Unmanned aerial vehicle accessory intelligent identification and storage method and system based on sensor positioning

Abstract

The application discloses a sensor positioning-based unmanned aerial vehicle accessory intelligent identification and storage method and system, and relates to the technical field of intelligent transportation.The method comprises the following steps: step one, obtaining accessory image data, accessory sensor positioning data and accessory storage management data of a target unmanned aerial vehicle accessory; step two, performing real-time positioning on the target unmanned aerial vehicle accessory by using a multi-sensor residual positioning algorithm; step three, determining a target image of the unmanned aerial vehicle accessory; step four, inputting the target image of the unmanned aerial vehicle accessory into an improved RTMDet identification network to perform accessory identification; step five, establishing a binding relationship between an accessory identity and a warehouse location; step six, performing ternary consistency verification on the target unmanned aerial vehicle accessory and generating an accessory intelligent storage strategy; and step seven, updating the accessory inventory and generating a traceability record.The application improves the positioning identification and storage accuracy of unmanned aerial vehicle accessories by using a multi-sensor residual positioning algorithm and an improved RTMDet identification network.

Description

Technical Field

[0001] This invention relates to the field of object recognition technology, and in particular to a method and system for intelligent identification and retrieval of drone accessories based on sensor positioning. Background Technology

[0002] With the large-scale application of drones in inspection, surveying, security, emergency rescue, and low-altitude operations, the types, models, and circulation frequency of drone accessories are continuously increasing. Intelligent identification and storage management technologies for accessories such as batteries, motors, propellers, gimbals, arms, and flight control boards have received widespread attention. Current drone accessory management methods mainly rely on manual registration, barcode scanning, RFID identification, ordinary shelf ledgers, or single image recognition devices for accessory warehousing, retrieval, and return management. However, these methods commonly suffer from the following problems in practical applications: Drone parts often exhibit high visual similarity and subtle model differences, especially batteries from the same series, propellers with different rotation directions, motors of similar specifications, and gimbal modules with similar interfaces. Traditional manual identification or ordinary target detection methods struggle to reliably distinguish part categories, models, and specifications, leading to frequent misidentification and incorrect retrieval. In warehouse areas, the storage locations of parts change frequently, and single-sensor positioning is susceptible to obstruction, signal attenuation, and interference from shelf structures, resulting in inaccurate determination of the shelf, storage compartment, and spatial coordinates of parts. This can cause discrepancies between inventory records and actual locations. Existing storage and retrieval management systems typically handle part identification, location positioning, and retrieval tasks separately, lacking consistency verification between part identity, storage location, and retrieval tasks. This makes it difficult to promptly detect misplaced, incorrectly retrieved, returned incorrectly, and abnormal states of parts. Furthermore, the lack of correlation between information on the person handling the parts, retrieval time, usage task, and compatible drone information during the storage and retrieval process hinders the formation of complete end-to-end traceability records, impacting the accuracy, efficiency, and operational safety of drone part management.

[0003] Therefore, how to provide a method and system for intelligent identification and retrieval of drone accessories based on sensor positioning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a sensor-based intelligent identification and storage method and system for drone parts. This invention fully utilizes a multi-sensor residual positioning algorithm, an improved RTMDet identification network, and a codebook quantization mechanism. It details the implementation process of real-time drone part positioning, target image extraction, part target detection, fine-grained model discrimination, part identity and storage location binding, ternary consistency verification, and full-process traceability. This enables collaborative management of drone part identity, storage compartment, shelf number, storage location coordinates, and storage / retrieval tasks. It boasts advantages such as high part identification accuracy, strong storage location reliability, high storage / retrieval verification efficiency, timely warnings of misplacement and misretrieval, and good traceability of part circulation.

[0005] The intelligent identification and retrieval method for UAV accessories based on sensor positioning according to an embodiment of the present invention includes the following steps: Step 1: Acquire the accessory image data, accessory sensor positioning data, and accessory storage and management data of the target drone accessories; Step 2: Based on the sensor positioning data of the accessory, the target drone accessory is located in real time using a multi-sensor residual positioning algorithm to obtain the real-time positioning result of the drone accessory; Step 3: Based on the accessory image data and the real-time positioning results of the drone accessories, determine the candidate image regions of the target drone accessories, and perform image preprocessing on the candidate image regions to obtain the target image of the drone accessories; Step 4: Input the target image of the drone accessory into the improved RTMDet recognition network, perform accessory target detection and fine-grained model discrimination, and obtain the drone accessory recognition result; the improved RTMDet recognition network introduces a codebook quantization mechanism in the accessory feature fusion process; Step 5: Based on the drone accessory identification results, drone accessory real-time positioning results, and accessory storage and retrieval management data, establish an accessory identity and storage location binding relationship; Step 6: Based on the accessory identity storage location binding relationship, perform a ternary consistency check on the target drone accessories, obtain the accessory access consistency check result, and generate an intelligent accessory access strategy based on the accessory access consistency check result; Step 7: Based on the intelligent storage and retrieval strategy for accessories, update the inventory status of drone accessories and generate a full-process traceability record for drone accessories.

[0006] Optionally, the accessory image data includes images of the target storage area, shelf area, storage compartment, and image acquisition time; The accessory sensor positioning data includes accessory positioning tag identification, positioning acquisition time, positioning signal strength, positioning distance value, positioning angle value, initial positioning coordinate value, and in-situ detection status; The spare parts storage and retrieval management data includes spare parts location information, spare parts inventory ledger data, spare parts storage and retrieval task data, and spare parts usage-related data.

[0007] Optionally, step two specifically includes: Determine the location data group corresponding to the target drone accessory based on the accessory location tag identification; The positioning data set includes the positioning signal strength, positioning distance value, positioning angle value, initial positioning coordinate value, and in-situ detection status of the target UAV accessory collected by multiple positioning sensors at the same positioning acquisition time. The location confidence weights are obtained by exponentially normalizing the strength of each location signal. The spatial coordinates of the target UAV component at the current positioning and data acquisition time are used as candidate spatial coordinates; Calculate the coordinate difference between the candidate spatial coordinates and the installation coordinates of the r-th positioning sensor to obtain the relative coordinate vector, and calculate the L2 norm of the relative coordinate vector to obtain the candidate ranging value. The difference between the candidate ranging value and the positioning distance value is used as the distance residual corresponding to the r-th positioning sensor. Convert each positioning angle value into a direction unit vector; Perform an inner product operation between the relative coordinate vector and the direction unit vector, and use the ratio of the inner product operation result to the candidate ranging value as the direction consistency value; calculate the difference between the constant 1 and the direction consistency value to obtain the angle residual corresponding to the r-th positioning sensor; The sum of the squares of the distance residual and the squares of the angle residual is multiplied by the positioning confidence weight to obtain the weighted observation residual term corresponding to the r-th positioning sensor. The weighted observation residuals corresponding to all positioning sensors are summed to obtain the positioning observation residuals. The squared coordinate deviation between the candidate spatial coordinates and the initial positioning coordinates is used as the initial value constraint term. The positioning observation residual term and the initial value constraint term are added together to construct the spatial coordinate optimization objective function for the target UAV parts; The spatial coordinate optimization objective function is minimized to obtain the spatial coordinates of the target UAV accessory at the current positioning and data acquisition time. Based on the parts storage location information, the spatial coordinates of the parts are matched with the storage location coordinates corresponding to each storage slot to determine the storage slot number corresponding to the target drone parts. Based on the storage compartment number, read the shelf number and storage location coordinates from the parts storage location information; The in-situ detection status is taken as the accessory in-situ status, and the accessory movement speed of the target UAV accessory is calculated based on the accessory's spatial coordinates at the current positioning and acquisition time and the accessory's spatial coordinates at the previous positioning and acquisition time. When the moving speed of the accessory is greater than the moving speed threshold, the moving state of the accessory of the target drone is determined to be moving; when the moving speed of the accessory is less than or equal to the moving speed threshold, the moving state of the accessory of the target drone is determined to be stationary. The spatial coordinates of the parts, the storage compartment number, the shelf number, the warehouse location coordinates, the location status of the parts, and the movement status of the parts are used as the real-time positioning results of the drone parts.

