Logistics transportation abnormal state inspection method based on unmanned aerial vehicle vision

Abstract

The application provides a logistics transportation abnormal state inspection method based on unmanned aerial vehicle vision, which comprises the following steps: reading target address information, target receiving area information, historical receiving image records and historical target receiving body verification records corresponding to a logistics transportation task, constructing a historical verification sample set, and generating an initial strategy record of approaching perturbation, a target receiving body reference reaction fingerprint and an inspection preparation record based on the historical verification sample set; controlling the unmanned aerial vehicle to fly to a preset inspection position of the target receiving area based on the inspection preparation record, collecting an effective real-time inspection image sequence, identifying candidate receiving bodies and forming a candidate observation sequence set; determining target perturbation action parameters based on the candidate observation sequence set and the initial strategy record of approaching perturbation, executing the target perturbation action, collecting a perturbation response image sequence, and performing fingerprint matching based on a current reaction fingerprint and the target receiving body reference reaction fingerprint.

Description

Technical Field

[0001] This invention relates to the field of UAV visual inspection technology, and in particular to a method for inspecting abnormal conditions in logistics transportation based on UAV vision. Background Technology

[0002] As drones are increasingly used in scenarios such as park delivery, community last-mile delivery, and emergency supplies transport, existing drone logistics technologies typically first use satellite positioning, inertial navigation, or preset routes to guide the drone to the vicinity of the target address, and then combine onboard vision to identify the landing area and delivery area to complete the last-mile delivery. Among these, visual recognition has become a key capability in the drone delivery process.

[0003] Existing methods typically detect and match target receiving locations based on the appearance, location, manual markings, cabinet outlines, platform edges, or preset landmarks of the receiving area. When the scene is relatively regular, the lighting is stable, and the receiving location is clearly marked, these methods can complete basic delivery tasks. However, in real-world scenarios such as residential building balconies, park cabinet entrances, and temporary receiving platforms, multiple candidate receivers with similar appearances, adjacent locations, and similar structures often exist. Furthermore, factors such as occlusion, backlighting, reflections, shadows, rain, fog, and changes in viewing angle can further exacerbate this, causing incorrect receivers to appear visually similar to the target receiver in the image. For these situations, existing methods often rely primarily on static appearance recognition or single-match results, easily misidentifying incorrect objects as correct ones. Therefore, this invention proposes a method for inspecting abnormal states in logistics transportation based on UAV vision.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may contain prior art information that is not common knowledge to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for inspecting abnormal states in logistics transportation based on UAV vision, thereby solving the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The method for inspecting abnormal conditions in logistics transportation based on UAV vision includes the following steps: S1. Read the target address information, target receiving area information, historical receiving image records and target receiver verification records corresponding to the logistics transportation task, construct a historical verification sample set, generate an approach perturbation initial strategy record based on the historical verification sample set through imitation learning, and generate a target receiver baseline response fingerprint and inspection preparation record. S2. Based on the inspection preparation record, control the UAV to fly to the preset inspection position in the target receiving area, collect real-time inspection image sequences, identify candidate receivers and form a set of candidate observation sequences. S3. Based on the candidate observation sequence set and the initial strategy record of the approach perturbation, the bat algorithm is used to determine the target perturbation action parameters of the candidate receiver, execute the target perturbation action and collect the perturbation response image sequence, extract the current response fingerprint and generate fingerprint record to be verified. S4. Based on the fingerprint to be verified record and the target receiver's baseline response fingerprint, perform fingerprint matching and uniqueness verification to determine the target receiver's location result and generate the target receiver verification result record. S5. Based on the target receiver verification results, record the package release position, package contact position and stable position after release, determine spatial consistency, and output the receiver camouflage abnormal state, misdelivery abnormal state or normal delivery inspection conclusion.

[0007] S1 specifically includes: reading the target address information, target receiving area information, historical receiving image records, and historical target receiver verification records corresponding to the logistics transportation task; performing field integrity verification, record alignment, and invalid deletion; and outputting a historical verification sample set; selecting expert approach samples based on the historical verification sample set; establishing the correspondence between observation status records and action parameter records; and generating an initial approach perturbation strategy record through imitation learning; establishing a target receiving area coordinate system based on the historical verification sample set; extracting edge displacement changes, local reflection changes, occlusion retreat changes, and surface texture changes in groups; generating a target receiver baseline response fingerprint; and outputting an inspection preparation record in conjunction with the initial approach perturbation strategy record.

[0008] S2 specifically includes: reading the inspection preparation record, verifying the target receiving area boundary, allowable approach height range, allowable hovering position range, target receiver baseline response fingerprint, and approach perturbation initial strategy record; determining the preset inspection position and acquiring a valid real-time inspection image sequence; performing target receiving area cropping, structural integrity verification, and cross-frame correlation based on the valid real-time inspection image sequence; identifying candidate receivers and generating candidate observation records; performing observability verification and perturbation executability verification based on the candidate observation records and approach perturbation initial strategy record; deleting candidate receivers that do not meet the verification conditions; and outputting a candidate observation sequence set.

[0009] S3 specifically includes: reading the candidate observation sequence set and the initial approach perturbation strategy record; generating initial perturbation action parameters for each candidate receiver; using the bat algorithm to determine the target perturbation action parameters under action boundaries and safety constraints; and outputting the perturbation execution record. Based on the perturbation execution record, the target perturbation action is executed in the order of action execution, while simultaneously acquiring the perturbation response image sequence corresponding to the candidate receiver and deleting failed segments and invalid image frames. Based on the perturbation response image sequence, edge displacement changes, local reflection changes, occlusion retreat changes, and surface texture changes are extracted to construct the current response fingerprint and generate a fingerprint verification record.

[0010] S4 specifically includes: reading the fingerprint to be verified record and the target receiver reference response fingerprint; selecting the corresponding target receiver reference response fingerprint subset according to the viewpoint grouping identifier and attitude grouping identifier; calculating the matching degree of edge displacement change, local reflection change, occlusion setback change and bearing surface texture change; and generating an initial verification result set; sorting, uniqueness difference determination, spatial continuity verification and imaging stability verification based on the initial verification result set; determining the target receiver positioning result and outputting the target receiver positioning record; and generating a target receiver verification result record based on the target receiver positioning record, which includes a unique verification passed status or a unique verification failed status.

