720 results about "Feature descriptor" patented technology

The invention discloses an unmanned aerial vehicle routing inspection line adaptive obstacle detection method and system based on a monocular camera. The method comprises the steps of obtaining initial flight parameters, evaluating and optimizing the flight parameters based on scene complexity, collecting an original image sequence and carrying out quality evaluation and exposure compensation. Constructing a multi-resolution feature pyramid, extracting feature descriptors, and performing feature matching and verification; performing multi-constraint pose optimization and depth estimation based on a feature matching result, and reconstructing a three-dimensional point cloud; calculating direction enhancement response to extract a power line candidate area, performing multi-view line segment matching and three-dimensional reconstruction, and analyzing dynamic characteristics of the power line; and based on the power line three-dimensional model and the depth map, carrying out obstacle segmentation, calculating an obstacle distance, evaluating a risk level, and generating early warning information and a monitoring scheme. According to the invention, different scenes can be adaptively processed, the obstacle risk can be accurately identified and evaluated, and the inspection efficiency and reliability are improved.

The invention provides a circuit board welding fault identification method and system based on machine vision, and relates to the technical field of visual identification, and the method comprises the steps: obtaining an image of a circuit board welding region through employing a structured light projection system and a multi-camera imaging system; realizing accurate positioning of a defect region in the region of interest by utilizing three-dimensional shape information and through morphological operation and an edge detection algorithm; for the accurately positioned defect area, obtaining high-precision three-dimensional point cloud data of the defect area, extracting a multi-scale curvature feature descriptor, and fusing the multi-scale curvature feature descriptor with visual features to form a multi-feature descriptor; inputting the multi-feature descriptors into a defect classification and recognition module based on ensemble learning to obtain the types and severity of the welding defects; and according to a defect identification result, the three-dimensional space distribution and severity of the welding defects are visually displayed.

The invention provides an unmanned aerial vehicle accessory intelligent identification and access method and system based on a sensor, and relates to the technical field of unmanned aerial vehicle accessory management, and the method comprises the steps: collecting the three-dimensional contour information of unmanned aerial vehicle accessories, generating standard point cloud data, carrying out the feature extraction, and carrying out the feature alignment and interaction, generating a high-dimensional fusion feature matrix of the target unmanned aerial vehicle accessories; based on the feature pyramid network, constructing multi-scale feature representation, generating hierarchical feature descriptors, performing similarity calculation, generating a similarity score matrix, performing optimal matching, establishing an adaptive verification rule base and performing accessory verification; calculating an optimal action sequence of an access path based on model information, mapping the optimal action sequence into a motion track parameter of a mechanical arm, collecting multi-point stress data, predicting the optimal clamping force, establishing an accessory service life prediction model, performing prediction in combination with historical data, and generating an accessory state evaluation report and a maintenance suggestion.

The invention relates to an image processing technology, and discloses an efficient panoramic image splicing method and system based on multi-view fusion, and the method comprises the steps: collecting a plurality of images from different views; performing multi-scale feature extraction on each image, generating a feature descriptor for each extracted feature point, and matching a corresponding feature point pair; performing multi-view geometric constraint screening on the feature point pairs; according to the screened feature point pairs, estimating a homography matrix between adjacent images, and carrying out global optimization on the homography matrix; aligning all the images into the same coordinate system; determining an overlapping region between adjacent images; and according to the pixel information in the overlapping areas, fusing the overlapping areas by adopting a self-adaptive weighted fusion algorithm so as to splice the plurality of images into a panoramic image. The invention further discloses a control device and a computer readable storage medium. The invention aims to improve the efficiency and accuracy of generating the multi-view fused panoramic image.

The invention belongs to the technical field of target detection and recognition, and particularly relates to a thinking chain-based large and small model collaborative target detection and recognition method. According to the method, the small model is responsible for most of easy-to-detect targets, the calculation pressure of the large model is reduced, the large model is responsible for suspected samples, vision and language multi-mode reasoning is combined, the overall false detection rate and the omission ratio are both reduced, confidence evaluation is conducted through the joint probability, automatic screening and manual rechecking of uncertain results are achieved, the reliability of key results is guaranteed, and the method is suitable for large-scale popularization and application. According to the'pseudo thinking chain + pseudo label 'method, by means of reasoning and labels generated by the model, data dependence on manual labeling is reduced, only low-confidence samples are manually confirmed, the human intervention range is narrowed, the human cost is remarkably saved, and semantic information with finer granularity is provided for the model by introducing phrase-level feature descriptors. And the identification capability of complex target attributes and states is improved.