[0008] Optionally, step three specifically includes: Based on the shelf number and storage compartment number, select the storage compartment image that matches both the shelf number and storage compartment number from the accessory image data, and use the storage compartment image with the smallest time difference between the image acquisition time and the current positioning acquisition time as the target accessory positioning associated image. Calculate the difference between the spatial coordinates of the accessory and the coordinates of the storage location to obtain the relative spatial coordinates of the target drone accessory relative to the corresponding storage slot; The relative spatial coordinates are converted into homogeneous coordinates and mapped onto the target accessory positioning association image through the projection matrix of the pre-calibrated image acquisition device to obtain the accessory projection center coordinates of the target UAV accessory; Using the coordinates of the component projection center as the center point of the region, and the projection width and projection height of the grid corresponding to the storage grid number as the region size, the region size is expanded outward according to the positioning error radius to obtain the candidate image region. Candidate image regions are extracted from the target component location association image to obtain candidate region images; Bilateral filtering is applied to the candidate region image to denoise it, resulting in a denoised candidate region image. The candidate region image for denoising is subjected to grayscale linear normalization to obtain a normalized candidate region image. Based on the radial and tangential distortion parameters of the image acquisition device, the Brown-Conrady distortion correction method is used to correct the distortion of the normalized candidate region image to obtain the corrected candidate region image. Background sample pixels are extracted from the edge ring region of the correction candidate region image. A background reference image with the same size as the correction candidate region image is constructed based on the pixel mean of the background sample pixels. A target foreground mask is constructed based on the pixel difference between the correction candidate region image and the background reference image. The target region boundary of the target UAV accessory is determined based on the target foreground mask, and the target region is cropped from the candidate region image according to the target region boundary to obtain the target image of the UAV accessory.

[0009] Optionally, the improved RTMDet recognition network includes a component backbone feature extraction module, a multi-scale component feature fusion module, a component detection output module, and a model fine-grained discrimination module; The component backbone feature extraction module performs five-stage convolutional feature extraction on the target image of UAV components. The output feature map of the third stage is used as the shallow component backbone feature map, the output feature map of the fourth stage is used as the middle component backbone feature map, and the output feature map of the fifth stage is used as the deep component backbone feature map. The multi-scale component feature fusion module performs channel mapping and scale alignment fusion on the shallow component backbone feature map, the middle component backbone feature map and the deep component backbone feature map respectively through the path aggregation pyramid fusion structure to obtain the shallow component fusion feature map, the middle component fusion feature map and the deep component fusion feature map. The multi-scale component feature fusion module introduces a codebook quantization mechanism to perform key structure prototype matching, discrete quantization, and residual back-injection enhancement on the shallow component fusion feature map, the middle component fusion feature map, and the deep component fusion feature map, thereby obtaining shallow component enhanced feature map, middle component enhanced feature map, and deep component enhanced feature map. The shallow, medium, and deep component enhancement feature maps are input into the component detection output module, and classification prediction and bounding box regression prediction are performed respectively to obtain candidate detection results for each feature level. The candidate detection results include candidate accessory bounding boxes, candidate accessory categories, and candidate recognition confidence scores. Construct a detection confidence threshold and a target box overlap threshold, delete candidate detection results whose candidate recognition confidence is less than the detection confidence threshold, and obtain the initial screening detection results; The initial screening results are processed by non-maximum suppression according to the candidate recognition confidence level from high to low. The initial screening results that belong to the same candidate accessory category as the retained initial screening results and whose target box intersection-union ratio is greater than the target box overlap threshold are deleted to obtain the accessory target detection results. The accessory target detection results include the accessory target bounding box position, accessory category, and recognition confidence level; The accessory target detection results are input into the model fine-grained discrimination module. Based on the position of the accessory target box, local visual features are extracted from the accessory enhanced feature map of the corresponding feature level. Global pooling and multi-branch classification are performed on the local visual features to obtain the accessory model, accessory specifications and compatible drone type. Based on accessory category, accessory model, accessory specifications, and compatible drone type, the corresponding accessory identification can be retrieved from the accessory inventory ledger data; The accessory identification result is determined by the accessory's identity identifier, accessory category, accessory model, accessory specifications, compatible drone type, accessory target box position, and recognition confidence level.

[0010] Optionally, the codebook quantization mechanism specifically includes: Construct a key structure prototype codebook, which includes multiple key structure prototype vectors. These key structure prototype vectors are used to characterize the battery contact structure, motor mounting hole structure, propeller edge structure, gimbal connection interface structure, arm snap-fit ​​structure, and flight control board plug-in interface structure of UAV accessories. The shallow component fusion feature map, the middle component fusion feature map, and the deep component fusion feature map are respectively mapped to the key structure feature space to obtain the shallow key structure feature map, the middle key structure feature map, and the deep key structure feature map. For any key structural feature map at any feature level, the distance between the structural feature vector corresponding to spatial location u and each key structural prototype vector in the key structural prototype codebook is calculated, and the key structural prototype vector with the smallest distance is taken as the quantized structural feature vector corresponding to spatial location u. The quantized structural feature vectors corresponding to each spatial location in the same feature level are combined to obtain a quantized structural feature map, and the residual between the key structural feature map and the quantized structural feature map is calculated to obtain a quantized residual feature map. The quantized structural feature map and the quantized residual feature map are concatenated and then back-injected to the component fusion feature map of the corresponding feature level through back-injection mapping parameters to obtain shallow component enhancement feature map, mid-level component enhancement feature map and deep component enhancement feature map.

[0011] Optionally, the accessory identity storage location binding relationship specifically includes: the binding relationship between accessory identity identifier and storage compartment number; the matching relationship between accessory category, accessory model, accessory specification and storage location bound accessory type; the position mapping relationship between accessory spatial coordinates and storage location coordinates; the status correspondence between accessory in-situ status and accessory inventory ledger data; and the association relationship between shelf number, storage compartment number and accessory identity identifier.

[0012] Optionally, the ternary consistency check includes: Perform a consistency check between the accessory identification identifier in the drone accessory identification result and the target accessory identification identifier in the target access task requirements; The target access task requirements include the target accessory identification, target accessory category, target accessory model, and target accessory specifications; The consistency of the accessory category, accessory model and accessory specifications in the drone accessory identification results with the accessory type bound to the target storage location information is verified. Perform a consistency check between the storage compartment number in the real-time location results of drone parts and the target storage compartment number in the target storage location information. The target storage location information includes the target shelf number, the target storage compartment number, the target storage location coordinates, and the type of storage location-bound accessories corresponding to the target storage location; Perform consistency verification between the real-time location results of drone parts and the location data of the parts inventory ledger. Perform a consistency check between the movement status of the drone accessories in the real-time location results and the current access operation type. The intelligent storage and retrieval strategy for spare parts includes a spare parts inbound control strategy, a spare parts outbound control strategy, a spare parts return verification strategy, a spare parts misplacement early warning strategy, and a spare parts mis-retrieval early warning strategy.

[0013] Optionally, the inventory status of the drone parts includes parts inbound status, parts outbound status, parts returned status, parts in place status, parts location binding status, and parts abnormal status. The full-process traceability record for drone accessories includes accessory identification, accessory category, accessory model, accessory specifications, shelf number, storage compartment number, storage location coordinates, accessory spatial coordinates, accessor information, access time, usage task, compatible drone information, identification confidence level, current access operation type, and abnormal warning information.