[0011] S5 specifically includes: reading the target receiver verification result record, acquiring image frames of the package release stage, package descent stage, and package placement stage, extracting the package release position, package contact position, and stable position after release, and generating a delivery status observation record; based on the delivery status observation record and the target receiver verification result record, performing spatial consistency determination on the package release position, package contact position, and stable position after release in sequence, and outputting a spatial consistency determination record; based on the spatial consistency determination record and the target receiver verification result record, outputting the receiver camouflage abnormal state, misdelivery abnormal state, or normal delivery inspection conclusion.

[0012] The beneficial effects of this invention are as follows: This invention addresses the problem of misidentifying receivers as genuine receivers by constructing a baseline response fingerprint of the target receiver and extracting the current response fingerprint for verification. Instead of relying solely on the static appearance of the receiver, it distinguishes between candidate receivers that are visually similar but have different physical characteristics. This solves the problem of existing technologies misidentifying disguised receivers as genuine receivers, significantly reducing the risk of misdelivery during final delivery. By generating an initial approach perturbation strategy record using a historical verification sample set and combining it with imitation learning to obtain initial perturbation action parameters that match the scene state, the UAV can proactively elicit receiver responses under near-real delivery conditions, improving the effectiveness and reproducibility of subsequent response feature extraction.

[0013] This invention introduces the Bat Algorithm into the determination of perturbation action parameters. Under the constraints of action boundaries and safety, it optimizes the adjustment of lateral swing amplitude, longitudinal swing amplitude, rise and fall amplitude, and yaw angle. This ensures that the perturbation action has sufficient distinguishability while avoiding impact on neighboring obstacles and flight safety, thus balancing recognition effectiveness and execution safety. By uniformly extracting, normalizing, and temporally stitching edge displacement changes, local reflection changes, occlusion retreat changes, and surface texture changes, a current response fingerprint that can be used for matching verification is formed. This allows receiver verification to be independent of single-frame images or single appearance features, improving robustness under complex lighting, occlusion, and viewing angle changes.

[0014] This invention, after verifying the target receiver, further combines the package release location, package contact location, and stable post-release location for spatial consistency determination. This enables a clear distinction between receiver camouflage anomalies, misdelivery anomalies, and normal delivery inspection conclusions, thereby improving the accuracy of anomaly type determination and the traceability of delivery results. A complete closed loop is formed from historical sample construction, candidate receiver screening, perturbation optimization execution, fingerprint verification to delivery consistency determination. The outputs of each stage are continued to be used in subsequent stages, thus providing a practical, verifiable, and scalable anomaly identification solution for UAV last-mile logistics inspection. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the logistics transportation anomaly inspection method based on UAV vision according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example: Figure 1 As shown, this embodiment provides a method for inspecting abnormal states in logistics transportation based on UAV vision, including the following steps: S1. Read the target address information, target receiving area information, historical receiving image records and target receiver verification records corresponding to the logistics transportation task, construct a historical verification sample set, generate an approach perturbation initial strategy record based on the historical verification sample set through imitation learning, and generate a target receiver baseline response fingerprint and inspection preparation record. S2. Based on the inspection preparation record, control the UAV to fly to the preset inspection position in the target receiving area, collect real-time inspection image sequences, identify candidate receivers and form a set of candidate observation sequences. S3. Based on the candidate observation sequence set and the initial strategy record of the approach perturbation, the bat algorithm is used to determine the target perturbation action parameters of the candidate receiver, execute the target perturbation action and collect the perturbation response image sequence, extract the current response fingerprint and generate fingerprint record to be verified. S4. Based on the fingerprint to be verified record and the target receiver's baseline response fingerprint, perform fingerprint matching and uniqueness verification to determine the target receiver's location result and generate the target receiver verification result record. S5. Based on the target receiver verification results, record the package release position, package contact position and stable position after release, determine spatial consistency, and output the receiver camouflage abnormal state, misdelivery abnormal state or normal delivery inspection conclusion.

[0018] S1 specifically includes the following sub-steps: S110: Read the target address information, target receiving area information, historical receiving image records, and historical target receiving body verification records corresponding to the logistics transportation task. First, verify the integrity of the fields, then perform record alignment and invalid deletion, and finally output the historical verification sample set. Specifically, first read the target address identifier, target receiving area boundary, allowable approach height range, and allowable hovering position range, and then read the historical receiving image records and historical target receiving body verification records associated with the target address identifier, respectively. The historical received image records include at least the task number, image frame number, acquisition timestamp, actual UAV pose information, image sharpness value, target receiving area boundary and image source identifier. The historical target receiver verification records include at least the task number, verification timestamp, target receiver positioning result, delivery conclusion, package release position, package contact position, stable position after release and manual verification mark.

[0019] Subsequently, the integrity of each of the two types of records is verified. If any field in the historical received image record is missing, such as task number, acquisition timestamp, actual UAV pose information, image sharpness value, or target receiving area boundary, the historical received image record is deleted. If any field in the historical target receiver verification record is missing, such as target receiver positioning result, delivery conclusion, package release position, package contact position, stable position after release, or manual verification mark, the historical target receiver verification record is deleted.

[0020] After field verification is completed, firstly, alignment is performed by task number, and then secondly, alignment is performed by the time difference between the collection timestamp and the verification timestamp. When the time difference is greater than a preset alignment threshold, the corresponding record pair is deleted. Preferably, the preset alignment threshold is 2 seconds. To avoid blurry samples contaminating subsequent imitation learning and benchmark construction, low-quality images are then deleted by image sharpness value. For example, when the image sharpness value is lower than 0.65, the corresponding historical received image record is deleted.

[0021] The records retained after verification, alignment, and deletion are reorganized into a historical verification sample set according to the task number. Each sample unit in the historical verification sample set includes at least the target address identifier, target receiving area boundary, continuous image frame fragments, actual UAV pose sequence, target receiver positioning result, delivery conclusion, package release position, package contact position, and stable position after release. The historical verification sample set serves as the common input for S120 and S130.

[0022] S120. Read the historical verification sample set, first filter the expert approach samples, then construct the correspondence between the observation state and the action parameters, then perform imitation learning, and finally output the initial strategy record of the approach perturbation.

[0023] Specifically, sample units that meet the following conditions are first selected from the historical verification sample set: the delivery conclusion is a normal delivery inspection conclusion, the target receiver positioning result is unique, the package release position is within the allowable release range, the package contact position falls within the boundary of the receiving area, and the stable position after release remains within the boundary of the receiving area within the subsequent observation window; only sample units that meet all the conditions are retained as expert approach samples, and the remaining sample units are deleted.

