92 results about "Spatial calibration" patented technology

The invention belongs to the technical field of intelligent agriculture, and particularly relates to a melon seedling feature recognition and analysis method based on image multi-dimensional features, which comprises the following steps of: firstly, synchronously acquiring two-dimensional image information and three-dimensional point cloud information of a melon crop seedling raising area to form multi-modal data, and performing spatial calibration and preprocessing; the method comprises the following steps of: extracting two-dimensional semantic features and three-dimensional geometric features of a target seedling, identifying a current growth stage of the target seedling based on the extracted features, dynamically adjusting a fusion weight of the two-dimensional semantic features and the three-dimensional geometric features, and performing feature fusion by adopting a cross-modal attention mechanism to generate a dynamic fusion feature vector; and inputting the dynamic fusion feature vector into a task identification model, and outputting the health state and quantitative growth parameters of the target seedling, thereby realizing high-precision, self-adaptive and multi-dimensional perception of the morphology and growth health state of the melon seedling, and playing an important role in promoting high-quality development of modern intelligent agriculture.

The invention relates to a multi-vehicle cooperative sensing method based on an attention mechanism, and belongs to the technical field of mobile communication. A multi-vehicle collaborative perception scene lacks selective attention to perception information of different values, so that perception characteristics of each vehicle are difficult to effectively utilize. At the same time, in a high mobility scene, it is difficult to align sensing information time delay and asynchronous features, resulting in low dynamic target sensing precision. In order to solve the problem, the method comprises the following steps: firstly, filtering redundant sensing information of cooperative vehicles, and constructing a screening model based on aerial view feature intersection-to-union ratio so as to select high-quality cooperative vehicles; spatial calibration is completed in combination with sparse consensus foreground features and geometric consistency verification, and feature fusion is realized through an attention mechanism; and finally, finishing time calibration of the dynamic target by adopting a time alignment strategy based on distillation learning. The method is suitable for multi-vehicle cooperative perception in an Internet of Vehicles scene, the focus on high-quality perception information is improved, and meanwhile, the real-time perception precision and robustness are enhanced.

The invention discloses a fusion detection and iron tower photoelectric tracking system and a fusion detection and iron tower photoelectric tracking method for a low-speed small target. The system comprises a multi-source detection module, a data fusion module and a photoelectric tracking module, and various heterogeneous data of a target are obtained through a radar detection unit, a frequency spectrum detection unit and a protocol cracking unit; after data preprocessing, performing multi-source data fusion by adopting a dynamic weight fusion algorithm and Kalman filtering to generate comprehensive target information; and based on the equipment space calibration parameters, calculating an optoelectronic equipment pointing instruction through a coordinate conversion model, and driving the optoelectronic equipment to accurately track a target. The system is further provided with a feedback optimization module which dynamically corrects the pointing instruction according to the tracking deviation. According to the invention, deep fusion and intelligent cooperation of multi-source detection data are realized, and the detection precision, tracking stability and system response speed of a low-speed small target are effectively improved.

The invention provides a power grid composite chain disaster multi-source monitoring damage assessment method and system, and relates to the technical field of data processing, and the method comprises the steps: dividing a reference geographic coverage area, generating a plurality of independent analysis units, enabling each three-dimensional discrete point to belong to the corresponding independent analysis unit, and carrying out the calculation to obtain a unit space calibration coefficient; performing regional correction on the fused disaster monitoring image data set by using the unit space calibration coefficient to obtain calibrated disaster monitoring image data; based on the calibrated disaster monitoring image data, disaster damage features are extracted through a deep learning semantic segmentation network; and based on the disaster damage characteristics, carrying out spatial diffusion path simulation and time evolution process analysis of disaster influence to obtain disaster spatio-temporal evolution characteristics, and carrying out comprehensive quantitative evaluation of damage degree and risk level on power grid facilities to generate a comprehensive damage evaluation result. According to the invention, the accuracy of power grid disaster damage monitoring and the timeliness of evaluation are improved, and the risk early warning and coping capability of power grid facilities is enhanced.