[0014] According to an embodiment of the present invention, a sensor-based intelligent identification and retrieval system for unmanned aerial vehicle (UAV) accessories includes: Data acquisition module: used to acquire accessory image data, accessory sensor positioning data, and accessory storage and management data of the target drone accessories; Multi-sensor residual positioning module: Used to locate the target drone parts in real time based on the positioning data of the parts' sensors and through the multi-sensor residual positioning algorithm, so as to obtain the real-time positioning result of the drone parts; Candidate image region processing module: Based on accessory image data and real-time positioning results of drone accessories, it determines the candidate image region of the target drone accessory, performs image preprocessing on the candidate image region, and obtains the target image of the drone accessory; Improved RTMDet Recognition Module: This module is used to input the target image of the UAV accessory into the improved RTMDet recognition network, perform accessory target detection and fine-grained model discrimination, and obtain the UAV accessory recognition result. The improved RTMDet recognition network introduces a codebook quantization mechanism in the accessory feature fusion process. Parts Identification and Storage Location Binding Module: Used to establish a parts identification and storage location binding relationship based on the drone parts identification results, drone parts real-time positioning results, and parts storage and retrieval management data; Verification and Strategy Generation Module: Based on the accessory identity storage location binding relationship, it performs a ternary consistency verification on the target UAV accessories, obtains the accessory access consistency verification result, and generates an intelligent accessory access strategy based on the accessory access consistency verification result. Inventory traceability module: Used to update the inventory status of drone parts and generate a full-process traceability record of drone parts based on the intelligent storage and retrieval strategy for parts.

[0015] The beneficial effects of this invention are: This invention fuses and solves the sensor positioning data of drone parts using a multi-sensor residual localization algorithm. It utilizes positioning signal strength, positioning distance, positioning angle, and initial positioning coordinates to jointly constrain the spatial coordinates of the target drone parts, reducing errors caused by single-sensor positioning due to occlusion, signal attenuation, and interference from rack structures, thus improving the reliability of real-time positioning results for drone parts. Furthermore, an improved RTMDet recognition network performs target detection and fine-grained model discrimination on the drone part target image. A codebook quantization mechanism is introduced during the part feature fusion process, mapping the fused features of parts at different scales to the key structural prototype codebook. Through key structural prototype matching, discrete quantization, and residual back-injection enhancement, the local structural representations of battery contact structures, motor mounting hole structures, propeller edge structures, gimbal connection interface structures, arm latching structures, and flight control board interface structures are strengthened, improving the accuracy of identifying the category, model, and specifications of similar drone parts. Meanwhile, this invention performs a ternary consistency check based on the drone accessory identification results, real-time drone accessory positioning results, and accessory storage and retrieval task data to determine whether there is consistency between accessory identity, storage slot, storage location binding type, and storage and retrieval task. Based on this, an intelligent accessory storage and retrieval strategy is generated, which can reduce problems such as misplacement, incorrect retrieval, and incorrect return location of accessories. Therefore, this invention can improve the positioning reliability, fine-grained identification accuracy, storage and retrieval verification accuracy, and end-to-end traceability capability of drone accessories in storage scenarios. Attached Figure Description

[0016] 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 schematic diagram of the intelligent identification and retrieval method and system for UAV accessories based on sensor positioning proposed in this invention; Figure 2 This is a flowchart of the intelligent identification and storage method for UAV accessories based on sensor positioning proposed in this invention and the multi-sensor residual positioning algorithm in the system. Figure 3 This is a flowchart of the structure of the improved RTMDet identification network in the sensor-based intelligent identification and retrieval method for UAV accessories proposed in this invention. Detailed Implementation

[0017] 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.

[0018] refer to Figures 1-3 A sensor-based intelligent identification and retrieval method for drone accessories includes the following steps: Step 1: Acquire the accessory image data, accessory sensor positioning data, and accessory storage and management data of the target drone accessories; Step 2: Based on the sensor positioning data of the accessory, the target drone accessory is located in real time using a multi-sensor residual positioning algorithm to obtain the real-time positioning result of the drone accessory; Step 3: Based on the accessory image data and the real-time positioning results of the drone accessories, determine the candidate image regions of the target drone accessories, and perform image preprocessing on the candidate image regions to obtain the target image of the drone accessories; Step 4: Input the target image of the drone accessory into the improved RTMDet recognition network, perform accessory target detection and fine-grained model discrimination, and obtain the drone accessory recognition result; the improved RTMDet recognition network introduces a codebook quantization mechanism in the accessory feature fusion process; Step 5: Based on the drone accessory identification results, drone accessory real-time positioning results, and accessory storage and retrieval management data, establish an accessory identity and storage location binding relationship; Step 6: Based on the accessory identity storage location binding relationship, perform a ternary consistency check on the target drone accessories, obtain the accessory access consistency check result, and generate an intelligent accessory access strategy based on the accessory access consistency check result; Step 7: Based on the intelligent storage and retrieval strategy for accessories, update the inventory status of drone accessories and generate a full-process traceability record for drone accessories.

[0019] In this embodiment, the accessory image data includes images of the target storage area, shelf area, storage compartment, and image acquisition time. The accessory sensor positioning data includes accessory positioning tag identification, positioning acquisition time, positioning signal strength, positioning distance value, positioning angle value, initial positioning coordinate value, and on-site detection status; The spare parts storage and retrieval management data includes spare parts location information, spare parts inventory ledger data, spare parts storage and retrieval task data, and spare parts usage-related data; The spare parts storage location information includes shelf number, storage compartment number, storage location coordinates, storage location availability status, and the spare parts type bound to the storage location; the spare parts inventory ledger data includes spare parts identification, spare parts category, spare parts model, spare parts specifications, current inventory quantity, spare parts location status, and spare parts abnormal status; the spare parts storage and retrieval task data includes inbound task data, outbound task data, return task data, and inventory task data; and the spare parts usage association data includes storage and retrieval information, storage and retrieval time, usage task, and compatible drone information.

[0020] In this embodiment, step two specifically includes: Determine the location data group corresponding to the target drone accessory based on the accessory location tag identification; The positioning data set includes the positioning signal strength, positioning distance value, positioning angle value, initial positioning coordinate value, and in-situ detection status of the target UAV accessory collected by multiple positioning sensors at the same positioning acquisition time. The location confidence weights are obtained by exponentially normalizing the strength of each location signal. The spatial coordinates of the target UAV component at the current positioning and data acquisition time are used as candidate spatial coordinates; Calculate the coordinate difference between the candidate spatial coordinates and the installation coordinates of the r-th positioning sensor to obtain the relative coordinate vector, and calculate the L2 norm of the relative coordinate vector to obtain the candidate ranging value. The difference between the candidate ranging value and the positioning distance value is used as the distance residual corresponding to the r-th positioning sensor. Convert each positioning angle value into a direction unit vector; Perform an inner product operation between the relative coordinate vector and the direction unit vector, and use the ratio of the inner product operation result to the candidate ranging value as the direction consistency value; calculate the difference between the constant 1 and the direction consistency value to obtain the angle residual corresponding to the r-th positioning sensor; The sum of the squares of the distance residual and the squares of the angle residual is multiplied by the positioning confidence weight to obtain the weighted observation residual term corresponding to the r-th positioning sensor. The weighted observation residuals corresponding to all positioning sensors are summed to obtain the positioning observation residuals. The squared coordinate deviation between the candidate spatial coordinates and the initial positioning coordinates is used as the initial value constraint term. The positioning observation residual term and the initial value constraint term are added together to construct the spatial coordinate optimization objective function for the target UAV parts; The spatial coordinate optimization objective function is minimized to obtain the spatial coordinates of the target UAV accessory at the current positioning and data acquisition time. Based on the parts storage location information, the spatial coordinates of the parts are matched with the storage location coordinates corresponding to each storage slot to determine the storage slot number corresponding to the target drone parts. Based on the storage compartment number, read the shelf number and storage location coordinates from the parts storage location information; The in-situ detection status is taken as the accessory in-situ status, and the accessory movement speed of the target UAV accessory is calculated based on the accessory's spatial coordinates at the current positioning and acquisition time and the accessory's spatial coordinates at the previous positioning and acquisition time. When the moving speed of the accessory is greater than the moving speed threshold, the moving state of the accessory of the target drone is determined to be moving; when the moving speed of the accessory is less than or equal to the moving speed threshold, the moving state of the accessory of the target drone is determined to be stationary. Among them, the moving speed threshold is a non-zero threshold constructed based on the positioning error radius and the time interval between adjacent positioning acquisitions; the positioning error radius is specifically constructed as follows: the positioning error of a single node is constructed based on the distance residual and angle residual corresponding to the r-th positioning sensor, and the positioning error of each single node is weighted and summed based on the positioning confidence weight corresponding to each positioning sensor to obtain the positioning error radius of the target UAV accessory; The spatial coordinates of the parts, storage compartment number, shelf number, warehouse location coordinates, part location status, and part movement status are used as the real-time positioning results of the drone parts.