[0024] Subsequently, the following parameters are extracted from each expert approach sample: candidate receiver center position offset, candidate receiver outer contour scale value, local occlusion ratio, local reflectivity level, current hovering height, current yaw angle, and distance between neighboring obstacles, forming an observation status record; at the same time, the lateral swing amplitude, longitudinal swing amplitude, rise and fall amplitude, yaw angle adjustment amount, and duration of a single action are extracted, forming an action parameter record.

[0025] To ensure reproducibility of training results, a one-to-one correspondence is established between each observation state record and its corresponding action parameter record according to the sample sequence number, forming an expert mapping sample set. Imitation learning is then performed based on this expert mapping sample set to train an initial mapping model from the current observation state to the action parameters. The training loss is calculated using the following formula: in, The training loss value is used to characterize the overall deviation between the predicted action parameters and the actual action parameters; For the number of samples approached by experts; For the first The predicted action parameter vector corresponding to each expert approach sample; For the first The vector of real action parameters corresponding to each expert's approach sample; This refers to the 2-norm operation.

[0026] When the training loss value decreases by less than the preset convergence threshold for three consecutive rounds, the current model is marked as a completed training model. Finally, the parameter set, applicable height range, applicable occlusion threshold, action safety threshold, observation state field table, action parameter field table, and training completion marker of the completed training model are written into the approach perturbation initial policy record. The approach perturbation initial policy record serves as the input to S130, S230, and S310.

[0027] S130. Read the historical verification sample set and the initial strategy record of approach perturbation. First, establish the target receiving area coordinate system. Then, extract edge displacement changes, local reflection changes, occlusion retreat changes and bearing surface texture changes in groups. Delete the off-center samples. Then, generate the target receiving body reference response fingerprint. Finally, output the inspection preparation record.

[0028] Specifically, the geometric center of the target receiving area boundary is first used as the origin of the coordinate system, the horizontal extension direction of the target receiving area boundary is used as the horizontal coordinate axis, and the direction perpendicular to the horizontal coordinate axis is used as the vertical coordinate axis. The target receiving area coordinate system is then established, and the target receiver position, candidate receiver contour and actual UAV pose information in the historical verification sample set are uniformly converted into the target receiving area coordinate system.

[0029] Subsequently, the historical verification samples are grouped according to the yaw angle range and the hovering height range. Preferably, they can be divided into groups of 15° yaw angle and 0.5m hovering height to avoid directly mixing approach conditions with too large differences.

[0030] For each set of samples, the edge displacement change, local reflection change, occlusion receding change, and surface texture change of the target receiver in consecutive image frames are extracted sequentially to form a set of response features for the same set. If the deviation of any historical sample from any of the above four types of features relative to the median value of the same set exceeds a preset deviation threshold, the historical sample is deleted. For example, if the overall local reflectance variation of a sample group is between 0.18 and 0.26, but a single sample reaches 0.41, that sample should be deleted to avoid strong reflection interfering with the baseline value. Weighted aggregation is performed on the retained samples to obtain the reference values ​​in the target receiver baseline response fingerprint, calculated as follows: in, The first in the target receiver reference response fingerprint Reference values ​​for each reaction characteristic; These correspond to the four types of reaction characteristics mentioned above. The number of samples to retain in the same group; For the first The sample weights of the retained samples; For the first The first retained sample Measurement values ​​of each reaction characteristic.

[0031] The sample weights are determined jointly based on the image sharpness value and the degree of viewpoint matching. The higher the image sharpness value and the smaller the viewpoint deviation, the greater the sample weight. Finally, the fingerprint number, viewpoint group identifier, attitude group identifier, edge displacement change reference value, local reflection change reference value, occlusion retreat change reference value, contact surface texture change reference value, and the allowable deviation range corresponding to each reference value are written into the target receiver reference response fingerprint. The target receiver reference response fingerprint, target receiving area boundary, allowable approach height range, allowable hovering position range, and approach perturbation initial strategy record are associated and encapsulated to output the inspection preparation record. The inspection preparation record serves as the input to S210.

[0032] S2 specifically includes the following sub-steps: S210. Read the inspection preparation record, first verify whether the target receiving area boundary, allowable approach height range, allowable hovering position range, target receiver reference response fingerprint and approach perturbation initial strategy record are complete, then generate preset inspection positions and acquire images, delete image frames that do not meet the validity conditions, and finally output a valid real-time inspection image sequence.

[0033] Specifically, when any of the above fields are missing from the inspection preparation record, the process is not terminated directly. Instead, an empty sequence record with a collection failure mark is generated and output as a valid real-time inspection image sequence so that subsequent steps can be uniformly concluded.

[0034] When all fields are complete, the system uses the geometric center of the target receiving area boundary as a reference to search for spatial points within the allowable hovering position range where the unobstructed line of sight condition is met. The spatial point closest to the target receiving area boundary and meeting the safety distance requirement is determined as the preset inspection position. The unobstructed line of sight condition means that the continuous obstruction length in the line of sight from the spatial point to the receptive area within the target receiving area does not exceed 20% of the total line of sight. The safety distance requirement means that the distance between the outer edge of the UAV and the fixed component outside the target receiving area is not less than 0.8m.

[0035] After determining the preset inspection position, control the UAV to fly to the preset inspection position and perform multi-view continuous imaging in the order from high to low and from front view to side view. Preferably, first, the first set of images is collected at the upper limit of the allowable approach height range, and then the UAV descends step by step with a height step of 0.5m, and a round of images is collected at each height step with a yaw angle step of 15° to form the original inspection image sequence.

[0036] Subsequently, the validity of the original inspection image sequence is verified frame by frame. If an image frame is missing any of the following fields: image frame number, acquisition timestamp, actual UAV pose information, current hovering altitude, or current yaw angle, the image frame is deleted. If the roll angle or pitch angle exceeds a preset attitude threshold, such as exceeding 8°, the corresponding image frame is deleted. If the image sharpness value is lower than a preset sharpness threshold, such as below 0.65, the corresponding image frame is deleted.

[0037] Finally, the retained image frames are arranged in ascending order by acquisition timestamp, and the image frame number, acquisition timestamp, actual UAV pose information, current hovering altitude, current yaw angle, image sharpness value, target receiving area cropping mark, and acquisition failure mark are written into the valid real-time inspection image sequence. The valid real-time inspection image sequence serves as the input to S220.