The invention provides a display screen dynamic control method and system, and relates to the technical field of data processing, the method comprises the following steps: constructing a space coordinate system comprising a plurality of entity nodes, the entity nodes comprising a display screen preset angular point, a user binocular vision center and a handheld interaction device; analyzing a relative pose relationship among the nodes by calculating a projection dot product of a space vector corresponding to each node; constructing a dynamic topology network based on the relative pose relationship, and synchronously collecting environment illumination intensity, user fixation point coordinates and user pose data; performing time sequence synchronization and Kalman filtering preprocessing on the acquired multi-modal data to obtain an initial feature set; and generating a spatial calibration coefficient according to the distribution characteristics of the dynamic topology network, and performing normalization compensation on the initial feature set through the calibration coefficient to obtain a calibrated multi-modal sensing feature vector. According to the invention, accurate and intelligent dynamic control is realized.

The invention discloses an artificial leather quality rapid detection method based on spectral analysis, and particularly relates to the field of quality detection.The method comprises the steps that component composition characteristics, surface and microcosmic characteristics and dynamic response characteristics of artificial leather are synchronously obtained through multi-dimensional original data acquisition, and a time-space correlation data set is formed through timestamp alignment and space calibration; constructing component analysis, surface association and dynamic response models, and eliminating space-time migration through a synchronous optimization model; after multi-dimensional features are extracted, optimized features are generated through cross-direction tensor fusion, attention enhancement and gating cycle unit time sequence fusion; and finally, quantizing the quality index based on a multi-task model, and outputting a quality grade and a traceable report in combination with a dynamic threshold value through hierarchical coding, graph neural network association enhancement and evidence fusion. The single sample collection duration is less than 3 minutes, the response time is less than 1 second, and full-dimension, rapid and accurate detection and quality traceability are realized.

In a virtual camera calibration method, a target object model of a physical camera device is obtained. The physical camera device includes a camera and a contact plane. The target object model with an operation controller is output for display through a head-mounted display device. The contact plane of the target object model and a contact plane of the operation controller are in an attached state. A camera position of the target object model corresponds to a camera position of the physical camera device. An indication that a camera of the target object model and the camera of the physical camera device are in an overlapping state is output for display. A virtual camera is calibrated based on spatial information of the operation controller and a spatial calibration operation for the virtual camera. The spatial information includes target position information and a target orientation.

The embodiment of the invention discloses an RTK-fused laser inertial navigation odometer mapping method and system, and the method comprises the steps: carrying out the time synchronization and space calibration of a laser radar, an inertial measurement unit and a double-mushroom-head antenna RTK system, and achieving the coordinate system alignment through lever arm compensation; multi-source constraints are fused in pose map optimization, and point cloud ghosting is eliminated by associating poses of a current frame and multiple historical frames; designing a dynamic fusion strategy for RTK positioning data, detecting the height of an environmental obstacle by using a laser radar, dynamically adjusting an RTK constraint weight, adding constraint factors in stages, carrying out intensive accumulation in an initialization stage to realize rapid alignment of a global coordinate system, and adding a suppression accumulative error in a subsequent low frequency; and a GTSAM optimizer is adopted to execute pose map optimization, and the optimized key frame point cloud is converted to a global coordinate system to generate a down-sampling map, so that the map consistency, precision and real-time performance of long-distance mapping are remarkably improved.

The invention relates to the technical field of video monitoring, in particular to a method for quickly searching a linkage monitoring camera in a panoramic video, which comprises the following steps of: obtaining monitoring equipment information and panoramic video picture information through a panoramic monitoring system; constructing a panoramic video picture correction mechanism to adjust the panoramic video picture information to obtain corrected panoramic video coordinate information; introducing a geographic coordinate and pixel rectangular coordinate transformation expression based on the basic imaging information, and analyzing a corresponding relation between a panoramic video picture and a physical space coordinate based on the geographic coordinate and pixel rectangular coordinate transformation expression and the corrected panoramic video coordinate information; combining monitoring equipment information, corrected panoramic video coordinate information and the corresponding relation to build a panoramic area and monitoring camera mapping library; and carrying out panoramic target area frame selection and panoramic video picture display according to the panoramic area and the monitoring camera mapping library. According to the invention, linkage display of a panoramic video and a camera local picture is realized through space calibration and an intelligent algorithm.

The invention belongs to the technical field of information technology, and discloses a moxibustion temperature control method and system based on multi-sensor fusion, and the method comprises the steps: obtaining a biopotential signal and position offset data from an acupoint region through a sensor, processing the biopotential signal through a classification algorithm, and obtaining a classified potential distribution mode; on the basis of the classified potential distribution modes, heat conduction influence data between adjacent acupoints are obtained, the position offset data are integrated through an information fusion algorithm, fused physiological state indexes are determined, and the problem of measurement distortion caused by acupoint movement and equipment offset is solved through multi-source data fusion and dynamic error compensation; and meanwhile, the system constructs an accurate potential distribution model by utilizing a spatial calibration and classification algorithm, and temperature control accuracy reaches a millimeter level in combination with real-time position tracking, so that the integrity and reliability of data are remarkably improved.