[0021] In this embodiment, step three specifically includes: Based on the shelf number and storage compartment number, select the storage compartment image that matches both the shelf number and storage compartment number from the accessory image data, and use the storage compartment image with the smallest time difference between the image acquisition time and the current positioning acquisition time as the target accessory positioning associated image. Calculate the difference between the spatial coordinates of the accessory and the coordinates of the storage location to obtain the relative spatial coordinates of the target drone accessory relative to the corresponding storage slot; The relative spatial coordinates are converted into homogeneous coordinates and mapped onto the target accessory positioning association image through the projection matrix of the pre-calibrated image acquisition device to obtain the accessory projection center coordinates of the target UAV accessory; Using the coordinates of the component projection center as the center point of the region, and the projection width and projection height of the grid corresponding to the storage grid number as the region size, the region size is expanded outward according to the positioning error radius to obtain the candidate image region. Among them, the grid projection width and grid projection height are the pixel occupancy size of the storage grid corresponding to the storage grid number in the target accessory positioning association image. The grid projection width is the width of the storage grid image area corresponding to the storage grid number in the horizontal pixel direction, and the grid projection height is the height of the storage grid image area corresponding to the storage grid number in the vertical pixel direction. The positioning error radius is the positioning error radius of the target UAV accessory calculated based on the distance residual, angle residual and positioning confidence weight corresponding to each positioning sensor. It is used to expand the candidate image area centered on the projection center coordinates of the accessory in both the horizontal and vertical directions. Candidate image regions are extracted from the target component location association image to obtain candidate region images; Bilateral filtering is applied to the candidate region image to denoise it, resulting in a denoised candidate region image. The candidate region image for denoising is subjected to grayscale linear normalization to obtain a normalized candidate region image. Based on the radial and tangential distortion parameters of the image acquisition device, the Brown-Conrady distortion correction method is used to correct the distortion of the normalized candidate region image to obtain the corrected candidate region image. Background sample pixels are extracted from the edge ring region of the correction candidate region image. A background reference image with the same size as the correction candidate region image is constructed based on the pixel mean of the background sample pixels. A target foreground mask is constructed based on the pixel difference between the correction candidate region image and the background reference image. Specifically, background sample pixels are extracted based on the edge annular region of the correction candidate region image. The process involves: using the outer boundary of the correction candidate region image as the outer boundary of the annular region, rounding up the smaller of the image width and height of the correction candidate region image to obtain the edge annular width; using the four image boundaries of the correction candidate region image as the outer boundaries of the annular region, shifting the left boundary to the right by [number] units, the right boundary to the left by [number] units, the upper boundary downwards by [number] units, and the lower boundary upwards by [number] units to obtain the inner boundaries of the annular region; defining the pixel region between the outer and inner boundaries of the annular region as the edge annular region; and extracting the pixel values ​​corresponding to all pixels within the edge annular region as background sample pixels. Specifically, the target foreground mask is constructed based on the pixel differences between the candidate correction region image and the background reference image. This involves: calculating the pixel difference value between each pixel in the candidate correction region image and the corresponding pixel in the background reference image to obtain a pixel difference image; counting the number of pixels corresponding to each pixel difference value in the pixel difference image to obtain a pixel difference histogram; calculating the inter-class variance corresponding to each candidate segmentation threshold based on the pixel difference histogram; determining the candidate segmentation threshold with the largest inter-class variance as the background separation threshold; identifying pixels with pixel differences greater than or equal to the background separation threshold as foreground pixels and assigning them a value of 1; identifying pixels with pixel differences less than the background separation threshold as background pixels and assigning them a value of 0; and generating the target foreground mask. The target region boundary of the target UAV accessory is determined based on the target foreground mask, and the target region is cropped from the correction candidate region image according to the target region boundary to obtain the target image of the UAV accessory. Specifically, determining the target region boundary of the target UAV accessory based on the target foreground mask involves: marking the foreground pixels with a value of 1 in the target foreground mask as connected components to obtain a set of foreground connected components; calculating the center coordinates of each foreground connected component; calculating the center distance between the center coordinates of each foreground connected component and the center coordinates of the accessory projection; determining the foreground connected component with the smallest center distance as the foreground region of the target accessory; and determining the left boundary coordinates, right boundary coordinates, upper boundary coordinates, and lower boundary coordinates of the target region based on the horizontal and vertical pixel coordinates of each foreground pixel in the foreground region of the target accessory, thus forming the target region boundary of the target UAV accessory.