[0038] S220: Read the valid real-time inspection image sequence, first perform target receiving area cropping and structural integrity verification, then identify candidate receivers and perform cross-frame association, delete short-term false targets and unstable structures, and finally output candidate observation records.

[0039] Specifically, when a failed acquisition in a valid real-time inspection image sequence is marked as valid, an empty candidate observation record with the reason for the missing candidate receiver is directly generated for unified processing in subsequent steps.

[0040] Under normal circumstances, the effective real-time inspection image sequence is first cropped frame by frame according to the boundary of the target receiving area, deleting image content that is completely outside the boundary of the target receiving area, and only retaining image areas that are inside the boundary of the target receiving area or intersect with the boundary of the target receiving area.

[0041] Subsequently, the receiving edges, surface textures, and opening contours are extracted from the cropped image region, and the structural integrity is verified on the extraction results. Only when a structure simultaneously satisfies the following conditions is it identified as a candidate receiver: it has continuous receiving boundaries, can accommodate the projection of the bounding rectangle, and has a stable edge boundary with the neighboring background.

[0042] Here, the outer rectangle projection of the package refers to the smallest enclosing rectangle formed by projecting the package onto the image plane according to the length and width dimensions of the package in the current logistics transportation task. For example, when the actual size of the package is 0.40m by 0.30m, the projection of the receiving area corresponding to the candidate receiver must not be less than 1.10 times the size of this rectangle.

[0043] After single-frame recognition is completed, candidate structures in adjacent image frames are correlated across frames. During correlation, the center position offset, outer contour overlap ratio, and scale change rate are calculated respectively. When all three indicators meet the preset correlation threshold, they are identified as the same candidate receiver and assigned a unified candidate receiver number.

[0044] To avoid reflective spots, birds, swaying branches and leaves, or short-term occlusion edges being mistakenly identified as candidate receivers, candidate structures that appear consecutively for fewer than a preset frame count threshold are directly deleted. Preferably, when the number of consecutive frames is less than 3, the candidate structure is deleted.

[0045] Finally, for each retained candidate receiver, a sequence of consecutive image frame segments, center position sequence, outer contour sequence, relative scale sequence, local occlusion ratio sequence, image sharpness sequence, neighborhood background structure sequence, neighborhood obstacle spacing, and corresponding actual UAV pose sequence are recorded, and the above information is written into the candidate observation record. The candidate observation record serves as the input to S230.

[0046] S230. Read the candidate observation records and the initial strategy record for approach perturbation. First, perform the observability verification, then perform the perturbation executability verification, delete the candidate receivers that do not meet any verification conditions, and finally output the candidate observation sequence set.

[0047] Specifically, when a candidate observation record is empty or contains a reason for missing candidate receivers, the set of candidate observation sequences marked as missing is directly output. Under normal circumstances, the center position sequence, outer contour sequence, local occlusion ratio sequence, image sharpness sequence, and neighborhood obstacle distance are first read from each candidate observation record, and the observability score of the candidate receiver is calculated as follows: in, For observability scoring; , and These are the rating weights, and the sum of the three rating weights is 1; This represents the average image sharpness. This represents the average proportion of local occlusion. For outer contour stability. When the observability score is lower than the preset observability threshold, the corresponding candidate observation record is deleted.

[0048] For candidate receivers that pass the observability verification, the applicable altitude range, applicable occlusion threshold, and action safety threshold in the initial approach perturbation strategy record are read again. The perturbation executability verification is then performed in combination with the current hovering altitude, the distance between neighboring obstacles, and the initial perturbation action parameters. If the current hovering altitude exceeds the applicable altitude range, the local occlusion ratio exceeds the applicable occlusion threshold, or it is determined based on the distance between neighboring obstacles that the UAV cannot safely perform perturbation actions within the allowed hovering position range, the corresponding candidate observation record is deleted.

[0049] After two rounds of verification, the retained records are organized into candidate observation sequences according to the candidate receiver number. The candidate receiver number, continuous image frame segments, center position sequence, outer contour sequence, local occlusion ratio sequence, image sharpness sequence, neighborhood obstacle spacing, observability score, perturbation feasibility verification result, and candidate receiver missing marker are then written into the candidate observation sequence set. The candidate observation sequence set serves as the input to S310.

[0050] S3 specifically includes the following sub-steps: S310. Read the candidate observation sequence set and the initial strategy record for approach perturbation. First, verify the missing markers and sequence field integrity of the candidate receivers. Then, generate initial perturbation action parameters for each candidate receiver. Use the bat algorithm to delete candidate action parameters that exceed the action boundary or do not meet the safety constraints and complete the optimization. Finally, output the perturbation execution record.

[0051] Specifically, when the candidate receiver missing marker is valid, a perturbation execution record with a marker indicating that no perturbation action will be performed is directly generated, and the reason for the missing candidate receiver is written into this record. Under normal circumstances, the candidate observation sequence is read one by one according to the candidate receiver number, and the initial mapping model in the approach perturbation initial strategy record is called. Based on the center position offset, outer contour scale, local occlusion ratio, current hovering height, and distance between neighboring obstacles, the initial perturbation action parameters consisting of lateral sway amplitude, longitudinal sway amplitude, rise and fall amplitude, yaw angle adjustment, and action duration are generated.

[0052] To avoid runaway optimization, action boundaries were set for each parameter; specifically, the lateral and longitudinal sway amplitudes did not exceed 50% of the allowable hovering position range width, the elevation range did not exceed 30% of the allowable approach height range span, the yaw angle adjustment did not exceed 30°, and the action duration did not exceed 2 seconds. Subsequently, the bat algorithm was used to perform online optimization of the initial perturbation action parameters, with the comprehensive fitness value as the optimization target, calculated as follows: in, This is the overall fitness value; , and These are the fitness weights, and the sum of the three fitness weights is 1. Characteristic separability of the reaction; To maintain image sharpness; For the degree of satisfaction of safety constraints.

[0053] During the optimization process, if a set of candidate action parameters causes the lateral sway amplitude, longitudinal sway amplitude, rise and fall amplitude, or yaw angle adjustment to exceed the corresponding action boundary, or causes the perturbation safety margin to be lower than the action safety threshold, the set of candidate action parameters should be deleted. For example, if the nearest obstacle distance is 1.20m, the action envelope distance is 1.05m, and the safety threshold is 0.20m, the corresponding candidate action parameters should be deleted because the safety margin is only 0.15m.