The invention relates to the technical field of chemical engineering and machine learning, and provides a method for predicting the solubility of carbon dioxide in a deep-eutectic solvent based on a machine learning model, and the method comprises the steps: collecting CO2 solubility data, carrying out the coding and feature descriptor conversion of a deep-eutectic solvent molecular structure through an open-source chemoinformatics software package RDKit, and constructing a data set. Four algorithms of XGBoost, random forest, a deep neural network and a convolutional neural network are adopted to construct a model, and hyper-parameters are optimized through grid search and k-fold cross validation. After the model is trained, the prediction performance is evaluated, and the optimal model is selected for CO2 solubility prediction. The method improves prediction efficiency and accuracy, has strong adaptability and generalization ability, and provides an effective tool for formula optimization of the eutectic solvent and improvement of CO2 capture efficiency.

The invention provides a modified metal surface defect detection method and device and a medium, and the method comprises the steps: obtaining a multi-angle reflection image sequence of a modified metal surface under the irradiation of a multi-spectral light source, and generating a defect sensitive parameter set based on a preset modified metal material characteristic database and in combination with the spectral reflectivity distribution information of the multi-angle reflection image sequence; performing cooperative feature enhancement on the multi-angle reflection image sequence and the defect sensitive parameter set, and enhancing the feature contrast of a defect area and a normal area through weight distribution to obtain an enhanced defect feature set; joint anomaly detection of a spatial domain and a spectral domain is carried out on the enhanced defect feature set, a potential defect region of the modified metal surface is obtained through identification, and a defect region feature descriptor is generated; and performing defect morphological quantitative analysis according to the defect region feature descriptors, and determining the type, position and severity level of the modified metal surface defect. According to the invention, the comprehensiveness and reliability of a defect detection result can be improved.

The invention discloses a track generation method and device for casting blank deburring and a medium, and relates to the technical field of grinding process.The track generation method mainly comprises the steps that according to FPFH feature descriptors of respective outer contours of an actual point cloud model and a standard point cloud model, key points matched with the standard point cloud model are searched through a KD tree, and the FPFH feature descriptors of the outer contours of the actual point cloud model and the standard point cloud model are obtained; performing multi-contour segmentation division of the actual point cloud model according to the distance between the key points; receiving polishing parameters input by the user based on the segmentation division results of the contours, and generating an initial track according to the polishing parameters; smoothing discrete trajectory points of the initial trajectory through B-spline interpolation, and generating a local obstacle avoidance path in combination with an industrial robot kinematics model and a fast expansion random tree star algorithm; outputting the final polishing track to a control terminal, and adjusting the track depth in real time according to the feedback of the force sensor. According to the method, regional differentiation track generation is supported, and the process adaptability and the personalized demand response capability of a complex contour are improved.

The invention discloses a method, device and system based on acoustic-thermal multi-mode fusion imaging, and relates to the technical field of artificial intelligence, and the method comprises the steps: firstly, collecting ultrasonic and thermal imaging data of target industrial equipment through an acoustic imager, and obtaining acoustic and thermal imaging features respectively; performing feature integration processing to obtain a first acoustic-thermal joint feature so as to determine a first modal integration unit corresponding to the acoustic feature, completing cross-modal fusion to obtain a first feature descriptor, similarly obtaining a second acoustic-thermal joint feature, a second modal integration unit and a second feature descriptor, and determining a fault evaluation result in combination with a fault type label. And then calibrating image data to construct a three-dimensional acoustic-thermal fusion substrate image and marking a result to obtain a diagnosis fusion image so as to realize accurate diagnosis of the target industrial equipment.

The invention discloses an unmanned aerial vehicle target intelligent identification and perception system and method based on deep learning, and relates to the technical field of artificial intelligence, and the method comprises the steps: extracting the importance feature of each target region, calculating a region importance score, generating a target region monitoring plan, and obtaining continuous frame images; an improved YOLOv5 model is adopted, continuous frames of images are input, a bounding box of a suspected target wild animal is output and mapped to an actual geographic coordinate system, and actual position coordinates are obtained; key points of the t-th frame image are detected, coordinates of the key points are extracted, visibility scores and confidence coefficients are calculated, and feature descriptors are formed; performing target detection and key point detection on the (t + 1) th frame of image to obtain actual position coordinates and feature descriptors of candidate targets, performing track matching, updating the actual position coordinates and the feature descriptors of the suspected target wild animals if the actual position coordinates and the feature descriptors are matched, and defining new target animals if the actual position coordinates and the feature descriptors are not matched; counting according to the track matching result to obtain the total number of the target wild animals.