[0022] In this embodiment, the improved RTMDet recognition network includes a component backbone feature extraction module, a multi-scale component feature fusion module, a component detection output module, and a model fine-grained discrimination module. The component backbone feature extraction module performs five-stage convolutional feature extraction on the target image of UAV components. The output feature map of the third stage is used as the shallow component backbone feature map, the output feature map of the fourth stage is used as the middle component backbone feature map, and the output feature map of the fifth stage is used as the deep component backbone feature map. Among them, the first and second stages of convolutional feature extraction use downsampling convolution with a kernel size of 3×3 and a stride of 2. The third, fourth and fifth stages of convolutional feature extraction all use depthwise separable convolution with a kernel size of 5×5 and channel mapping convolution with a kernel size of 1×1. The multi-scale component feature fusion module performs channel mapping and scale alignment fusion on the shallow component backbone feature map, the middle component backbone feature map and the deep component backbone feature map respectively through the path aggregation pyramid fusion structure to obtain the shallow component fusion feature map, the middle component fusion feature map and the deep component fusion feature map. The path aggregation pyramid fusion structure is specifically as follows: The mid-layer component backbone feature map is upsampled and scaled to be aligned with the shallow component backbone feature map to obtain the mid-layer upsampled feature map; the shallow component backbone feature map and the mid-layer upsampled feature map are concatenated, and channel compression, local structure fusion and channel reshaping are performed through fusion convolutional units to obtain the shallow component fused feature map. The deep component backbone feature map is upsampled and scaled with the mid-layer component backbone feature map to obtain the deep upsampled feature map; the mid-layer component backbone feature map and the deep upsampled feature map are concatenated, and channel compression, local structure fusion and channel reshaping are performed through fusion convolutional units to obtain the mid-layer initial fused feature map; The shallow component fusion feature map is scaled by downsampling and the initial fusion feature map of the middle layer to obtain the shallow downsampled feature map; the initial fusion feature map of the middle layer and the shallow downsampled feature map are concatenated, and channel compression, local structure fusion and channel reshaping are performed through fusion convolution units to obtain the middle component fusion feature map. The mid-layer component fusion feature map is downsampled and scaled to be aligned with the deep component backbone feature map to obtain the mid-layer downsampled feature map; the deep component backbone feature map and the mid-layer downsampled feature map are concatenated, and channel compression, local structure fusion and channel reshaping are performed through fusion convolution units to obtain the deep component fusion feature map; The fusion convolutional unit sequentially includes a channel compression convolution with a kernel size of 1×1, a depthwise separable convolution with a kernel size of 5×5, and a channel mapping convolution with a kernel size of 1×1. The multi-scale component feature fusion module introduces a codebook quantization mechanism to perform key structure prototype matching, discrete quantization, and residual back-injection enhancement on the shallow component fusion feature map, the middle component fusion feature map, and the deep component fusion feature map, thereby obtaining shallow component enhanced feature map, middle component enhanced feature map, and deep component enhanced feature map. The shallow, medium, and deep component enhancement feature maps are input into the component detection output module, and classification prediction and bounding box regression prediction are performed respectively to obtain candidate detection results for each feature level. The candidate detection results include candidate accessory bounding boxes, candidate accessory categories, and candidate recognition confidence scores. Construct a detection confidence threshold and a target box overlap threshold, delete candidate detection results whose candidate recognition confidence is less than the detection confidence threshold, and obtain the initial screening detection results; In this invention, the detection confidence threshold specifically involves: obtaining the confidence scores of all candidate identifications in the same batch of candidate detection results; grouping and statistically analyzing the candidate identification confidence scores according to the candidate accessory categories to obtain the mean confidence score and confidence score dispersion for each candidate accessory category; and determining the detection confidence threshold for the corresponding candidate accessory category based on the mean confidence score and confidence score dispersion. The target box overlap threshold is specifically determined as follows: obtain the target boxes of each candidate accessory under the same candidate accessory category; then calculate the intersection-union ratio (IU) between any two candidate accessory target boxes; statistically analyze the distribution of the IU and determine the target box overlap threshold by combining the area difference and center distance of the candidate accessory target boxes. The initial screening results are processed by non-maximum suppression according to the candidate recognition confidence level from high to low. The initial screening results that belong to the same candidate accessory category as the retained initial screening results and whose target box intersection-union ratio is greater than the target box overlap threshold are deleted to obtain the accessory target detection results. The accessory target detection results include the accessory target bounding box location, accessory category, and recognition confidence level; The accessory target detection results are input into the model fine-grained discrimination module. Based on the position of the accessory target box, local visual features are extracted from the accessory enhanced feature map of the corresponding feature level. Global pooling and multi-branch classification are performed on the local visual features to obtain the accessory model, accessory specifications and compatible drone type. Based on accessory category, accessory model, accessory specifications, and compatible drone type, the corresponding accessory identification can be retrieved from the accessory inventory ledger data; The accessory identification result is determined by the accessory's identity identifier, accessory category, accessory model, accessory specifications, compatible drone type, accessory target box position, and recognition confidence level.

[0023] In this embodiment, the codebook quantization mechanism specifically includes: Construct a key structure prototype codebook, which includes multiple key structure prototype vectors. These key structure prototype vectors are used to characterize the battery contact structure, motor mounting hole structure, propeller edge structure, gimbal connection interface structure, arm snap-fit ​​structure, and flight control board plug-in interface structure of UAV accessories. The key structure prototype codebook is a trainable codebook parameter in the improved RTMDet recognition network. When constructing the key structure prototype codebook, firstly, key structure feature vectors corresponding to the battery contact structure, motor mounting hole structure, propeller edge structure, gimbal connection interface structure, arm snap-fit ​​structure, and flight control board interface structure are extracted based on training image samples of UAV accessories. Then, multiple key structure prototype vectors are initialized based on these key structure feature vectors. During the training process of the improved RTMDet recognition network, these key structure prototype vectors are used as trainable parameters and updated through backpropagation to obtain the trained key structure prototype codebook. The shallow component fusion feature map, the middle component fusion feature map, and the deep component fusion feature map are respectively mapped to the key structure feature space to obtain the shallow key structure feature map, the middle key structure feature map, and the deep key structure feature map. The key structural feature space refers to the structural feature vector space formed by channel mapping of the component fusion feature map using structural feature projection parameters. For shallow, mid-level, and deep component fusion feature maps, mapping is performed using structural feature projection parameters corresponding to the feature level, so that the local visual features at each spatial location are converted into structural feature vectors with the same dimension as the key structural prototype vector in the key structural prototype codebook.

[0024] For any key structural feature map at any feature level, the distance between the structural feature vector corresponding to spatial location u and each key structural prototype vector in the key structural prototype codebook is calculated, and the key structural prototype vector with the smallest distance is taken as the quantized structural feature vector corresponding to spatial location u. The quantized structural feature vectors corresponding to each spatial location in the same feature level are combined to obtain a quantized structural feature map, and the residual between the key structural feature map and the quantized structural feature map is calculated to obtain a quantized residual feature map. The quantized structural feature map and the quantized residual feature map are concatenated and then back-injected to the component fusion feature map of the corresponding feature level through back-injection mapping parameters to obtain shallow component enhancement feature map, mid-level component enhancement feature map and deep component enhancement feature map.

[0025] The back-injection mapping parameters are 1×1 convolution parameters set for the three feature layers of shallow, middle and deep layers respectively. The back-injection mapping parameters are constructed based on the number of channels of the quantized enhanced splicing feature map and the number of channels of the accessory fusion feature map of the corresponding feature layer, and are updated through backpropagation during the training of the improved RTMDet recognition network.

[0026] In this invention, the improved RTMDet recognition network inherits the single-stage target detection framework of the standard RTMDet recognition network, that is, it still retains the basic processing flow of backbone feature extraction, multi-scale feature fusion and detection output.

[0027] The main improvement of the improved RTMDet recognition network lies in the introduction of a codebook quantization mechanism during the accessory feature fusion process. Specifically, after obtaining shallow, mid-level, and deep accessory fusion feature maps, a key structure prototype codebook is constructed to characterize the key local structures of UAV accessories. This key structure prototype codebook includes multiple key structure prototype vectors representing battery contact structures, motor mounting hole structures, propeller edge structures, gimbal connection interface structures, arm snap-fit ​​structures, and flight control board plug-in interfaces. Subsequently, the shallow, mid-level, and deep accessory fusion feature maps are mapped to the key structure feature space, respectively. Through key structure prototype matching, discrete quantization, and residual backinjection enhancement, enhanced shallow, mid-level, and deep accessory feature maps are obtained.

[0028] The improvements described above are effective in the following ways: By introducing a key structural prototype codebook into the accessory feature fusion process, local structural features in the multi-scale accessory fusion features are mapped onto discrete key structural prototypes, thereby enhancing the expressive power of key details such as battery contact structures, motor mounting hole structures, propeller edge structures, gimbal connection interface structures, arm snap-fit ​​structures, and flight control board interface structures. Through discrete quantization and residual backinjection enhancement, the interference of shelf background, lighting changes, occlusion, and similar appearances on the recognition results can be reduced, improving the ability to distinguish between similar UAV accessories based on model, specifications, and compatible UAV types. Therefore, the improved RTMDet recognition network improves the accuracy and reliability of fine-grained UAV accessory recognition results and provides a more accurate recognition foundation for subsequent accessory identity location binding and ternary consistency verification.

[0029] In this embodiment, the binding relationship between the accessory identity and the storage location specifically includes: the binding relationship between the accessory identity identifier and the storage compartment number; the matching relationship between the accessory category, accessory model, accessory specification and the accessory type bound to the storage location; the position mapping relationship between the spatial coordinates of the accessory and the coordinates of the storage location; the status correspondence between the accessory's in-situ status and the accessory inventory ledger data; and the association relationship between the shelf number, the storage compartment number and the accessory identity identifier.