[0054] The candidate action parameters that have not been deleted are continuously iterated and updated. When the comprehensive fitness value increases by less than 0.01 for three consecutive generations, or when the maximum number of iterations of 20 is reached, the optimization is stopped, and the target perturbation action parameter with the highest comprehensive fitness value is retained.

[0055] Finally, the candidate receiver number, target perturbation action parameters, action execution order, action start time, action end time, expected actual UAV pose range, non-perturbation action flag, and reason for missing candidate receiver are all written into the perturbation execution record. The perturbation execution record serves as the input to S320.

[0056] S320: Read the perturbation execution record, execute the target perturbation actions one by one in the order of action execution and collect images simultaneously, then verify the correspondence between the image frames and the actual UAV pose information, delete the failed execution segments and invalid image frames, and finally output the perturbation response image sequence.

[0057] Specifically, when the non-perturbation action in the perturbation execution record is marked as valid, a perturbation response image sequence with empty image segments and the reason for not executing the perturbation action is directly output. Under normal circumstances, the candidate receiver number, target perturbation action parameters, and expected actual UAV pose range in each perturbation execution record are read in the order of action execution. After the actual UAV pose enters the expected actual UAV pose range, the corresponding target perturbation action is initiated, and continuous imaging of the corresponding candidate receiver and its neighboring region is performed during the action duration.

[0058] During continuous imaging, the image frame number, acquisition timestamp, actual UAV pose information, current hovering altitude, current yaw angle, and candidate receiver number are recorded simultaneously to form the original disturbance response image fragment; preferably, the sampling frequency is not less than 10 frames per second to ensure that edge displacement changes, local reflection changes, occlusion retreat changes, and surface texture changes are continuously traceable in time.

[0059] Subsequently, the validity of each frame of the original disturbance response image segment is verified. If any of the following fields are missing from the image frame number, acquisition timestamp, actual UAV pose information, current hovering altitude, current yaw angle, or candidate receiver number, the image frame is deleted. If a sudden change in the UAV's attitude occurs during the execution, the candidate receiver completely goes out of the frame, the image clarity is continuously lower than the preset clarity threshold, or the actual UAV pose deviates from the expected actual UAV pose range, the corresponding abnormal image frame is deleted.

[0060] To ensure consistency between the image and the pose, the acquisition timestamp is used as the primary key to match each image frame with the actual UAV pose information with the smallest time difference. When the time difference between the acquisition timestamp of the image frame and the timestamp of the actual UAV pose information exceeds a preset alignment threshold, such as exceeding 0.10s, the image frame is deleted.

[0061] After deletion, count the number of valid image frames corresponding to each target perturbation action; when the number of valid image frames is lower than the preset frame threshold, such as less than 8 frames, mark the entire image segment corresponding to the target perturbation action as invalid.

[0062] Finally, all valid image segments are rearranged according to candidate receiver number and action execution order to form a disturbance response image sequence. Each record in the disturbance response image sequence includes at least the candidate receiver number, action execution order, consecutive valid image frame segments, acquisition timestamp sequence, actual UAV pose sequence, current hovering altitude sequence, current yaw angle sequence, and reason for not performing the perturbation action. The disturbance response image sequence serves as the input to S330.

[0063] S330. Read the disturbance response image sequence, first extract edge displacement changes, local reflection changes, occlusion retreat changes and surface texture changes, then uniformly map them to the target receiving area coordinate system and perform normalization processing, delete sequences with missing features or abnormal temporal jumps, and finally output the fingerprint record to be verified.

[0064] Specifically, when the disturbance response image sequence contains a reason for not performing the perturbation action, a fingerprint record to be verified with an invalid verification mark is directly generated.

[0065] Under normal circumstances, for each candidate receiver corresponding to the perturbation response image sequence, four types of response features are first extracted from consecutive valid image frames. Among them, edge displacement change is defined as the change in position of the receiving edge of the candidate receiver in consecutive image frames, local reflection change is defined as the change in brightness and distribution position of the local bright area of ​​the candidate receiver, occlusion retreat change is defined as the retreat of the edge of the occluder in the neighborhood of the candidate receiver in consecutive image frames, and receiving surface texture change is defined as the texture response difference of the receiving surface texture of the candidate receiver in consecutive image frames.

[0066] After extraction, the feature point positions in each image frame are uniformly transformed to the target receiving area coordinate system using the corresponding actual UAV pose information, with the geometric center of the target receiving area boundary as the unified reference point. To eliminate the influence of different candidate receiver scale differences on the matching results, the above four types of features are normalized based on the outer contour scale value corresponding to the candidate receiver, resulting in normalized response feature vectors.

[0067] Subsequently, the normalized reaction feature vectors are concatenated temporally according to the order of action execution and the order of data acquisition to form the current reaction fingerprint, which is expressed as follows: in, This is the current reaction fingerprint; For the first The normalized response feature vectors corresponding to each valid image frame, and ; The number of valid image frames. If a disturbance response image sequence has continuous feature loss, continuous feature jumps exceeding a preset jump threshold, or the number of valid image frames is less than a preset frame number threshold, then the current response fingerprint corresponding to the candidate receiver is deleted; for example, if the normalized difference of the edge displacement change of the same candidate receiver in two adjacent frames continuously exceeds 0.35, and the corresponding local reflection change directions are opposite, it can be determined as a temporal jump anomaly.

[0068] Finally, the candidate receiver number, current response fingerprint, target perturbation action parameters, action execution order, number of valid image frames, normalization marker, center position sequence, outer contour sequence, image sharpness sequence, corresponding actual UAV pose sequence, and verification invalidation marker are all written into the fingerprint verification record. The fingerprint verification record serves as the input to S410.

[0069] S4 specifically includes the following sub-steps: S410. Read the fingerprint record to be verified and the target receiver reference response fingerprint. First, verify whether the record fields and grouping identifiers are complete. Then, select the corresponding target receiver reference response fingerprint subset according to the view grouping identifier and attitude grouping identifier. Subsequently, calculate the edge displacement change matching degree, local reflection change matching degree, occlusion retreat change matching degree and bearing surface texture change matching degree respectively, and perform temporal consistency verification. Delete invalid records that lack core feature sequences or corresponding reference subsets. Finally, output the initial verification result set.