The invention provides an intervention control method for aided driving and an automobile, and the method comprises the steps: obtaining a panoramic stitching image of a test environment, and recognizing a calibration object set in the panoramic stitching image; generating a semantic object according to the calibration object set; obtaining pose data of the vehicle, and determining a spatial relationship parameter between the vehicle and the semantic object according to the semantic object and the pose data; performing intention classification according to the semantic object and the driving state information of the user to obtain a driving intention identifier; and determining whether the vehicle is in danger according to the semantic object, the space calibration data and the driving intention identifier, and intervening in controlling the vehicle when the vehicle is in danger.

The invention relates to the technical field of space calibration, in particular to a non-overlapping field-of-view dual-camera space calibration method based on an ArUco calibration board, and the method comprises the steps: carrying out the single-target calibration of two cameras; fixing the calibration plate at a plurality of preset postures and spatial positions, carrying out image acquisition, and obtaining an image frame of the calibration plate under the current spatial pose; extracting an ArUco marker ID and a two-dimensional angular point coordinate of the ArUco marker ID in the image frame, and recovering three-dimensional information of the calibration plate in each frame of image under a camera coordinate system; three-dimensional space points corresponding to the same marker ID in the two camera coordinate systems are recovered, and a plurality of space point pairs with the known corresponding relation are constructed; calculating a rotation matrix and a translation vector from the first camera coordinate system to the second camera coordinate system; according to the method, marker data in respective visual fields are collected under a public world coordinate system, and the unique coding characteristic of the marker data is combined, so that cross-visual-field space point pair matching and accurate transformation estimation are realized.

The invention provides a sea target identification method and device based on AIS and SAR image fusion, and the method comprises the steps: obtaining the AIS data of a sea target through an AIS system, obtaining the SAR image of the sea target through an SAR satellite, and carrying out the preprocessing of the data; after time calibration and space calibration are carried out on the preprocessed data, the fusion weight of the AIS data and the SAR image in Kalman filtering is determined according to the confidence score of the AIS data and the SAR image; determining a dynamic measurement noise covariance matrix according to the fusion weight, calculating a Kalman gain according to the dynamic measurement noise covariance matrix and an error covariance matrix, and updating the state of the marine target and the error covariance matrix by using the Kalman gain; and inputting the SAR image and the final state output by the Kalman filtering into the deep learning model to obtain a sea target recognition result output by the deep learning model. According to the invention, the precision, robustness and real-time performance of sea target identification are improved.

The invention belongs to the technical field of building information model and cultural heritage digital protection, and particularly relates to a historical building BIM refined modeling method based on multi-source data fusion, and the method comprises the steps: collecting multi-source heterogeneous data such as three-dimensional laser point cloud, unmanned aerial vehicle oblique photography and surveying and mapping drawings; spatial calibration is carried out under the unified geographic coordinates and time reference, and a structured metadatabase is constructed; based on building component classification and fineness grading standards, parameterized BIM modeling is executed, and details such as a platform base, a bucket arch and a roof are truly restored; parameterized reconstruction and environment integration of the vegetation system are synchronously carried out; and finally, importing an unreal engine, and constructing a high-fidelity and strong-immersion digital humanity scene in combination with physical rendering, dynamic weather, day and night circulation and AI interaction logic. According to the method, total-factor, high-precision and interactive digital twinning of historical buildings is realized, and the protection, display and activation utilization level of cultural heritage is remarkably improved.

The invention belongs to the technical field of image processing, and relates to an egg surface microcrack detection method and system based on machine vision. The method comprises the following steps: acquiring a transmission image of an egg, performing spatial calibration, determining a geometric center and a long axis scale of an egg region, and establishing a polar coordinate mapping relation of each pixel point; according to the linear distance from each pixel point to the geometric center and the included angle relative to the long axis direction, the geometric sensitivity weight of each pixel point is evaluated, and the optical path distortion degree of the edge high-curvature area is represented; extracting contrast basic energy of the pixel points by using a local space window, and carrying out nonlinear evolution on a local gray gradient by combining with a geometric sensitivity weight to obtain an adaptive contrast energy evolution result; and coupling the local background variance and the geometric sensitivity weight, and calculating the crack identification confidence of each pixel point so as to carry out identification decision of the egg surface microcracks. According to the invention, extremely fine crack signals can be captured, and accurate detection of the microcracks on the surface of the egg is realized.