[0030] In this embodiment, the ternary consistency check includes: Perform a consistency check between the accessory identification identifier in the drone accessory identification result and the target accessory identification identifier in the target access task requirements; The target access task requirements include the target accessory identification, target accessory category, target accessory model, and target accessory specifications; The consistency of the accessory category, accessory model and accessory specifications in the drone accessory identification results with the accessory type bound to the target storage location information is verified. Perform a consistency check between the storage compartment number in the real-time location results of drone parts and the target storage compartment number in the target storage location information. The target storage location information includes the target shelf number, the target storage compartment number, the target storage location coordinates, and the type of storage location-bound accessories corresponding to the target storage location; Perform consistency verification between the real-time location results of drone parts and the location data of the parts inventory ledger. Perform a consistency check between the movement status of the drone accessories in the real-time location results and the current access operation type. The intelligent storage and retrieval strategy for spare parts includes a spare parts inbound control strategy, a spare parts outbound control strategy, a spare parts return verification strategy, a spare parts misplacement early warning strategy, and a spare parts mis-retrieval early warning strategy. The accessory warehousing control strategy is as follows: When the storage operation type is warehousing operation type, based on the accessory category, accessory model and accessory specifications in the drone accessory identification results, the storage slot number that is bound to the storage location with the same accessory type and whose storage location availability status is idle is selected from the accessory storage location information, and the selected storage slot number and its corresponding shelf number and storage location coordinates are determined as the target storage location where the target drone accessory should be placed. The parts outbound control strategy is as follows: When the storage and retrieval operation type is outbound operation type, according to the target parts identification in the target storage and retrieval task requirements, the corresponding shelf number, storage compartment number and storage location coordinates are queried from the parts identification storage location binding relationship, and the queried shelf number, storage compartment number and storage location coordinates are determined as the outbound prompt location of the target drone parts; The accessory return verification strategy is as follows: when the storage operation type is the return operation type, the storage cell number in the real-time positioning result of the drone accessory is compared with the storage cell number bound to the accessory identity identifier in the accessory identity storage location binding relationship. If the two are consistent, the return storage location is determined to be correct; if the two are inconsistent, the return storage location is determined to be incorrect. The accessory misplacement warning strategy is as follows: when the accessory category, accessory model and accessory specification in the drone accessory identification result are inconsistent with the accessory type bound to the target storage location information, or when the storage cell number in the drone accessory real-time positioning result is inconsistent with the target storage cell number in the target storage location information, an accessory misplacement warning message is generated. The accessory mis-acquisition warning strategy is as follows: when the accessory identification in the UAV accessory identification result is inconsistent with the target accessory identification in the target access task requirements, an accessory mis-acquisition warning message is generated.

[0031] In this embodiment, the inventory status of drone parts includes parts in stock, parts out of stock, parts returned, parts in place, parts bound to storage location, and parts abnormal status. The entire process traceability record for drone accessories includes accessory identification, accessory category, accessory model, accessory specifications, shelf number, storage compartment number, storage location coordinates, accessory spatial coordinates, accessor information, access time, usage task, compatible drone information, identification confidence level, current access operation type, and abnormal warning information.

[0032] A sensor-based intelligent identification and retrieval system for drone accessories includes: Data acquisition module: used to acquire accessory image data, accessory sensor positioning data, and accessory storage and management data of the target drone accessories; Multi-sensor residual positioning module: Used to locate the target drone parts in real time based on the positioning data of the parts' sensors and through the multi-sensor residual positioning algorithm, so as to obtain the real-time positioning result of the drone parts; Candidate image region processing module: Based on accessory image data and real-time positioning results of drone accessories, it determines the candidate image region of the target drone accessory, performs image preprocessing on the candidate image region, and obtains the target image of the drone accessory; Improved RTMDet Recognition Module: This module is used to input the target image of the UAV accessory into the improved RTMDet recognition network, perform accessory target detection and fine-grained model discrimination, and obtain the UAV accessory recognition result. The improved RTMDet recognition network introduces a codebook quantization mechanism in the accessory feature fusion process. Parts Identification and Storage Location Binding Module: Used to establish a parts identification and storage location binding relationship based on the drone parts identification results, drone parts real-time positioning results, and parts storage and retrieval management data; Verification and Strategy Generation Module: Based on the accessory identity storage location binding relationship, it performs a ternary consistency verification on the target UAV accessories, obtains the accessory access consistency verification result, and generates an intelligent accessory access strategy based on the accessory access consistency verification result. Inventory traceability module: Used to update the inventory status of drone parts and generate a full-process traceability record of drone parts based on the intelligent storage and retrieval strategy for parts.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to the parts storage management scenario of a power inspection drone maintenance center. This maintenance center has 48 inspection drones of various models, and manages a total of 1860 drone parts daily, including batteries, motors, propellers, gimbals, arms, flight control boards, and camera modules. These parts are stored in 6 sets of intelligent shelves and 216 storage compartments. Because batteries of the same series have similar appearances, propellers with different rotation directions have similar shapes, and motors of different specifications have only minor differences in mounting hole positions, the original manual ledger and barcode scanning methods were prone to problems such as parts being placed in the wrong compartment, the wrong model being retrieved, and the back-end ledger not being updated in a timely manner during the warehousing, outbound, and return processes. Especially when multiple drones simultaneously perform inspection tasks and parts are returned in batches, warehouse personnel need to repeatedly verify the part model, shelf number, and storage compartment number, resulting in low management efficiency.

[0034] In this embodiment of the invention, positioning sensors are deployed near each storage compartment, and accessory positioning tags are configured for the target drone accessories. Images of the target storage area, shelf area, and storage compartments are acquired using an image acquisition device. The system first acquires accessory image data, accessory sensor positioning data, and accessory storage and retrieval management data. Then, based on the accessory positioning tag, it determines the positioning data group corresponding to the target drone accessory. A multi-sensor residual positioning algorithm is used to solve for the accessory's spatial coordinates, and the corresponding shelf number, storage compartment number, storage location coordinates, accessory location status, and accessory movement status are determined. Subsequently, the system selects the corresponding storage compartment image based on the shelf number and storage compartment number, and combines the accessory's spatial coordinates and storage location coordinates to determine candidate image regions. These candidate image regions are then subjected to denoising, normalization, distortion correction, background separation, and cropping to obtain the target image of the drone accessory.

[0035] After the target image of a UAV accessory is input into the improved RTMDet recognition network, the network performs key structural prototype matching, discrete quantization, and residual back-injection enhancement on the battery contact structure, motor mounting hole structure, propeller edge structure, gimbal connection interface structure, arm snap-fit ​​structure, and flight control board plug interface structure through a codebook quantization mechanism during accessory feature fusion. The output includes accessory category, accessory model, accessory specifications, compatible UAV type, accessory target box position, and recognition confidence. The system further combines accessory inventory ledger data to query accessory identification, establishes accessory identity-location binding relationships, and performs ternary consistency checks on target UAV accessories. Based on the check results, it generates intelligent accessory storage and retrieval strategies, completing tasks such as inbound storage location determination, outbound location prompts, return location verification, accessory misplacement warnings, and accessory mis-retrieval warnings.

[0036] To further verify the effectiveness of the present invention, it is compared and analyzed with four comparative schemes. Among them, comparative scheme 1 is a manual ledger and barcode scanning scheme; comparative scheme 2 is a single-point RFID positioning and ordinary image detection scheme; comparative scheme 3 is a multi-sensor residual positioning and standard RTMDet identification network scheme; and comparative scheme 4 is a single-point RFID positioning and improved RTMDet identification network scheme.

[0037] Table 1. Comparison of the Implementation Effects of Different UAV Parts Storage and Retrieval Management Schemes As shown in Table 1, the solution of this invention outperforms the four comparative solutions in all indicators. Compared with comparative solutions 1 and 2, the average positioning error of the solution of this invention is reduced to 7.8 cm, and the storage compartment matching accuracy is increased to 98.6%. This indicates that the multi-sensor residual positioning algorithm can effectively reduce the positional deviation caused by manual recording, barcode scanning, and single-point RFID positioning in scenarios with shelf obstruction, signal attenuation, and dense compartments.