The invention relates to the technical field of target detection, and discloses a pet target detection method and device and a camera, and the method comprises the steps: carrying out the motion triggering collection and image enhancement preprocessing of a front end region of a feeder, and obtaining an enhanced image frame sequence; performing feature extraction of dynamic receptive field adjustment on the enhanced image frame sequence to obtain pet feature descriptors and position information; behavior time sequence feature analysis is executed, and pet behavior sequence feature vectors are obtained; constructing a state transition diagram according to the pet behavior sequence feature vector, and performing time sequence consistency analysis to obtain a pet state judgment result; power management and decision execution are carried out on the feeder based on the pet state judgment result, feeding control under the low-power-consumption condition is achieved, behavior misjudgment caused by posture fluctuation is effectively avoided, the behavior recognition accuracy is improved, the accurate feeding control problem in a multi-pet family is solved, and the user experience is improved.

The invention discloses an operation and maintenance manipulator intelligent control method and system based on visual identification, and relates to the technical field of intelligent manipulator control, and the method comprises the steps: collecting RGB image data and depth image data of an operation and maintenance operation area, and obtaining a standardized image matrix and a mapping relation matrix; inputting the standardized image matrix into an improved ResNet residual network model, generating a comprehensive feature descriptor, and calculating a spatial position coordinate and an attitude angle of the target equipment based on the mapping relation matrix; based on the current joint angle state of the manipulator, an improved Jacobian matrix inverse kinematics algorithm is used for solving a target angle sequence of each joint, a preset operation mode library is matched based on the comprehensive feature descriptor, and a grabbing force parameter and a motion speed parameter are determined; and converting the target angle sequence into a control instruction, and sending the control instruction to each joint driver of the manipulator to drive the manipulator to complete action planning. According to the invention, full-process automation from environment perception to task execution is realized.

The invention provides an industrial robot welding track real-time optimization method and system based on deep learning, and relates to the technical field of industrial robots, and the method comprises the steps: collecting three-dimensional point cloud data, carrying out the multi-scale processing to generate a high-precision curved surface, constructing a welding seam feature descriptor, and carrying out the real-time optimization of the welding track of an industrial robot; and the mapping relation between the technological parameters and the welding posture is optimized on the basis of a multi-target reward function, the mapping relation is converted into an initial movement track, the track is dynamically optimized through a segmented self-adaptive optimization model and a variable structure filtering algorithm, and self-adaptive adjustment of the welding process is achieved. The welding precision and quality of the complex curved surface are improved, and the adaptability and stability of robot welding are enhanced.

The invention discloses a power battery health state on-line monitoring and early warning method and system based on multi-data fusion. The method comprises the following steps: synchronously acquiring multi-dimensional heterogeneous data in a battery operation process through a multi-sensor data acquisition module; carrying out standardization processing and preprocessing on the multi-dimensional heterogeneous data, extracting a comprehensive feature vector and carrying out multi-sensor data fusion; a multi-sensor data fusion algorithm is adopted to carry out weight distribution on the fused comprehensive feature vector, a dynamic weight coefficient is calculated according to the sensitivity and reliability of each parameter to the health state of the battery, and a fused battery state feature descriptor is obtained; and performing mode recognition and state classification on the fused battery state feature descriptors by using a pre-established evaluation model to realize quantitative evaluation of the health degree of the battery. According to the invention, real-time monitoring, evaluation and early warning of the health state of the battery in the whole life cycle are realized, and a reliable basis is provided for safe operation and maintenance decision of a battery system.

The invention relates to a curved surface contour analysis method and related equipment for laser multi-axis linkage machining, and the method comprises the following steps: carrying out the curved surface reconstruction of a target curved surface workpiece based on a three-dimensional point cloud data set, and obtaining a digital curved surface model; performing shape analysis on the digital curved surface model to obtain a curved surface feature descriptor, and if the curved surface feature descriptor does not meet an expected geometric characteristic standard, performing spectral decomposition on the digital curved surface model based on the curved surface feature descriptor to obtain a contour feature spectrum; performing path planning on the contour characteristic spectrum based on a geodesic theory to obtain a processing track sequence; and performing kinematics solution on the processing track sequence to obtain linkage control parameters of each axis, and adjusting the linkage control parameters of each axis based on the expected geometric characteristic standard, thereby solving the technical problem that the processed curved surface workpiece cannot reach the expected geometric characteristic standard due to lack of an effective curved surface contour analysis method.