The invention relates to the technical field of image detection, and provides a multi-mode wafer defect detection system. The system comprises a guided wave acquisition module, a stress reconstruction module, a crack database building module, a type identification module and a multi-modal data fusion module, stress distribution information of a crack area is extracted through guided wave propagation parameters, and depth positioning and space calibration of cracks in a multi-layer structure are realized in combination with correction of propagation velocity by temperature. A crack database is constructed by utilizing wave velocity characteristics generated by standardized indentations, a clear mapping relation is established between propagation characteristics and crack types, so that the accuracy of crack property judgment is improved, and the accuracy of crack property judgment is improved by performing spatial registration on a crack recognition result and a stripe abnormal region in an optical interference pattern. Recognition blind areas existing in image detection can be effectively covered, guided wave detection and optical detection results are fused, and the defect coverage range and recognition integrity are enhanced.

The invention belongs to the related technical field of fluid temperature field measurement, and discloses a Scanning LIF technology-based three-dimensional temperature field measurement system and method, which comprises a sheet light scanning module and a double-color image synchronous acquisition module, and is characterized in that a parallel sheet light source is generated through double-camera double-color synchronous imaging and galvanometer scanning; according to the system, space calibration images, temperature calibration images and experiment images at different positions can be obtained through testing, the experiment images collected by the two cameras are subjected to ratio processing after space calibration, and a fluorescence intensity ratio-temperature relation curve is generated through temperature calibration at different positions. And a temperature field image corresponding to each position can be calculated. The system can generate parallel sheet light sources, and can change the number of scanning sheet light, the deflection angle of the galvanometer, the motion mode of the galvanometer, the residence time and the moving time of the galvanometer in a period, the exposure time of a camera and the like according to measurement requirements so as to meet actual use requirements.

The embodiment of the invention provides a road surface feature recognition method and system, and the method comprises the steps: detecting the road surface data of a road on which a vehicle runs through different sensors in a multi-mode sensor, and obtaining various types of initial road surface data; time synchronization and space calibration are carried out on the multiple kinds of initial road surface data, multiple kinds of target road surface data are obtained, the multiple kinds of target road surface data correspond to the multiple kinds of initial road surface data, and time synchronization errors among the multiple kinds of target road surface data are within a first error threshold value range; the spatial calibration error among the various target road surface data is within a second error threshold range; determining target weights corresponding to different sensors based on the environment perception information of the environment where the road is located, and fusing the various target road surface data according to the target weights corresponding to the different sensors to obtain a data fusion result; and based on the data fusion result, identifying the road surface features of the road. The technical problem of low accuracy of pavement feature recognition is solved.

This invention provides a method for fish passage behavior recognition based on a multimodal large model. First, imaging sonar image sequences and underwater optical video sequences of fish passing through fish passages or facilities are simultaneously acquired, and the multimodal data are aligned through time stamp synchronization and spatial calibration. Then, fish target detection, segmentation, and feature extraction are performed on the sonar image sequences and optical video sequences respectively to obtain the fish's spatial location, body length, water depth, motion state, morphological structure, and posture features. The obtained acoustic and optical features are input into a multimodal coding and fusion model based on the Transformer architecture, and a unified fish behavior representation vector is constructed through a cross-modal attention mechanism. During the training phase, model parameters are adjusted through a multi-task learning approach using cross-entropy loss, mean squared error loss, and contrastive learning loss, ultimately achieving fish passage behavior category recognition and prediction of individual and group passage difficulty scores.

The application provides a method and system for spatial calibration of a radar device. The method comprises: obtaining an image and radar measurement data at the same time, the radar measurement data comprising radar coordinates of a plurality of landmark points in a geographic region; obtaining geographic coordinates of the landmark points recorded in the image according to pixel coordinates of the landmark points recorded in the image; determining at least two landmark point matching pairs, each landmark point matching pair comprising a geographic coordinate of a landmark point and a radar coordinate of the landmark point; and obtaining a calibration relationship between the radar measurement data and the geographic region according to the at least two landmark point matching pairs. The method can improve the accuracy and efficiency of spatial calibration.