[0038] Compared to Comparative Solution 3, the present invention improves the model and specification recognition accuracy from 92.8% to 96.7%, indicating that the introduction of a codebook quantization mechanism in the accessory feature fusion process of the improved RTMDet recognition network enhances the expressive ability of key local structures such as battery contact structures, motor mounting hole structures, propeller edge structures, and gimbal connection interface structures, thereby improving the fine-grained recognition accuracy of similar UAV accessories. Compared to Comparative Solution 4, the storage grid matching accuracy of the present invention improves from 92.1% to 98.6%, and the average search time for a single item is reduced from 38.4s to 22.4s. This shows that improving the recognition network without improving the positioning method still makes it susceptible to single-point positioning errors. However, the present invention, through the collaborative processing of a multi-sensor residual positioning algorithm and the improved RTMDet recognition network, can simultaneously improve the reliability of accessory positioning and the accuracy of recognition.

[0039] Furthermore, the misplacement recognition rate of the present invention reaches 97.4%, the misretrieval recognition rate reaches 98.1%, and the inventory ledger consistency rate reaches 98.8%. This indicates that performing a ternary consistency check based on the accessory identity storage location binding relationship can effectively detect inconsistencies between accessory identity, storage slot, and storage / retrieval tasks, reducing the problems of misplacement, misretrieval, and inventory ledger lag in drone accessories.

[0040] 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 method for intelligent identification and retrieval of UAV accessories based on sensor positioning, characterized in that, Includes the following steps: Step 1: Acquire the accessory image data, accessory sensor positioning data, and accessory storage and management data of the target drone accessories; Step 2: Based on the sensor positioning data of the accessory, the target drone accessory is located in real time using a multi-sensor residual positioning algorithm to obtain the real-time positioning result of the drone accessory; Step 3: Based on the accessory image data and the real-time positioning results of the drone accessories, determine the candidate image regions of the target drone accessories, and perform image preprocessing on the candidate image regions to obtain the target image of the drone accessories; Step 4: Input the target image of the drone accessory into the improved RTMDet recognition network, perform accessory target detection and fine-grained model discrimination, and obtain the drone accessory recognition result; the improved RTMDet recognition network introduces a codebook quantization mechanism in the accessory feature fusion process; Step 5: Based on the drone accessory identification results, drone accessory real-time positioning results, and accessory storage and retrieval management data, establish an accessory identity and storage location binding relationship; Step 6: Based on the accessory identity storage location binding relationship, perform a ternary consistency check on the target drone accessories, obtain the accessory access consistency check result, and generate an intelligent accessory access strategy based on the accessory access consistency check result; Step 7: Based on the intelligent storage and retrieval strategy for accessories, update the inventory status of drone accessories and generate a full-process traceability record for drone accessories.

2. The intelligent identification and retrieval method for UAV accessories based on sensor positioning according to claim 1, characterized in that, The accessory image data includes images of the target storage area, shelf area, storage compartment, and image acquisition time. The accessory sensor positioning data includes accessory positioning tag identification, positioning acquisition time, positioning signal strength, positioning distance value, positioning angle value, initial positioning coordinate value, and in-situ detection status; The spare parts storage and retrieval management data includes spare parts location information, spare parts inventory ledger data, spare parts storage and retrieval task data, and spare parts usage-related data.

3. The intelligent identification and retrieval method for UAV accessories based on sensor positioning according to claim 1, characterized in that, Step two specifically includes: Determine the location data group corresponding to the target drone accessory based on the accessory location tag identification; The positioning data set includes the positioning signal strength, positioning distance value, positioning angle value, initial positioning coordinate value, and in-situ detection status of the target UAV accessory collected by multiple positioning sensors at the same positioning acquisition time. The location confidence weights are obtained by exponentially normalizing the strength of each location signal. The spatial coordinates of the target UAV component at the current positioning and data acquisition time are used as candidate spatial coordinates; Calculate the coordinate difference between the candidate spatial coordinates and the installation coordinates of the r-th positioning sensor to obtain the relative coordinate vector, and calculate the L2 norm of the relative coordinate vector to obtain the candidate ranging value. The difference between the candidate ranging value and the positioning distance value is used as the distance residual corresponding to the r-th positioning sensor. Convert each positioning angle value into a direction unit vector; Perform an inner product operation between the relative coordinate vector and the direction unit vector, and use the ratio of the inner product operation result to the candidate ranging value as the direction consistency value; calculate the difference between the constant 1 and the direction consistency value to obtain the angle residual corresponding to the r-th positioning sensor; The sum of the squares of the distance residual and the squares of the angle residual is multiplied by the positioning confidence weight to obtain the weighted observation residual term corresponding to the r-th positioning sensor. The weighted observation residuals corresponding to all positioning sensors are summed to obtain the positioning observation residuals. The squared coordinate deviation between the candidate spatial coordinates and the initial positioning coordinates is used as the initial value constraint term. The positioning observation residual term and the initial value constraint term are added together to construct the spatial coordinate optimization objective function for the target UAV parts; The spatial coordinate optimization objective function is minimized to obtain the spatial coordinates of the target UAV accessory at the current positioning and data acquisition time. Based on the parts storage location information, the spatial coordinates of the parts are matched with the storage location coordinates corresponding to each storage slot to determine the storage slot number corresponding to the target drone parts. Based on the storage compartment number, read the shelf number and storage location coordinates from the parts storage location information; The in-situ detection status is taken as the accessory in-situ status, and the accessory movement speed of the target UAV accessory is calculated based on the accessory's spatial coordinates at the current positioning and acquisition time and the accessory's spatial coordinates at the previous positioning and acquisition time. When the moving speed of the accessory is greater than the moving speed threshold, the moving state of the accessory of the target drone is determined to be moving; when the moving speed of the accessory is less than or equal to the moving speed threshold, the moving state of the accessory of the target drone is determined to be stationary. The spatial coordinates of the parts, the storage compartment number, the shelf number, the warehouse location coordinates, the location status of the parts, and the movement status of the parts are used as the real-time positioning results of the drone parts.

4. The intelligent identification and retrieval method for UAV accessories based on sensor positioning according to claim 1, characterized in that, Step three specifically includes: Based on the shelf number and storage compartment number, select the storage compartment image that matches both the shelf number and storage compartment number from the accessory image data, and use the storage compartment image with the smallest time difference between the image acquisition time and the current positioning acquisition time as the target accessory positioning associated image. Calculate the difference between the spatial coordinates of the accessory and the coordinates of the storage location to obtain the relative spatial coordinates of the target drone accessory relative to the corresponding storage slot; The relative spatial coordinates are converted into homogeneous coordinates and mapped onto the target accessory positioning association image through the projection matrix of the pre-calibrated image acquisition device to obtain the accessory projection center coordinates of the target UAV accessory; Using the coordinates of the component projection center as the center point of the region, and the projection width and projection height of the grid corresponding to the storage grid number as the region size, the region size is expanded outward according to the positioning error radius to obtain the candidate image region. Candidate image regions are extracted from the target component location association image to obtain candidate region images; Bilateral filtering is applied to the candidate region image to denoise it, resulting in a denoised candidate region image. The candidate region image for denoising is subjected to grayscale linear normalization to obtain a normalized candidate region image. Based on the radial and tangential distortion parameters of the image acquisition device, the Brown-Conrady distortion correction method is used to correct the distortion of the normalized candidate region image to obtain the corrected candidate region image. Background sample pixels are extracted from the edge ring region of the correction candidate region image. A background reference image with the same size as the correction candidate region image is constructed based on the pixel mean of the background sample pixels. A target foreground mask is constructed based on the pixel difference between the correction candidate region image and the background reference image. The target region boundary of the target UAV accessory is determined based on the target foreground mask, and the target region is cropped from the candidate region image according to the target region boundary to obtain the target image of the UAV accessory.