[0070] Specifically, when the invalidation flag in a fingerprint verification record is valid, an initial verification result with the reason for invalidation is directly generated. Under normal circumstances, the candidate receiver number, current response fingerprint, action execution order, center position sequence, outer contour sequence, image sharpness sequence, and corresponding actual UAV pose sequence are read one by one; if any record is missing any field of candidate receiver number, current response fingerprint, or action execution order, the fingerprint verification record is deleted.

[0071] Subsequently, the fingerprint number, view group identifier, attitude group identifier, edge displacement change reference value, local reflection change reference value, occlusion retreat change reference value, surface texture change reference value, and corresponding allowable deviation range in the target receiver reference response fingerprint are read. The corresponding target receiver reference response fingerprint subset is selected according to the view group identifier and attitude group identifier to which the current response fingerprint belongs. If no corresponding subset exists, the record corresponding to the candidate receiver is deleted.

[0072] After subset selection, the four feature sequences in the current reaction fingerprint are compared with the corresponding reference sequences in the target receiver's baseline reaction fingerprint subset, and the matching degree is calculated based on the allowable deviation range. To ensure that the comparison results of the four feature types can be uniformly ranked, the matching degrees are weighted and merged into a fingerprint matching score, which is calculated as follows: in, Fingerprint matching score; , , and For matching weights, the sum of the four matching weights is 1; For edge displacement variation matching degree; For matching degree of local reflection changes; To compensate for changes in matching degree due to occlusion; To match the texture variation of the receiving surface.

[0073] Preferably, when the edge displacement change and occlusion retreat change have a higher ability to distinguish the receiver, the efficiency can be improved. and The value of is determined. After obtaining the fingerprint matching score, a temporal consistency check is performed on the current reaction fingerprint. Specifically, the peak occurrence order, change direction, and duration interval of the four types of features are checked sequentially to see if they are consistent with the corresponding baseline subset. If the peak occurrence order of any core feature is reversed, or the change direction of two consecutive core features is opposite to the baseline direction, the temporal consistency check result is recorded as failing.

[0074] Finally, the candidate receiver number, the target receiver baseline response fingerprint subset number, fingerprint matching score, four types of feature matching degree, temporal consistency verification result, center position sequence, outer contour sequence, image sharpness sequence, current response fingerprint, corresponding perturbation response image sequence index, and verification validity marker are written into the initial verification result set. The initial verification result set serves as the input to S420.

[0075] S420. Read the initial verification result set, first filter the candidate receivers that are valid and pass the temporal consistency verification, then sort them according to the fingerprint matching score, and combine the uniqueness difference, spatial continuity and imaging stability to determine whether there is a unique target receiver that has passed the verification, delete the candidate receivers that do not meet the inclusion conditions, and finally output the target receiver positioning record.

[0076] Specifically, the candidate receiver number, fingerprint matching score, timing consistency verification result and verification validity mark are first read from the initial verification result set. Only candidate receivers with a valid verification validity mark and a passing timing consistency verification result are retained; the remaining candidate receivers do not participate in the subsequent sorting.

[0077] If no valid candidate recipients are found after screening, a unique verification mark is generated directly, and the reason for failure is recorded as missing candidate recipients or invalid verification.

[0078] If valid candidate receivers exist, they are sorted from highest to lowest fingerprint matching score to obtain the top-ranked and second-ranked candidate receivers. To fix the uniqueness determination path, the uniqueness difference between the top-ranked and second-ranked scores is calculated using the following formula: in, This is a unique difference; The fingerprint matching score of the first candidate receiver is used to rank them. The fingerprint matching score is the next highest-ranking candidate receiver; when only one candidate receiver remains after filtering, it will be... Recorded as 0. After completing the uniqueness difference calculation, the center position sequence, outer contour sequence, and image sharpness sequence corresponding to the first candidate receiver are read, and spatial continuity verification and imaging stability verification are performed. Among them, the spatial continuity verification is used to determine whether the displacement of the center position sequence in consecutive image frames is within the allowable fluctuation range, and the imaging stability verification is used to determine whether the outer contour sequence is continuously closed and whether the image sharpness sequence is continuously higher than the preset sharpness threshold.

[0079] The first candidate receiver in the ranking is determined to be the only target receiver that has passed the verification only when the score of the first ranking is not lower than the preset verification threshold, the uniqueness difference is not lower than the preset uniqueness difference threshold, the spatial continuity verification is passed, and the imaging stability verification is passed. For example, when the score of the first ranking is 0.86, the score of the second ranking is 0.61, the uniqueness difference is 0.25, and the displacement fluctuation of the center position sequence for 5 consecutive frames is less than 8% of the width of the receiving area, it can be determined to be the only target receiver that has passed the verification.

[0080] If all the above conditions are not met, the current task is determined to have failed the unique verification. For cases where the unique verification is passed, the center position sequence, outer contour sequence, and corresponding actual UAV pose sequence of the candidate receiver in consecutive image frames are read to calculate the target receiver center position, receiving area boundary, and allowable release range, forming the target receiver positioning result. For cases where the unique verification is not passed, the target receiver center position is not generated; instead, a unique verification failure marker, a corresponding candidate receiver number set, a current response fingerprint set, and a corresponding disturbance response image sequence index set are generated.

[0081] Finally, the task number, verification status, target receiver number or candidate receiver number set, target receiver center location, receiving area boundary, allowed release range, first-rank score, second-rank score, uniqueness difference, current response fingerprint set, and corresponding disturbance response image sequence index set are written into the target receiver location record. The target receiver location record serves as a common input for both S430 and S510.

[0082] S430. Read the target receiver location record, first convert it into a target receiver verification status that can be directly called in the subsequent delivery link according to the verification status, then complete the record fields according to different verification statuses, delete the result records with incomplete fields, and finally output the target receiver verification result record.

[0083] Specifically, the task number, verification status, target receiver number or candidate receiver number set, target receiver center location, receiving area boundary, allowed release range, first ranking score, second ranking score, uniqueness difference, current response fingerprint set and corresponding disturbance response image sequence index set are read from the target receiver location record. The verification status is limited to two enumerated states: unique verification passed and unique verification failed. No intermediate states are set.