The invention provides a brain aneurysm intelligent detection and positioning method and system based on multi-feature fusion, and relates to the technical field of image processing, and the method comprises the steps: receiving and preprocessing a head angiography image; utilizing a multi-scale feature extraction network to obtain feature maps and fusing the feature maps; constructing anatomical candidate regions by using density clustering; extracting a tumor contour through a graph cut energy function; constructing a vascular network topological graph based on probability feature mapping, and calculating a position feature descriptor; and finally carrying out classification discrimination and marking and displaying a detection result. According to the method, various characteristics are fused, the accuracy and sensitivity of cerebral aneurysm detection are improved, and the misdiagnosis rate is reduced.

The invention discloses a slope cutting house building deformation monitoring method based on unmanned aerial vehicle remote sensing, and the method comprises the steps: carrying out the periodic observation of an artificial slope through an unmanned aerial vehicle carrying a three-dimensional laser scanner, and obtaining slope three-dimensional point clouds in at least two different periods; performing denoising, down-sampling and coordinate normalization on the point cloud of each period to obtain a preprocessed data set; generating a global feature descriptor of each period by using a PointNet feature extraction unit; on the basis of a Lucas-Kanade iterative optimization algorithm, the descriptors of all periods are matched with the first period as the reference, and optimal rigid body transformation parameters are obtained; completing registration of the cross-period point clouds in a unified coordinate system according to the parameters; calculating displacement by adopting point-by-point difference and forming a differential deformation field; and comparing the displacement with a preset safety threshold, positioning an abnormal area and giving a risk level. According to the method, centimeter order registration and deformation identification of the cross-period point cloud under the unified stable reference are realized, and early warning and hidden danger assessment of a slope cutting and house building scene are supported.

The invention relates to the technical field of image recognition, in particular to an online real-time image recognition anti-counterfeiting verification method and system based on gold transaction. The method comprises the following steps: acquiring a micro texture image of the gold product; generating a unique digital fingerprint based on the physical unclonable feature of the microscopic texture image, and performing encryption in combination with a UTC timestamp to generate a dynamic encryption tag; extracting a feature descriptor of the micro texture image and performing texture comparison with a pre-stored database to obtain a texture matching degree; decrypting the dynamic encryption label to generate a dynamic code aging state; and according to the texture matching degree and the dynamic code aging state, performing anti-counterfeiting discrimination on the micro texture image of the gold product to obtain a genuine product discrimination result and a counterfeit discrimination result. According to the method, the micro texture image, the near infrared spectrum data, the dynamic encryption label and the block chain technology are combined, so that the accuracy, the safety and the efficiency of gold anti-counterfeiting are improved.

The invention provides a binocular inertial navigation SLAM positioning method fused with deep learning, and the method specifically comprises the steps: receiving an image of a binocular camera and IMU data through a system, and extracting image feature points and feature descriptors through a Superpoint network; feature matching and three-dimensional reconstruction are carried out through front and back frames and left and right eye images, and a dynamic object is detected through a YoV8-seg network to carry out dynamic point elimination; calculating a camera pose based on IMU pre-integration, screening key frames by using a sliding window mechanism, and performing BA optimization to complete binocular inertial navigation initialization; carrying out loopback detection on the feature descriptor binary coding training DBoW3 bag-of-word model; and a LightGlue network is applied to match the current key frame and the candidate key frame, and pose map optimization is carried out through an RANSAC optimization matching result. According to the method provided by the invention, the robustness of the SLAM front end in a dark and weak texture environment can be effectively improved, dynamic scene interference is reduced, the accuracy of pose map optimization is improved, and high precision and stability of real-time positioning are ensured.

The invention discloses a BIM-based building engineering data processing method and system, and belongs to the technical field of digital construction of building engineering. The method comprises the steps that construction scanning point cloud and BIM model data are acquired, and a feature skeleton is extracted; constructing a dynamic local coordinate system to establish a skeleton-model mapping relation; carrying out space division by adopting a locality sensitive hash index and executing ICP registration calculation; reversely deriving an error propagation path to generate a compensation instruction; adjusting a feature descriptor and reconstructing a mapping relation; and finally, model difference comparison and dynamic updating are realized. According to the method, automatic calibration of the construction point cloud and the BIM model is realized, and the problems of low efficiency and large error of traditional manual calibration are solved.