5. The intelligent identification and retrieval method for UAV accessories based on sensor positioning according to claim 1, characterized in that, The improved RTMDet recognition network includes a component backbone feature extraction module, a multi-scale component feature fusion module, a component detection output module, and a model fine-grained discrimination module; The component backbone feature extraction module performs five-stage convolutional feature extraction on the target image of UAV components. The output feature map of the third stage is used as the shallow component backbone feature map, the output feature map of the fourth stage is used as the middle component backbone feature map, and the output feature map of the fifth stage is used as the deep component backbone feature map. The multi-scale component feature fusion module performs channel mapping and scale alignment fusion on the shallow component backbone feature map, the middle component backbone feature map and the deep component backbone feature map respectively through the path aggregation pyramid fusion structure to obtain the shallow component fusion feature map, the middle component fusion feature map and the deep component fusion feature map. The multi-scale component feature fusion module introduces a codebook quantization mechanism to perform key structure prototype matching, discrete quantization, and residual back-injection enhancement on the shallow component fusion feature map, the middle component fusion feature map, and the deep component fusion feature map, thereby obtaining shallow component enhanced feature map, middle component enhanced feature map, and deep component enhanced feature map. The shallow, medium, and deep component enhancement feature maps are input into the component detection output module, and classification prediction and bounding box regression prediction are performed respectively to obtain candidate detection results for each feature level. The candidate detection results include candidate accessory bounding boxes, candidate accessory categories, and candidate recognition confidence scores. Construct a detection confidence threshold and a target box overlap threshold, delete candidate detection results whose candidate recognition confidence is less than the detection confidence threshold, and obtain the initial screening detection results; The initial screening results are processed by non-maximum suppression according to the candidate recognition confidence level from high to low. The initial screening results that belong to the same candidate accessory category as the retained initial screening results and whose target box intersection-union ratio is greater than the target box overlap threshold are deleted to obtain the accessory target detection results. The accessory target detection results include the accessory target bounding box position, accessory category, and recognition confidence level; The accessory target detection results are input into the model fine-grained discrimination module. Based on the position of the accessory target box, local visual features are extracted from the accessory enhanced feature map of the corresponding feature level. Global pooling and multi-branch classification are performed on the local visual features to obtain the accessory model, accessory specifications and compatible drone type. Based on accessory category, accessory model, accessory specifications, and compatible drone type, the corresponding accessory identification can be retrieved from the accessory inventory ledger data; The accessory identification result is determined by the accessory's identity identifier, accessory category, accessory model, accessory specifications, compatible drone type, accessory target box position, and recognition confidence level.

6. The intelligent identification and retrieval method for UAV accessories based on sensor positioning according to claim 1, characterized in that, The codebook quantization mechanism specifically includes: Construct a key structure prototype codebook, which includes multiple key structure prototype vectors. These key structure prototype vectors are used to characterize the battery contact structure, motor mounting hole structure, propeller edge structure, gimbal connection interface structure, arm snap-fit ​​structure, and flight control board plug-in interface structure of UAV accessories. The shallow component fusion feature map, the middle component fusion feature map, and the deep component fusion feature map are respectively mapped to the key structure feature space to obtain the shallow key structure feature map, the middle key structure feature map, and the deep key structure feature map. For any key structural feature map at any feature level, the distance between the structural feature vector corresponding to spatial location u and each key structural prototype vector in the key structural prototype codebook is calculated, and the key structural prototype vector with the smallest distance is taken as the quantized structural feature vector corresponding to spatial location u. The quantized structural feature vectors corresponding to each spatial location in the same feature level are combined to obtain a quantized structural feature map, and the residual between the key structural feature map and the quantized structural feature map is calculated to obtain a quantized residual feature map. The quantized structural feature map and the quantized residual feature map are concatenated and then back-injected to the component fusion feature map of the corresponding feature level through back-injection mapping parameters to obtain shallow component enhancement feature map, mid-level component enhancement feature map and deep component enhancement feature map.

7. The intelligent identification and retrieval method for UAV accessories based on sensor positioning according to claim 1, characterized in that, The specific binding relationship between the accessory identity and storage location includes: the binding relationship between the accessory identity identifier and the storage compartment number; the matching relationship between the accessory category, accessory model, accessory specification and the accessory type bound to the storage location; the location mapping relationship between the spatial coordinates of the accessory and the coordinates of the storage location; the status correspondence between the accessory's in-situ status and the accessory inventory ledger data; and the association relationship between the shelf number, storage compartment number and accessory identity identifier.

8. The intelligent identification and retrieval method for UAV accessories based on sensor positioning according to claim 1, characterized in that, The ternary consistency check includes: Perform a consistency check between the accessory identification identifier in the drone accessory identification result and the target accessory identification identifier in the target access task requirements; The target access task requirements include the target accessory identification, target accessory category, target accessory model, and target accessory specifications; The consistency of the accessory category, accessory model and accessory specifications in the drone accessory identification results with the accessory type bound to the target storage location information is verified. Perform a consistency check between the storage compartment number in the real-time location results of drone parts and the target storage compartment number in the target storage location information. The target storage location information includes the target shelf number, the target storage compartment number, the target storage location coordinates, and the type of storage location-bound accessories corresponding to the target storage location; Perform consistency verification between the real-time location results of drone parts and the location data of the parts inventory ledger. Perform a consistency check between the movement status of the drone accessories in the real-time location results and the current access operation type. The intelligent storage and retrieval strategy for spare parts includes a spare parts inbound control strategy, a spare parts outbound control strategy, a spare parts return verification strategy, a spare parts misplacement early warning strategy, and a spare parts mis-retrieval early warning strategy.

9. The intelligent identification and retrieval method for UAV accessories based on sensor positioning according to claim 1, characterized in that, The inventory status of drone parts includes parts inbound status, parts outbound status, parts returned status, parts in place status, parts location binding status, and parts abnormal status. The full-process traceability record for drone accessories includes accessory identification, accessory category, accessory model, accessory specifications, shelf number, storage compartment number, storage location coordinates, accessory spatial coordinates, accessor information, access time, usage task, compatible drone information, identification confidence level, current access operation type, and abnormal warning information.

10. A sensor-based intelligent identification and retrieval system for drone accessories, comprising the sensor-based intelligent identification and retrieval method for drone accessories as described in any one of claims 1 to 9, characterized in that, include: Data acquisition module: used to acquire accessory image data, accessory sensor positioning data, and accessory storage and management data of the target drone accessories; Multi-sensor residual positioning module: Used to locate the target drone parts in real time based on the positioning data of the parts' sensors and through the multi-sensor residual positioning algorithm, so as to obtain the real-time positioning result of the drone parts; Candidate image region processing module: Based on accessory image data and real-time positioning results of drone accessories, it determines the candidate image region of the target drone accessory, performs image preprocessing on the candidate image region, and obtains the target image of the drone accessory; Improved RTMDet Recognition Module: This module is used to input the target image of the UAV accessory into the improved RTMDet recognition network, perform accessory target detection and fine-grained model discrimination, and obtain the UAV accessory recognition result. The improved RTMDet recognition network introduces a codebook quantization mechanism in the accessory feature fusion process. Parts Identification and Storage Location Binding Module: Used to establish a parts identification and storage location binding relationship based on the drone parts identification results, drone parts real-time positioning results, and parts storage and retrieval management data; Verification and Strategy Generation Module: Based on the accessory identity storage location binding relationship, it performs a ternary consistency verification on the target UAV accessories, obtains the accessory access consistency verification result, and generates an intelligent accessory access strategy based on the accessory access consistency verification result. Inventory traceability module: Used to update the inventory status of drone parts and generate a full-process traceability record of drone parts based on the intelligent storage and retrieval strategy for parts.

Images