The invention discloses an automatic driving risk assessment method and device, a vehicle, equipment and a medium, and belongs to the technical field of automatic driving. The method comprises the following steps: respectively acquiring a visual image and a radar reflectivity graph of a target road surface; after time synchronization and space calibration are carried out on the visual image and the radar reflectivity graph, visual features of an area where a lane line is located are extracted from the visual image by adopting a convolutional neural network, a target reflectivity area higher than a reflectivity threshold in the radar reflectivity graph is identified, and reflectivity distribution features of the target reflectivity area are extracted; fusing the visual features and the reflectivity distribution features according to time synchronization and space calibration results to obtain a lane line comprehensive feature set; generating a lane line sensing result according to the lane line comprehensive feature set; and matching is performed from the risk level relationship according to the aging degree of the lane line and the complexity of the road surface, and the target risk level corresponding to the lane line modeling information is obtained, so that the safety of automatic driving is improved.

The invention discloses a vehicle obstacle detection method, device and equipment, a medium and a program product. The method comprises the steps of executing time synchronization operation and space calibration operation of a camera and a radar through a hardware-level timestamp synchronization module and a dynamic calibration compensation algorithm, and obtaining visual data and radar data after execution is completed; through a backbone network of the obstacle detection model, feature extraction is carried out on the visual data and the radar data by utilizing mixing precision quantification and combining an attention module, and a feature extraction result is determined; through a neck network of an obstacle detection model, feature fusion is carried out by combining environmental sensor data on the basis of the feature extraction result, and a feature fusion result is determined; and performing obstacle detection on the basis of the feature fusion result through the head network pruned by the obstacle detection model through a dynamic pruning algorithm, and determining an obstacle detection result. The defects of multi-sensor space-time alignment, extreme weather robustness and edge computing resource limitation in the prior art are overcome.

The utility model discloses a three-dimensional spatial calibration module for realizing spatial three-dimensional positioning and camera calibration, which comprises a cubic framework, calibration target surfaces are respectively arranged on six surfaces of the cubic framework, and laser etching checkerboards, dot matrixes or ArUco coding patterns are respectively arranged on the six calibration target surfaces. And the six surfaces of the cubic skeleton are respectively provided with an identification surface coding module. According to the utility model, multi-surface pattern arrangement is adopted, and multi-angle three-dimensional calibration can be realized by a single module; the structural cubic design meets spatial analytic geometry constraints, and the reconstruction stability is improved; a built-in attitude sensing system records calibration attitudes and improves reproduction and error analysis capability; the modular quick-change base design supports quick plugging and mobile calibration; a wireless tag and visual coding combined recognition mechanism enhances the recognition and matching efficiency of an automatic system.

The embodiment of the invention discloses an infrared band multi-mode calibration method and device. The method comprises the following steps: calculating a point spread function (PSF) of an original infrared image, and generating a spatial response matching filter function (SRMF) according to the PSF; carrying out two-dimensional convolution calculation on the SRMF to obtain an infrared image after effective space calibration; generating a possible spectral response function of ice pollution, and evaluating an optimal ice pollution spectral response through a cross calibration method; according to the optimal ice pollution spectral response, a spectral response function after possible spectral offset is generated, the optimal possible spectral offset is evaluated through a cross calibration method, and a spectral response function after spectral calibration is obtained through optimal sampling; calculating a linear coefficient, a nonlinear coefficient and a constant coefficient of radiometric calibration; and according to each coefficient of radiometric calibration, calculating the radiation quantity of the infrared image after multi-modal calibration. The method can overcome the existing problems and promote the quantification and multi-source fusion application of the satellite infrared data.

The invention relates to a method for fusing a monitoring video and a 3D virtual scene, and belongs to the technical field of image processing and computer vision. The method comprises the following steps: firstly, constructing a three-dimensional digital scene and completing space calibration of a camera; then, camera offset is detected and corrected through real-time video analysis, and meanwhile, a target in a picture is recognized and tracked; secondly, the 2D pixel coordinates of the target are converted into space coordinates in a 3D scene through inverse projection calculation, and the positioning precision is improved through the multi-view intersection technology; and finally, based on the converted 3D coordinate data, performing fusion rendering on a video picture in a rendering engine, and outputting an interactive three-dimensional scene which can be used for situation awareness and behavior analysis. According to the method, the problems of inaccurate virtual and real space registration, dynamic error accumulation and unnatural rendering in the prior art are effectively solved, and high-precision and dynamic fusion of the monitoring video and the virtual scene is realized.

