mission area boundary and historical reconnaissance records, an image patch region projection index is established. The image patch sequence is rearranged according to the image patch region projection index to generate an ordered image patch sequence and a region historical state vector. The ordered image patch sequence, the historical state vector of the region, and the local reconnaissance features are input into the region bias injection layer. Occlusion response and weak target response are extracted from the local reconnaissance features. The region bias sequence is generated based on the historical state vector of the region, the occlusion response, and the weak target response. The region bias sequence is added to the ordered image patch sequence of the region to generate the bias image patch sequence. The biased image patch sequence is input into the global foldback fusion layer, and global encoding and image patch foldback are performed to generate global detection features. Shallow detection features, local detection features and global detection features are then fused to generate fused detection features. The fused reconnaissance features are input into the reconnaissance decoding layer to generate region discrimination attributes and region spatial coordinates. The region discrimination attributes are written into the region grid according to the task region boundary to generate a reconnaissance region map containing the region discrimination attributes. An improved MobileViT model was trained using reconnaissance training samples with region discrimination annotations. Region classification error, occlusion recognition error, weak target recognition error, verification discrimination error, and repeated reconnaissance recognition error were used as joint training objectives to update the parameters of the lightweight local coding layer, reconnaissance region reconstruction layer, region bias injection layer, global foldback fusion layer, and reconnaissance decoding layer, thus obtaining the trained improved MobileViT model.

6. The UAV reconnaissance method based on visual intelligence according to claim 1, characterized in that, The construction of the reconnaissance waypoint map and the writing of region discrimination attributes into the reconnaissance waypoints include: Read the region grid, region spatial coordinates and region discrimination attributes in the reconnaissance area map, convert the region grid center point, boundary turning point, starting waypoint and ending waypoint into reconnaissance waypoints, and generate reconnaissance waypoint coordinates; Establish candidate route connection relationships based on reconnaissance waypoint coordinates and regional grid adjacency relationships, delete candidate routes that cross mission area boundaries and impassable areas, and generate an initial reconnaissance waypoint map; The region's spatial coordinates are used to match the region's discrimination attributes to the reconnaissance waypoints, and the starting waypoint marker, ending waypoint marker, region affiliation marker, and candidate segment connection relationship are written to generate a reconnaissance waypoint map.

7. The UAV reconnaissance method based on visual intelligence according to claim 1, characterized in that, The generated candidate reconnaissance tracks include: Read the reconnaissance waypoint map and the original reconnaissance data of the UAV, extract the reconnaissance waypoint status from the area discrimination attributes, filter the reconnaissance waypoints that have the verification status, the obstruction status or the weak target status but do not have the repeated reconnaissance status, and generate a set of waypoints to be inserted for verification. Taking candidate segments as the processing object, the feasibility of insertion is judged by combining the set of waypoints to be inserted for verification, the heading angle and the minimum turning radius. Insertable waypoints for verification are retained, and the verification waypoint insertion operator is constructed by insertable verification waypoints, corresponding candidate segments and segment reconnection order. The verification waypoint insertion operator is introduced into the Dubins algorithm. The Dubins algorithm expands candidate waypoints from the starting waypoint. The verification waypoint insertion operator inserts insertable verification waypoints into the candidate segments and splits the candidate segments to generate a sequence of verification-expanded waypoints. The Dubins algorithm connects the starting waypoint, the extended waypoint sequence, and the ending waypoint based on the heading angle and minimum turning radius to generate an initial curvature segment set. It then deletes curvature-restricted segments that cross repeated reconnaissance status areas or cross area ownership marker boundaries, generating a candidate curvature segment set. Candidate reconnaissance tracks are generated by connecting the starting waypoint, the extended waypoint sequence, and the ending waypoint according to the set of candidate curvature segments.

8. The UAV reconnaissance method based on visual intelligence according to claim 1, characterized in that, The generated and updated candidate reconnaissance tracks include: Read the reconnaissance waypoints, candidate curvature segments, verification extended waypoint sequences and area discrimination attributes from the candidate reconnaissance tracks, and form waypoint attribute vectors according to verification status, occlusion status, weak target status, repeated reconnaissance status and area spatial coordinates; Based on the waypoint attribute vector, perform itemized corrections on the basic waypoint cost, reduce the passage cost corresponding to the verification state, the occlusion state and the weak target state, increase the passage cost corresponding to the repeated reconnaissance state, and generate the corrected waypoint cost; Write the corrected waypoint cost back to the corresponding reconnaissance waypoint, summarize the corrected waypoint costs of adjacent reconnaissance waypoints based on the candidate curvature segment connectivity, delete candidate reconnaissance tracks that lack candidate curvature segment connectivity, and generate updated candidate reconnaissance tracks.

9. The UAV reconnaissance method based on visual intelligence according to claim 1, characterized in that, The process of determining and outputting the updated area reconnaissance situation results with spatial coordinates and the updated UAV reconnaissance track includes: Read and update the reconnaissance waypoints, corrected waypoint costs, candidate curvature segment connections, area discrimination attributes, and area spatial coordinates from the candidate reconnaissance tracks. Summarize the corrected waypoint costs and segment lengths in the connection order from the starting waypoint to the ending waypoint to generate track evaluation values. Sort the tracks from low to high according to their evaluation values, filter out the continuous and matching start and end waypoints of the updated candidate reconnaissance tracks, and determine the target reconnaissance track. Bind the reconnaissance waypoints, curvature-restricted segments, area discrimination attributes, and area spatial coordinates in the target reconnaissance track to generate updated area reconnaissance situation results with spatial coordinates, and output the updated UAV reconnaissance track.

10. A visual intelligence-based unmanned aerial vehicle (UAV) reconnaissance system, comprising executing the visual intelligence-based UAV reconnaissance method according to any one of claims 1 to 9, characterized in that, Includes the following modules: The raw reconnaissance data acquisition module is used to receive data from the airborne visual payload, flight control system, mission configuration terminal and historical record database, and generate raw reconnaissance data of UAV; The spatiotemporal registration module is used to perform temporal matching and spatial correlation on the raw UAV reconnaissance data to generate aligned reconnaissance visual sequences; The reconnaissance region discrimination module is used to construct an improved MobileViT model and perform image patch region reconstruction, region bias injection and reconnaissance decoding to generate a reconnaissance region map; The reconnaissance waypoint map construction module is used to extract reconnaissance waypoints and candidate flight segments, write them into the area discrimination attributes, and generate a reconnaissance waypoint map; The verification track generation module is used to introduce the verification waypoint insertion operator into the Dubins algorithm to generate candidate reconnaissance tracks; The waypoint cost correction module is used to correct waypoint costs based on area discrimination attributes and generate updated candidate reconnaissance tracks. The reconnaissance result output module is used to determine the updated reconnaissance trajectory of the UAV based on waypoint costs and segment connectivity, and generate updated area reconnaissance situation results with spatial coordinates.

Images