[0084] If the verification status is "uniquely passed", the target receiver number, target receiver center location, receiving area boundary, allowed release range, fingerprint matching score, and uniqueness difference will be written into the result record, and the result record will be marked as a valid verification result that can be directly used for S510 to collect delivery status information. If the verification status is "uniquely failed", the candidate receiver number set, the fingerprint matching score corresponding to each candidate receiver, the failure reason mark, the current response fingerprint set, the corresponding disturbance response image sequence index set, and the abnormal review mark will be written into the result record. Among them, the failure reason mark includes at least the following: candidate receiver missing, first score lower than the verification pass threshold, uniqueness difference lower than the uniqueness difference threshold, spatial continuity verification failed, and imaging stability verification failed.

[0085] Subsequently, a field integrity check is performed on the result record. If any field is missing from the target receiver center location, receiving area boundary, or allowed release range in the unique verification state, the result record is deleted and the verification anomaly flag is written back. If any field is missing from the candidate receiver number set, failure reason flag, current response fingerprint set, or corresponding disturbance response image sequence index set in the unique verification state, the result record is deleted and the verification anomaly flag is written back.

[0086] Finally, the verification results are compiled into a target receiver verification result record. Specifically, when the target receiver verification status is "uniquely passed," S510 retrieves the package release position, package contact position, and stable position after release from the target receiver's center location, receiving area boundary, and allowed release range. When the target receiver verification status is "uniquely failed," S510 directly includes the current delivery process in the anomaly review record based on the anomaly review flag. The target receiver verification result record serves as input for S510 and S530.

[0087] S5 specifically includes the following sub-steps: S510. Read the target receiver verification result record. First, verify whether the verification status, target receiver center position, receiving area boundary, allowable release range and abnormal review mark are complete. Then, collect delivery status information according to the verification status. Delete invalid image frames that are missing key fields or insufficient to locate the release time, contact time and stabilization time. Finally, output the delivery status observation record.

[0088] Specifically, when the target receiver verification status is uniquely verified, the image frames corresponding to the package release stage, package descent stage, and package placement stage are continuously read, and the image frame number, acquisition timestamp, actual UAV pose information, and package outer contour are read simultaneously; if any image frame is missing any field of the image frame number, acquisition timestamp, actual UAV pose information, or package outer contour, the image frame is deleted.

[0089] Subsequently, the separation relationship between the outer contour of the package and the drone mounting position is detected. When the outer contour of the package first separates from the mounting position and no longer overlaps with the mounting position in the next two consecutive frames, the center position of the package at that moment is recorded as the package release position.

[0090] The contact relationship between the outer contour of the package and the boundary of the receiving area is continuously monitored. When the outer contour of the package first contacts the boundary of the receiving area or the inner surface of the receiving area, and the movement direction of the lower edge of the package changes from downward to attachment or rebound, the center position of the package at that moment is recorded as the package contact position. To determine the stable position after release, the displacement of the package center position between adjacent image frames is calculated as follows: in, For the first Frame and the The displacement of the wrap center position between frames; For the first The center position of the package in the frame. This represents the center position of the package in frame t-1. This is a 2-norm operation. When the displacement within a consecutive preset number of frames is less than a preset stability threshold, the center position of the package in the last frame of that consecutive frame segment is recorded as the stable position after release; for example, when the displacement in 5 consecutive frames is less than 0.03m, the package is determined to have entered a stable state.

[0091] When the target recipient's verification status is "failed unique verification", the package center position and outer contour are still collected in the above manner, but all collection results are written into the abnormal review record and are not entered into the normal consistency judgment.

[0092] Finally, the mission number, target receiver verification status, target receiver center position, receiving area boundary, allowable release range, image frame number sequence, acquisition timestamp sequence, actual UAV pose sequence, package outer contour sequence, package release position, package contact position, stable position after release, and anomaly verification marker are written into the delivery status observation record. The delivery status observation record serves as input to the S520.

[0093] S520: Read the delivery status observation record and the target receiver verification result record. First, verify whether the package release position, package contact position and stable position after release are complete. Then, perform spatial consistency determination according to the release order, delete invalid records with missing position fields or insufficient observation windows, and finally output the spatial consistency determination record.

[0094] Specifically, when the target receiver's verification status is uniquely verified, the following checks are performed: first, whether the package release position is within the allowed release range; second, whether the package contact position falls within the receiving area boundary; and finally, whether the stable position after release remains within the receiving area boundary. "Remaining within the receiving area boundary" means that from the moment of stabilization, the package's center position remains within the receiving area boundary throughout the subsequent preset observation window, for example, no boundary crossing occurs within the subsequent 1-second observation window. To ensure that the three verification results can be recorded uniformly, a spatial consistency score is constructed, calculated as follows: in, Spatial consistency score; , and The three decision weights are set to a weight, and the sum of the three decision weights is 1. To ensure consistent release position results; The result is consistent with the contact position; The results are consistent with the stable position. The results of release position consistency, contact position consistency, and stable position consistency are all represented by binary values, with 1 for consistency and 0 for inconsistency.

[0095] When all three results are 1, spatial consistency is determined to be valid; when any result is 0, spatial consistency is determined to be invalid. When the target receiver verification status is "failed unique verification," the above three position comparisons are no longer performed, and the spatial consistency determination result is directly recorded as "spatial consistency invalid," and the "failed unique verification" flag is written to the record as the upstream cause. If the delivery status observation record lacks any of the fields for package release position, package contact position, or stable position after release, or if the subsequent observation window is insufficient to verify whether the stable position after release continues to remain within the boundary of the receiving area, the normal determination record is deleted and transferred to the abnormal review record.

[0096] Finally, the task number, target receiver verification status, release position consistency result, contact position consistency result, stable position consistency result, spatial consistency score, spatial consistency judgment result, and failure reason flag are written into the spatial consistency judgment record. The spatial consistency judgment record serves as the input to S530.

[0097] S530: Read the spatial consistency judgment record and the target receiver verification result record. First, perform conclusion classification according to the verification status and spatial consistency judgment result. Then, verify whether the conclusion fields are complete. Delete invalid results that are missing the corresponding necessary fields for the conclusion. Finally, output the inspection conclusion record.

[0098] Specifically, when the target receiver verification status is "failed unique verification", the receiver spoofing abnormal status is directly output, and the candidate receiver number set, the fingerprint matching score corresponding to each candidate receiver, the current response fingerprint set, the corresponding disturbance response image sequence index set, and the reason for failing unique verification are written into the inspection conclusion record for verification of the target receiver's authenticity. When the upstream reason is that the candidate receiver is missing or the acquisition failed, it is also written into the upstream failure reason field of the inspection conclusion record.