The invention discloses an infrared and visible light image registration method based on an improved SIFT algorithm, and the method comprises the steps: constructing a Gaussian difference pyramid of infrared and visible light images, so as to effectively capture the multi-scale features in the images; performing local adaptive FAST key point extraction on each layer of the pyramid, and detecting significant feature points of the image on a multi-scale level; based on gradient direction distribution of images in key point neighborhoods, main direction distribution is carried out on key points, so that the key points can still keep consistent under different rotation angles; coding the key points based on the gradient information features and generating corresponding feature descriptors to be applied to subsequent registration operation; and performing initial registration by using the similarity between the descriptors, verifying and optimizing an initial registration result by using an optimized RANSAC algorithm, and eliminating mismatching points. According to the method, the number of infrared and visible light image feature point detection and matching is increased, the time complexity is reduced, and the matching precision is ensured.

The invention relates to the technical field of distributed storage verification, and discloses a storage information verification system based on a consensus mechanism. The system comprises a consensus network communication module, a storage feature analysis module, a verification decision core module, a time sequence prediction module, a decision compensation module and a verification execution engine module. The consensus network communication module is connected with a distributed storage node and receives a storage verification request; the storage feature analysis module extracts multi-dimensional feature vectors of the target data to form a feature descriptor set; the verification decision core module is combined with an embedded consensus verification space and dynamic confidence coefficient constraint condition operation to output an initial verification decision scheme; the time sequence prediction module divides a time window and generates a prediction feature descriptor set; the decision compensation module carries out deviation compensation according to the result to obtain an optimization scheme; and the verification execution engine module sends an instruction to trigger the distributed nodes to execute verification operation. A dynamic verification system constructed by the system adapts to dynamic characteristics and heterogeneous environments of distributed storage.

The invention relates to the technical field of image processing, in particular to a remote sensing image registration method based on convolution attention, which comprises the following steps: inputting an original remote sensing image pair into a convolution layer of a modal adapter to obtain original features, and processing the original features by adopting a convolution attention method to obtain target features; inputting the target feature as a modal adaptive feature into a feature encoder to obtain a feature descriptor; inputting the feature descriptors into a feature detector, and estimating global detection feature points by using conditional probability; and performing preliminary matching on the feature descriptors and the feature points by using an FLANN algorithm, and performing outer point elimination in combination with an RANSAC algorithm to obtain a registration result. According to the method, richer feature representation can be provided for remote sensing images of sparse and repeated texture areas, and the registration effect is improved.

The invention discloses a traffic monitoring video rapid target extraction method for edge equipment, and relates to the technical field of intelligent traffic video processing and edge calculation target detection, and the method comprises the steps: carrying out the adaptive downsampling processing of original video frame data, and generating downsampling video frame data; extracting a foreground target candidate region, and constructing a traffic region-of-interest mask in combination with a lane line detection result; carrying out pixel AND operation on the traffic region-of-interest mask and the foreground target candidate region to generate accurate candidate target region data, and extracting a target feature vector; carrying out weighted fusion on the target feature vector through a lightweight attention mechanism, generating a fusion feature descriptor, and calculating a target confidence score; and carrying out screening and duplicate removal processing on the accurate candidate target area data, and outputting traffic target extraction result data. According to the invention, the target in the traffic video can be rapidly and accurately extracted and processed in a low-delay manner on the edge equipment with limited computing resources.

The invention discloses a cancer image contrast enhancement method and system based on deep learning, belongs to the technical field of medical image processing, and adopts an adaptive histogram equalization algorithm for preprocessing to enhance the contrast of a local region. Texture, shape and edge features are extracted through multi-scale analysis, and feature descriptors are constructed. And based on the feature descriptor, segmenting the image by using a region growing segmentation algorithm to obtain a focus region boundary. A self-adaptive contrast enhancement technology is adopted for the segmented focus area, and the focus display effect is highlighted. In consideration of noise interference, a denoising method based on wavelet transform is introduced, noise is suppressed in a high-frequency sub-band, and details are reserved in a low-frequency sub-band. Through the organic combination of the above technical schemes, the contrast and definition of the cancer image can be effectively enhanced, the focus area is highlighted, and noise interference is inhibited at the same time.