The application provides a plant disease small sample classification method and system based on feature space calibration, and belongs to the field of data processing. To solve the problem that the existing deep learning algorithm needs a large number of annotated data sets when classifying images, but the incidence of some plant diseases is low, the image data is small, and the deep learning algorithm is limited in large-scale application in plant disease identification. The FAS score is used to score and analyze each feature output layer of the Swin-Transformer V2 network after removing the full connection layer, the layer with the highest score is selected as the feature vector, the Swin-Transformer V2F6 network is obtained, and then the plant disease feature space calibration is carried out. The accuracy of the small sample disease classification can be effectively improved, and compared with the use of a large number of annotated data sets, the accuracy of identification can be improved on the basis of a small amount of data sets.

The invention discloses a magnetic resonance diffusion-weighted imaging image reconstruction method and device, equipment and a storage medium. The method comprises the following steps: acquiring k-space calibration data and each time of undersampling k-space data; obtaining to-be-input data according to the undersampled k-space data of each time and the K-space calibration data, wherein the to-be-input data comprises an initial reconstruction image, a background phase diagram and a calibration image; inputting the to-be-input data into a conditional fraction diffusion model for reconstruction to obtain a DWI image; and if the DWI image does not meet the preset requirement, taking the DWI image as a new initial reconstruction image, and returning to the step of obtaining the DWI image according to the to-be-input data to carry out iterative operation until it is determined that the DWI image meets the preset requirement, and taking the DWI image meeting the requirement as a target DWI image. The method effectively solves the problems of low signal-to-noise ratio and multiple artifacts in DWI image reconstruction through the combination of the conditional fractional diffusion model and iterative optimization, and improves the image reconstruction quality and efficiency.

According to the machining system and the using method thereof, the spatial position of the fixing piece is converted into quantifiable data through the spatial calibration unit, a rough mode depending on naked eye estimation is replaced from the source, and objective and accurate initial input is ensured; the position acquisition unit is used for accurately acquiring three-dimensional coordinate data; furthermore, the data processing unit automatically generates a three-dimensional curve and a processing instruction conforming to an anatomical structure through a preset algorithm based on the three-dimensional coordinate data, so that a design link depending on spatial imagination and subjective experience of a doctor is thoroughly abandoned, and accurate and quantitative conversion from data to a scheme is realized; and finally, the target processing equipment performs processing under the direct driving or guidance of the instruction, and the uncertainty and repeated trial and error of manual operation are eliminated, so that the personal error accumulation of each link is gradually eliminated in a full-link digital closed loop, and the core technical problem of poor rod bending precision is systematically solved finally.

The invention discloses a full-automatic double-sided synchronous labeling system based on multi-axis robot cooperative control, and relates to the technical field of production supervision. The production line image data of the production line are obtained through the multi-view structure optical camera, the scene mapping space is established, the production line image data are mapped to the scene mapping space for space calibration to obtain the real-time production scene data, and then the production line three-dimensional structure model is established according to the real-time production scene data. A standard product labeling position is set in the three-dimensional structure model, and then according to the spatial position relation between the standard product labeling position and the product model, the position or shape of the machined product is adjusted through the industrial robot; and real-time surface light transmittance of the two sides of the processed product is collected, the two sides of the processed product are labeled according to the real-time surface light transmittance when labeling countdown is finished, a non-shadow label image of the labeling position of the product is collected through a line-scan digital camera, and whether labeling is qualified or not is judged according to the non-shadow label image.

The invention relates to a smoke detection method, device and equipment based on multi-modal fusion. The method comprises the following steps: controlling an acquisition unit to acquire image data and point cloud data of a scene to be monitored through a synchronization unit mounted on a rail-mounted inspection robot; based on a pre-established initial pixel point cloud mapping relationship, performing spatial calibration and dynamic association on the image data and the point cloud data, and performing multilevel feature extraction on the image data and the point cloud data after spatial calibration and dynamic association to obtain an image feature map and a point cloud feature map; based on a preset cross-modal feature interaction mechanism, fusing the image feature map and the point cloud feature map to obtain a fused feature map; and performing smoke detection on the fused feature map and outputting a detection report. The problems that in the prior art, accurate space alignment and an effective feature interaction mechanism are lacked between point cloud and image data, feature information fusion is poor, and detection accuracy is low are solved.