[0099] When the local target receiver verification status is uniquely passed and the spatial consistency judgment result is invalid, the misdelivery abnormal status is output, and the target receiver number, target receiver center position, receiving area boundary, allowable release range, package release position, package contact position, stable position after release, spatial consistency score and failure reason mark are written into the inspection conclusion record for review of package spatial landing point deviation.

[0100] When the local target receiver verification status is uniquely verified and the spatial consistency judgment result is valid, a normal delivery inspection conclusion is output, and the target receiver number, target receiver positioning result, package release position, package contact position, stable position after release, and delivery completion mark are written into the inspection conclusion record to generate a delivery completion certificate.

[0101] If any inspection conclusion lacks a required field, the final inspection conclusion is not output directly. Instead, an incomplete field marker is written back, and the corresponding record is included in the anomaly review record. This completes the full-chain closed-loop processing of the logistics transportation anomaly inspection workflow based on UAV vision.

[0102] All the above formulas are performed using dimensionless numerical calculations; the relevant formulas are based on empirical models that approximate the real situation, obtained through extensive data collection and software simulation fitting. The preset parameters and thresholds involved in the formulas can be conventionally set and adjusted by those skilled in the art according to the physical constraints of the actual application scenario.

[0103] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0104] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0105] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for inspecting abnormal conditions in logistics transportation based on UAV vision, characterized in that, Includes the following steps: S1. Read the target address information, target receiving area information, historical receiving image records and target receiver verification records corresponding to the logistics transportation task, construct a historical verification sample set, generate an approach perturbation initial strategy record based on the historical verification sample set through imitation learning, and generate a target receiver baseline reaction fingerprint and inspection preparation record. S2. Based on the inspection preparation record, control the UAV to fly to the preset inspection position in the target receiving area, collect real-time inspection image sequences, identify candidate receivers and form a set of candidate observation sequences. S3. Based on the candidate observation sequence set and the initial strategy record of the approach perturbation, the bat algorithm is used to determine the target perturbation action parameters of the candidate receiver, execute the target perturbation action and collect the perturbation response image sequence, extract the current response fingerprint and generate fingerprint record to be verified. S4. Based on the fingerprint to be verified record and the target receiver's baseline response fingerprint, perform fingerprint matching and uniqueness verification to determine the target receiver's location result and generate the target receiver verification result record.

2. The method for inspecting abnormal states in logistics transportation based on UAV vision according to claim 1, characterized in that, Also includes: S5. Based on the target receiver verification results, record the package release position, package contact position and stable position after release, determine spatial consistency, and output the receiver camouflage abnormal state, misdelivery abnormal state or normal delivery inspection conclusion.

3. The method for inspecting abnormal states in logistics transportation based on UAV vision according to claim 1, characterized in that, S1 specifically includes: Read the target address information, target receiving area information, historical receiving image records and historical target receiving body verification records corresponding to the logistics transportation task, perform field integrity verification, record alignment and invalid deletion, and output the historical verification sample set; Expert approach samples were selected based on historical verification sample sets, a correspondence between observation state records and action parameter records was established, and initial approach perturbation strategy records were generated through imitation learning.

4. The method for inspecting abnormal states in logistics transportation based on UAV vision according to claim 3, characterized in that, Also includes: Based on the historical verification sample set, a target receiving area coordinate system is established. The edge displacement change, local reflection change, occlusion retreat change and bearing surface texture change are extracted in groups to generate the target receiving body reference response fingerprint. Combined with the approach perturbation initial strategy, the inspection preparation record is recorded and output.

5. The method for inspecting abnormal states in logistics transportation based on UAV vision according to claim 1, characterized in that, S2 specifically includes: Read the inspection preparation record, verify the target receiving area boundary, allowable approach height range, allowable hovering position range, target receiver baseline response fingerprint and approach perturbation initial strategy record, determine the preset inspection position and collect effective real-time inspection image sequence; Based on the effective real-time inspection image sequence, target receiving area cropping, structural integrity verification and cross-frame correlation are performed to identify candidate receivers and generate candidate observation records; Based on the candidate observation records and the initial strategy records for approach perturbation, perform observability verification and perturbation executability verification, delete candidate receivers that do not meet the verification conditions, and output a set of candidate observation sequences.

6. The method for inspecting abnormal states in logistics transportation based on UAV vision according to claim 1, characterized in that, S3 specifically includes: Read the candidate observation sequence set and the initial strategy record for approach perturbation, generate initial perturbation action parameters for each candidate receiver, and use the bat algorithm to determine the target perturbation action parameters under action boundaries and safety constraints, and output the perturbation execution record; Based on the perturbation execution record, the target perturbation action is executed in the order of action execution, and the perturbation response image sequence corresponding to the candidate receiver is collected simultaneously, and the execution failure segment and invalid image frame are deleted; Based on the perturbation response image sequence, edge displacement changes, local reflection changes, occlusion and receding changes, and surface texture changes are extracted to construct the current response fingerprint and generate fingerprint records to be verified.

7. The method for inspecting abnormal states in logistics transportation based on UAV vision according to claim 1, characterized in that, S4 specifically includes: Read the fingerprint record to be verified and the reference response fingerprint of the target receiver, select the corresponding subset of the reference response fingerprint of the target receiver according to the view grouping identifier and the attitude grouping identifier, calculate the matching degree of edge displacement change, local reflection change, occlusion retreat change and bearing surface texture change, and generate an initial verification result set; Based on the initial verification result set, sorting, uniqueness difference determination, spatial continuity verification, and imaging stability verification are performed to determine the target receiver positioning result and output the target receiver positioning record.

8. The method for inspecting abnormal states in logistics transportation based on UAV vision according to claim 7, characterized in that, Also includes: A target receiver verification result record is generated based on the target receiver location record. The target receiver verification result record includes a unique verification status or a unique verification failure status.

9. The method for inspecting abnormal states in logistics transportation based on UAV vision according to claim 1, characterized in that, S5 specifically includes: Read the target receiver verification result record, collect image frames of the package release stage, package falling stage and package landing stage, extract the package release position, package contact position and stable position after release, and generate delivery status observation record; Based on the delivery status observation record and the target receiver verification result record, spatial consistency determination is performed sequentially on the package release position, package contact position and stable position after release, and spatial consistency determination record is output. Based on the spatial consistency determination record and the target receiver verification result record, the system outputs the receiver camouflage abnormal state, misdelivery abnormal state, or normal delivery inspection conclusion.

Images