The invention provides a multi-scale feature fused textile quality detection method, which comprises the following steps of: aiming at multi-scale feature representation, obtaining a torsion angle and a torsion direction feature descriptor which are kept unchanged under image rotation transformation by calculating an invariant moment of a feature vector, and generating a multi-scale torsion feature vector; calculating a correlation coefficient between the torsion frequency and the torsion density in the multi-scale torsion feature vector and the brightness gradient and the brightness change rate in the brightness feature vector, and determining the correlation degree of the torsion feature and the brightness feature; according to the comprehensive feature vector, a support vector machine classifier is adopted to classify the textile image, and whether a defect area with an abnormal torsion area or an abnormal brightness dynamic range exists in the textile is judged; and in the defect area, local features of a torsion angle, a torsion amplitude, average brightness and brightness contrast are extracted, the torsion degree and the brightness abnormal degree of the defect area are judged through threshold comparison, and quantitative description of the defect is obtained.

The invention provides a robot production line article grabbing method based on visual localization, which comprises the following steps: preprocessing a multi-modal image to obtain an original image; performing grid mapping on the original image to obtain a multi-level feature descriptor; based on the multi-level feature descriptors, a dynamic mapping relation among the three coordinate systems is established through a non-rigid coordinate system alignment algorithm, and coordinate offset of movement of the conveyor belt is compensated in real time; performing hierarchical feature matching on the multi-level feature descriptors and an article template library, identifying article categories and extracting contour geometric features; based on the contour geometric features and the surface curvature distribution, the three-dimensional pose and candidate grabbing points of the object are obtained through a geometric constraint optimization algorithm; and generating a robot obstacle avoidance track according to the candidate grabbing points and the robot motion model. According to the method, accurate coordinate compensation is realized through multi-modal data fusion and dynamic adaptive grid mapping, and the article positioning accuracy is improved in combination with hierarchical feature matching.

The invention discloses a gynaecology and obstetrics emergency medical data security sharing system based on a block chain, and relates to the technical field of medical information processing. The method is used for solving the security and timeliness problems of multi-mechanism data sharing in an emergency scene. The method comprises the following steps: firstly, performing grading processing and desensitization on personal identifiers and medical data of patients through a data grading and desensitization module, and outputting data which can be safely shared; then, an on-chain evidence storage module performs windowing processing on the continuous monitoring data flow, extracts key physiological features, generates feature value Hash, constructs data feature descriptors and submits the data feature descriptors to a block chain network; the emergency access token management module realizes dynamic authorization through a smart contract, and generates a temporary access token associated with the data feature descriptor; and finally, the secure decryption and data sharing module verifies the authority based on the hierarchical decryption key, decrypts the data and compares the eigenvalue hash, and shares and records an operation log with an authorization party after ensuring the data integrity, thereby realizing efficient, secure and traceable data sharing.

Methods, systems, and computer-readable storage media for providing a set of column pairs, each column pair including a column of a bank statement table, and a column of a super invoice table, each column pair corresponding to a modality, the super invoice table including at least one row including data associated with multiple invoices, for each column pair, determining a feature descriptor based on an operator, a feature vector being provided based on feature descriptors of the set of column pairs, inputting the feature vector to a ML model that processes the feature vector to determine a probability of a match between the bank statement, and a super invoice represented by the super invoice table, and outputting a binary output representing one of a match and no match between the bank statement, and the super invoice based on the probability.

The invention discloses a multilevel feature fusion adaptive point cloud registration method, and relates to the field of point cloud processing technology and industrial visual inspection technology, and the method comprises the steps: carrying out the preprocessing of a point cloud pair; multi-scale feature extraction is carried out on the preprocessed point cloud pair, a feature pyramid is constructed through the extracted multi-scale features, and a top-down guide mechanism is adopted between layers of the feature pyramid; performing preliminary feature fusion on feature points of the point cloud pair based on the extracted multi-scale features; performing multi-level registration on the point cloud pair; the multiple layers comprise first-layer registration, second-layer registration and third-layer registration, a transformation matrix of each layer is obtained after each layer of registration is completed, and registration is carried out on the to-be-registered point cloud through the transformation matrix of each layer; according to the method, a multi-scale feature extraction strategy is adopted, three feature descriptors with different scales, namely ESF, SHOT and FPFH feature descriptors, are selected for the degradation problem in point cloud registration, and feature extraction and parameter adaptive adjustment from macroscopic to microscopic are achieved.